diff --git a/.devops/cloud-v-pipeline b/.devops/cloud-v-pipeline new file mode 100644 index 000000000..f3a4944f8 --- /dev/null +++ b/.devops/cloud-v-pipeline @@ -0,0 +1,22 @@ +node('x86_runner1'){ // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries + stage('Cleanup'){ + cleanWs() // Cleaning previous CI build in workspace + } + stage('checkout repo'){ + retry(5){ // Retry if the cloning fails due to some reason + checkout scm // Clone the repo on Runner + } + } + stage('Compiling llama.cpp'){ + sh'''#!/bin/bash + make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V + ''' + } + stage('Running llama.cpp'){ + sh'''#!/bin/bash + module load gnu-bin2/0.1 # loading latest versions of vector qemu and vector gcc + qemu-riscv64 -L /softwares/gnu-bin2/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt # Running llama.cpp on vector qemu-riscv64 + cat llama_log.txt # Printing results + ''' + } +} diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile index e5fcb37d6..360602d65 100644 --- a/.devops/full-cuda.Dockerfile +++ b/.devops/full-cuda.Dockerfile @@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build ARG CUDA_DOCKER_ARCH=all RUN apt-get update && \ - apt-get install -y build-essential python3 python3-pip + apt-get install -y build-essential python3 python3-pip git COPY requirements.txt requirements.txt diff --git a/.devops/main-cuda.Dockerfile b/.devops/main-cuda.Dockerfile index 30c01196a..2b7faf7c1 100644 --- a/.devops/main-cuda.Dockerfile +++ b/.devops/main-cuda.Dockerfile @@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build ARG CUDA_DOCKER_ARCH=all RUN apt-get update && \ - apt-get install -y build-essential + apt-get install -y build-essential git WORKDIR /app diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 9d0a6c222..8b869f688 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -27,7 +27,7 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 - name: Dependencies id: depends @@ -52,7 +52,7 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 - name: Dependencies id: depends @@ -87,7 +87,7 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 - name: Dependencies id: depends @@ -121,7 +121,7 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 - name: Dependencies id: depends @@ -149,7 +149,7 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 - name: Dependencies id: depends @@ -174,7 +174,7 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 - name: Dependencies id: depends @@ -197,6 +197,62 @@ jobs: cd build ctest --verbose --timeout 900 + macOS-latest-cmake-ios: + runs-on: macos-latest + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v1 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=iOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 + cmake --build . --config Release + + macOS-latest-cmake-tvos: + runs-on: macos-latest + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v1 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=tvOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 + cmake --build . --config Release + windows-latest-cmake: runs-on: windows-latest @@ -224,7 +280,9 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 + with: + fetch-depth: 0 - name: Download OpenCL SDK id: get_opencl @@ -334,20 +392,21 @@ jobs: strategy: matrix: - cuda: ['12.1.0', '11.7.1'] + cuda: ['12.2.0', '11.7.1'] build: ['cublas'] steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 + with: + fetch-depth: 0 - - uses: Jimver/cuda-toolkit@v0.2.10 + - uses: Jimver/cuda-toolkit@v0.2.11 id: cuda-toolkit with: cuda: ${{ matrix.cuda }} - # TODO(green-sky): _dev seems to fail, and non dev are not enought - #sub-packages: '["nvcc", "cudart", "cublas", "cudart_dev", "cublas_dev"]' + sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]' - name: Build id: cmake_build @@ -384,27 +443,11 @@ jobs: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip - name: Copy and pack Cuda runtime - if: ${{ matrix.cuda == '12.1.0' }} - # TODO(green-sky): paths are cuda 12 specific run: | echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}" - mkdir '.\build\bin\cudart\' - cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_12.dll" '.\build\bin\cudart\' - cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_12.dll" '.\build\bin\cudart\' - cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_12.dll" '.\build\bin\cudart\' - 7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\* - - - name: Copy and pack Cuda runtime - if: ${{ matrix.cuda == '11.7.1' }} - # TODO(green-sky): paths are cuda 11 specific - run: | - echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}" - mkdir '.\build\bin\cudart\' - ls "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" - cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_110.dll" '.\build\bin\cudart\' - cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_11.dll" '.\build\bin\cudart\' - cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_11.dll" '.\build\bin\cudart\' - 7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\* + $dst='.\build\bin\cudart\' + robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll + 7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\* - name: Upload Cuda runtime if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} @@ -413,6 +456,22 @@ jobs: path: | cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip + freeBSD-latest: + runs-on: macos-12 + steps: + - name: Clone + uses: actions/checkout@v3 + + - name: Build + uses: cross-platform-actions/action@v0.19.0 + with: + operating_system: freebsd + version: '13.2' + run: | + sudo pkg update + sudo pkg install -y gmake automake autoconf pkgconf llvm15 clinfo clover opencl clblast openblas + gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 + release: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} @@ -429,7 +488,9 @@ jobs: steps: - name: Clone id: checkout - uses: actions/checkout@v1 + uses: actions/checkout@v3 + with: + fetch-depth: 0 - name: Determine tag name id: tag @@ -487,7 +548,7 @@ jobs: # # steps: # - name: Clone -# uses: actions/checkout@v1 +# uses: actions/checkout@v3 # # - name: Dependencies # run: | @@ -511,7 +572,7 @@ jobs: # # steps: # - name: Clone -# uses: actions/checkout@v1 +# uses: actions/checkout@v3 # # - name: Dependencies # run: | @@ -535,7 +596,7 @@ jobs: # # steps: # - name: Clone -# uses: actions/checkout@v1 +# uses: actions/checkout@v3 # # - name: Dependencies # run: | @@ -565,7 +626,7 @@ jobs: # # steps: # - name: Clone -# uses: actions/checkout@v1 +# uses: actions/checkout@v3 # # - name: Add msbuild to PATH # uses: microsoft/setup-msbuild@v1 @@ -604,7 +665,7 @@ jobs: # # steps: # - name: Clone -# uses: actions/checkout@v1 +# uses: actions/checkout@v3 # # - name: Add msbuild to PATH # uses: microsoft/setup-msbuild@v1 @@ -650,7 +711,7 @@ jobs: # # steps: # - name: Clone -# uses: actions/checkout@v1 +# uses: actions/checkout@v3 # # - name: Dependencies # run: | diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 379fbd7ad..9c90c77ac 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -26,8 +26,15 @@ jobs: strategy: matrix: config: - - { tag: "light", dockerfile: ".devops/main.Dockerfile" } - - { tag: "full", dockerfile: ".devops/full.Dockerfile" } + - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" } + # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I + # have disabled them for now until the reason why + # is understood. + - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" } + - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" } + - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } steps: - name: Check out the repo uses: actions/checkout@v3 @@ -51,7 +58,7 @@ jobs: with: context: . push: true - platforms: linux/amd64,linux/arm64 + platforms: ${{ matrix.config.platforms }} tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}" file: ${{ matrix.config.dockerfile }} @@ -60,6 +67,6 @@ jobs: with: context: . push: ${{ github.event_name == 'push' }} - platforms: linux/amd64,linux/arm64 + platforms: ${{ matrix.config.platforms }} tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}" file: ${{ matrix.config.dockerfile }} diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml index a6289e335..e61bfc6c3 100644 --- a/.github/workflows/gguf-publish.yml +++ b/.github/workflows/gguf-publish.yml @@ -24,7 +24,7 @@ jobs: runs-on: ubuntu-latest steps: - - uses: actions/checkout@v2 + - uses: actions/checkout@v3 - name: Set up Python uses: actions/setup-python@v2 with: diff --git a/CMakeLists.txt b/CMakeLists.txt index d4fa5c261..abecd684b 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -135,6 +135,7 @@ set(CMAKE_C_STANDARD 11) set(CMAKE_C_STANDARD_REQUIRED true) set(THREADS_PREFER_PTHREAD_FLAG ON) find_package(Threads REQUIRED) +include(CheckCXXCompilerFlag) if (NOT MSVC) if (LLAMA_SANITIZE_THREAD) @@ -171,8 +172,8 @@ if (LLAMA_METAL) find_library(METALKIT_FRAMEWORK MetalKit REQUIRED) message(STATUS "Metal framework found") - - set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h) + set(GGML_HEADERS_METAL ggml-metal.h) + set(GGML_SOURCES_METAL ggml-metal.m) add_compile_definitions(GGML_USE_METAL) if (LLAMA_METAL_NDEBUG) @@ -191,7 +192,6 @@ if (LLAMA_METAL) ${METALKIT_FRAMEWORK} ) endif() - if (LLAMA_BLAS) if (LLAMA_STATIC) set(BLA_STATIC ON) @@ -268,7 +268,8 @@ if (LLAMA_BLAS) endif() if (LLAMA_K_QUANTS) - set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h) + set(GGML_HEADERS_EXTRA k_quants.h) + set(GGML_SOURCES_EXTRA k_quants.c) add_compile_definitions(GGML_USE_K_QUANTS) if (LLAMA_QKK_64) add_compile_definitions(GGML_QKK_64) @@ -284,7 +285,8 @@ if (LLAMA_CUBLAS) enable_language(CUDA) - set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h) + set(GGML_HEADERS_CUDA ggml-cuda.h) + set(GGML_SOURCES_CUDA ggml-cuda.cu) add_compile_definitions(GGML_USE_CUBLAS) # if (LLAMA_CUDA_CUBLAS) @@ -332,6 +334,7 @@ if (LLAMA_MPI) find_package(MPI) if (MPI_C_FOUND) message(STATUS "MPI found") + set(GGML_HEADERS_MPI ggml-mpi.h) set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h) add_compile_definitions(GGML_USE_MPI) add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS}) @@ -354,7 +357,8 @@ if (LLAMA_CLBLAST) if (CLBlast_FOUND) message(STATUS "CLBlast found") - set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h) + set(GGML_HEADERS_OPENCL ggml-opencl.h) + set(GGML_SOURCES_OPENCL ggml-opencl.cpp) add_compile_definitions(GGML_USE_CLBLAST) @@ -382,13 +386,15 @@ if (LLAMA_HIPBLAS) message(STATUS "HIP and hipBLAS found") add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS) add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h) + if (BUILD_SHARED_LIBS) + set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON) + endif() if (LLAMA_CUDA_FORCE_DMMV) target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV) endif() target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X}) target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y}) target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER}) - target_compile_definitions(ggml-rocm PRIVATE CC_TURING=1000000000) set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX) target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas) @@ -421,6 +427,7 @@ if (LLAMA_ALL_WARNINGS) -Wextra -Wpedantic -Wcast-qual + -Wmissing-declarations -Wno-unused-function -Wno-multichar ) @@ -439,7 +446,7 @@ if (LLAMA_ALL_WARNINGS) endif() -if (MSVC) +if (WIN32) add_compile_definitions(_CRT_SECURE_NO_WARNINGS) if (BUILD_SHARED_LIBS) @@ -461,6 +468,13 @@ endif() # TODO: probably these flags need to be tweaked on some architectures # feel free to update the Makefile for your architecture and send a pull request or issue message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}") +if (MSVC) + string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR) + message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}") +else () + set(CMAKE_GENERATOR_PLATFORM_LWR "") +endif () + if (NOT MSVC) if (LLAMA_STATIC) add_link_options(-static) @@ -476,25 +490,33 @@ if (NOT MSVC) endif() endif() -if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") +if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64")) message(STATUS "ARM detected") if (MSVC) - # TODO: arm msvc? + add_compile_definitions(__ARM_NEON) + add_compile_definitions(__ARM_FEATURE_FMA) + add_compile_definitions(__ARM_FEATURE_DOTPROD) + # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16 + add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead else() + check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E) + if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") + add_compile_options(-mfp16-format=ieee) + endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") # Raspberry Pi 1, Zero - add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access) + add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7") # Raspberry Pi 2 - add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations) + add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8") # Raspberry Pi 3, 4, Zero 2 (32-bit) - add_compile_options(-mfp16-format=ieee -mno-unaligned-access) + add_compile_options(-mno-unaligned-access) endif() endif() -elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$") +elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" ) message(STATUS "x86 detected") if (MSVC) if (LLAMA_AVX512) @@ -551,27 +573,84 @@ else() message(STATUS "Unknown architecture") endif() +# +# POSIX conformance +# + +# clock_gettime came in POSIX.1b (1993) +# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional +# posix_memalign came in POSIX.1-2001 / SUSv3 +# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985) +add_compile_definitions(_XOPEN_SOURCE=600) + +# Somehow in OpenBSD whenever POSIX conformance is specified +# some string functions rely on locale_t availability, +# which was introduced in POSIX.1-2008, forcing us to go higher +if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD") + remove_definitions(-D_XOPEN_SOURCE=600) + add_compile_definitions(_XOPEN_SOURCE=700) +endif() + +# Data types, macros and functions related to controlling CPU affinity and +# some memory allocation are available on Linux through GNU extensions in libc +if (CMAKE_SYSTEM_NAME MATCHES "Linux") + add_compile_definitions(_GNU_SOURCE) +endif() + +# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1, +# and on macOS its availability depends on enabling Darwin extensions +# similarly on DragonFly, enabling BSD extensions is necessary +if ( + CMAKE_SYSTEM_NAME MATCHES "Darwin" OR + CMAKE_SYSTEM_NAME MATCHES "iOS" OR + CMAKE_SYSTEM_NAME MATCHES "tvOS" OR + CMAKE_SYSTEM_NAME MATCHES "DragonFly" +) + add_compile_definitions(_DARWIN_C_SOURCE) +endif() + +# alloca is a non-standard interface that is not visible on BSDs when +# POSIX conformance is specified, but not all of them provide a clean way +# to enable it in such cases +if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD") + add_compile_definitions(__BSD_VISIBLE) +endif() +if (CMAKE_SYSTEM_NAME MATCHES "NetBSD") + add_compile_definitions(_NETBSD_SOURCE) +endif() +if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD") + add_compile_definitions(_BSD_SOURCE) +endif() + # # libraries # # ggml +if (GGML_USE_CPU_HBM) + add_definitions(-DGGML_USE_CPU_HBM) + find_library(memkind memkind REQUIRED) +endif() + add_library(ggml OBJECT ggml.c ggml.h ggml-alloc.c ggml-alloc.h - ${GGML_SOURCES_CUDA} - ${GGML_SOURCES_OPENCL} - ${GGML_SOURCES_METAL} - ${GGML_SOURCES_MPI} - ${GGML_SOURCES_EXTRA} + ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA} + ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL} + ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL} + ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI} + ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA} ) target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES}) target_compile_features(ggml PUBLIC c_std_11) # don't bump target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS}) +if (GGML_USE_CPU_HBM) + target_link_libraries(ggml PUBLIC memkind) +endif() add_library(ggml_static STATIC $) if (BUILD_SHARED_LIBS) @@ -601,14 +680,53 @@ if (BUILD_SHARED_LIBS) if (LLAMA_METAL) set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal") endif() - install(TARGETS llama LIBRARY) 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_BUILD_NUMBER ${BUILD_NUMBER}) +set(LLAMA_BUILD_COMMIT ${BUILD_COMMIT}) +set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER}) + +configure_package_config_file( + ${CMAKE_CURRENT_SOURCE_DIR}/scripts/LlamaConfig.cmake.in + ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.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}/LlamaConfigVersion.cmake + VERSION ${LLAMA_INSTALL_VERSION} + COMPATIBILITY SameMajorVersion) + +install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake + ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake + DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama) + +set(GGML_PUBLIC_HEADERS "ggml.h" + "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}" + "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}") + +set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}") +install(TARGETS ggml PUBLIC_HEADER) + +set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/llama.h) +install(TARGETS llama LIBRARY PUBLIC_HEADER) + install( FILES convert.py PERMISSIONS diff --git a/Makefile b/Makefile index 86e36ba52..a1438b80d 100644 --- a/Makefile +++ b/Makefile @@ -2,7 +2,7 @@ BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative tests/test-c.o # Binaries only useful for tests -TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1 +TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama # Code coverage output files COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report @@ -49,7 +49,7 @@ test: $(TEST_TARGETS) ./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \ elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \ continue; \ - elif [ "$$test_target" = "tests/test-tokenizer-1" ]; then \ + elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; then \ continue; \ else \ echo "Running test $$test_target..."; \ @@ -106,6 +106,48 @@ MK_CFLAGS = $(OPT) -std=c11 -fPIC MK_CXXFLAGS = $(OPT) -std=c++11 -fPIC MK_LDFLAGS = +# clock_gettime came in POSIX.1b (1993) +# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional +# posix_memalign came in POSIX.1-2001 / SUSv3 +# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985) +MK_CPPFLAGS += -D_XOPEN_SOURCE=600 + +# Somehow in OpenBSD whenever POSIX conformance is specified +# some string functions rely on locale_t availability, +# which was introduced in POSIX.1-2008, forcing us to go higher +ifeq ($(UNAME_S),OpenBSD) + MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700 +endif + +# Data types, macros and functions related to controlling CPU affinity and +# some memory allocation are available on Linux through GNU extensions in libc +ifeq ($(UNAME_S),Linux) + MK_CPPFLAGS += -D_GNU_SOURCE +endif + +# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1, +# and on macOS its availability depends on enabling Darwin extensions +# similarly on DragonFly, enabling BSD extensions is necessary +ifeq ($(UNAME_S),Darwin) + MK_CPPFLAGS += -D_DARWIN_C_SOURCE +endif +ifeq ($(UNAME_S),DragonFly) + MK_CPPFLAGS += -D__BSD_VISIBLE +endif + +# alloca is a non-standard interface that is not visible on BSDs when +# POSIX conformance is specified, but not all of them provide a clean way +# to enable it in such cases +ifeq ($(UNAME_S),FreeBSD) + MK_CPPFLAGS += -D__BSD_VISIBLE +endif +ifeq ($(UNAME_S),NetBSD) + MK_CPPFLAGS += -D_NETBSD_SOURCE +endif +ifeq ($(UNAME_S),OpenBSD) + MK_CPPFLAGS += -D_BSD_SOURCE +endif + ifdef LLAMA_DEBUG MK_CFLAGS += -O0 -g MK_CXXFLAGS += -O0 -g @@ -130,9 +172,16 @@ endif # LLAMA_DISABLE_LOGS # warnings MK_CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ -Wmissing-prototypes -Werror=implicit-int -Wno-unused-function -MK_CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar +MK_CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wmissing-declarations -Wno-unused-function -Wno-multichar -ifeq '' '$(findstring clang++,$(CXX))' +# TODO(cebtenzzre): remove this once PR #2632 gets merged +TTFS_CXXFLAGS = $(CXXFLAGS) -Wno-missing-declarations + +ifneq '' '$(findstring clang,$(shell $(CXX) --version))' + # clang++ only + MK_CXXFLAGS += -Wmissing-prototypes + TTFS_CXXFLAGS += -Wno-missing-prototypes +else # g++ only MK_CXXFLAGS += -Wno-format-truncation -Wno-array-bounds endif @@ -358,7 +407,6 @@ ifdef LLAMA_HIPBLAS HIPFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) HIPFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER) - HIPFLAGS += -DCC_TURING=1000000000 ifdef LLAMA_CUDA_FORCE_DMMV HIPFLAGS += -DGGML_CUDA_FORCE_DMMV endif # LLAMA_CUDA_FORCE_DMMV @@ -451,22 +499,22 @@ main: examples/main/main.cpp build-info.h ggml. @echo '==== Run ./main -h for help. ====' @echo -simple: examples/simple/simple.cpp build-info.h ggml.o llama.o common.o $(OBJS) +simple: examples/simple/simple.cpp ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS) +quantize: examples/quantize/quantize.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o $(OBJS) +quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o common.o $(OBJS) +perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o common.o $(OBJS) +embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS) +save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS) @@ -483,7 +531,7 @@ gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + $(CXX) $(TTFS_CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) @@ -506,7 +554,7 @@ metal: examples/metal/metal.cpp ggml.o $(OBJS) endif build-info.h: $(wildcard .git/index) scripts/build-info.sh - @sh scripts/build-info.sh > $@.tmp + @sh scripts/build-info.sh $(CC) > $@.tmp @if ! cmp -s $@.tmp $@; then \ mv $@.tmp $@; \ else \ @@ -519,7 +567,7 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh tests: $(TEST_TARGETS) -benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS) +benchmark-matmult: examples/benchmark/benchmark-matmult.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) ./$@ @@ -556,7 +604,7 @@ tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h gg tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS) +tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-c.o: tests/test-c.c llama.h diff --git a/Package.swift b/Package.swift index 96f52c4f0..fb95ef7eb 100644 --- a/Package.swift +++ b/Package.swift @@ -2,8 +2,30 @@ import PackageDescription +#if arch(arm) || arch(arm64) +let platforms: [SupportedPlatform]? = [ + .macOS(.v11), + .iOS(.v14), + .watchOS(.v4), + .tvOS(.v14) +] +let exclude: [String] = [] +let additionalSources: [String] = ["ggml-metal.m"] +let additionalSettings: [CSetting] = [ + .unsafeFlags(["-fno-objc-arc"]), + .define("GGML_SWIFT"), + .define("GGML_USE_METAL") +] +#else +let platforms: [SupportedPlatform]? = nil +let exclude: [String] = ["ggml-metal.metal"] +let additionalSources: [String] = [] +let additionalSettings: [CSetting] = [] +#endif + let package = Package( name: "llama", + platforms: platforms, products: [ .library(name: "llama", targets: ["llama"]), ], @@ -11,23 +33,23 @@ let package = Package( .target( name: "llama", path: ".", - exclude: ["ggml-metal.metal"], + exclude: exclude, sources: [ "ggml.c", "llama.cpp", "ggml-alloc.c", - "k_quants.c" - ], + "k_quants.c", + ] + additionalSources, publicHeadersPath: "spm-headers", cSettings: [ .unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_K_QUANTS"), .define("GGML_USE_ACCELERATE") - ], + ] + additionalSettings, linkerSettings: [ .linkedFramework("Accelerate") ] - ), + ) ], cxxLanguageStandard: .cxx11 ) diff --git a/README.md b/README.md index 17a5c2cbf..b3845afd7 100644 --- a/README.md +++ b/README.md @@ -11,21 +11,9 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics -- #### IMPORTANT: Tokenizer fixes and API change (developers and projects using `llama.cpp` built-in tokenization must read): https://github.com/ggerganov/llama.cpp/pull/2810 +- Local Falcon 180B inference on Mac Studio -- GGUFv2 adds support for 64-bit sizes + backwards compatible: https://github.com/ggerganov/llama.cpp/pull/2821 - -- Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717 - -- A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398) - - Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa) - - ### Current `master` should be considered in Beta - expect some issues for a few days! - - ### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up! - - ### Issues with non-GGUF models will be considered with low priority! + https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e ---- @@ -413,7 +401,7 @@ Building the program with BLAS support may lead to some performance improvements - #### hipBLAS - This provide BLAS acceleation on HIP supported GPU like AMD GPU. + This provides BLAS acceleration on HIP-supported AMD GPUs. Make sure to have ROCm installed. You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html). Windows support is coming soon... @@ -737,12 +725,12 @@ python3 convert.py pygmalion-7b/ --outtype q4_1 - Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data. - Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including: - - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGML) - - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGML) - - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGML) - - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML) - - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML) - - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML) + - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF) + - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF) + - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF) + - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF) + - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF) + - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF) ### Verifying the model files @@ -856,8 +844,17 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th #### Images We have two Docker images available for this project: -1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. -2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. +1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`) +2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`) + +Additionally, there the following images, similar to the above: + +- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) +- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) + +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 Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now). #### Usage diff --git a/common/common.cpp b/common/common.cpp index 6e5d5b4d5..02ec0f8d0 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -78,7 +78,7 @@ int32_t get_num_physical_cores() { return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4; } -void process_escapes(std::string& input) { +static void process_escapes(std::string& input) { std::size_t input_len = input.length(); std::size_t output_idx = 0; @@ -374,6 +374,17 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { #else fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n"); fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); +#endif + } else if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") { + if (++i >= argc) { + invalid_param = true; + break; + } +#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD + params.n_gpu_layers_draft = std::stoi(argv[i]); +#else + fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n"); + fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); #endif } else if (arg == "--main-gpu" || arg == "-mg") { if (++i >= argc) { @@ -423,8 +434,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { #endif // GGML_USE_CUBLAS } else if (arg == "--no-mmap") { params.use_mmap = false; - } else if (arg == "--mtest") { - params.mem_test = true; } else if (arg == "--numa") { params.numa = true; } else if (arg == "--export") { @@ -664,6 +673,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD printf(" -ngl N, --n-gpu-layers N\n"); printf(" number of layers to store in VRAM\n"); + printf(" -ngld N, --n-gpu-layers-draft N\n"); + printf(" number of layers to store in VRAM for the draft model\n"); printf(" -ts SPLIT --tensor-split SPLIT\n"); printf(" how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); printf(" -mg i, --main-gpu i the GPU to use for scratch and small tensors\n"); @@ -674,7 +685,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" Not recommended since this is both slower and uses more VRAM.\n"); #endif // GGML_USE_CUBLAS #endif - printf(" --mtest compute maximum memory usage\n"); printf(" --export export the computation graph to 'llama.ggml'\n"); printf(" --verbose-prompt print prompt before generation\n"); fprintf(stderr, " --simple-io use basic IO for better compatibility in subprocesses and limited consoles\n"); @@ -1212,7 +1222,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false"); fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str()); fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str()); - fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false"); fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false"); fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers); fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict); diff --git a/common/common.h b/common/common.h index 012bf5e13..f9dfd4a2c 100644 --- a/common/common.h +++ b/common/common.h @@ -3,6 +3,7 @@ #pragma once #include "llama.h" +#include "build-info.h" #define LOG_NO_FILE_LINE_FUNCTION #include "log.h" @@ -20,8 +21,13 @@ #define DIRECTORY_SEPARATOR '/' #endif // _WIN32 -#define die(msg) do { fputs("error: " msg "\n", stderr); exit(1); } while (0) -#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", ##__VA_ARGS__); exit(1); } while (0) +#define die(msg) do { fputs("error: " msg "\n", stderr); exit(1); } while (0) +#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0) + +#define print_build_info() do { \ + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); \ + fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET); \ +} while(0) // // CLI argument parsing @@ -38,6 +44,7 @@ struct gpt_params { int32_t n_draft = 16; // number of tokens to draft during speculative decoding int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited) int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default) + int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default) int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens. @@ -109,7 +116,6 @@ struct gpt_params { bool perplexity = false; // compute perplexity over the prompt bool use_mmap = true; // use mmap for faster loads bool use_mlock = false; // use mlock to keep model in memory - bool mem_test = false; // compute maximum memory usage bool numa = false; // attempt optimizations that help on some NUMA systems bool export_cgraph = false; // export the computation graph bool verbose_prompt = false; // print prompt tokens before generation diff --git a/common/console.cpp b/common/console.cpp index 23545e5be..f65cbc6ed 100644 --- a/common/console.cpp +++ b/common/console.cpp @@ -158,7 +158,7 @@ namespace console { } } - char32_t getchar32() { + static char32_t getchar32() { #if defined(_WIN32) HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE); wchar_t high_surrogate = 0; @@ -212,7 +212,7 @@ namespace console { #endif } - void pop_cursor() { + static void pop_cursor() { #if defined(_WIN32) if (hConsole != NULL) { CONSOLE_SCREEN_BUFFER_INFO bufferInfo; @@ -233,7 +233,7 @@ namespace console { putc('\b', out); } - int estimateWidth(char32_t codepoint) { + static int estimateWidth(char32_t codepoint) { #if defined(_WIN32) (void)codepoint; return 1; @@ -242,7 +242,7 @@ namespace console { #endif } - int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) { + static int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) { #if defined(_WIN32) CONSOLE_SCREEN_BUFFER_INFO bufferInfo; if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) { @@ -303,7 +303,7 @@ namespace console { #endif } - void replace_last(char ch) { + static void replace_last(char ch) { #if defined(_WIN32) pop_cursor(); put_codepoint(&ch, 1, 1); @@ -312,7 +312,7 @@ namespace console { #endif } - void append_utf8(char32_t ch, std::string & out) { + static void append_utf8(char32_t ch, std::string & out) { if (ch <= 0x7F) { out.push_back(static_cast(ch)); } else if (ch <= 0x7FF) { @@ -333,7 +333,7 @@ namespace console { } // Helper function to remove the last UTF-8 character from a string - void pop_back_utf8_char(std::string & line) { + static void pop_back_utf8_char(std::string & line) { if (line.empty()) { return; } @@ -349,7 +349,7 @@ namespace console { line.erase(pos); } - bool readline_advanced(std::string & line, bool multiline_input) { + static bool readline_advanced(std::string & line, bool multiline_input) { if (out != stdout) { fflush(stdout); } @@ -452,7 +452,7 @@ namespace console { return has_more; } - bool readline_simple(std::string & line, bool multiline_input) { + static bool readline_simple(std::string & line, bool multiline_input) { #if defined(_WIN32) std::wstring wline; if (!std::getline(std::wcin, wline)) { diff --git a/common/grammar-parser.cpp b/common/grammar-parser.cpp index 177d1e3a8..5a545a807 100644 --- a/common/grammar-parser.cpp +++ b/common/grammar-parser.cpp @@ -9,7 +9,7 @@ namespace grammar_parser { // NOTE: assumes valid utf8 (but checks for overrun) // copied from llama.cpp - std::pair decode_utf8(const char * src) { + static std::pair decode_utf8(const char * src) { static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; uint8_t first_byte = static_cast(*src); uint8_t highbits = first_byte >> 4; @@ -24,19 +24,19 @@ namespace grammar_parser { return std::make_pair(value, pos); } - uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) { + static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) { uint32_t next_id = static_cast(state.symbol_ids.size()); auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id)); return result.first->second; } - uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) { + static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) { uint32_t next_id = static_cast(state.symbol_ids.size()); state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id; return next_id; } - void add_rule( + static void add_rule( parse_state & state, uint32_t rule_id, const std::vector & rule) { @@ -46,11 +46,11 @@ namespace grammar_parser { state.rules[rule_id] = rule; } - bool is_word_char(char c) { + static bool is_word_char(char c) { return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9'); } - std::pair parse_hex(const char * src, int size) { + static std::pair parse_hex(const char * src, int size) { const char * pos = src; const char * end = src + size; uint32_t value = 0; @@ -73,7 +73,7 @@ namespace grammar_parser { return std::make_pair(value, pos); } - const char * parse_space(const char * src, bool newline_ok) { + static const char * parse_space(const char * src, bool newline_ok) { const char * pos = src; while (*pos == ' ' || *pos == '\t' || *pos == '#' || (newline_ok && (*pos == '\r' || *pos == '\n'))) { @@ -88,7 +88,7 @@ namespace grammar_parser { return pos; } - const char * parse_name(const char * src) { + static const char * parse_name(const char * src) { const char * pos = src; while (is_word_char(*pos)) { pos++; @@ -99,7 +99,7 @@ namespace grammar_parser { return pos; } - std::pair parse_char(const char * src) { + static std::pair parse_char(const char * src) { if (*src == '\\') { switch (src[1]) { case 'x': return parse_hex(src + 2, 2); @@ -129,7 +129,7 @@ namespace grammar_parser { uint32_t rule_id, bool is_nested); - const char * parse_sequence( + static const char * parse_sequence( parse_state & state, const char * src, const std::string & rule_name, @@ -247,7 +247,7 @@ namespace grammar_parser { return pos; } - const char * parse_rule(parse_state & state, const char * src) { + static const char * parse_rule(parse_state & state, const char * src) { const char * name_end = parse_name(src); const char * pos = parse_space(name_end, false); size_t name_len = name_end - src; @@ -285,7 +285,7 @@ namespace grammar_parser { } } - void print_grammar_char(FILE * file, uint32_t c) { + static void print_grammar_char(FILE * file, uint32_t c) { if (0x20 <= c && c <= 0x7f) { fprintf(file, "%c", static_cast(c)); } else { @@ -294,7 +294,7 @@ namespace grammar_parser { } } - bool is_char_element(llama_grammar_element elem) { + static bool is_char_element(llama_grammar_element elem) { switch (elem.type) { case LLAMA_GRETYPE_CHAR: return true; case LLAMA_GRETYPE_CHAR_NOT: return true; @@ -304,7 +304,7 @@ namespace grammar_parser { } } - void print_rule_binary(FILE * file, const std::vector & rule) { + static void print_rule_binary(FILE * file, const std::vector & rule) { for (auto elem : rule) { switch (elem.type) { case LLAMA_GRETYPE_END: fprintf(file, "END"); break; @@ -334,7 +334,7 @@ namespace grammar_parser { fprintf(file, "\n"); } - void print_rule( + static void print_rule( FILE * file, uint32_t rule_id, const std::vector & rule, diff --git a/convert-baichuan-hf-to-gguf.py b/convert-baichuan-hf-to-gguf.py new file mode 100755 index 000000000..8bd34dc44 --- /dev/null +++ b/convert-baichuan-hf-to-gguf.py @@ -0,0 +1,304 @@ +#!/usr/bin/env python3 +# HF baichuan --> gguf conversion + +from __future__ import annotations + +import argparse +import json +import os +import struct +import sys +from pathlib import Path +from typing import TYPE_CHECKING, Any +import itertools +import gguf +import numpy as np +import torch +from sentencepiece import SentencePieceProcessor # type: ignore[import] + + +if TYPE_CHECKING: + from typing import TypeAlias + +NDArray: TypeAlias = 'np.ndarray[Any, Any]' + +# reverse HF permute back to original pth layout + + +def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray: + if n_kv_head is not None and n_head != n_kv_head: + n_head //= n_kv_head + + return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape)) + +def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray: + r = weights.shape[0] // 3 + return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv)) + +def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray: + r = weights.shape[0] // 3 + return weights[r * n_part : r * n_part + r, ...] + +def count_model_parts(dir_model: str) -> int: + num_parts = 0 + + for filename in os.listdir(dir_model): + if filename.startswith("pytorch_model-"): + num_parts += 1 + + if num_parts > 0: + print("gguf: found " + str(num_parts) + " model parts") + + return num_parts + + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file") + parser.add_argument( + "--vocab-only", action="store_true", + help="extract only the vocab", + ) + parser.add_argument( + "--outfile", type=Path, + help="path to write to; default: based on input", + ) + parser.add_argument( + "model", type=Path, + help="directory containing model file, or model file itself (*.bin)", + ) + parser.add_argument( + "ftype", type=int, choices=[0, 1], default=1, nargs='?', + help="output format - use 0 for float32, 1 for float16", + ) + return parser.parse_args() + +args = parse_args() + +dir_model = args.model +ftype = args.ftype +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file = sys.stderr) + sys.exit(1) + +# possible tensor data types +# ftype == 0 -> float32 +# ftype == 1 -> float16 + +# map from ftype to string +ftype_str = ["f32", "f16"] + +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' + +print("gguf: loading model "+dir_model.name) + +with open(dir_model / "config.json", "r", encoding="utf-8") as f: + hparams = json.load(f) +print("hello print: ",hparams["architectures"][0]) +if hparams["architectures"][0] != "BaichuanForCausalLM": + print("Model architecture not supported: " + hparams["architectures"][0]) + + sys.exit() + +# get number of model parts +num_parts = count_model_parts(dir_model) +print(f"num_parts:{num_parts}\n") +ARCH=gguf.MODEL_ARCH.BAICHUAN +gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) + +print("gguf: get model metadata") + +block_count = hparams["num_hidden_layers"] +head_count = hparams["num_attention_heads"] + +if "num_key_value_heads" in hparams: + head_count_kv = hparams["num_key_value_heads"] +else: + head_count_kv = head_count + +if "_name_or_path" in hparams: + hf_repo = hparams["_name_or_path"] +else: + hf_repo = "" + +if "max_sequence_length" in hparams: + ctx_length = hparams["max_sequence_length"] +elif "max_position_embeddings" in hparams: + ctx_length = hparams["max_position_embeddings"] +elif "model_max_length" in hparams: + ctx_length = hparams["model_max_length"] +else: + print("gguf: can not find ctx length parameter.") + + sys.exit() + + +gguf_writer.add_name(dir_model.name) +gguf_writer.add_source_hf_repo(hf_repo) +gguf_writer.add_tensor_data_layout("Meta AI original pth") +gguf_writer.add_context_length(ctx_length) +gguf_writer.add_embedding_length(hparams["hidden_size"]) +gguf_writer.add_block_count(block_count) +gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) +gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) +gguf_writer.add_head_count(head_count) +gguf_writer.add_head_count_kv(head_count_kv) +gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"]) + +if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]: + if "type" in hparams["rope_scaling"]: + if hparams["rope_scaling"]["type"] == "linear": + gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"]) + + +# TOKENIZATION + +print("gguf: get tokenizer metadata") + +tokens: list[bytes] = [] +scores: list[float] = [] +toktypes: list[int] = [] + +tokenizer_model_file = dir_model / 'tokenizer.model' +if not tokenizer_model_file.is_file(): + print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr) + sys.exit(1) + +# vocab type sentencepiece +print("gguf: get sentencepiece tokenizer vocab, scores and token types") + +tokenizer = SentencePieceProcessor(str(tokenizer_model_file)) + +for i in range(tokenizer.vocab_size()): + text: bytes + score: float + + piece = tokenizer.id_to_piece(i) + text = piece.encode("utf-8") + score = tokenizer.get_score(i) + + toktype = 1 # defualt to normal token type + if tokenizer.is_unknown(i): + toktype = 2 + if tokenizer.is_control(i): + toktype = 3 + + # toktype = 4 is user-defined = tokens from added_tokens.json + + if tokenizer.is_unused(i): + toktype = 5 + if tokenizer.is_byte(i): + toktype = 6 + + tokens.append(text) + scores.append(score) + toktypes.append(toktype) + +added_tokens_file = dir_model / 'added_tokens.json' +if added_tokens_file.is_file(): + with open(added_tokens_file, "r", encoding="utf-8") as f: + addtokens_json = json.load(f) + + print("gguf: get added tokens") + + for key in addtokens_json: + tokens.append( key.encode("utf-8") ) + scores.append(-1000.0) + toktypes.append(4) # user-defined token type + + +gguf_writer.add_tokenizer_model("llama") +gguf_writer.add_token_list(tokens) +gguf_writer.add_token_scores(scores) +gguf_writer.add_token_types(toktypes) + +special_vocab = gguf.SpecialVocab(dir_model) +special_vocab.add_to_gguf(gguf_writer) + +# TENSORS + +tensor_map = gguf.get_tensor_name_map(ARCH,block_count) + +# tensor info +print("gguf: get tensor metadata") + +if num_parts == 0: + part_names = iter(("pytorch_model.bin",)) +else: + part_names = ( + f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) + ) + + +for part_name in part_names: + if args.vocab_only: + break + print("gguf: loading model part '" + part_name + "'") + model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") + + tmp=model_part + for i in range(block_count): + if f"model.layers.{i}.self_attn.W_pack.weight" in model_part: + print(f"Unpacking and permuting layer {i}") + tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count) + tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv) + tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2) + del tmp[f"model.layers.{i}.self_attn.W_pack.weight"] + + for name in model_part.keys(): + data = model_part[name] + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue + + old_dtype = data.dtype + + # convert any unsupported data types to float32 + if data.dtype != torch.float16 and data.dtype != torch.float32: + data = data.to(torch.float32) + + data = data.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: + print("Can not map tensor '" + name + "'") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(name + " -> " + new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + gguf_writer.add_tensor(new_name, data) + + +print("gguf: write header") +gguf_writer.write_header_to_file() +print("gguf: write metadata") +gguf_writer.write_kv_data_to_file() +if not args.vocab_only: + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() + +gguf_writer.close() + +print(f"gguf: model successfully exported to '{fname_out}'") +print("") diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index 6ed2b88c6..88338d823 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -55,10 +55,22 @@ def count_model_parts(dir_model: Path) -> int: def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") - parser.add_argument("ftype", type=int, help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1) + parser.add_argument( + "--vocab-only", action="store_true", + help="extract only the vocab", + ) + parser.add_argument( + "--outfile", type=Path, + help="path to write to; default: based on input", + ) + parser.add_argument( + "model", type=Path, + help="directory containing model file, or model file itself (*.bin)", + ) + parser.add_argument( + "ftype", type=int, choices=[0, 1], default=1, nargs='?', + help="output format - use 0 for float32, 1 for float16", + ) return parser.parse_args() args = parse_args() @@ -137,7 +149,9 @@ with open(tokenizer_json_file, "r", encoding="utf-8") as f: print("gguf: get gpt2 tokenizer vocab") -vocab_size = len(tokenizer_json["model"]["vocab"]) +# The number of tokens in tokenizer.json can differ from the expected vocab size. +# This causes downstream issues with mismatched tensor sizes when running the inference +vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"]) # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py tokenizer = AutoTokenizer.from_pretrained(dir_model) diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py index b9c8b4607..782410e44 100755 --- a/convert-gptneox-hf-to-gguf.py +++ b/convert-gptneox-hf-to-gguf.py @@ -56,10 +56,22 @@ def count_model_parts(dir_model: Path) -> int: def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") - parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) + parser.add_argument( + "--vocab-only", action="store_true", + help="extract only the vocab", + ) + parser.add_argument( + "--outfile", type=Path, + help="path to write to; default: based on input", + ) + parser.add_argument( + "model", type=Path, + help="directory containing model file, or model file itself (*.bin)", + ) + parser.add_argument( + "ftype", type=int, choices=[0, 1], default=1, nargs='?', + help="output format - use 0 for float32, 1 for float16", + ) return parser.parse_args() args = parse_args() diff --git a/convert-starcoder-hf-to-gguf.py b/convert-starcoder-hf-to-gguf.py new file mode 100755 index 000000000..331e84e98 --- /dev/null +++ b/convert-starcoder-hf-to-gguf.py @@ -0,0 +1,248 @@ +#!/usr/bin/env python3 +# HF starcoder --> gguf conversion + +from __future__ import annotations + +import argparse +import json +import os +import struct +import sys +from pathlib import Path +from typing import Any + +import numpy as np +import torch +from transformers import AutoTokenizer # type: ignore[import] + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + + +def bytes_to_unicode(): + # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a significant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8+n) + n += 1 + return dict(zip(bs, (chr(n) for n in cs))) + + +def count_model_parts(dir_model: Path) -> int: + num_parts = 0 + for filename in os.listdir(dir_model): + if filename.startswith("pytorch_model-"): + num_parts += 1 + + if num_parts > 0: + print("gguf: found " + str(num_parts) + " model parts") + return num_parts + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") + parser.add_argument("ftype", type=int, help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1) + return parser.parse_args() + +args = parse_args() + +dir_model = args.model +ftype = args.ftype +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file = sys.stderr) + sys.exit(1) + +# possible tensor data types +# ftype == 0 -> float32 +# ftype == 1 -> float16 + +# map from ftype to string +ftype_str = ["f32", "f16"] + +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' + +print("gguf: loading model "+dir_model.name) + +with open(dir_model / "config.json", "r", encoding="utf-8") as f: + hparams = json.load(f) + +if hparams["architectures"][0] != "GPTBigCodeForCausalLM": + print("Model architecture not supported: " + hparams["architectures"][0]) + + sys.exit(1) + +# get number of model parts +num_parts = count_model_parts(dir_model) + +ARCH=gguf.MODEL_ARCH.STARCODER +gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) + +print("gguf: get model metadata") + +block_count = hparams["n_layer"] + +gguf_writer.add_name("StarCoder") +gguf_writer.add_context_length(hparams["n_positions"]) +gguf_writer.add_embedding_length(hparams["n_embd"]) +gguf_writer.add_feed_forward_length(4 * hparams["n_embd"]) +gguf_writer.add_block_count(block_count) +gguf_writer.add_head_count(hparams["n_head"]) +gguf_writer.add_head_count_kv(1) +gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) +gguf_writer.add_file_type(ftype) + +# TOKENIZATION + +print("gguf: get tokenizer metadata") + +tokens: list[bytearray] = [] +scores: list[float] = [] +toktypes: list[int] = [] + +tokenizer_json_file = dir_model / 'tokenizer.json' +if not tokenizer_json_file.is_file(): + print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr) + sys.exit(1) + +# gpt2 tokenizer +gguf_writer.add_tokenizer_model("gpt2") + +with open(tokenizer_json_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) + +print("gguf: get gpt2 tokenizer vocab") + +# The number of tokens in tokenizer.json can differ from the expected vocab size. +# This causes downstream issues with mismatched tensor sizes when running the inference +vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"]) + +# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py +tokenizer = AutoTokenizer.from_pretrained(dir_model) + +reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} +byte_encoder = bytes_to_unicode() +byte_decoder = {v: k for k, v in byte_encoder.items()} + +for i in range(vocab_size): + if i in reverse_vocab: + try: + text = bytearray([byte_decoder[c] for c in reverse_vocab[i]]) + except KeyError: + text = bytearray() + for c in reverse_vocab[i]: + if ord(c) < 256: # single byte character + text.append(byte_decoder[ord(c)]) + else: # multibyte special token character + text.extend(c.encode('utf-8')) + else: + print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") + pad_token = f"[PAD{i}]".encode("utf8") + text = bytearray(pad_token) + + tokens.append(text) + scores.append(0.0) # dymmy + toktypes.append(gguf.TokenType.NORMAL) # dummy + +gguf_writer.add_token_list(tokens) +gguf_writer.add_token_scores(scores) +gguf_writer.add_token_types(toktypes) + +special_vocab = gguf.SpecialVocab(dir_model, load_merges = True) +special_vocab.add_to_gguf(gguf_writer) + +# TENSORS + +tensor_map = gguf.get_tensor_name_map(ARCH,block_count) + +# params for qkv transform +n_head = hparams["n_head"] +n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1 + +head_dim = hparams["n_embd"] // n_head + +# tensor info +print("gguf: get tensor metadata") + +if num_parts == 0: + part_names = iter(("pytorch_model.bin",)) +else: + part_names = ( + f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) + ) + +for part_name in part_names: + if args.vocab_only: + break + print("gguf: loading model part '" + part_name + "'") + model_part = torch.load(dir_model / part_name, map_location="cpu") + + for name in model_part.keys(): + data = model_part[name] + + old_dtype = data.dtype + + # convert any unsupported data types to float32 + if data.dtype != torch.float16 and data.dtype != torch.float32: + data = data.to(torch.float32) + + data = data.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: + print("Can not map tensor '" + name + "'") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + + gguf_writer.add_tensor(new_name, data) + + +print("gguf: write header") +gguf_writer.write_header_to_file() +print("gguf: write metadata") +gguf_writer.write_kv_data_to_file() +if not args.vocab_only: + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() + +gguf_writer.close() + +print(f"gguf: model successfully exported to '{fname_out}'") +print("") diff --git a/convert.py b/convert.py index 79a7cd52b..4ac5030db 100755 --- a/convert.py +++ b/convert.py @@ -145,7 +145,6 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { class Params: n_vocab: int n_embd: int - n_mult: int n_layer: int n_ctx: int n_ff: int @@ -161,15 +160,6 @@ class Params: # path to the directory containing the model files path_model: Path | None = None - @staticmethod - def find_n_mult(n_ff: int, n_embd: int) -> int: - # hardcoded magic range - for n_mult in range(8192, 1, -1): - calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult - if calc_ff == n_ff: - return n_mult - raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).") - @staticmethod def guessed(model: LazyModel) -> Params: # try transformer naming first @@ -197,7 +187,6 @@ class Params: return Params( n_vocab = n_vocab, n_embd = n_embd, - n_mult = n_mult, n_layer = n_layer, n_ctx = -1, n_ff = n_ff, @@ -225,8 +214,6 @@ class Params: else: f_rope_scale = None - n_mult = Params.find_n_mult(n_ff, n_embd) - if "max_sequence_length" in config: n_ctx = config["max_sequence_length"] elif "max_position_embeddings" in config: @@ -238,7 +225,6 @@ class Params: return Params( n_vocab = n_vocab, n_embd = n_embd, - n_mult = n_mult, n_layer = n_layer, n_ctx = n_ctx, n_ff = n_ff, @@ -250,7 +236,7 @@ class Params: ) # LLaMA v2 70B params.json - # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1 + # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1} @staticmethod def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) @@ -258,7 +244,6 @@ class Params: n_vocab = config["vocab_size"] if "vocab_size" in config else -1 n_embd = config["dim"] n_layer = config["n_layers"] - n_mult = config["multiple_of"] n_ff = -1 n_head = config["n_heads"] n_head_kv = config["n_kv_heads"] if "n_kv_heads" in config else n_head @@ -285,7 +270,6 @@ class Params: return Params( n_vocab = n_vocab, n_embd = n_embd, - n_mult = n_mult, n_layer = n_layer, n_ctx = n_ctx, n_ff = n_ff, diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp index a99ece9a6..ed61125ea 100644 --- a/examples/baby-llama/baby-llama.cpp +++ b/examples/baby-llama/baby-llama.cpp @@ -9,12 +9,12 @@ #endif #ifdef LLAMA_DEFAULT_RMS_EPS -static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS; +constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS; #else -static const float rms_norm_eps = 5e-6f; +constexpr float rms_norm_eps = 5e-6f; #endif -float frand() { +static float frand() { return (float)rand()/(float)RAND_MAX; } @@ -25,19 +25,21 @@ struct random_normal_distribution { float max; }; -void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) { +static void init_random_normal_distribution( + struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max +) { rnd->gen = std::mt19937(seed); rnd->nd = std::normal_distribution{mean, std}; rnd->min = min; rnd->max = max; } -float frand_normal(struct random_normal_distribution * rnd) { +static float frand_normal(struct random_normal_distribution * rnd) { const float r = rnd->nd(rnd->gen); return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r); } -void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { +static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); if (plan.work_size > 0) { @@ -48,13 +50,9 @@ void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, ggml_graph_compute(graph, &plan); } -struct ggml_tensor * randomize_tensor( - struct ggml_tensor * tensor, - int ndims, - const int64_t ne[], - float fmin, - float fmax) { - +static struct ggml_tensor * randomize_tensor( + struct ggml_tensor * tensor, int ndims, const int64_t ne[], float fmin, float fmax +) { switch (ndims) { case 1: for (int i0 = 0; i0 < ne[0]; i0++) { @@ -95,11 +93,9 @@ struct ggml_tensor * randomize_tensor( return tensor; } -struct ggml_tensor * randomize_tensor_normal( - struct ggml_tensor * tensor, - int ndims, - const int64_t ne[], - struct random_normal_distribution * rnd) { +static struct ggml_tensor * randomize_tensor_normal( + struct ggml_tensor * tensor, int ndims, const int64_t ne[], struct random_normal_distribution * rnd +) { float scale = 1.0; // xavier switch (ndims) { case 1: @@ -159,7 +155,7 @@ struct llama_hparams { } }; -uint32_t get_n_ff(const struct llama_hparams* hparams) { +static uint32_t get_n_ff(const struct llama_hparams* hparams) { const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult; return n_ff; } @@ -260,7 +256,7 @@ struct llama_model_lora { std::vector layers; }; -void init_model(struct llama_model * model) { +static void init_model(struct llama_model * model) { const auto & hparams = model->hparams; const uint32_t n_embd = hparams.n_embd; @@ -297,7 +293,7 @@ void init_model(struct llama_model * model) { } -void init_model_lora(struct llama_model_lora * model) { +static void init_model_lora(struct llama_model_lora * model) { const auto & hparams = model->hparams; const uint32_t n_embd = hparams.n_embd; @@ -340,7 +336,7 @@ void init_model_lora(struct llama_model_lora * model) { } } -void set_param_model(struct llama_model * model) { +static void set_param_model(struct llama_model * model) { const auto& hparams = model->hparams; const uint32_t n_layer = hparams.n_layer; @@ -366,7 +362,7 @@ void set_param_model(struct llama_model * model) { } } -void set_param_model_lora(struct llama_model_lora * model) { +static void set_param_model_lora(struct llama_model_lora * model) { const auto& hparams = model->hparams; const uint32_t n_layer = hparams.n_layer; @@ -397,7 +393,7 @@ void set_param_model_lora(struct llama_model_lora * model) { } } -void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) { +static void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) { const auto & hparams = model->hparams; const uint32_t n_layer = hparams.n_layer; @@ -426,7 +422,9 @@ void randomize_model(struct llama_model * model, int seed, float mean, float std } -void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, float std, float min, float max) { +static void randomize_model_lora( + struct llama_model_lora * model, int seed, float mean, float std, float min, float max +) { const auto & hparams = model->hparams; const uint32_t n_layer = hparams.n_layer; @@ -459,7 +457,7 @@ void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, } } -bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) { +static bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) { const auto & hparams = model->hparams; const uint32_t n_ctx = hparams.n_ctx; @@ -495,7 +493,7 @@ bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int return true; } -bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) { +static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) { const auto & hparams = model->hparams; const uint32_t n_ctx = hparams.n_ctx; @@ -531,15 +529,15 @@ bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * return true; } -struct ggml_tensor * forward( - struct llama_model * model, - struct llama_kv_cache * cache, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_past) { - +static struct ggml_tensor * forward( + struct llama_model * model, + struct llama_kv_cache * cache, + struct ggml_context * ctx0, + struct ggml_cgraph * gf, + struct ggml_tensor * tokens_input, + const int n_tokens, + const int n_past +) { const int N = n_tokens; struct llama_kv_cache& kv_self = *cache; @@ -756,25 +754,25 @@ struct ggml_tensor * forward( return inpL; } -void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) { +static void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) { GGML_ASSERT(tensor->n_dims == 1); GGML_ASSERT(tensor->ne[0] == ne0); } -void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) { +static void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) { GGML_ASSERT(tensor->n_dims == 2); GGML_ASSERT(tensor->ne[0] == ne0); GGML_ASSERT(tensor->ne[1] == ne1); } -void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) { +static void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) { GGML_ASSERT(tensor->n_dims == 3); GGML_ASSERT(tensor->ne[0] == ne0); GGML_ASSERT(tensor->ne[1] == ne1); GGML_ASSERT(tensor->ne[2] == ne2); } -void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) { +static void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) { GGML_ASSERT(tensor->n_dims == 4); GGML_ASSERT(tensor->ne[0] == ne0); GGML_ASSERT(tensor->ne[1] == ne1); @@ -782,16 +780,16 @@ void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int6 GGML_ASSERT(tensor->ne[3] == ne3); } -struct ggml_tensor * forward_batch( - struct llama_model * model, - struct llama_kv_cache * cache, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_past, - const int n_batch) { - +static struct ggml_tensor * forward_batch( + struct llama_model * model, + struct llama_kv_cache * cache, + struct ggml_context * ctx0, + struct ggml_cgraph * gf, + struct ggml_tensor * tokens_input, + const int n_tokens, + const int n_past, + const int n_batch +) { const int N = n_tokens; struct llama_kv_cache& kv_self = *cache; @@ -1073,16 +1071,15 @@ struct ggml_tensor * forward_batch( return inpL; } - -struct ggml_tensor * forward_lora( - struct llama_model_lora * model, - struct llama_kv_cache * cache, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_past) { - +static struct ggml_tensor * forward_lora( + struct llama_model_lora * model, + struct llama_kv_cache * cache, + struct ggml_context * ctx0, + struct ggml_cgraph * gf, + struct ggml_tensor * tokens_input, + const int n_tokens, + const int n_past +) { const int N = n_tokens; struct llama_kv_cache& kv_self = *cache; @@ -1328,7 +1325,7 @@ struct ggml_tensor * forward_lora( return inpL; } -void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) { +static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) { assert(logits->n_dims == 2); assert(probs->n_dims == 2); assert(best_samples->n_dims == 1); @@ -1359,7 +1356,10 @@ void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, str } } -void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) { +static void sample_softmax_batch( + struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, + struct ggml_tensor * best_samples +) { GGML_ASSERT(best_samples->n_dims == 2); GGML_ASSERT(logits->n_dims == 3); GGML_ASSERT(probs->n_dims == 3); @@ -1393,7 +1393,7 @@ void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits } } -void print_row(struct ggml_tensor * probs, int i) { +static void print_row(struct ggml_tensor * probs, int i) { for (int k = 0; k < probs->ne[0]; ++k) { float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k); printf(" %.2f", p); @@ -1401,7 +1401,7 @@ void print_row(struct ggml_tensor * probs, int i) { printf("\n"); } -void print_matrix(struct ggml_tensor * probs) { +static void print_matrix(struct ggml_tensor * probs) { assert(probs->n_dims == 2); for (int i = 0; i < probs->ne[1]; ++i) { for (int k = 0; k < probs->ne[0]; ++k) { @@ -1412,7 +1412,7 @@ void print_matrix(struct ggml_tensor * probs) { } } -void print_token(int token, int n_vocab) { +static void print_token(int token, int n_vocab) { for (int k = 0; k < token; ++k) { printf(" "); } @@ -1423,14 +1423,14 @@ void print_token(int token, int n_vocab) { printf("\n"); } -void print_tokens(struct ggml_tensor * tokens, int n_vocab) { +static void print_tokens(struct ggml_tensor * tokens, int n_vocab) { for (int i=0; ine[0]; ++i) { int token = ggml_get_i32_1d(tokens, i); print_token(token, n_vocab); } } -void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) { +static void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) { int n_tokens = tokens_input->ne[0]; int n_vocab = targets->ne[0]; float randomness = 0.0f; @@ -1451,7 +1451,9 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru } } -void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) { +static void get_example_targets_batch( + struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets +) { GGML_ASSERT(tokens_input->n_dims == 2); GGML_ASSERT( targets->n_dims == 3); int n_tokens = tokens_input->ne[0]; @@ -1474,7 +1476,7 @@ void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct } } -void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) { +static void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) { int n_tokens = tokens_input->ne[0]; int n_vocab = targets->ne[0]; for (int i=0; i #include @@ -34,7 +29,8 @@ struct ostream_beam_view { llama_context * ctx; llama_beam_view beam_view; }; -std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) { + +static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) { os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens("; for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) { os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]); @@ -50,7 +46,7 @@ struct beam_search_callback_data { // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same. // For example, eob can be flagged due to maximum token length, stop words, etc. -bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const size_t n_tokens) { +static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) { return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx); } @@ -60,7 +56,7 @@ bool is_at_eob(const beam_search_callback_data & callback_data, const llama_toke // * When all beams converge to a common prefix, they are made available in beams_state.beams[0]. // This is also called when the stop condition is met. // Collect tokens into std::vector response which is pointed to by callback_data. -void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) { +static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) { auto& callback_data = *static_cast(callback_data_ptr); // Mark beams as EOS as needed. for (size_t i = 0 ; i < beams_state.n_beams ; ++i) { diff --git a/examples/benchmark/CMakeLists.txt b/examples/benchmark/CMakeLists.txt index 3f3415350..14916d831 100644 --- a/examples/benchmark/CMakeLists.txt +++ b/examples/benchmark/CMakeLists.txt @@ -1,7 +1,8 @@ set(TARGET benchmark) add_executable(${TARGET} benchmark-matmult.cpp) install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_include_directories(${TARGET} PRIVATE ../../common) target_compile_features(${TARGET} PRIVATE cxx_std_11) if(TARGET BUILD_INFO) add_dependencies(${TARGET} BUILD_INFO) diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index f7215f43b..561309acb 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -1,5 +1,5 @@ +#include "common.h" #include "ggml.h" -#include "build-info.h" #include #include @@ -99,7 +99,7 @@ int main(int argc, char ** argv) { exit(1); } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); printf("Starting Test\n"); // create the ggml context diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 293b455d0..c291f0adf 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -115,7 +115,7 @@ struct TransformerWeights { } }; -void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) { +static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) { // we calloc instead of malloc to keep valgrind happy w->token_embedding_table = new float[p->vocab_size * p->dim](); printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); @@ -158,7 +158,7 @@ void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) { } } -int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) { +static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) { if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast(p->vocab_size * p->dim)) return 1; if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast(p->n_layers * p->dim)) return 1; if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast(p->n_layers * p->dim * p->dim)) return 1; @@ -189,7 +189,7 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shar return 0; } -void print_sample_weights(TransformerWeights *w){ +static void print_sample_weights(TransformerWeights *w){ printf("----- Quick print of first of the weight vales of all the variables\n"); printf("%f\n", w->token_embedding_table[0]); printf("%f\n", w->rms_att_weight[0]); @@ -324,7 +324,7 @@ struct train_params { int mem_compute1_gb; }; -void print_params(struct my_llama_hparams * params) { +static void print_params(struct my_llama_hparams * params) { printf("%s: n_vocab: %d\n", __func__, params->n_vocab); printf("%s: n_ctx: %d\n", __func__, params->n_ctx); printf("%s: n_embd: %d\n", __func__, params->n_embd); @@ -335,7 +335,7 @@ void print_params(struct my_llama_hparams * params) { printf("%s: n_rot: %d\n", __func__, params->n_rot); } -void init_model(struct my_llama_model * model) { +static void init_model(struct my_llama_model * model) { const auto & hparams = model->hparams; const uint32_t n_embd = hparams.n_embd; @@ -408,17 +408,17 @@ void init_model(struct my_llama_model * model) { } } -float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) { +static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) { float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]); return *ptr; } -int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) { +static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) { int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]); return *ptr; } -void print_row(struct ggml_tensor * probs, int i) { +static void print_row(struct ggml_tensor * probs, int i) { for (int k = 0; k < probs->ne[0]; ++k) { float p = get_f32_2d(probs, k, i); printf(" %f", p); @@ -426,7 +426,7 @@ void print_row(struct ggml_tensor * probs, int i) { printf("\n"); } -void print_matrix(struct ggml_tensor * probs) { +static void print_matrix(struct ggml_tensor * probs) { assert(probs->n_dims == 2); for (int i = 0; i < probs->ne[1]; ++i) { for (int k = 0; k < probs->ne[0]; ++k) { @@ -531,7 +531,7 @@ struct llama_file { } }; -bool is_ggml_file(const char *filename) { +static bool is_ggml_file(const char * filename) { llama_file file(filename, "rb"); if (file.size < 4) { return false; @@ -540,7 +540,7 @@ bool is_ggml_file(const char *filename) { return magic == GGUF_MAGIC; } -static std::string llama_escape_whitespaces(const std::string& text) { +static std::string llama_escape_whitespaces(const std::string & text) { std::ostringstream out; for (char c : text) { if (c == ' ') out << "\xe2\x96\x81"; @@ -549,7 +549,7 @@ static std::string llama_escape_whitespaces(const std::string& text) { return out.str(); } -void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) { +static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) { if (is_ggml_file(filename)) { struct ggml_context * ctx_data = NULL; @@ -637,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) } } -void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) { +static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) { int ct; switch (gg_weights->n_dims){ case 1: @@ -673,7 +673,9 @@ void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * kar } } -void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) { +static void save_as_llama_model( + struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename +) { // convert AK weights into GG weights one by one. // w->token_embedding_table -> model->tok_embeddings // float* -> struct ggml_tensor @@ -785,7 +787,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod gguf_free(ctx); } -struct train_params get_default_train_params() { +static struct train_params get_default_train_params() { struct train_params params; params.fn_vocab_model = "models/7B/ggml-model-f16.gguf"; params.fn_llama2c_output_model = "ak_llama_model.bin"; @@ -835,7 +837,7 @@ struct train_params get_default_train_params() { return params; } -void print_usage(int /*argc*/, char ** argv, const struct train_params * params) { +static void print_usage(int /*argc*/, char ** argv, const struct train_params * params) { fprintf(stderr, "usage: %s [options]\n", argv[0]); fprintf(stderr, "\n"); fprintf(stderr, "options:\n"); @@ -846,7 +848,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params) fprintf(stderr, "\n"); } -bool params_parse(int argc, char ** argv, struct train_params * params) { +static bool params_parse(int argc, char ** argv, struct train_params * params) { bool invalid_param = false; bool reqd_param_found = false; std::string arg; @@ -901,7 +903,7 @@ bool params_parse(int argc, char ** argv, struct train_params * params) { return true; } -std::string basename(const std::string &path) { +static std::string basename(const std::string &path) { size_t pos = path.find_last_of("/\\"); if (pos == std::string::npos) { return path; diff --git a/examples/embd-input/embd-input-lib.cpp b/examples/embd-input/embd-input-lib.cpp index 87aac3479..fc6e44eb2 100644 --- a/examples/embd-input/embd-input-lib.cpp +++ b/examples/embd-input/embd-input-lib.cpp @@ -1,8 +1,4 @@ -// Defines sigaction on msys: -#ifndef _GNU_SOURCE -#define _GNU_SOURCE -#endif - +#include "common.h" #include "embd-input.h" #include @@ -27,7 +23,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) { return nullptr; } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = uint32_t(time(NULL)); diff --git a/examples/embd-input/embd-input.h b/examples/embd-input/embd-input.h index efb5ba5e2..eff5e3b84 100644 --- a/examples/embd-input/embd-input.h +++ b/examples/embd-input/embd-input.h @@ -3,7 +3,6 @@ #include "common.h" #include "llama.h" -#include "build-info.h" extern "C" { diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 49ab3e063..0788f362c 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -1,6 +1,5 @@ #include "common.h" #include "llama.h" -#include "build-info.h" #include @@ -17,12 +16,7 @@ int main(int argc, char ** argv) { params.embedding = true; - if (params.n_ctx > 2048) { - fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);" - "expect poor results\n", __func__, params.n_ctx); - } - - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -47,6 +41,12 @@ int main(int argc, char ** argv) { return 1; } + const int n_ctx_train = llama_n_ctx_train(ctx); + if (params.n_ctx > n_ctx_train) { + fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n", + __func__, n_ctx_train, params.n_ctx); + } + // print system information { fprintf(stderr, "\n"); diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp index a34010f10..9ab63a293 100644 --- a/examples/gguf/gguf.cpp +++ b/examples/gguf/gguf.cpp @@ -13,14 +13,14 @@ #define MIN(a, b) ((a) < (b) ? (a) : (b)) #define MAX(a, b) ((a) > (b) ? (a) : (b)) -template +template static std::string to_string(const T & val) { std::stringstream ss; ss << val; return ss.str(); } -bool gguf_ex_write(const std::string & fname) { +static bool gguf_ex_write(const std::string & fname) { struct gguf_context * ctx = gguf_init_empty(); gguf_set_val_u8 (ctx, "some.parameter.uint8", 0x12); @@ -85,7 +85,7 @@ bool gguf_ex_write(const std::string & fname) { } // just read tensor info -bool gguf_ex_read_0(const std::string & fname) { +static bool gguf_ex_read_0(const std::string & fname) { struct gguf_init_params params = { /*.no_alloc = */ false, /*.ctx = */ NULL, @@ -143,7 +143,7 @@ bool gguf_ex_read_0(const std::string & fname) { } // read and create ggml_context containing the tensors and their data -bool gguf_ex_read_1(const std::string & fname) { +static bool gguf_ex_read_1(const std::string & fname) { struct ggml_context * ctx_data = NULL; struct gguf_init_params params = { diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index dedaa34fd..34ddfde39 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -74,14 +74,6 @@ static T stdev(const std::vector & v) { return stdev; } -static bool ggml_cpu_has_metal() { -#if defined(GGML_USE_METAL) - return true; -#else - return false; -#endif -} - static std::string get_cpu_info() { std::string id; #ifdef __linux__ diff --git a/examples/main-cmake-pkg/.gitignore b/examples/main-cmake-pkg/.gitignore new file mode 100644 index 000000000..e32c11c7f --- /dev/null +++ b/examples/main-cmake-pkg/.gitignore @@ -0,0 +1,51 @@ +# Prerequisites +*.d + +# Compiled Object files +*.slo +*.lo +*.o +*.obj + +# Precompiled Headers +*.gch +*.pch + +# Compiled Dynamic libraries +*.so +*.dylib +*.dll + +# Fortran module files +*.mod +*.smod + +# Compiled Static libraries +*.lai +*.la +*.a +*.lib + +# Executables +*.exe +*.out +*.app + +*.gguf + +*.log +.DS_Store +.build/ +.cache/ +.direnv/ +.envrc +.swiftpm +.venv +.clang-tidy +.vs/ +.vscode/ + +build*/ +out/ +tmp/ + diff --git a/examples/main-cmake-pkg/CMakeLists.txt b/examples/main-cmake-pkg/CMakeLists.txt new file mode 100644 index 000000000..473738719 --- /dev/null +++ b/examples/main-cmake-pkg/CMakeLists.txt @@ -0,0 +1,36 @@ +cmake_minimum_required(VERSION 3.12) +project("main-cmake-pkg" C CXX) +set(TARGET main-cmake-pkg) + +find_package(Llama 0.0.1 REQUIRED) + +# Bake common functionality in with target. Because applications +# using the relocatable Llama package should be outside of the +# source tree, main-cmake-pkg pretends the dependencies are built-in. + +set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common") +add_library(common OBJECT + ${_common_path}/common.h + ${_common_path}/common.cpp + ${_common_path}/console.h + ${_common_path}/console.cpp + ${_common_path}/grammar-parser.h + ${_common_path}/grammar-parser.cpp + ) + +# WARNING: because build-info.h is auto-generated, it will only +# be available after the user has built the llama.cpp sources. +# +configure_file(${_common_path}/../build-info.h + ${CMAKE_CURRENT_BINARY_DIR}/build-info.h + COPYONLY) + +target_include_directories(common PUBLIC ${LLAMA_INCLUDE_DIR} + ${CMAKE_CURRENT_BINARY_DIR}) + +add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp) +target_include_directories(${TARGET} PRIVATE ${_common_path}) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) + diff --git a/examples/main-cmake-pkg/README.md b/examples/main-cmake-pkg/README.md new file mode 100644 index 000000000..6d665f28f --- /dev/null +++ b/examples/main-cmake-pkg/README.md @@ -0,0 +1,37 @@ +# llama.cpp/example/main-cmake-pkg + +This program builds the [main](../main) application using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree. + +## Building + +Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions. + +### Considerations + +When hardware acceleration libraries are used (e.g. CUBlas, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_. + +### Build llama.cpp and install to C:\LlamaCPP directory + +In this case, CLBlast was already installed so the CMake package is referenced in `CMAKE_PREFIX_PATH`. + +```cmd +git clone https://github.com/ggerganov/llama.cpp +cd llama.cpp +mkdir build +cd build +cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64 +cmake --build . --config Release +cmake --install . --prefix C:/LlamaCPP +``` + +### Build main-cmake-pkg + + +```cmd +cd ..\examples\main-cmake-pkg +mkdir build +cd build +cmake .. -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64 +cmake --build . --config Release +cmake --install . --prefix C:/MyLlamaApp +``` diff --git a/examples/main/README.md b/examples/main/README.md index 2773fe976..26e1e28dd 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -144,7 +144,7 @@ The `--ctx-size` option allows you to set the size of the prompt context used by Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8. -- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model. +- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model. ### Keep Prompt @@ -274,7 +274,7 @@ These options help improve the performance and memory usage of the LLaMA models. ### NUMA support -- `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop\_caches' as root. +- `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root. ### Memory Float 32 @@ -302,7 +302,6 @@ These options provide extra functionality and customization when running the LLa - `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated. - `--verbose-prompt`: Print the prompt before generating text. -- `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly. - `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance. - `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS. - `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS. diff --git a/examples/main/main.cpp b/examples/main/main.cpp index c9ca7719b..d78112260 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -1,8 +1,3 @@ -// Defines sigaction on msys: -#ifndef _GNU_SOURCE -#define _GNU_SOURCE -#endif - #include "common.h" #include "console.h" @@ -46,7 +41,8 @@ static std::ostringstream * g_output_ss; static std::vector * g_output_tokens; static bool is_interacting = false; -void write_logfile( + +static void write_logfile( const llama_context * ctx, const gpt_params & params, const llama_model * model, const std::vector & input_tokens, const std::string & output, const std::vector & output_tokens @@ -91,7 +87,7 @@ void write_logfile( } #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) -void sigint_handler(int signo) { +static void sigint_handler(int signo) { if (signo == SIGINT) { if (!is_interacting) { is_interacting = true; @@ -153,6 +149,7 @@ int main(int argc, char ** argv) { } LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -187,8 +184,10 @@ int main(int argc, char ** argv) { return 1; } - if (params.n_ctx > llama_n_ctx(ctx)) { - LOG_TEE("%s: warning: base model only supports context sizes no greater than %d tokens (%d specified)\n", __func__, llama_n_ctx(ctx), params.n_ctx); + const int n_ctx_train = llama_n_ctx_train(ctx); + if (params.n_ctx > n_ctx_train) { + LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n", + __func__, n_ctx_train, params.n_ctx); } else if (params.n_ctx < 8) { LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__); params.n_ctx = 8; @@ -201,23 +200,6 @@ int main(int argc, char ** argv) { params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info()); } - // determine the maximum memory usage needed to do inference for the given n_batch and n_ctx parameters - // uncomment the "used_mem" line in llama.cpp to see the results - if (params.mem_test) { - { - LOG_TEE("%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx); - - const std::vector tmp(params.n_batch, llama_token_bos(ctx)); - llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads); - } - - llama_print_timings(ctx); - llama_free(ctx); - llama_free_model(model); - - return 0; - } - // export the cgraph and exit if (params.export_cgraph) { llama_eval_export(ctx, "llama.ggml"); diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 1b760683b..4958cdfb9 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1,6 +1,5 @@ #include "common.h" #include "llama.h" -#include "build-info.h" #include #include @@ -28,9 +27,10 @@ struct results_log_softmax { float prob; }; -void write_logfile(const llama_context * ctx, const gpt_params & params, - const llama_model * model, const struct results_perplexity & results) { - +static void write_logfile( + const llama_context * ctx, const gpt_params & params, const llama_model * model, + const struct results_perplexity & results +) { if (params.logdir.empty()) { return; } @@ -76,7 +76,7 @@ void write_logfile(const llama_context * ctx, const gpt_params & params, fclose(logfile); } -std::vector softmax(const std::vector& logits) { +static std::vector softmax(const std::vector& logits) { std::vector probs(logits.size()); float max_logit = logits[0]; for (float v : logits) max_logit = std::max(max_logit, v); @@ -92,7 +92,7 @@ std::vector softmax(const std::vector& logits) { return probs; } -results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) { +static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) { float max_logit = logits[0]; for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]); double sum_exp = 0.0; @@ -100,9 +100,10 @@ results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) { return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp}; } -void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, - double & nll, double & nll2, float * logit_history, float * prob_history) { - +static void process_logits( + int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, + double & nll, double & nll2, float * logit_history, float * prob_history +) { std::mutex mutex; int counter = 0; auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () { @@ -130,7 +131,7 @@ void process_logits(int n_vocab, const float * logits, const int * tokens, int n } -results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) { +static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) { // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` // Output: `perplexity: 13.5106 [114/114]` @@ -260,8 +261,7 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) return {tokens, std::exp(nll / count), logit_history, prob_history}; } -results_perplexity perplexity(llama_context * ctx, const gpt_params & params) { - +static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) { if (params.ppl_stride > 0) { return perplexity_v2(ctx, params); } @@ -400,8 +400,9 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) { return {tokens, ppl, logit_history, prob_history}; } -std::vector hellaswag_evaluate_tokens(llama_context * ctx, const std::vector& tokens, int n_past, int n_batch, - int n_vocab, int n_thread) { +static std::vector hellaswag_evaluate_tokens( + llama_context * ctx, const std::vector& tokens, int n_past, int n_batch, int n_vocab, int n_thread +) { std::vector result; result.reserve(tokens.size() * n_vocab); size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch; @@ -421,7 +422,7 @@ std::vector hellaswag_evaluate_tokens(llama_context * ctx, const std::vec return result; } -void hellaswag_score(llama_context * ctx, const gpt_params & params) { +static void hellaswag_score(llama_context * ctx, const gpt_params & params) { // Calculates hellaswag score (acc_norm) from prompt // // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl @@ -668,7 +669,7 @@ int main(int argc, char ** argv) { params.n_ctx += params.ppl_stride/2; } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -693,9 +694,10 @@ int main(int argc, char ** argv) { return 1; } - if (params.n_ctx > llama_n_ctx(ctx)) { - fprintf(stderr, "%s: warning: model might not support context sizes greater than %d tokens (%d specified);" - "expect poor results\n", __func__, llama_n_ctx(ctx), params.n_ctx); + const int n_ctx_train = llama_n_ctx_train(ctx); + if (params.n_ctx > n_ctx_train) { + fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n", + __func__, n_ctx_train, params.n_ctx); } // print system information diff --git a/examples/quantize-stats/CMakeLists.txt b/examples/quantize-stats/CMakeLists.txt index c5c394058..db182e263 100644 --- a/examples/quantize-stats/CMakeLists.txt +++ b/examples/quantize-stats/CMakeLists.txt @@ -2,4 +2,5 @@ set(TARGET quantize-stats) add_executable(${TARGET} quantize-stats.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_include_directories(${TARGET} PRIVATE ../../common) target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index 6ce03ba7b..9f930dede 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -1,7 +1,6 @@ -#include "ggml.h" -#include "build-info.h" - #define LLAMA_API_INTERNAL +#include "common.h" +#include "ggml.h" #include "llama.h" #include @@ -34,8 +33,8 @@ struct quantize_stats_params { std::vector include_types; }; -const size_t HISTOGRAM_BUCKETS = 150; -const double HISTOGRAM_RANGE = 0.03; +constexpr size_t HISTOGRAM_BUCKETS = 150; +constexpr double HISTOGRAM_RANGE = 0.03; struct error_stats { size_t num_samples; @@ -44,8 +43,7 @@ struct error_stats { uint64_t error_histogram[HISTOGRAM_BUCKETS]; }; - -void quantize_stats_print_usage(int /*argc*/, char ** argv) { +static void quantize_stats_print_usage(int /*argc*/, char ** argv) { quantize_stats_params params; fprintf(stderr, "usage: %s [options]\n", argv[0]); fprintf(stderr, "\n"); @@ -71,7 +69,7 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) { } // Check if a layer is included/excluded by command line -bool layer_included(const quantize_stats_params & params, const std::string & layer) { +static bool layer_included(const quantize_stats_params & params, const std::string & layer) { for (const auto& excluded : params.exclude_layers) { if (std::regex_search(layer, std::regex(excluded))) { return false; @@ -86,7 +84,7 @@ bool layer_included(const quantize_stats_params & params, const std::string & la } // Update error statistics given vectors with the before/after result of quantization -void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) { +static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) { for (int64_t i = 0; i < nelements; i++) { double diff = input[i] - output[i]; stats.total_error += diff * diff; @@ -96,14 +94,14 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou stats.num_samples += nelements; } -void combine_error_stats(error_stats & into, const error_stats & from) { +static void combine_error_stats(error_stats & into, const error_stats & from) { into.num_samples += from.num_samples; into.total_error += from.total_error; if (from.max_error > into.max_error) into.max_error = from.max_error; for (size_t i=0; inb[3] == tensor->nb[2]*tensor->ne[2]; } -void test_roundtrip_on_chunk( - const ggml_tensor * layer, - int64_t offset, - int64_t chunk_size, - const ggml_type_traits_t & qfns, - bool use_reference, - float * input_scratch, - char * quantized_scratch, - float * output_scratch, - error_stats & stats) { - +static void test_roundtrip_on_chunk( + const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference, + float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats +) { if (layer->type == GGML_TYPE_F16) { for (int i = 0; i < chunk_size; i++) { input_scratch[i] = ggml_get_f32_1d(layer, i + offset); @@ -174,18 +165,11 @@ void test_roundtrip_on_chunk( // Run quantization function for a single layer and update error stats -void test_roundtrip_on_layer( - std::string & name, - bool print_layer_stats, - const ggml_type_traits_t & qfns, - bool use_reference, - const ggml_tensor * layer, - std::vector & input_scratch, - std::vector & quantized_scratch, - std::vector & output_scratch, - error_stats & total_error, - int max_thread = 0) { - +static void test_roundtrip_on_layer( + std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference, + const ggml_tensor * layer, std::vector & input_scratch, std::vector & quantized_scratch, + std::vector & output_scratch, error_stats & total_error, int max_thread = 0 +) { assert(tensor_is_contiguous(layer)); error_stats layer_error {}; uint64_t nelements = ggml_nelements(layer); @@ -314,7 +298,7 @@ int main(int argc, char ** argv) { return 1; } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); // load the model fprintf(stderr, "Loading model\n"); diff --git a/examples/quantize/CMakeLists.txt b/examples/quantize/CMakeLists.txt index 47d0be72e..4a8eed544 100644 --- a/examples/quantize/CMakeLists.txt +++ b/examples/quantize/CMakeLists.txt @@ -2,6 +2,7 @@ set(TARGET quantize) add_executable(${TARGET} quantize.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_include_directories(${TARGET} PRIVATE ../../common) target_compile_features(${TARGET} PRIVATE cxx_std_11) if(TARGET BUILD_INFO) add_dependencies(${TARGET} BUILD_INFO) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 1bf182482..acb79e690 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -1,5 +1,4 @@ -#include "build-info.h" - +#include "common.h" #include "llama.h" #include @@ -40,7 +39,7 @@ static const std::vector QUANT_OPTIONS = { }; -bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) { +static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) { std::string ftype_str; for (auto ch : ftype_str_in) { @@ -72,7 +71,7 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std: // usage: // ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] // -void usage(const char * executable) { +static void usage(const char * executable) { printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); 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"); @@ -161,7 +160,7 @@ int main(int argc, char ** argv) { } } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str()); if (params.nthread > 0) { diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index 14e9501ca..eac307904 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -1,6 +1,5 @@ #include "common.h" #include "llama.h" -#include "build-info.h" #include #include @@ -17,7 +16,7 @@ int main(int argc, char ** argv) { return 1; } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + print_build_info(); if (params.n_predict < 0) { params.n_predict = 16; diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 3f3c64650..1bb8e92c0 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1083,8 +1083,9 @@ static json format_final_response(llama_server_context &llama, const std::string return res; } -static json format_partial_response(llama_server_context &llama, const std::string &content, const std::vector &probs) -{ +static json format_partial_response( + llama_server_context &llama, const std::string &content, const std::vector &probs +) { json res = json{ {"content", content}, {"stop", false}, @@ -1215,7 +1216,7 @@ static void log_server_request(const Request &req, const Response &res) }); } -bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) { +static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) { return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx); } @@ -1225,7 +1226,7 @@ bool is_at_eob(llama_server_context & server_context, const llama_token * tokens // * When all beams converge to a common prefix, they are made available in beams_state.beams[0]. // This is also called when the stop condition is met. // Collect tokens into std::vector response which is pointed to by callback_data. -void beam_search_callback(void * callback_data, llama_beams_state beams_state) { +static void beam_search_callback(void *callback_data, llama_beams_state beams_state) { auto & llama = *static_cast(callback_data); // Mark beams as EOS as needed. for (size_t i = 0 ; i < beams_state.n_beams ; ++i) { @@ -1258,7 +1259,8 @@ struct token_translator { std::string operator()(const completion_token_output & cto) const { return (*this)(cto.tok); } }; -void append_to_generated_text_from_generated_token_probs(llama_server_context & llama) { +static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama) +{ auto & gtps = llama.generated_token_probs; auto translator = token_translator{llama.ctx}; auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); }; diff --git a/examples/simple/CMakeLists.txt b/examples/simple/CMakeLists.txt index 0ac9cb03a..7da5ff6f3 100644 --- a/examples/simple/CMakeLists.txt +++ b/examples/simple/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} simple.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp index 4ee85faca..440d22ecf 100644 --- a/examples/simple/simple.cpp +++ b/examples/simple/simple.cpp @@ -1,9 +1,3 @@ -#ifndef _GNU_SOURCE -#define _GNU_SOURCE -#endif - -#include "build-info.h" - #include "common.h" #include "llama.h" diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index c6211ac79..aa904183f 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -1,7 +1,3 @@ -#ifndef _GNU_SOURCE -#define _GNU_SOURCE -#endif - #include "build-info.h" #include "common.h" @@ -46,6 +42,7 @@ int main(int argc, char ** argv) { // load the draft model params.model = params.model_draft; + params.n_gpu_layers = params.n_gpu_layers_draft; std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params); // tokenize the prompt @@ -85,7 +82,7 @@ int main(int argc, char ** argv) { //GGML_ASSERT(n_vocab == llama_n_vocab(ctx_dft)); // how many tokens to draft each time - const int n_draft = params.n_draft; + int n_draft = params.n_draft; int n_predict = 0; int n_drafted = 0; @@ -134,6 +131,7 @@ int main(int argc, char ** argv) { LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted)); int i_dft = 0; + while (true) { // sample from the target model const llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft); @@ -177,6 +175,27 @@ int main(int argc, char ** argv) { llama_eval(ctx_dft, &id, 1, n_past_dft, params.n_threads); ++n_past_dft; + // heuristic for n_draft + { + const int n_draft_cur = (int) drafted.size(); + const bool all_accepted = i_dft == n_draft_cur; + + LOG("n_draft = %d\n", n_draft); + LOG("n_draft_cur = %d\n", n_draft_cur); + LOG("i_dft = %d\n", i_dft); + LOG("all_accepted = %d\n", all_accepted); + + if (all_accepted && n_draft == n_draft_cur) { + LOG(" - max drafted tokens accepted - n_draft += 8\n"); + n_draft = std::min(30, n_draft + 8); + } else if (all_accepted) { + LOG(" - partially drafted tokens accepted - no change\n"); + } else { + LOG(" - drafted token rejected - n_draft -= 1\n"); + n_draft = std::max(2, n_draft - 1); + } + } + drafted.clear(); drafted.push_back(id); diff --git a/flake.nix b/flake.nix index 02095411e..b0fb8642c 100644 --- a/flake.nix +++ b/flake.nix @@ -34,7 +34,7 @@ with pkgs; [ openblas ] ); pkgs = import nixpkgs { inherit system; }; - nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ]; + nativeBuildInputs = with pkgs; [ cmake ninja pkg-config ]; llama-python = pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]); postPatch = '' @@ -45,6 +45,8 @@ postInstall = '' mv $out/bin/main $out/bin/llama mv $out/bin/server $out/bin/llama-server + mkdir -p $out/include + cp ${src}/llama.h $out/include/ ''; cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ]; in @@ -93,6 +95,10 @@ type = "app"; program = "${self.packages.${system}.default}/bin/quantize"; }; + apps.train-text-from-scratch = { + type = "app"; + program = "${self.packages.${system}.default}/bin/train-text-from-scratch"; + }; apps.default = self.apps.${system}.llama; devShells.default = pkgs.mkShell { buildInputs = [ llama-python ]; diff --git a/ggml-alloc.c b/ggml-alloc.c index a896601d1..304964be4 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -1,8 +1,3 @@ -// defines MAP_ANONYMOUS -#ifndef _GNU_SOURCE -#define _GNU_SOURCE -#endif - #include "ggml-alloc.h" #include "ggml.h" #include @@ -136,6 +131,10 @@ static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_ten return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size; } +static bool ggml_is_view(struct ggml_tensor * t) { + return t->view_src != NULL; +} + void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { #ifdef GGML_ALLOCATOR_DEBUG GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources @@ -316,7 +315,11 @@ static void * alloc_vmem(size_t size) { #if defined(_WIN32) return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS); #elif defined(_POSIX_MAPPED_FILES) - return mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0); + void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0); + if (ptr == MAP_FAILED) { + return NULL; + } + return ptr; #else // use a fixed address for other platforms uintptr_t base_addr = (uintptr_t)-size - 0x100; @@ -339,8 +342,8 @@ static void free_vmem(void * base_addr, size_t size) { // allocate uncommitted virtual memory to measure the size of the graph static void alloc_measure_vmem(void ** base_addr, size_t * size) { - // 1TB for 64-bit, 1GB for 32-bit - *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<40; + // 128GB for 64-bit, 1GB for 32-bit + *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37; do { *base_addr = alloc_vmem(*size); if (*base_addr != NULL) { @@ -400,10 +403,6 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) { //////////// compute graph allocator -static bool ggml_is_view(struct ggml_tensor * t) { - return t->view_src != NULL; -} - static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) { if (a->type != b->type) { return false; diff --git a/ggml-cuda.cu b/ggml-cuda.cu index d2dbf824e..fe7332b2a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -13,7 +13,7 @@ #ifdef __HIP_PLATFORM_AMD__ // for rocblas_initialize() #include "rocblas/rocblas.h" -#endif +#endif // __HIP_PLATFORM_AMD__ #define CUBLAS_COMPUTE_32F HIPBLAS_R_32F #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT @@ -68,19 +68,29 @@ #include #include #include -#endif +#endif // defined(GGML_USE_HIPBLAS) #include "ggml-cuda.h" #include "ggml.h" -#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products -#ifndef CC_TURING -#define CC_TURING 700 -#endif +#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products +#define CC_TURING 700 +#define CC_OFFSET_AMD 1000000 +#define CC_RDNA2 CC_OFFSET_AMD + 1030 #if defined(GGML_USE_HIPBLAS) #define __CUDA_ARCH__ 1300 +#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \ + defined(__gfx1150__) || defined(__gfx1151__) +#define RDNA3 +#endif + +#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \ + defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__) +#define RDNA2 +#endif + #ifndef __has_builtin #define __has_builtin(x) 0 #endif @@ -132,7 +142,7 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { #endif return c; } -#endif +#endif // defined(GGML_USE_HIPBLAS) #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -144,8 +154,11 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); do { \ cudaError_t err_ = (err); \ if (err_ != cudaSuccess) { \ - fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \ + int id; \ + cudaGetDevice(&id); \ + fprintf(stderr, "\nCUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \ cudaGetErrorString(err_)); \ + fprintf(stderr, "current device: %d\n", id); \ exit(1); \ } \ } while (0) @@ -155,8 +168,11 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); do { \ cublasStatus_t err_ = (err); \ if (err_ != CUBLAS_STATUS_SUCCESS) { \ + int id; \ + cudaGetDevice(&id); \ fprintf(stderr, "\ncuBLAS error %d at %s:%d: %s\n", \ err_, __FILE__, __LINE__, cublasGetStatusString(err_)); \ + fprintf(stderr, "current device: %d\n", id); \ exit(1); \ } \ } while (0) @@ -165,7 +181,10 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); do { \ cublasStatus_t err_ = (err); \ if (err_ != CUBLAS_STATUS_SUCCESS) { \ + int id; \ + cudaGetDevice(&id); \ fprintf(stderr, "\ncuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \ + fprintf(stderr, "current device: %d\n", id); \ exit(1); \ } \ } while (0) @@ -212,10 +231,13 @@ typedef void (*to_fp32_cuda_t)(const void * __restrict__ x, float * __restrict__ typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v); typedef void (*cpy_kernel_t)(const char * cx, char * cdst); typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); -typedef void (*ggml_cuda_op_t)( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, float * src0_ddf_i, - float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main); +typedef void (*ggml_cuda_op_mul_mat_t)( + 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, const cudaStream_t & stream); +typedef void (*ggml_cuda_op_flatten_t)( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream); // QK = number of values after dequantization // QR = QK / number of values before dequantization @@ -396,11 +418,29 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); #endif +#define MUL_MAT_SRC1_COL_STRIDE 128 + +#define MAX_STREAMS 8 +static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { nullptr }; + struct ggml_tensor_extra_gpu { void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors - cudaEvent_t events[GGML_CUDA_MAX_DEVICES]; // events for synchronizing multiple GPUs + cudaEvent_t events[GGML_CUDA_MAX_DEVICES][MAX_STREAMS]; // events for synchronizing multiple GPUs }; +// this is faster on Windows +// probably because the Windows CUDA libraries forget to make this check before invoking the drivers +inline cudaError_t ggml_cuda_set_device(const int device) { + int current_device; + CUDA_CHECK(cudaGetDevice(¤t_device)); + + if (device == current_device) { + return cudaSuccess; + } + + return cudaSetDevice(device); +} + static int g_device_count = -1; static int g_main_device = 0; static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES]; @@ -413,8 +453,6 @@ static size_t g_scratch_offset = 0; static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr}; -static cudaStream_t g_cudaStreams_main[GGML_CUDA_MAX_DEVICES] = { nullptr }; - static __global__ void add_f32(const float * x, const float * y, float * dst, const int kx, const int ky) { const int i = blockDim.x*blockIdx.x + threadIdx.x; @@ -3444,6 +3482,12 @@ static __device__ __forceinline__ void mul_mat_q( } } +#define MMQ_X_Q4_0_RDNA2 64 +#define MMQ_Y_Q4_0_RDNA2 128 +#define NWARPS_Q4_0_RDNA2 8 +#define MMQ_X_Q4_0_RDNA1 64 +#define MMQ_Y_Q4_0_RDNA1 64 +#define NWARPS_Q4_0_RDNA1 8 #define MMQ_X_Q4_0_AMPERE 64 #define MMQ_Y_Q4_0_AMPERE 128 #define NWARPS_Q4_0_AMPERE 4 @@ -3451,11 +3495,32 @@ static __device__ __forceinline__ void mul_mat_q( #define MMQ_Y_Q4_0_PASCAL 64 #define NWARPS_Q4_0_PASCAL 8 -template static __global__ void mul_mat_q4_0( +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + mul_mat_q4_0( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q4_0_RDNA2; + const int mmq_y = MMQ_Y_Q4_0_RDNA2; + const int nwarps = NWARPS_Q4_0_RDNA2; +#else + const int mmq_x = MMQ_X_Q4_0_RDNA1; + const int mmq_y = MMQ_Y_Q4_0_RDNA1; + const int nwarps = NWARPS_Q4_0_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q4_0_AMPERE; const int mmq_y = MMQ_Y_Q4_0_AMPERE; const int nwarps = NWARPS_Q4_0_AMPERE; @@ -3478,6 +3543,12 @@ template static __global__ void mul_mat_q4_0( #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q4_1_RDNA2 64 +#define MMQ_Y_Q4_1_RDNA2 128 +#define NWARPS_Q4_1_RDNA2 8 +#define MMQ_X_Q4_1_RDNA1 64 +#define MMQ_Y_Q4_1_RDNA1 64 +#define NWARPS_Q4_1_RDNA1 8 #define MMQ_X_Q4_1_AMPERE 64 #define MMQ_Y_Q4_1_AMPERE 128 #define NWARPS_Q4_1_AMPERE 4 @@ -3486,14 +3557,33 @@ template static __global__ void mul_mat_q4_0( #define NWARPS_Q4_1_PASCAL 8 template static __global__ void -#if __CUDA_ARCH__ < CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_TURING __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2) #endif // __CUDA_ARCH__ < CC_TURING mul_mat_q4_1( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q4_1_RDNA2; + const int mmq_y = MMQ_Y_Q4_1_RDNA2; + const int nwarps = NWARPS_Q4_1_RDNA2; +#else + const int mmq_x = MMQ_X_Q4_1_RDNA1; + const int mmq_y = MMQ_Y_Q4_1_RDNA1; + const int nwarps = NWARPS_Q4_1_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q4_1_AMPERE; const int mmq_y = MMQ_Y_Q4_1_AMPERE; const int nwarps = NWARPS_Q4_1_AMPERE; @@ -3516,6 +3606,12 @@ template static __global__ void #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q5_0_RDNA2 64 +#define MMQ_Y_Q5_0_RDNA2 128 +#define NWARPS_Q5_0_RDNA2 8 +#define MMQ_X_Q5_0_RDNA1 64 +#define MMQ_Y_Q5_0_RDNA1 64 +#define NWARPS_Q5_0_RDNA1 8 #define MMQ_X_Q5_0_AMPERE 128 #define MMQ_Y_Q5_0_AMPERE 64 #define NWARPS_Q5_0_AMPERE 4 @@ -3523,11 +3619,32 @@ template static __global__ void #define MMQ_Y_Q5_0_PASCAL 64 #define NWARPS_Q5_0_PASCAL 8 -template static __global__ void mul_mat_q5_0( +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + mul_mat_q5_0( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q5_0_RDNA2; + const int mmq_y = MMQ_Y_Q5_0_RDNA2; + const int nwarps = NWARPS_Q5_0_RDNA2; +#else + const int mmq_x = MMQ_X_Q5_0_RDNA1; + const int mmq_y = MMQ_Y_Q5_0_RDNA1; + const int nwarps = NWARPS_Q5_0_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q5_0_AMPERE; const int mmq_y = MMQ_Y_Q5_0_AMPERE; const int nwarps = NWARPS_Q5_0_AMPERE; @@ -3550,6 +3667,12 @@ template static __global__ void mul_mat_q5_0( #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q5_1_RDNA2 64 +#define MMQ_Y_Q5_1_RDNA2 128 +#define NWARPS_Q5_1_RDNA2 8 +#define MMQ_X_Q5_1_RDNA1 64 +#define MMQ_Y_Q5_1_RDNA1 64 +#define NWARPS_Q5_1_RDNA1 8 #define MMQ_X_Q5_1_AMPERE 128 #define MMQ_Y_Q5_1_AMPERE 64 #define NWARPS_Q5_1_AMPERE 4 @@ -3557,11 +3680,32 @@ template static __global__ void mul_mat_q5_0( #define MMQ_Y_Q5_1_PASCAL 64 #define NWARPS_Q5_1_PASCAL 8 -template static __global__ void mul_mat_q5_1( +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +mul_mat_q5_1( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q5_1_RDNA2; + const int mmq_y = MMQ_Y_Q5_1_RDNA2; + const int nwarps = NWARPS_Q5_1_RDNA2; +#else + const int mmq_x = MMQ_X_Q5_1_RDNA1; + const int mmq_y = MMQ_Y_Q5_1_RDNA1; + const int nwarps = NWARPS_Q5_1_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q5_1_AMPERE; const int mmq_y = MMQ_Y_Q5_1_AMPERE; const int nwarps = NWARPS_Q5_1_AMPERE; @@ -3584,6 +3728,12 @@ template static __global__ void mul_mat_q5_1( #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q8_0_RDNA2 64 +#define MMQ_Y_Q8_0_RDNA2 128 +#define NWARPS_Q8_0_RDNA2 8 +#define MMQ_X_Q8_0_RDNA1 64 +#define MMQ_Y_Q8_0_RDNA1 64 +#define NWARPS_Q8_0_RDNA1 8 #define MMQ_X_Q8_0_AMPERE 128 #define MMQ_Y_Q8_0_AMPERE 64 #define NWARPS_Q8_0_AMPERE 4 @@ -3591,11 +3741,32 @@ template static __global__ void mul_mat_q5_1( #define MMQ_Y_Q8_0_PASCAL 64 #define NWARPS_Q8_0_PASCAL 8 -template static __global__ void mul_mat_q8_0( +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + mul_mat_q8_0( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q8_0_RDNA2; + const int mmq_y = MMQ_Y_Q8_0_RDNA2; + const int nwarps = NWARPS_Q8_0_RDNA2; +#else + const int mmq_x = MMQ_X_Q8_0_RDNA1; + const int mmq_y = MMQ_Y_Q8_0_RDNA1; + const int nwarps = NWARPS_Q8_0_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q8_0_AMPERE; const int mmq_y = MMQ_Y_Q8_0_AMPERE; const int nwarps = NWARPS_Q8_0_AMPERE; @@ -3618,6 +3789,12 @@ template static __global__ void mul_mat_q8_0( #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q2_K_RDNA2 64 +#define MMQ_Y_Q2_K_RDNA2 128 +#define NWARPS_Q2_K_RDNA2 8 +#define MMQ_X_Q2_K_RDNA1 128 +#define MMQ_Y_Q2_K_RDNA1 32 +#define NWARPS_Q2_K_RDNA1 8 #define MMQ_X_Q2_K_AMPERE 64 #define MMQ_Y_Q2_K_AMPERE 128 #define NWARPS_Q2_K_AMPERE 4 @@ -3625,11 +3802,32 @@ template static __global__ void mul_mat_q8_0( #define MMQ_Y_Q2_K_PASCAL 64 #define NWARPS_Q2_K_PASCAL 8 -template static __global__ void mul_mat_q2_K( +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +mul_mat_q2_K( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q2_K_RDNA2; + const int mmq_y = MMQ_Y_Q2_K_RDNA2; + const int nwarps = NWARPS_Q2_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q2_K_RDNA1; + const int mmq_y = MMQ_Y_Q2_K_RDNA1; + const int nwarps = NWARPS_Q2_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q2_K_AMPERE; const int mmq_y = MMQ_Y_Q2_K_AMPERE; const int nwarps = NWARPS_Q2_K_AMPERE; @@ -3652,6 +3850,12 @@ template static __global__ void mul_mat_q2_K( #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q3_K_RDNA2 128 +#define MMQ_Y_Q3_K_RDNA2 64 +#define NWARPS_Q3_K_RDNA2 8 +#define MMQ_X_Q3_K_RDNA1 32 +#define MMQ_Y_Q3_K_RDNA1 128 +#define NWARPS_Q3_K_RDNA1 8 #define MMQ_X_Q3_K_AMPERE 128 #define MMQ_Y_Q3_K_AMPERE 128 #define NWARPS_Q3_K_AMPERE 4 @@ -3660,14 +3864,33 @@ template static __global__ void mul_mat_q2_K( #define NWARPS_Q3_K_PASCAL 8 template static __global__ void -#if __CUDA_ARCH__ < CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_TURING __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2) #endif // __CUDA_ARCH__ < CC_TURING mul_mat_q3_K( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q3_K_RDNA2; + const int mmq_y = MMQ_Y_Q3_K_RDNA2; + const int nwarps = NWARPS_Q3_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q3_K_RDNA1; + const int mmq_y = MMQ_Y_Q3_K_RDNA1; + const int nwarps = NWARPS_Q3_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q3_K_AMPERE; const int mmq_y = MMQ_Y_Q3_K_AMPERE; const int nwarps = NWARPS_Q3_K_AMPERE; @@ -3690,6 +3913,12 @@ template static __global__ void #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q4_K_RDNA2 64 +#define MMQ_Y_Q4_K_RDNA2 128 +#define NWARPS_Q4_K_RDNA2 8 +#define MMQ_X_Q4_K_RDNA1 32 +#define MMQ_Y_Q4_K_RDNA1 64 +#define NWARPS_Q4_K_RDNA1 8 #define MMQ_X_Q4_K_AMPERE 64 #define MMQ_Y_Q4_K_AMPERE 128 #define NWARPS_Q4_K_AMPERE 4 @@ -3698,14 +3927,33 @@ template static __global__ void #define NWARPS_Q4_K_PASCAL 8 template static __global__ void -#if __CUDA_ARCH__ < CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_TURING __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2) #endif // __CUDA_ARCH__ < CC_TURING mul_mat_q4_K( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q4_K_RDNA2; + const int mmq_y = MMQ_Y_Q4_K_RDNA2; + const int nwarps = NWARPS_Q4_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q4_K_RDNA1; + const int mmq_y = MMQ_Y_Q4_K_RDNA1; + const int nwarps = NWARPS_Q4_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q4_K_AMPERE; const int mmq_y = MMQ_Y_Q4_K_AMPERE; const int nwarps = NWARPS_Q4_K_AMPERE; @@ -3728,6 +3976,12 @@ template static __global__ void #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q5_K_RDNA2 64 +#define MMQ_Y_Q5_K_RDNA2 128 +#define NWARPS_Q5_K_RDNA2 8 +#define MMQ_X_Q5_K_RDNA1 32 +#define MMQ_Y_Q5_K_RDNA1 64 +#define NWARPS_Q5_K_RDNA1 8 #define MMQ_X_Q5_K_AMPERE 64 #define MMQ_Y_Q5_K_AMPERE 128 #define NWARPS_Q5_K_AMPERE 4 @@ -3735,11 +3989,32 @@ template static __global__ void #define MMQ_Y_Q5_K_PASCAL 64 #define NWARPS_Q5_K_PASCAL 8 -template static __global__ void mul_mat_q5_K( +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +mul_mat_q5_K( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q5_K_RDNA2; + const int mmq_y = MMQ_Y_Q5_K_RDNA2; + const int nwarps = NWARPS_Q5_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q5_K_RDNA1; + const int mmq_y = MMQ_Y_Q5_K_RDNA1; + const int nwarps = NWARPS_Q5_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q5_K_AMPERE; const int mmq_y = MMQ_Y_Q5_K_AMPERE; const int nwarps = NWARPS_Q5_K_AMPERE; @@ -3762,6 +4037,12 @@ template static __global__ void mul_mat_q5_K( #endif // __CUDA_ARCH__ >= CC_TURING } +#define MMQ_X_Q6_K_RDNA2 64 +#define MMQ_Y_Q6_K_RDNA2 128 +#define NWARPS_Q6_K_RDNA2 8 +#define MMQ_X_Q6_K_RDNA1 32 +#define MMQ_Y_Q6_K_RDNA1 64 +#define NWARPS_Q6_K_RDNA1 8 #define MMQ_X_Q6_K_AMPERE 64 #define MMQ_Y_Q6_K_AMPERE 64 #define NWARPS_Q6_K_AMPERE 4 @@ -3770,14 +4051,33 @@ template static __global__ void mul_mat_q5_K( #define NWARPS_Q6_K_PASCAL 8 template static __global__ void -#if __CUDA_ARCH__ < CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_TURING __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2) #endif // __CUDA_ARCH__ < CC_TURING mul_mat_q6_K( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { -#if __CUDA_ARCH__ >= CC_TURING +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q6_K_RDNA2; + const int mmq_y = MMQ_Y_Q6_K_RDNA2; + const int nwarps = NWARPS_Q6_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q6_K_RDNA1; + const int mmq_y = MMQ_Y_Q6_K_RDNA1; + const int nwarps = NWARPS_Q6_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_TURING const int mmq_x = MMQ_X_Q6_K_AMPERE; const int mmq_y = MMQ_Y_Q6_K_AMPERE; const int nwarps = NWARPS_Q6_K_AMPERE; @@ -4086,7 +4386,8 @@ static __global__ void rope_neox_f32(const float * x, float * dst, const int nco dst[i + ncols/2] = x0*sin_theta + x1*cos_theta; } -static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p, const float block_p, const float theta_scale) { +static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p0, + const float p_delta, const int p_delta_rows, const float theta_scale, const int n_ctx) { const int col = blockDim.x*blockIdx.x + threadIdx.x; const int half_n_dims = ncols/4; @@ -4098,8 +4399,9 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol const int i = row*ncols + col; const float col_theta_scale = powf(theta_scale, col); + const float p = p0 + p_delta*(row/p_delta_rows); - const float theta = p*col_theta_scale; + const float theta = min(p, p_delta*(n_ctx - 2))*col_theta_scale; const float sin_theta = sinf(theta); const float cos_theta = cosf(theta); @@ -4109,7 +4411,7 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol dst[i + 0] = x0*cos_theta - x1*sin_theta; dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta; - const float block_theta = block_p*col_theta_scale; + const float block_theta = max(p - p_delta*(n_ctx - 2), 0.f)*col_theta_scale; const float sin_block_theta = sinf(block_theta); const float cos_block_theta = cosf(block_theta); @@ -4558,7 +4860,15 @@ static void ggml_mul_mat_q4_0_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q4_0_RDNA2; + mmq_y = MMQ_Y_Q4_0_RDNA2; + nwarps = NWARPS_Q4_0_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q4_0_RDNA1; + mmq_y = MMQ_Y_Q4_0_RDNA1; + nwarps = NWARPS_Q4_0_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q4_0_AMPERE; mmq_y = MMQ_Y_Q4_0_AMPERE; nwarps = NWARPS_Q4_0_AMPERE; @@ -4595,7 +4905,15 @@ static void ggml_mul_mat_q4_1_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q4_1_RDNA2; + mmq_y = MMQ_Y_Q4_1_RDNA2; + nwarps = NWARPS_Q4_1_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q4_1_RDNA1; + mmq_y = MMQ_Y_Q4_1_RDNA1; + nwarps = NWARPS_Q4_1_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q4_1_AMPERE; mmq_y = MMQ_Y_Q4_1_AMPERE; nwarps = NWARPS_Q4_1_AMPERE; @@ -4632,7 +4950,15 @@ static void ggml_mul_mat_q5_0_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q5_0_RDNA2; + mmq_y = MMQ_Y_Q5_0_RDNA2; + nwarps = NWARPS_Q5_0_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q5_0_RDNA1; + mmq_y = MMQ_Y_Q5_0_RDNA1; + nwarps = NWARPS_Q5_0_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q5_0_AMPERE; mmq_y = MMQ_Y_Q5_0_AMPERE; nwarps = NWARPS_Q5_0_AMPERE; @@ -4669,7 +4995,15 @@ static void ggml_mul_mat_q5_1_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q5_1_RDNA2; + mmq_y = MMQ_Y_Q5_1_RDNA2; + nwarps = NWARPS_Q5_1_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q5_1_RDNA1; + mmq_y = MMQ_Y_Q5_1_RDNA1; + nwarps = NWARPS_Q5_1_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q5_1_AMPERE; mmq_y = MMQ_Y_Q5_1_AMPERE; nwarps = NWARPS_Q5_1_AMPERE; @@ -4706,7 +5040,15 @@ static void ggml_mul_mat_q8_0_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q8_0_RDNA2; + mmq_y = MMQ_Y_Q8_0_RDNA2; + nwarps = NWARPS_Q8_0_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q8_0_RDNA1; + mmq_y = MMQ_Y_Q8_0_RDNA1; + nwarps = NWARPS_Q8_0_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q8_0_AMPERE; mmq_y = MMQ_Y_Q8_0_AMPERE; nwarps = NWARPS_Q8_0_AMPERE; @@ -4743,7 +5085,15 @@ static void ggml_mul_mat_q2_K_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q2_K_RDNA2; + mmq_y = MMQ_Y_Q2_K_RDNA2; + nwarps = NWARPS_Q2_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q2_K_RDNA1; + mmq_y = MMQ_Y_Q2_K_RDNA1; + nwarps = NWARPS_Q2_K_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q2_K_AMPERE; mmq_y = MMQ_Y_Q2_K_AMPERE; nwarps = NWARPS_Q2_K_AMPERE; @@ -4782,7 +5132,15 @@ static void ggml_mul_mat_q3_K_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q3_K_RDNA2; + mmq_y = MMQ_Y_Q3_K_RDNA2; + nwarps = NWARPS_Q3_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q3_K_RDNA1; + mmq_y = MMQ_Y_Q3_K_RDNA1; + nwarps = NWARPS_Q3_K_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q3_K_AMPERE; mmq_y = MMQ_Y_Q3_K_AMPERE; nwarps = NWARPS_Q3_K_AMPERE; @@ -4820,7 +5178,15 @@ static void ggml_mul_mat_q4_K_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q4_K_RDNA2; + mmq_y = MMQ_Y_Q4_K_RDNA2; + nwarps = NWARPS_Q4_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q4_K_RDNA1; + mmq_y = MMQ_Y_Q4_K_RDNA1; + nwarps = NWARPS_Q4_K_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q4_K_AMPERE; mmq_y = MMQ_Y_Q4_K_AMPERE; nwarps = NWARPS_Q4_K_AMPERE; @@ -4857,7 +5223,15 @@ static void ggml_mul_mat_q5_K_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q5_K_RDNA2; + mmq_y = MMQ_Y_Q5_K_RDNA2; + nwarps = NWARPS_Q5_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q5_K_RDNA1; + mmq_y = MMQ_Y_Q5_K_RDNA1; + nwarps = NWARPS_Q5_K_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q5_K_AMPERE; mmq_y = MMQ_Y_Q5_K_AMPERE; nwarps = NWARPS_Q5_K_AMPERE; @@ -4894,7 +5268,15 @@ static void ggml_mul_mat_q6_K_q8_1_cuda( const int compute_capability = g_compute_capabilities[id]; int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_TURING) { + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q6_K_RDNA2; + mmq_y = MMQ_Y_Q6_K_RDNA2; + nwarps = NWARPS_Q6_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q6_K_RDNA1; + mmq_y = MMQ_Y_Q6_K_RDNA1; + nwarps = NWARPS_Q6_K_RDNA1; + } else if (compute_capability >= CC_TURING) { mmq_x = MMQ_X_Q6_K_AMPERE; mmq_y = MMQ_Y_Q6_K_AMPERE; nwarps = NWARPS_Q6_K_AMPERE; @@ -4984,12 +5366,13 @@ static void rope_neox_f32_cuda(const float * x, float * dst, const int ncols, co rope_neox_f32<<>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale); } -static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p, const float block_p, const float theta_scale, cudaStream_t stream) { - GGML_ASSERT(nrows % 4 == 0); - const dim3 block_dims(4*CUDA_ROPE_BLOCK_SIZE, 1, 1); - const int num_blocks_x = (ncols + 4*CUDA_ROPE_BLOCK_SIZE - 1) / (4*CUDA_ROPE_BLOCK_SIZE); +static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0, + const float p_delta, const int p_delta_rows, const float theta_scale, const int n_ctx, cudaStream_t stream) { + GGML_ASSERT(ncols % 4 == 0); + const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1); + const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE; const dim3 block_nums(num_blocks_x, nrows, 1); - rope_glm_f32<<>>(x, dst, ncols, p, block_p, theta_scale); + rope_glm_f32<<>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale, n_ctx); } static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, @@ -5127,25 +5510,30 @@ void ggml_init_cublas() { GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES); int64_t total_vram = 0; fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count); - for (int id = 0; id < g_device_count; ++id) { + for (int64_t id = 0; id < g_device_count; ++id) { cudaDeviceProp prop; CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); - fprintf(stderr, " Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor); + fprintf(stderr, " Device %ld: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor); g_tensor_split[id] = total_vram; total_vram += prop.totalGlobalMem; - +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + g_compute_capabilities[id] = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD; +#else g_compute_capabilities[id] = 100*prop.major + 10*prop.minor; +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) } - for (int id = 0; id < g_device_count; ++id) { + for (int64_t id = 0; id < g_device_count; ++id) { g_tensor_split[id] /= total_vram; } - for (int id = 0; id < g_device_count; ++id) { - CUDA_CHECK(cudaSetDevice(id)); + for (int64_t id = 0; id < g_device_count; ++id) { + CUDA_CHECK(ggml_cuda_set_device(id)); - // create main stream - CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id], cudaStreamNonBlocking)); + // create cuda streams + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking)); + } // create cublas handle CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id])); @@ -5214,7 +5602,8 @@ static cudaError_t ggml_cuda_cpy_tensor_2d( if (src->backend == GGML_BACKEND_CPU) { kind = cudaMemcpyHostToDevice; src_ptr = (char *) src->data; - } else if (src->backend == GGML_BACKEND_GPU) { + } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) { + GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1])); kind = cudaMemcpyDeviceToDevice; struct ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra; int id; @@ -5253,236 +5642,205 @@ static cudaError_t ggml_cuda_cpy_tensor_2d( } inline void ggml_cuda_op_add( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddq_i != nullptr || src0_ddf_i != nullptr); - GGML_ASSERT(src1_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); - - const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + GGML_ASSERT(src1->type == GGML_TYPE_F32); const int64_t ne10 = src1->ne[0]; const int64_t ne11 = src1->ne[1]; - // compute if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - add_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne00*i01_diff, ne10*ne11, cudaStream_main); + add_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream); } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { - add_f16_f32_f16_cuda((half *) src0_ddq_i, src1_ddf_i, (half *) dst_ddf_i, ne00*i01_diff, cudaStream_main); + add_f16_f32_f16_cuda((const half *) src0_dd, src1_dd, (half *) dst_dd, ggml_nelements(src0), main_stream); } else { GGML_ASSERT(false); } (void) src1; (void) dst; - (void) src0_ddq_i; - (void) i02; - (void) i1; } inline void ggml_cuda_op_mul( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(src1_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); - - const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne10 = src1->ne[0]; const int64_t ne11 = src1->ne[1]; - mul_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne00*i01_diff, ne10*ne11, cudaStream_main); + mul_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream); (void) dst; - (void) src0_ddq_i; - (void) i02; - (void) i1; } inline void ggml_cuda_op_gelu( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); - const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; - - // compute - gelu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main); + gelu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_silu( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); - const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; - - // compute - silu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main); + silu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_norm( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t nrows = ggml_nrows(src0); - // compute - norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main); + norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, main_stream); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_rms_norm( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t nrows = ggml_nrows(src0); float eps; memcpy(&eps, dst->op_params, sizeof(float)); - // compute - rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, eps, cudaStream_main); + rms_norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_mul_mat_q( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ - - GGML_ASSERT(src0_ddq_i != nullptr); - GGML_ASSERT(src1_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + 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, const cudaStream_t & stream) { const int64_t ne00 = src0->ne[0]; const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; GGML_ASSERT(ne10 % QK8_1 == 0); const int64_t ne0 = dst->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t row_diff = row_high - row_low; int id; CUDA_CHECK(cudaGetDevice(&id)); // the main device has a larger memory buffer to hold the results from all GPUs // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into - const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : i01_diff; - - const int64_t padded_row_size = ne10 % MATRIX_ROW_PADDING == 0 ? - ne10 : ne10 - ne10 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING; - size_t as; - void * src1_q8_1 = ggml_cuda_pool_malloc(padded_row_size*ne11*sizeof(block_q8_1)/QK8_1, &as); - quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne10, ne11, padded_row_size, cudaStream_main); + const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff; switch (src0->type) { case GGML_TYPE_Q4_0: - ggml_mul_mat_q4_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q4_1: - ggml_mul_mat_q4_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q5_0: - ggml_mul_mat_q5_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q5_1: - ggml_mul_mat_q5_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q8_0: - ggml_mul_mat_q8_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q2_K: - ggml_mul_mat_q2_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q3_K: - ggml_mul_mat_q3_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q4_K: - ggml_mul_mat_q4_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q5_K: - ggml_mul_mat_q5_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; case GGML_TYPE_Q6_K: - ggml_mul_mat_q6_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main); + ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); break; default: GGML_ASSERT(false); break; } - ggml_cuda_pool_free(src1_q8_1, as); - (void) src1; (void) dst; - (void) src0_ddf_i; - (void) i02; - (void) i1; + (void) src1_ddf_i; } static int64_t get_row_rounding(ggml_type type) { - int max_compute_capability = INT_MIN; - for (int id = 0; id < g_device_count; ++id) { - if (max_compute_capability < g_compute_capabilities[id] - && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { - max_compute_capability = g_compute_capabilities[id]; + int64_t min_compute_capability = INT_MAX; + int64_t max_compute_capability = INT_MIN; + for (int64_t id = 0; id < g_device_count; ++id) { + if (g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { + if (min_compute_capability > g_compute_capabilities[id]) { + min_compute_capability = g_compute_capabilities[id]; + } + if (max_compute_capability < g_compute_capabilities[id]) { + max_compute_capability = g_compute_capabilities[id]; + } } } +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + switch(type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + return max_compute_capability >= CC_RDNA2 ? 128 : 64; + case GGML_TYPE_F16: + return 1; + case GGML_TYPE_Q2_K: + return max_compute_capability >= CC_RDNA2 ? 128 : 32; + case GGML_TYPE_Q3_K: + return min_compute_capability < CC_RDNA2 ? 128 : 64; + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + return max_compute_capability >= CC_RDNA2 ? 128 : 64; + default: + GGML_ASSERT(false); + } +#else switch(type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: @@ -5503,170 +5861,147 @@ static int64_t get_row_rounding(ggml_type type) { default: GGML_ASSERT(false); } +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) } -inline void ggml_cuda_op_mul_mat_vec( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ - - GGML_ASSERT(src0_ddq_i != nullptr); - GGML_ASSERT(src1_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); +inline void ggml_cuda_op_mul_mat_vec_q( + 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, const cudaStream_t & stream) { const int64_t ne00 = src0->ne[0]; - const int64_t nrows = i01_high - i01_low; + const int64_t row_diff = row_high - row_low; -#ifdef GGML_CUDA_FORCE_DMMV - const bool use_mul_mat_vec_q = false; - (void) g_compute_capabilities[0]; -#else - int id; - CUDA_CHECK(cudaGetDevice(&id)); - - bool mul_mat_vec_q_implemented = - src0->type == GGML_TYPE_Q4_0 || - src0->type == GGML_TYPE_Q4_1 || - src0->type == GGML_TYPE_Q5_0 || - src0->type == GGML_TYPE_Q5_1 || - src0->type == GGML_TYPE_Q8_0; -#if QK_K == 256 - mul_mat_vec_q_implemented = mul_mat_vec_q_implemented || - src0->type == GGML_TYPE_Q2_K || - src0->type == GGML_TYPE_Q3_K || - src0->type == GGML_TYPE_Q4_K || - src0->type == GGML_TYPE_Q5_K || - src0->type == GGML_TYPE_Q6_K; -#endif // QK_K == 256 - - const bool use_mul_mat_vec_q = g_compute_capabilities[id] >= MIN_CC_DP4A && mul_mat_vec_q_implemented; -#endif - - if (use_mul_mat_vec_q) { - const int64_t padded_row_size = ne00 % MATRIX_ROW_PADDING == 0 ? - ne00 : ne00 - ne00 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING; - size_t as; - void * src1_q8_1 = ggml_cuda_pool_malloc(padded_row_size*sizeof(block_q8_1)/QK8_1, &as); - quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne00, 1, padded_row_size, cudaStream_main); - - switch (src0->type) { - case GGML_TYPE_Q4_0: - mul_mat_vec_q4_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q4_1: - mul_mat_vec_q4_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q5_0: - mul_mat_vec_q5_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q5_1: - mul_mat_vec_q5_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q8_0: - mul_mat_vec_q8_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q2_K: - mul_mat_vec_q2_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q3_K: - mul_mat_vec_q3_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q4_K: - mul_mat_vec_q4_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q5_K: - mul_mat_vec_q5_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q6_K: - mul_mat_vec_q6_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - default: - GGML_ASSERT(false); - break; - } - - ggml_cuda_pool_free(src1_q8_1, as); - } else { - // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics -#ifdef GGML_CUDA_F16 - size_t ash; - dfloat * src1_dfloat = nullptr; // dfloat == half - - bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || - src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || - src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; - - if (src1_convert_f16) { - src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash); - ggml_cpy_f32_f16_cuda((char *) src1_ddf_i, (char *) src1_dfloat, ne00, - ne00, 1, sizeof(float), 0, 0, - ne00, 1, sizeof(half), 0, 0, cudaStream_main); - } -#else - dfloat * src1_dfloat = src1_ddf_i; // dfloat == float, no conversion -#endif // GGML_CUDA_F16 - - switch (src0->type) { - case GGML_TYPE_Q4_0: - dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q4_1: - dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q5_0: - dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q5_1: - dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q8_0: - dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q2_K: - dequantize_mul_mat_vec_q2_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q3_K: - dequantize_mul_mat_vec_q3_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q4_K: - dequantize_mul_mat_vec_q4_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q5_K: - dequantize_mul_mat_vec_q5_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_Q6_K: - dequantize_mul_mat_vec_q6_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - case GGML_TYPE_F16: - convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main); - break; - default: - GGML_ASSERT(false); - break; - } - -#ifdef GGML_CUDA_F16 - if (src1_convert_f16) { - ggml_cuda_pool_free(src1_dfloat, ash); - } -#endif // GGML_CUDA_F16 + switch (src0->type) { + case GGML_TYPE_Q4_0: + mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + default: + GGML_ASSERT(false); + break; } (void) src1; (void) dst; - (void) src0_ddf_i; - (void) i02; - (void) i1; + (void) src1_ddf_i; + (void) src1_ncols; + (void) src1_padded_row_size; +} + +inline void ggml_cuda_op_dequantize_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, const cudaStream_t & stream) { + + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics +#ifdef GGML_CUDA_F16 + size_t ash; + dfloat * src1_dfloat = nullptr; // dfloat == half + + bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || + src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || + src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; + + if (src1_convert_f16) { + src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash); + ggml_cpy_f32_f16_cuda((const char *) src1_ddf_i, (char *) src1_dfloat, ne00, + ne00, 1, sizeof(float), 0, 0, + ne00, 1, sizeof(half), 0, 0, stream); + } +#else + const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion +#endif // GGML_CUDA_F16 + + switch (src0->type) { + case GGML_TYPE_Q4_0: + dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_F16: + convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + default: + GGML_ASSERT(false); + break; + } + +#ifdef GGML_CUDA_F16 + if (src1_convert_f16) { + ggml_cuda_pool_free(src1_dfloat, ash); + } +#endif // GGML_CUDA_F16 + + (void) src1; + (void) dst; + (void) src1_ddq_i; + (void) src1_ncols; + (void) src1_padded_row_size; } inline void ggml_cuda_op_mul_mat_cublas( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + 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, const cudaStream_t & stream) { - GGML_ASSERT(src0_ddf_i != nullptr); + GGML_ASSERT(src0_dd_i != nullptr); GGML_ASSERT(src1_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(dst_dd_i != nullptr); const float alpha = 1.0f; const float beta = 0.0f; @@ -5674,43 +6009,54 @@ inline void ggml_cuda_op_mul_mat_cublas( const int64_t ne00 = src0->ne[0]; const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; const int64_t ne0 = dst->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t row_diff = row_high - row_low; + + float * src0_ddq_as_f32; + size_t src0_as = 0; + + if (src0->type != GGML_TYPE_F32) { + const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type); + src0_ddq_as_f32 = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_as); // NOLINT + to_fp32_cuda(src0_dd_i, src0_ddq_as_f32, row_diff*ne00, stream); + } + const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32; int id; CUDA_CHECK(cudaGetDevice(&id)); // the main device has a larger memory buffer to hold the results from all GPUs // ldc == nrows of the matrix that cuBLAS writes into - int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : i01_diff; + int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff; - CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], cudaStream_main)); + CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream)); CUBLAS_CHECK( cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, - i01_diff, ne11, ne10, + row_diff, src1_ncols, ne10, &alpha, src0_ddf_i, ne00, - src1_ddf_i, ne10, - &beta, dst_ddf_i, ldc)); + src1_ddf_i, ne10, + &beta, dst_dd_i, ldc)); + + if (src0_as > 0) { + ggml_cuda_pool_free(src0_ddq_as_f32, src0_as); + } (void) dst; - (void) src0_ddq_i; - (void) i02; - (void) i1; + (void) src1_ddq_i; + (void) src1_padded_row_size; } inline void ggml_cuda_op_rope( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t nrows = ggml_nrows(src0); const int n_past = ((int32_t *) dst->op_params)[0]; const int n_dims = ((int32_t *) dst->op_params)[1]; @@ -5723,44 +6069,37 @@ inline void ggml_cuda_op_rope( memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale; const bool is_neox = mode & 2; const bool is_glm = mode & 4; // compute if (is_glm) { - const float p = (((mode & 1) == 0 ? n_past + i02 : i02)) * freq_scale; - const float id_p = min(p, n_ctx - 2.f); - const float block_p = max(p - (n_ctx - 2.f), 0.f); - rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main); + rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, p0, freq_scale, ne01, theta_scale, n_ctx, main_stream); } else if (is_neox) { GGML_ASSERT(ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet"); - const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale; - rope_neox_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main); + rope_neox_f32_cuda(src0_dd, dst_dd, ne00, nrows, p0, freq_scale, ne01, theta_scale, main_stream); } else { - const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale; - rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main); + rope_f32_cuda(src0_dd, dst_dd, ne00, nrows, p0, freq_scale, ne01, theta_scale, main_stream); } (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_alibi( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; const int64_t ne02 = src0->ne[2]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t nrows = ggml_nrows(src0); const int n_past = ((int32_t *) dst->op_params)[0]; const int n_head = ((int32_t *) dst->op_params)[1]; @@ -5775,334 +6114,354 @@ inline void ggml_cuda_op_alibi( const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor); const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor); - // compute - alibi_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_heads_log2_floor, m0, m1, cudaStream_main); + alibi_f32_cuda(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream); (void) src1; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_diag_mask_inf( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; - const int64_t i01_diff = i01_high - i01_low; + const int nrows0 = ggml_nrows(src0); const int n_past = ((int32_t *) dst->op_params)[0]; - // compute - diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main); + diag_mask_inf_f32_cuda(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_soft_max( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; + const int64_t nrows = ggml_nrows(src0); - // compute - soft_max_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main); + soft_max_f32_cuda(src0_dd, dst_dd, ne00, nrows, main_stream); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } inline void ggml_cuda_op_scale( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, - float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1, - cudaStream_t & cudaStream_main){ + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { - GGML_ASSERT(src0_ddf_i != nullptr); - GGML_ASSERT(dst_ddf_i != nullptr); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); const float scale = ((float *) src1->data)[0]; - const int64_t ne00 = src0->ne[0]; - const int64_t i01_diff = i01_high - i01_low; - - // compute - scale_f32_cuda(src0_ddf_i, dst_ddf_i, scale, ne00*i01_diff, cudaStream_main); + scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream); CUDA_CHECK(cudaGetLastError()); (void) src1; (void) dst; - (void) src0_ddq_i; - (void) src1_ddf_i; - (void) i02; - (void) i1; + (void) src1_dd; } -static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - ggml_cuda_op_t op, bool src0_needs_f32, bool flatten_rows) { +static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) { + const int64_t nrows0 = ggml_nrows(src0); + + const bool use_src1 = src1 != nullptr; + const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1; + + GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT( dst->backend != GGML_BACKEND_GPU_SPLIT); + + struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + struct ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; + struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + + const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT; + const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU; + const bool dst_on_device = dst->backend == GGML_BACKEND_GPU; + + const bool src1_stays_on_host = use_src1 && dst->op == GGML_OP_SCALE; + + // dd = data device + float * src0_ddf = nullptr; + float * src1_ddf = nullptr; + float * dst_ddf = nullptr; + + // as = actual size + size_t src0_asf = 0; + size_t src1_asf = 0; + size_t dst_asf = 0; + + ggml_cuda_set_device(g_main_device); + const cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + + if (src0_on_device) { + src0_ddf = (float *) src0_extra->data_device[g_main_device]; + } else { + src0_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_asf); + CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream)); + } + + if (use_src1 && !src1_stays_on_host) { + if (src1_on_device) { + src1_ddf = (float *) src1_extra->data_device[g_main_device]; + } else { + src1_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src1), &src1_asf); + CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream)); + } + } + if (dst_on_device) { + dst_ddf = (float *) dst_extra->data_device[g_main_device]; + } else { + dst_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(dst), &dst_asf); + } + + // do the computation + op(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); + CUDA_CHECK(cudaGetLastError()); + + // copy dst to host if necessary + if (!dst_on_device) { + CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream)); + } + + if (src0_asf > 0) { + ggml_cuda_pool_free(src0_ddf, src0_asf); + } + if (src1_asf > 0) { + ggml_cuda_pool_free(src1_ddf, src1_asf); + } + if (dst_asf > 0) { + ggml_cuda_pool_free(dst_ddf, dst_asf); + } + + if (dst->backend == GGML_BACKEND_CPU) { + CUDA_CHECK(cudaDeviceSynchronize()); + } +} + +static void ggml_cuda_op_mul_mat( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op, + const bool convert_src1_to_q8_1) { + const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; const int64_t nrows0 = ggml_nrows(src0); - const bool use_src1 = src1 != nullptr; - const int64_t ne10 = use_src1 ? src1->ne[0] : 1; - const int64_t ne11 = use_src1 ? src1->ne[1] : 1; - const int64_t ne12 = use_src1 ? src1->ne[2] : 1; - const int64_t ne13 = use_src1 ? src1->ne[3] : 1; - const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1; + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + const int64_t nrows1 = ggml_nrows(src1); GGML_ASSERT(ne03 == ne13); const int64_t ne0 = dst->ne[0]; const int64_t ne1 = dst->ne[1]; - const int nb2 = dst->nb[2]; - const int nb3 = dst->nb[3]; + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT); - GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT); - // strides for iteration over dims 3 and 2 - const int64_t num_iters_0 = ne02 >= ne12 ? ne02*ne03 : ne12*ne13; - const int64_t num_iters = flatten_rows ? 1 : num_iters_0; - const int64_t stride_mod = flatten_rows ? num_iters_0 : 1; - const int64_t src0_stride = ne00 * ne01 * stride_mod; - const int64_t src1_stride = ne10 * ne11 * stride_mod; - const int64_t dst_stride = ne0 * ne1 * stride_mod; + GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0); - const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01; - const int64_t i03_max = flatten_rows ? 1 : ne03; - const int64_t i02_max = flatten_rows ? 1 : (ne02 >= ne12 ? ne02 : ne12); - const int64_t i02_divisor = ne02 >= ne12 ? 1 : ne12 / ne02; - GGML_ASSERT(!(flatten_rows && ne02 < ne12)); + const int64_t i02_divisor = ne12 / ne02; const size_t src0_ts = ggml_type_size(src0->type); const size_t src0_bs = ggml_blck_size(src0->type); + const size_t q8_1_ts = sizeof(block_q8_1); + const size_t q8_1_bs = QK8_1; - struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - struct ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; - struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT; const bool src0_is_contiguous = ggml_is_contiguous(src0); - const bool src0_is_f32 = src0->type == GGML_TYPE_F32; - const bool src1_is_contiguous = use_src1 && ggml_is_contiguous(src1); - const bool src1_stays_on_host = use_src1 && ( - dst->op == GGML_OP_SCALE || dst->op == GGML_OP_DIAG_MASK_INF || dst->op == GGML_OP_ROPE); + const bool src1_is_contiguous = ggml_is_contiguous(src1); + const int64_t src1_padded_col_size = ne10 % MATRIX_ROW_PADDING == 0 ? + ne10 : ne10 - ne10 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING; const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT; + GGML_ASSERT(!(split && ne02 > 1)); + GGML_ASSERT(!(split && ne03 > 1)); GGML_ASSERT(!(split && ne02 < ne12)); - const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type); - // dd = data device - char * src0_ddq[GGML_CUDA_MAX_DEVICES] = {nullptr}; // quantized - float * src0_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; // float - float * src1_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; - float * dst_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; + char * src0_dd[GGML_CUDA_MAX_DEVICES] = {nullptr}; + float * src1_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; // float + char * src1_ddq[GGML_CUDA_MAX_DEVICES] = {nullptr}; // q8_1 + float * dst_dd[GGML_CUDA_MAX_DEVICES] = {nullptr}; - // asq = actual size quantized, asf = actual size float - size_t src0_asq[GGML_CUDA_MAX_DEVICES] = {0}; - size_t src0_asf[GGML_CUDA_MAX_DEVICES] = {0}; + // as = actual size + size_t src0_as[GGML_CUDA_MAX_DEVICES] = {0}; size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0}; - size_t dst_asf[GGML_CUDA_MAX_DEVICES] = {0}; + size_t src1_asq[GGML_CUDA_MAX_DEVICES] = {0}; + size_t dst_as[GGML_CUDA_MAX_DEVICES] = {0}; - // if multiple devices are used they need to wait for the main device - // here an event is recorded that signifies that the main device has finished calculating the input data - if (split && g_device_count > 1) { - CUDA_CHECK(cudaSetDevice(g_main_device)); - CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device], g_cudaStreams_main[g_main_device])); - } + int64_t row_low[GGML_CUDA_MAX_DEVICES]; + int64_t row_high[GGML_CUDA_MAX_DEVICES]; - for (int id = 0; id < g_device_count; ++id) { - if (!split && id != g_main_device) { - continue; - } + for (int64_t id = 0; id < g_device_count; ++id) { + // by default, use all rows + row_low[id] = 0; + row_high[id] = ne01; - const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU && id == g_main_device; - const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device; - - int64_t row_low, row_high; + // for multi GPU, get the row boundaries from tensor split + // and round to mul_mat_q tile sizes if (split) { const int64_t rounding = get_row_rounding(src0->type); - row_low = id == 0 ? 0 : nrows0*g_tensor_split[id]; - row_low -= row_low % rounding; - - if (id == g_device_count - 1) { - row_high = nrows0; - } else { - row_high = nrows0*g_tensor_split[id + 1]; - row_high -= row_high % rounding; + if (id != 0) { + row_low[id] = ne01*g_tensor_split[id]; + row_low[id] -= row_low[id] % rounding; + } + + if (id != g_device_count - 1) { + row_high[id] = ne01*g_tensor_split[id + 1]; + row_high[id] -= row_high[id] % rounding; } - } else { - row_low = 0; - row_high = nrows0*i02_divisor; } - if (row_low == row_high) { + } + + for (int64_t id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device) || row_low[id] == row_high[id]) { continue; } - int64_t row_diff = row_high - row_low; + const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device; + const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device; - cudaSetDevice(id); - cudaStream_t cudaStream_main = g_cudaStreams_main[id]; - - // wait for main GPU data if necessary - if (split && id != g_main_device) { - CUDA_CHECK(cudaStreamWaitEvent(cudaStream_main, src0_extra->events[g_main_device])); - } + ggml_cuda_set_device(id); + const cudaStream_t stream = g_cudaStreams[id][0]; if (src0_on_device && src0_is_contiguous) { - if (src0_is_f32) { - src0_ddf[id] = (float *) src0_extra->data_device[id]; - } else { - src0_ddq[id] = (char *) src0_extra->data_device[id]; - } + src0_dd[id] = (char *) src0_extra->data_device[id]; } else { - if (src0_is_f32) { - src0_ddf[id] = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_asf[id]); - } else { - src0_ddq[id] = (char *) ggml_cuda_pool_malloc(row_diff*ne00 * src0_ts/src0_bs, &src0_asq[id]); + const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0); + src0_dd[id] = (char *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_as[id]); + } + + if (src1_on_device && src1_is_contiguous) { + src1_ddf[id] = (float *) src1_extra->data_device[id]; + } else { + src1_ddf[id] = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src1), &src1_asf[id]); + } + + if (convert_src1_to_q8_1) { + src1_ddq[id] = (char *) ggml_cuda_pool_malloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs, &src1_asq[id]); + + if (split && src1_on_device && src1_is_contiguous) { + quantize_row_q8_1_cuda(src1_ddf[id], src1_ddq[id], ne10, nrows1, src1_padded_col_size, stream); + CUDA_CHECK(cudaGetLastError()); } } - if (src0_needs_f32 && !src0_is_f32) { - src0_ddf[id] = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_asf[id]); - } - - if (use_src1 && !src1_stays_on_host) { - if (src1_on_device && src1_is_contiguous) { - src1_ddf[id] = (float *) src1_extra->data_device[id]; - } else { - src1_ddf[id] = (float *) ggml_cuda_pool_malloc(num_iters*src1_stride * sizeof(float), &src1_asf[id]); - } - } if (dst_on_device) { - dst_ddf[id] = (float *) dst_extra->data_device[id]; + dst_dd[id] = (float *) dst_extra->data_device[id]; } else { - size_t size_dst_ddf = split ? row_diff*ne1 * sizeof(float) : num_iters*dst_stride * sizeof(float); - dst_ddf[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_asf[id]); + const size_t size_dst_ddf = split ? (row_high[id]-row_low[id])*ne1*sizeof(float) : ggml_nbytes(dst); + dst_dd[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_as[id]); } + } - for (int64_t i03 = 0; i03 < i03_max; i03++) { - const int64_t i13 = i03 % ne13; - for (int64_t i02 = 0; i02 < i02_max; i02++) { - const int64_t i12 = i02 % ne12; + // if multiple devices are used they need to wait for the main device + // here an event is recorded that signals that the main device has finished calculating the input data + if (split && g_device_count > 1) { + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0])); + } - const int64_t i0 = i03*i02_max + i02; + const int64_t src1_col_stride = split && g_device_count > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11; + for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) { + const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0; + const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; - // i0 values that contain the lower/upper rows for a split tensor when using multiple GPUs - const int64_t i0_offset_low = row_low/rows_per_iter; - const int64_t i0_offset_high = row_high/rows_per_iter; + for (int64_t id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device) || row_low[id] == row_high[id]) { + continue; + } - int64_t i01_low = 0; - int64_t i01_high = rows_per_iter; - if (split) { - if (i0 < i0_offset_low || i0 > i0_offset_high) { - continue; - } - if (i0 == i0_offset_low) { - i01_low = row_low % rows_per_iter; - } - if (i0 == i0_offset_high) { - i01_high = row_high % rows_per_iter; - } - } + const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device; + const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device; + const int64_t row_diff = row_high[id] - row_low[id]; - // There is possibly a bug in the Windows nvcc compiler regarding instruction reordering or optimizing out local variables. - // Removing the first assert or changing the order of the arguments causes the second assert to fail. - // Removing both asserts results in i01_high becoming 0 which in turn results in garbage output. - // The root cause seems to be a problem with i0_offset_high becoming 0 when it should always be >0 (for single GPU). - GGML_ASSERT(i01_low == 0 || g_device_count > 1); - GGML_ASSERT(i01_high == rows_per_iter || g_device_count > 1); + ggml_cuda_set_device(id); + const cudaStream_t stream = g_cudaStreams[id][is]; - const int64_t i01_diff = i01_high - i01_low; - if (i01_diff == 0) { - continue; - } - const int64_t i11 = i13*ne12 + i12; + // wait for main GPU data if necessary + if (split && (id != g_main_device || is != 0)) { + CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[g_main_device][0])); + } + + for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) { + const int64_t i03 = i0 / ne12; + const int64_t i02 = i0 % ne12; + + const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs; // for split tensors the data begins at i0 == i0_offset_low - char * src0_ddq_i = src0_ddq[id] + (i0/i02_divisor - i0_offset_low)*src0_stride*src0_ts/src0_bs; - float * src0_ddf_i = src0_ddf[id] + (i0/i02_divisor - i0_offset_low)*src0_stride; - float * src1_ddf_i = src1_ddf[id] + i11*src1_stride; - float * dst_ddf_i = dst_ddf[id] + (i0 - i0_offset_low)*dst_stride; - - // for split tensors the data pointer needs to be rounded down - // to the bin edge for i03, i02 bins beyond the first - if (i0 - i0_offset_low > 0) { - GGML_ASSERT(!flatten_rows); - src0_ddq_i -= (row_low % ne01)*ne00 * src0_ts/src0_bs; - src0_ddf_i -= (row_low % ne01)*ne00; - dst_ddf_i -= (row_low % ne0)*ne1; - } + char * src0_dd_i = src0_dd[id] + (i0/i02_divisor) * ne01*ne00*src0_ts/src0_bs; + float * src1_ddf_i = src1_ddf[id] + (i0*ne11 + src1_col_0) * ne10; + char * src1_ddq_i = src1_ddq[id] + src1_ddq_i_offset; + float * dst_dd_i = dst_dd[id] + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff); // the main device memory buffer can be on VRAM scratch, with space for all partial results // in that case an offset on dst_ddf_i is needed if (dst->backend == GGML_BACKEND_GPU && id == g_main_device) { - dst_ddf_i += i01_low; // offset is 0 if no tensor split + dst_dd_i += row_low[id]; // offset is 0 if no tensor split } // copy src0, src1 to device if necessary - if (use_src1 && !src1_stays_on_host) { - if (src1->backend == GGML_BACKEND_CPU) { - GGML_ASSERT(!flatten_rows || nrows0 == ggml_nrows(src1)); - int64_t nrows1 = flatten_rows ? nrows0 : ne11; - CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, nrows1, cudaStream_main)); - } else if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) { - if (id != g_main_device) { - GGML_ASSERT(!flatten_rows); + if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) { + if (id != g_main_device) { + if (convert_src1_to_q8_1) { + char * src1_ddq_i_source = src1_ddq[g_main_device] + src1_ddq_i_offset; + CUDA_CHECK(cudaMemcpyAsync(src1_ddq_i, src1_ddq_i_source, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, + cudaMemcpyDeviceToDevice, stream)); + } else { float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device]; - src1_ddf_i_source += i11*src1_stride; - CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_stride*sizeof(float), - cudaMemcpyDeviceToDevice, cudaStream_main)); + src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10; + CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_ncols*ne10*sizeof(float), + cudaMemcpyDeviceToDevice, stream)); } - } else if (src1_on_device && !src1_is_contiguous) { - GGML_ASSERT(!split); - CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, ne11, cudaStream_main)); - } else { - GGML_ASSERT(false); } + } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) { + CUDA_CHECK(ggml_cuda_cpy_tensor_2d( + src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream)); + } else { + GGML_ASSERT(false); } - if ((!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) { - if (src0_is_f32) { - CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main)); - } else { - CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main)); - } - } - - // convert src0 to f32 if it is necessary for the ggml_cuda_op - if (src0_needs_f32 && !src0_is_f32) { - to_fp32_cuda(src0_ddq_i, src0_ddf_i, i01_diff*ne00, cudaStream_main); + if (convert_src1_to_q8_1 && src1->backend == GGML_BACKEND_CPU) { + quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream); CUDA_CHECK(cudaGetLastError()); } + if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) { + CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, row_low[id], row_high[id], stream)); + } + // do the computation - op(src0, src1, dst, src0_ddq_i, src0_ddf_i, src1_ddf_i, dst_ddf_i, i02, i01_low, i01_high, i11, cudaStream_main); + op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, + row_low[id], row_high[id], src1_ncols, src1_padded_col_size, stream); CUDA_CHECK(cudaGetLastError()); // copy dst to host or other device if necessary @@ -6124,95 +6483,86 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU. // Instead they need to be copied to the correct slice in ne0 = dst row index. // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results. - float * dhf_dst_i = (float *) ((char *) dst_off_device + i01_low*sizeof(float) + i02*nb2 + i03*nb3); - CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), dst_ddf_i, i01_diff*sizeof(float), - i01_diff*sizeof(float), ne1, kind, cudaStream_main)); + float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); + GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); + dhf_dst_i += src1_col_0*ne0 + row_low[id]; + CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), dst_dd_i, row_diff*sizeof(float), + row_diff*sizeof(float), src1_ncols, kind, stream)); } else { float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); - CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i, dst_stride*sizeof(float), kind, cudaStream_main)); + GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); + dhf_dst_i += src1_col_0*ne0; + CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), kind, stream)); } } - // signify to main device that other device is done - if (split && g_device_count > 1 && id != g_main_device) { - CUDA_CHECK(cudaEventRecord(src0_extra->events[id], cudaStream_main)); + // add event for the main device to wait on until other device is done + if (split && (id != g_main_device || is != 0)) { + CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream)); } } } } - // wait until each device is finished, then free their buffers - for (int id = 0; id < g_device_count; ++id) { - if (src0_asq[id] == 0 && src0_asf[id] == 0 && src1_asf[id] == 0 && dst_asf[id] == 0) { - continue; - } + for (int64_t id = 0; id < g_device_count; ++id) { + CUDA_CHECK(ggml_cuda_set_device(id)); - CUDA_CHECK(cudaSetDevice(id)); - - if (src0_asq[id] > 0) { - ggml_cuda_pool_free(src0_ddq[id], src0_asq[id]); - } - if (src0_asf[id] > 0) { - ggml_cuda_pool_free(src0_ddf[id], src0_asf[id]); + // free buffers again when done + if (src0_as[id] > 0) { + ggml_cuda_pool_free(src0_dd[id], src0_as[id]); } if (src1_asf[id] > 0) { ggml_cuda_pool_free(src1_ddf[id], src1_asf[id]); } - if (dst_asf[id] > 0) { - ggml_cuda_pool_free(dst_ddf[id], dst_asf[id]); + if (src1_asq[id] > 0) { + ggml_cuda_pool_free(src1_ddq[id], src1_asq[id]); + } + if (dst_as[id] > 0) { + ggml_cuda_pool_free(dst_dd[id], dst_as[id]); } } // main device waits for all other devices to be finished if (split && g_device_count > 1) { - CUDA_CHECK(cudaSetDevice(g_main_device)); - for (int id = 0; id < g_device_count; ++id) { - if (id != g_main_device && src0_extra->events[id]) { - CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams_main[g_main_device], src0_extra->events[id])); + int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE; + is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS; + + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + for (int64_t id = 0; id < g_device_count; ++id) { + for (int64_t is = 0; is < is_max; ++is) { + CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is])); } } } if (dst->backend == GGML_BACKEND_CPU) { - CUDA_CHECK(cudaSetDevice(g_main_device)); + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); CUDA_CHECK(cudaDeviceSynchronize()); } } void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - // ggml_cuda_add permits f16 dst even though this could in theory cause problems with the pointer arithmetic in ggml_cuda_op. - // Due to flatten_rows == true this does in practice not make a difference however. - // Better solution would be nice but right now that would require disproportionate changes. - GGML_ASSERT( - (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) && - src1->type == GGML_TYPE_F32 && - (dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16)); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_add, false, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add); } void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul, true, false); // TODO ggml_cuda_op needs modification for flatten + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul); } void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_gelu, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu); } void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_silu, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu); } void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_norm, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm); } void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rms_norm, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm); } bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { @@ -6246,8 +6596,8 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr const int64_t ne12 = src1->ne[2]; - CUDA_CHECK(cudaSetDevice(g_main_device)); - cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device]; + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; void * src0_ddq = src0_extra->data_device[g_main_device]; @@ -6258,7 +6608,7 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; - ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, cudaStream_main); + ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream); } void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){ @@ -6277,8 +6627,8 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1 const int64_t nb01 = src0->nb[1]; const int64_t nb02 = src0->nb[2]; - CUDA_CHECK(cudaSetDevice(g_main_device)); - cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device]; + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; void * src0_ddq = src0_extra->data_device[g_main_device]; @@ -6289,38 +6639,49 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1 struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; - const int row_stride_x = nb01 / sizeof(half); - const int channel_stride_x = nb02 / sizeof(half); + const int64_t row_stride_x = nb01 / sizeof(half); + const int64_t channel_stride_x = nb02 / sizeof(half); - ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, cudaStream_main); + ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream); } void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { bool all_on_device = (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) && src1->backend == GGML_BACKEND_GPU && dst->backend == GGML_BACKEND_GPU; + int64_t min_compute_capability = INT_MAX; + for (int64_t id = 0; id < g_device_count; ++id) { + if (min_compute_capability > g_compute_capabilities[id] + && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { + min_compute_capability = g_compute_capabilities[id]; + } + } + if (all_on_device && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { ggml_cuda_mul_mat_vec_p021(src0, src1, dst); } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) { ggml_cuda_mul_mat_vec_nc(src0, src1, dst); }else if (src0->type == GGML_TYPE_F32) { - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false); + ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) { if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) { - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_vec, false, false); - } else { - int min_compute_capability = INT_MAX; - for (int id = 0; id < g_device_count; ++id) { - if (min_compute_capability > g_compute_capabilities[id] - && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { - min_compute_capability = g_compute_capabilities[id]; - } - } - if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) { - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_q, false, false); +#ifdef GGML_CUDA_FORCE_DMMV + const bool use_mul_mat_vec_q = false; +#else + const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type); +#endif // GGML_CUDA_FORCE_DMMV + + if (use_mul_mat_vec_q) { + ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true); } else { - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false); + ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false); + } + } else { + if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) { + ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true); + } else { + ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); } } } else { @@ -6329,8 +6690,7 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_ } void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_scale, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale); } void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -6359,8 +6719,8 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens const int64_t nb11 = src1->nb[1]; const int64_t nb12 = src1->nb[2]; - CUDA_CHECK(cudaSetDevice(g_main_device)); - cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device]; + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); + cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; const struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; const struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; @@ -6370,10 +6730,10 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { ggml_cpy_f32_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, - ne10, ne11, nb10, nb11, nb12, cudaStream_main); + ne10, ne11, nb10, nb11, nb12, main_stream); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, - ne10, ne11, nb10, nb11, nb12, cudaStream_main); + ne10, ne11, nb10, nb11, nb12, main_stream); } else { GGML_ASSERT(false); } @@ -6387,28 +6747,20 @@ void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens } void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_diag_mask_inf, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_diag_mask_inf); } void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_soft_max, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_soft_max); } void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented - - const int mode = ((int32_t *) dst->op_params)[2]; - const bool is_glm = mode & 4; - - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, !is_glm); // flatten support not implemented for glm + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rope); } void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); - ggml_cuda_op(src0, src1, dst, ggml_cuda_op_alibi, true, true); + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi); } void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -6418,7 +6770,7 @@ void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens } void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) { - int nrows = ggml_nrows(tensor); + const int64_t nrows = ggml_nrows(tensor); const int64_t ne0 = tensor->ne[0]; @@ -6428,14 +6780,14 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) { struct ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu; memset(extra, 0, sizeof(*extra)); - for (int id = 0; id < g_device_count; ++id) { + for (int64_t id = 0; id < g_device_count; ++id) { if (backend == GGML_BACKEND_GPU && id != g_main_device) { continue; } - cudaSetDevice(id); + ggml_cuda_set_device(id); - int row_low, row_high; + int64_t row_low, row_high; if (backend == GGML_BACKEND_GPU) { row_low = 0; row_high = nrows; @@ -6485,7 +6837,9 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) { extra->data_device[id] = buf; if (backend == GGML_BACKEND_GPU_SPLIT) { - CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id], cudaEventDisableTiming)); + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming)); + } } } @@ -6499,15 +6853,17 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - for (int id = 0; id < g_device_count; ++id) { + for (int64_t id = 0; id < g_device_count; ++id) { if (extra->data_device[id] != nullptr) { - CUDA_CHECK(cudaSetDevice(id)); + CUDA_CHECK(ggml_cuda_set_device(id)); CUDA_CHECK(cudaFree(extra->data_device[id])); } - if (extra->events[id] != nullptr) { - CUDA_CHECK(cudaSetDevice(id)); - CUDA_CHECK(cudaEventDestroy(extra->events[id])); + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + if (extra->events[id][is] != nullptr) { + CUDA_CHECK(ggml_cuda_set_device(id)); + CUDA_CHECK(cudaEventDestroy(extra->events[id][is])); + } } } @@ -6559,7 +6915,7 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo force_inplace; const size_t size = ggml_nbytes(tensor); - CUDA_CHECK(cudaSetDevice(g_main_device)); + CUDA_CHECK(ggml_cuda_set_device(g_main_device)); if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) { struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra; char * src0_ddc = (char *) src0_extra->data_device[g_main_device]; diff --git a/ggml-metal.m b/ggml-metal.m index 7e2355ce6..1139ee311 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -63,7 +63,10 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(relu); GGML_METAL_DECL_KERNEL(gelu); GGML_METAL_DECL_KERNEL(soft_max); + GGML_METAL_DECL_KERNEL(soft_max_4); GGML_METAL_DECL_KERNEL(diag_mask_inf); + GGML_METAL_DECL_KERNEL(diag_mask_inf_8); + GGML_METAL_DECL_KERNEL(get_rows_f32); GGML_METAL_DECL_KERNEL(get_rows_f16); GGML_METAL_DECL_KERNEL(get_rows_q4_0); GGML_METAL_DECL_KERNEL(get_rows_q4_1); @@ -75,8 +78,10 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(get_rows_q6_K); GGML_METAL_DECL_KERNEL(rms_norm); GGML_METAL_DECL_KERNEL(norm); + GGML_METAL_DECL_KERNEL(mul_mat_f32_f32); GGML_METAL_DECL_KERNEL(mul_mat_f16_f32); GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row); + GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_l4); GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32); GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32); GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32); @@ -85,6 +90,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32); GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32); GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32); + GGML_METAL_DECL_KERNEL(mul_mm_f32_f32); GGML_METAL_DECL_KERNEL(mul_mm_f16_f32); GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32); GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32); @@ -117,14 +123,17 @@ static NSString * const msl_library_source = @"see metal.metal"; struct ggml_metal_context * ggml_metal_init(int n_cb) { metal_printf("%s: allocating\n", __func__); - // Show all the Metal device instances in the system - NSArray * devices = MTLCopyAllDevices(); id device; NSString * s; + +#if TARGET_OS_OSX + // Show all the Metal device instances in the system + NSArray * devices = MTLCopyAllDevices(); for (device in devices) { s = [device name]; metal_printf("%s: found device: %s\n", __func__, [s UTF8String]); } +#endif // Pick and show default Metal device device = MTLCreateSystemDefaultDevice(); @@ -139,14 +148,22 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { ctx->n_buffers = 0; ctx->concur_list_len = 0; - ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT); + ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); -#if 0 - // compile from source string and show compile log +#ifdef GGML_SWIFT + // load the default.metallib file { NSError * error = nil; - ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error]; + NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; + NSString * llamaBundlePath = [bundle pathForResource:@"llama_llama" ofType:@"bundle"]; + NSBundle * llamaBundle = [NSBundle bundleWithPath:llamaBundlePath]; + NSString * libPath = [llamaBundle pathForResource:@"default" ofType:@"metallib"]; + NSURL * libURL = [NSURL fileURLWithPath:libPath]; + + // Load the metallib file into a Metal library + ctx->library = [ctx->device newLibraryWithURL:libURL error:&error]; + if (error) { metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; @@ -161,7 +178,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"]; NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; - NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]); NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error]; @@ -207,7 +224,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(relu); GGML_METAL_ADD_KERNEL(gelu); GGML_METAL_ADD_KERNEL(soft_max); + GGML_METAL_ADD_KERNEL(soft_max_4); GGML_METAL_ADD_KERNEL(diag_mask_inf); + GGML_METAL_ADD_KERNEL(diag_mask_inf_8); + GGML_METAL_ADD_KERNEL(get_rows_f32); GGML_METAL_ADD_KERNEL(get_rows_f16); GGML_METAL_ADD_KERNEL(get_rows_q4_0); GGML_METAL_ADD_KERNEL(get_rows_q4_1); @@ -219,8 +239,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(get_rows_q6_K); GGML_METAL_ADD_KERNEL(rms_norm); GGML_METAL_ADD_KERNEL(norm); + GGML_METAL_ADD_KERNEL(mul_mat_f32_f32); GGML_METAL_ADD_KERNEL(mul_mat_f16_f32); GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row); + GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_l4); GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32); GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32); GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32); @@ -229,6 +251,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32); GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32); GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32); + GGML_METAL_ADD_KERNEL(mul_mm_f32_f32); GGML_METAL_ADD_KERNEL(mul_mm_f16_f32); GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32); GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32); @@ -247,13 +270,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { #undef GGML_METAL_ADD_KERNEL } - metal_printf("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); metal_printf("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); +#if TARGET_OS_OSX + metal_printf("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); if (ctx->device.maxTransferRate != 0) { metal_printf("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); } else { metal_printf("%s: maxTransferRate = built-in GPU\n", __func__); } +#endif return ctx; } @@ -273,7 +298,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(relu); GGML_METAL_DEL_KERNEL(gelu); GGML_METAL_DEL_KERNEL(soft_max); + GGML_METAL_DEL_KERNEL(soft_max_4); GGML_METAL_DEL_KERNEL(diag_mask_inf); + GGML_METAL_DEL_KERNEL(diag_mask_inf_8); + GGML_METAL_DEL_KERNEL(get_rows_f32); GGML_METAL_DEL_KERNEL(get_rows_f16); GGML_METAL_DEL_KERNEL(get_rows_q4_0); GGML_METAL_DEL_KERNEL(get_rows_q4_1); @@ -285,8 +313,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(get_rows_q6_K); GGML_METAL_DEL_KERNEL(rms_norm); GGML_METAL_DEL_KERNEL(norm); + GGML_METAL_DEL_KERNEL(mul_mat_f32_f32); GGML_METAL_DEL_KERNEL(mul_mat_f16_f32); GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row); + GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_l4); GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32); GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32); GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32); @@ -295,6 +325,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32); GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32); GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32); + GGML_METAL_DEL_KERNEL(mul_mm_f32_f32); GGML_METAL_DEL_KERNEL(mul_mm_f16_f32); GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32); GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32); @@ -365,6 +396,7 @@ static id ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru for (int i = 0; i < ctx->n_buffers; ++i) { const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data; + //metal_printf("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name); if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) { *offs = (size_t) ioffs; @@ -454,6 +486,7 @@ bool ggml_metal_add_buffer( } } +#if TARGET_OS_OSX metal_printf(", (%8.2f / %8.2f)", ctx->device.currentAllocatedSize / 1024.0 / 1024.0, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); @@ -463,6 +496,9 @@ bool ggml_metal_add_buffer( } else { metal_printf("\n"); } +#else + metal_printf(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0); +#endif } return true; @@ -698,6 +734,7 @@ void ggml_metal_graph_compute( case GGML_OP_ADD: { GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); // utilize float4 GGML_ASSERT(ne00 % 4 == 0); @@ -705,6 +742,7 @@ void ggml_metal_graph_compute( if (ggml_nelements(src1) == ne10) { // src1 is a row + GGML_ASSERT(ne11 == 1); [encoder setComputePipelineState:ctx->pipeline_add_row]; } else { [encoder setComputePipelineState:ctx->pipeline_add]; @@ -721,6 +759,7 @@ void ggml_metal_graph_compute( case GGML_OP_MUL: { GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); // utilize float4 GGML_ASSERT(ne00 % 4 == 0); @@ -728,6 +767,7 @@ void ggml_metal_graph_compute( if (ggml_nelements(src1) == ne10) { // src1 is a row + GGML_ASSERT(ne11 == 1); [encoder setComputePipelineState:ctx->pipeline_mul_row]; } else { [encoder setComputePipelineState:ctx->pipeline_mul]; @@ -743,6 +783,8 @@ void ggml_metal_graph_compute( } break; case GGML_OP_SCALE: { + GGML_ASSERT(ggml_is_contiguous(src0)); + const float scale = *(const float *) src1->data; [encoder setComputePipelineState:ctx->pipeline_scale]; @@ -750,7 +792,7 @@ void ggml_metal_graph_compute( [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; [encoder setBytes:&scale length:sizeof(scale) atIndex:2]; - const int64_t n = ggml_nelements(dst); + const int64_t n = ggml_nelements(dst)/4; [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -762,7 +804,7 @@ void ggml_metal_graph_compute( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - const int64_t n = ggml_nelements(dst); + const int64_t n = ggml_nelements(dst)/4; [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -782,7 +824,7 @@ void ggml_metal_graph_compute( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - const int64_t n = ggml_nelements(dst); + const int64_t n = ggml_nelements(dst)/4; [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -796,13 +838,16 @@ void ggml_metal_graph_compute( { const int nth = 32; - [encoder setComputePipelineState:ctx->pipeline_soft_max]; + if (ne00%4 == 0) { + [encoder setComputePipelineState:ctx->pipeline_soft_max_4]; + } else { + [encoder setComputePipelineState:ctx->pipeline_soft_max]; + } [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; @@ -810,14 +855,23 @@ void ggml_metal_graph_compute( { const int n_past = ((int32_t *)(dst->op_params))[0]; - [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf]; + if (ne00%8 == 0) { + [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf_8]; + } else { + [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf]; + } [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; [encoder setBytes:&n_past length:sizeof(int) atIndex:4]; - [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + if (ne00%8 == 0) { + [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } + else { + [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } } break; case GGML_OP_MUL_MAT: { @@ -830,13 +884,14 @@ void ggml_metal_graph_compute( // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel - if (ggml_is_contiguous(src0) && - ggml_is_contiguous(src1) && + if (!ggml_is_transposed(src0) && + !ggml_is_transposed(src1) && src1t == GGML_TYPE_F32 && [ctx->device supportsFamily:MTLGPUFamilyApple7] && ne00%32 == 0 && ne11 > 1) { switch (src0->type) { + case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32]; break; case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break; case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break; case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break; @@ -856,25 +911,38 @@ void ggml_metal_graph_compute( [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5]; [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6]; [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9]; - [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:8]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:9]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:10]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:11]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:12]; + [encoder setBytes:&gqa length:sizeof(gqa) atIndex:13]; [encoder setThreadgroupMemoryLength:8192 atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; } else { int nth0 = 32; int nth1 = 1; + int nrows = 1; // use custom matrix x vector kernel switch (src0t) { + case GGML_TYPE_F32: + { + [encoder setComputePipelineState:ctx->pipeline_mul_mat_f32_f32]; + nrows = 4; + } break; case GGML_TYPE_F16: { nth0 = 32; nth1 = 1; if (ne11 * ne12 < 4) { [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row]; + } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { + [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_l4]; + nrows = ne11; } else { [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32]; + nrows = 4; } } break; case GGML_TYPE_Q4_0: @@ -995,7 +1063,7 @@ void ggml_metal_graph_compute( else if (src0t == GGML_TYPE_Q6_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else { - int64_t ny = (ne11 + 3)/4; + int64_t ny = (ne11 + nrows - 1)/nrows; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } } @@ -1003,6 +1071,7 @@ void ggml_metal_graph_compute( case GGML_OP_GET_ROWS: { switch (src0->type) { + case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_get_rows_f32]; break; case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break; case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break; case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break; @@ -1018,9 +1087,9 @@ void ggml_metal_graph_compute( [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:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4]; - [encoder setBytes:&(dst->nb[1]) length:sizeof(uint64_t) atIndex:5]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:5]; const int64_t n = ggml_nelements(src1); diff --git a/ggml-metal.metal b/ggml-metal.metal index 5070561fb..7f1c3d9ea 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -38,7 +38,7 @@ kernel void kernel_add_row( device const float4 * src0, device const float4 * src1, device float4 * dst, - constant int64_t & nb, + constant int64_t & nb, uint tpig[[thread_position_in_grid]]) { dst[tpig] = src0[tpig] + src1[tpig % nb]; } @@ -63,18 +63,18 @@ kernel void kernel_mul_row( } kernel void kernel_scale( - device const float * src0, - device float * dst, + device const float4 * src0, + device float4 * dst, constant float & scale, uint tpig[[thread_position_in_grid]]) { dst[tpig] = src0[tpig] * scale; } kernel void kernel_silu( - device const float * src0, - device float * dst, + device const float4 * src0, + device float4 * dst, uint tpig[[thread_position_in_grid]]) { - float x = src0[tpig]; + device const float4 & x = src0[tpig]; dst[tpig] = x / (1.0f + exp(-x)); } @@ -89,10 +89,10 @@ constant float GELU_COEF_A = 0.044715f; constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; kernel void kernel_gelu( - device const float * src0, - device float * dst, + device const float4 * src0, + device float4 * dst, uint tpig[[thread_position_in_grid]]) { - float x = src0[tpig]; + device const float4 & x = src0[tpig]; // BEWARE !!! // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs! @@ -107,7 +107,6 @@ kernel void kernel_soft_max( constant int64_t & ne00, constant int64_t & ne01, constant int64_t & ne02, - threadgroup float * buf [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -119,64 +118,70 @@ kernel void kernel_soft_max( device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; // parallel max - buf[tpitg[0]] = -INFINITY; - for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { - buf[tpitg[0]] = MAX(buf[tpitg[0]], psrc0[i00]); + float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY; + for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) { + lmax = MAX(lmax, psrc0[i00]); } - - // reduce - threadgroup_barrier(mem_flags::mem_threadgroup); - for (uint i = ntg[0]/2; i > 0; i /= 2) { - if (tpitg[0] < i) { - buf[tpitg[0]] = MAX(buf[tpitg[0]], buf[tpitg[0] + i]); - } - threadgroup_barrier(mem_flags::mem_threadgroup); - } - - //// broadcast - not needed. There is a threadgroup barrier above in the last iteration of - // the loop, and when that is done, buf[0] has the correct (synchronized) value - //if (tpitg[0] == 0) { - // buf[0] = buf[0]; - //} - - //threadgroup_barrier(mem_flags::mem_threadgroup); - - const float max = buf[0]; + const float max = simd_max(lmax); // parallel sum - buf[tpitg[0]] = 0.0f; + float lsum = 0.0f; for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { const float exp_psrc0 = exp(psrc0[i00] - max); - buf[tpitg[0]] += exp_psrc0; + lsum += exp_psrc0; // Remember the result of exp here. exp is expensive, so we really do not // whish to compute it twice. pdst[i00] = exp_psrc0; } - // reduce - threadgroup_barrier(mem_flags::mem_threadgroup); - for (uint i = ntg[0]/2; i > 0; i /= 2) { - if (tpitg[0] < i) { - buf[tpitg[0]] += buf[tpitg[0] + i]; - } - threadgroup_barrier(mem_flags::mem_threadgroup); - } - - // broadcast - not needed, see above - //// broadcast - //if (tpitg[0] == 0) { - // buf[0] = buf[0]; - //} - - //threadgroup_barrier(mem_flags::mem_threadgroup); - - const float sum = buf[0]; + const float sum = simd_sum(lsum); for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { pdst[i00] /= sum; } } +kernel void kernel_soft_max_4( + device const float * src0, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t i03 = tgpig[2]; + const int64_t i02 = tgpig[1]; + const int64_t i01 = tgpig[0]; + + device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + + // parallel max + float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY; + for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) { + lmax4 = fmax(lmax4, psrc4[i00]); + } + float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3])); + + const float max = simd_max(lmax); + + // parallel sum + float4 lsum4 = 0.0f; + for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) { + const float4 exp_psrc4 = exp(psrc4[i00] - max); + lsum4 += exp_psrc4; + pdst4[i00] = exp_psrc4; + } + float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3]; + + const float sum = simd_sum(lsum); + + for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) { + pdst4[i00] /= sum; + } +} + kernel void kernel_diag_mask_inf( device const float * src0, device float * dst, @@ -192,6 +197,33 @@ kernel void kernel_diag_mask_inf( dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY; } else { dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00]; + } +} + +kernel void kernel_diag_mask_inf_8( + device const float4 * src0, + device float4 * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int & n_past, + uint3 tpig[[thread_position_in_grid]]) { + + const int64_t i = 2*tpig[0]; + + dst[i+0] = src0[i+0]; + dst[i+1] = src0[i+1]; + int64_t i4 = 4*i; + const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01; + const int64_t i01 = i4/(ne00); i4 -= i01*ne00; + const int64_t i00 = i4; + for (int k = 3; k >= 0; --k) { + if (i00 + 4 + k <= n_past + i01) { + break; + } + dst[i+1][k] = -INFINITY; + if (i00 + k > n_past + i01) { + dst[i][k] = -INFINITY; + } } } @@ -491,6 +523,79 @@ kernel void kernel_mul_mat_q8_0_f32( } } +#define N_F32_F32 4 + +kernel void kernel_mul_mat_f32_f32( + device const char * src0, + device const char * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]]) { + + const int64_t r0 = tgpig.x; + const int64_t rb = tgpig.y*N_F32_F32; + const int64_t im = tgpig.z; + + device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); + + if (ne00 < 128) { + for (int row = 0; row < N_F32_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + + float sumf = 0; + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + device const float4 * x4 = (device const float4 *)x; + for (int row = 0; row < N_F32_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + device const float4 * y4 = (device const float4 *) y; + + float sumf = 0; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } +} + kernel void kernel_mul_mat_f16_f32_1row( device const char * src0, device const char * src1, @@ -616,6 +721,49 @@ kernel void kernel_mul_mat_f16_f32( } } +// Assumes row size (ne00) is a multiple of 4 +kernel void kernel_mul_mat_f16_f32_l4( + device const char * src0, + device const char * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]]) { + + const int nrows = ne11; + const int64_t r0 = tgpig.x; + const int64_t im = tgpig.z; + + device const half4 * x4 = (device const half4 *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); + + for (int r1 = 0; r1 < nrows; ++r1) { + device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12); + + float sumf = 0; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } +} + kernel void kernel_alibi_f32( device const float * src0, device float * dst, @@ -1123,31 +1271,40 @@ kernel void kernel_mul_mat_q3_K_f32( device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0; device const float * yy = (device const float *) src1 + r1*ne10 + r2*ne00*ne1; - float yl[16]; + float yl[32]; - const uint16_t kmask1 = 0x0303; + const uint16_t kmask1 = 0x3030; const uint16_t kmask2 = 0x0f0f; - const int tid = tiisg/2; - const int ix = tiisg%2; - const int ip = tid/8; // 0 or 1 - const int il = tid/2 - 4*ip; // 0...3 + const int tid = tiisg/4; + const int ix = tiisg%4; + const int ip = tid/4; // 0 or 1 + const int il = 2*((tid%4)/2); // 0 or 2 const int ir = tid%2; const int n = 8; const int l0 = n*ir; - const uint16_t m1 = 1 << (4*ip + il); - const uint16_t m2 = m1 << 8; + // One would think that the Metal compiler would figure out that ip and il can only have + // 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it + // with these two tales. + // + // Possible masks for the high bit + const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200}, // ip = 0, il = 0 + {0x0004, 0x0400, 0x0008, 0x0800}, // ip = 0, il = 2 + {0x0010, 0x1000, 0x0020, 0x2000}, // ip = 1, il = 0 + {0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2 + + // Possible masks for the low 2 bits + const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}}; + + const ushort4 hm = mm[2*ip + il/2]; const int shift = 2*il; - const uint16_t qm1 = 0x0003 << shift; - const uint16_t qm2 = 0x0300 << shift; - const int32_t v1 = 4 << shift; - const int32_t v2 = 1024 << shift; + const float v1 = il == 0 ? 4.f : 64.f; + const float v2 = 4.f * v1; const uint16_t s_shift1 = 4*ip; - const uint16_t s_shift2 = s_shift1 + 2*(il/2); - const int ik = 4 + (il%2); + const uint16_t s_shift2 = s_shift1 + il; const int q_offset = 32*ip + l0; const int y_offset = 128*ip + 32*il + l0; @@ -1156,12 +1313,19 @@ kernel void kernel_mul_mat_q3_K_f32( device const float * y1 = yy + ix*QK_K + y_offset; - float sumf1[2] = {0.f}, sumf2[2] = {0.f}; - for (int i = ix; i < nb; i += 2) { + uint32_t scales32, aux32; + thread uint16_t * scales16 = (thread uint16_t *)&scales32; + thread const int8_t * scales = (thread const int8_t *)&scales32; + + float sumf1[2] = {0.f}; + float sumf2[2] = {0.f}; + for (int i = ix; i < nb; i += 4) { for (int l = 0; l < 8; ++l) { - yl[l+0] = y1[l+ 0]; - yl[l+8] = y1[l+16]; + yl[l+ 0] = y1[l+ 0]; + yl[l+ 8] = y1[l+16]; + yl[l+16] = y1[l+32]; + yl[l+24] = y1[l+48]; } device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset); @@ -1172,27 +1336,43 @@ kernel void kernel_mul_mat_q3_K_f32( for (int row = 0; row < 2; ++row) { const float d_all = (float)dh[0]; - const char2 scales = as_type((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4))); - float s1 = 0, s2 = 0; - for (int l = 0; l < n; l += 2) { - const uint16_t qs = q[l/2]; - s1 += yl[l+0] * ((int32_t)(qs & qm1) - ((h[l/2] & m1) ? 0 : v1)); - s2 += yl[l+1] * ((int32_t)(qs & qm2) - ((h[l/2] & m2) ? 0 : v2)); - } - float d = d_all * (s1 + 1.f/256.f * s2); - sumf1[row] += d * scales[0]; - sumf2[row] += d; + scales16[0] = a[4]; + scales16[1] = a[5]; + aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030; + scales16[0] = a[il+0]; + scales16[1] = a[il+1]; + scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32; - s1 = s2 = 0; + float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0; for (int l = 0; l < n; l += 2) { - const uint16_t qs = q[l/2+8]; - s1 += yl[l+8] * ((int32_t)(qs & qm1) - ((h[l/2+8] & m1) ? 0 : v1)); - s2 += yl[l+9] * ((int32_t)(qs & qm2) - ((h[l/2+8] & m2) ? 0 : v2)); + const int32_t qs = q[l/2]; + s1 += yl[l+0] * (qs & qm[il/2][0]); + s2 += yl[l+1] * (qs & qm[il/2][1]); + s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]); + s4 += yl[l+16] * (qs & qm[il/2][2]); + s5 += yl[l+17] * (qs & qm[il/2][3]); + s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]); } - d = d_all * (s1 + 1.f/256.f * s2); - sumf1[row] += d * scales[1]; - sumf2[row] += d; + float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1); + float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2); + sumf1[row] += d1 * (scales[0] - 32); + sumf2[row] += d2 * (scales[2] - 32); + + s1 = s2 = s3 = s4 = s5 = s6 = 0; + for (int l = 0; l < n; l += 2) { + const int32_t qs = q[l/2+8]; + s1 += yl[l+8] * (qs & qm[il/2][0]); + s2 += yl[l+9] * (qs & qm[il/2][1]); + s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]); + s4 += yl[l+24] * (qs & qm[il/2][2]); + s5 += yl[l+25] * (qs & qm[il/2][3]); + s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]); + } + d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1); + d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2); + sumf1[row] += d1 * (scales[1] - 32); + sumf2[row] += d2 * (scales[3] - 32); q += step; h += step; @@ -1201,15 +1381,17 @@ kernel void kernel_mul_mat_q3_K_f32( } - y1 += 2 * QK_K; + y1 += 4 * QK_K; } for (int row = 0; row < 2; ++row) { - const float sumf = (sumf1[row] - 32.f*sumf2[row]) / (1 << shift); - const float tot = simd_sum(sumf); - if (tiisg == 0) { - dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = tot; + const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift); + sumf1[row] = simd_sum(sumf); + } + if (tiisg == 0) { + for (int row = 0; row < 2; ++row) { + dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = sumf1[row]; } } } @@ -1290,13 +1472,13 @@ kernel void kernel_mul_mat_q4_K_f32( device const float * src1, device float * dst, constant int64_t & ne00, - constant int64_t & ne01[[buffer(4)]], - constant int64_t & ne02[[buffer(5)]], - constant int64_t & ne10[[buffer(9)]], - constant int64_t & ne12[[buffer(11)]], - constant int64_t & ne0[[buffer(15)]], - constant int64_t & ne1[[buffer(16)]], - constant uint & gqa[[buffer(17)]], + constant int64_t & ne01 [[buffer(4)]], + constant int64_t & ne02 [[buffer(5)]], + constant int64_t & ne10 [[buffer(9)]], + constant int64_t & ne12 [[buffer(11)]], + constant int64_t & ne0 [[buffer(15)]], + constant int64_t & ne1 [[buffer(16)]], + constant uint & gqa [[buffer(17)]], uint3 tgpig[[threadgroup_position_in_grid]], uint tiisg[[thread_index_in_simdgroup]], uint sgitg[[simdgroup_index_in_threadgroup]]) { @@ -1564,17 +1746,25 @@ kernel void kernel_mul_mat_q5_K_f32( sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2); sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2); - float4 acc = {0.f, 0.f, 0.f, 0.f}; + float4 acc1 = {0.f}; + float4 acc2 = {0.f}; for (int l = 0; l < n; ++l) { uint8_t h = qh[l]; - acc[0] += yl[l+0] * ((uint16_t)(q1[l] & 0x0F) + (h & hm1 ? 16 : 0)); - acc[1] += yl[l+8] * ((uint16_t)(q1[l] & 0xF0) + (h & hm2 ? 256 : 0)); - acc[2] += yh[l+0] * ((uint16_t)(q2[l] & 0x0F) + (h & hm3 ? 16 : 0)); - acc[3] += yh[l+8] * ((uint16_t)(q2[l] & 0xF0) + (h & hm4 ? 256 : 0)); + acc1[0] += yl[l+0] * (q1[l] & 0x0F); + acc1[1] += yl[l+8] * (q1[l] & 0xF0); + acc1[2] += yh[l+0] * (q2[l] & 0x0F); + acc1[3] += yh[l+8] * (q2[l] & 0xF0); + acc2[0] += h & hm1 ? yl[l+0] : 0.f; + acc2[1] += h & hm2 ? yl[l+8] : 0.f; + acc2[2] += h & hm3 ? yh[l+0] : 0.f; + acc2[3] += h & hm4 ? yh[l+8] : 0.f; } const float dall = dh[0]; const float dmin = dh[1]; - sumf[row] += dall * (acc[0] * sc8[0] + acc[1] * sc8[1] * 1.f/16.f + acc[2] * sc8[4] + acc[3] * sc8[5] * 1.f/16.f) - + sumf[row] += dall * (sc8[0] * (acc1[0] + 16.f*acc2[0]) + + sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) + + sc8[4] * (acc1[2] + 16.f*acc2[2]) + + sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) - dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]); q1 += step; @@ -1747,6 +1937,15 @@ kernel void kernel_mul_mat_q6_K_f32( //============================= templates and their specializations ============================= +// NOTE: this is not dequantizing - we are simply fitting the template +template +void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) { + float4x4 temp = *(((device float4x4 *)src)); + for (int i = 0; i < 16; i++){ + reg[i/4][i%4] = temp[i/4][i%4]; + } +} + template void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) { half4x4 temp = *(((device half4x4 *)src)); @@ -1758,28 +1957,30 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) template void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 1); - const half d = il ? (xb->d / 16.h) : xb->d; - const half m = il ? ( -8.h * 16.h) : -8.h; + const float d1 = il ? (xb->d / 16.h) : xb->d; + const float d2 = d1 / 256.f; + const float md = -8.h * xb->d; const ushort mask0 = il ? 0x00F0 : 0x000F; - const ushort mask1 = il ? 0xF000 : 0x0F00; + const ushort mask1 = mask0 << 8; for (int i=0;i<8;i++) { - reg[i/2][2*(i%2)] = (((qs[i] & mask0) ) + m) * d; - reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) + m) * d; + reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md; + reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md; } } template 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); - const half d = il ? (xb->d / 16.h) : xb->d; - const half m = xb->m; + const float d1 = il ? (xb->d / 16.h) : xb->d; + const float d2 = d1 / 256.f; + const float m = xb->m; const ushort mask0 = il ? 0x00F0 : 0x000F; - const ushort mask1 = il ? 0xF000 : 0x0F00; + const ushort mask1 = mask0 << 8; for (int i=0;i<8;i++) { - reg[i/2][2*(i%2)] = (((qs[i] & mask0) ) * d) + m; - reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) * d) + m; + reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m; + reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m; } } @@ -1815,7 +2016,7 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg template void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) { - const float d_all = (float)(xb->d); + const half d_all = xb->d; device const uint8_t * q = (device const uint8_t *)xb->qs; device const uint8_t * h = (device const uint8_t *)xb->hmask; device const int8_t * scales = (device const int8_t *)xb->scales; @@ -1828,16 +2029,18 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg ((il/4)>0 ? 12 : 3); uint16_t kmask2 = il/8 ? 0xF0 : 0x0F; uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4]; - int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2) : \ - (scale_2&kmask2) | ((scale_1&kmask1) << 4); - float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f); + int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2) + : (scale_2&kmask2) | ((scale_1&kmask1) << 4); + half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h); + const half ml = 4.h * dl; - il = (il/2)%4; - float coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h); - uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); + il = (il/2) & 3; + const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h); + const uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); + dl *= coef; for (int i = 0; i < 16; ++i) { - reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i] & m) ? 0 : 4.f/coef)); + reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml); } #else float kcoef = il&1 ? 1.f/16.f : 1.f; @@ -1852,26 +2055,31 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg #endif } +static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) { + return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)} + : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))}; +} + template void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) { - device const uint8_t * q = xb->qs; + device const uchar * q = xb->qs; #if QK_K == 256 - const float d = (float)(xb->d); - const float min = (float)(xb->dmin); short is = (il/4) * 2; q = q + (il/4) * 32 + 16 * (il&1); - il = il%4; - const uchar4 sc = get_scale_min_k4(is, xb->scales); - const float dl = il<2 ? d * sc[0] : d * sc[2]/16.h; - const float ml = il<2 ? min * sc[1] : min * sc[3]; + il = il & 3; + const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales); + const half d = il < 2 ? xb->d : xb->d / 16.h; + const half min = xb->dmin; + const half dl = d * sc[0]; + const half ml = min * sc[1]; #else q = q + 16 * (il&1); device const uint8_t * s = xb->scales; device const half2 * dh = (device const half2 *)xb->d; const float2 d = (float2)dh[0]; const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h; - const float ml = il<2 ? d[1] * (s[0]>>4) : d[1 ]* (s[1]>>4); + const float ml = il<2 ? d[1] * (s[0]>>4) : d[1] * (s[1]>>4); #endif const ushort mask = il<2 ? 0x0F : 0xF0; for (int i = 0; i < 16; ++i) { @@ -1885,19 +2093,19 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg device const uint8_t * qh = xb->qh; #if QK_K == 256 - const float d = (float)(xb->d); - const float min = (float)(xb->dmin); short is = (il/4) * 2; q = q + 32 * (il/4) + 16 * (il&1); qh = qh + 16 * (il&1); uint8_t ul = 1 << (il/2); - il = il%4; - const uchar4 sc = get_scale_min_k4(is, xb->scales); - const float dl = il<2 ? d * sc[0] : d * sc[2]/16.h; - const float ml = il<2 ? min * sc[1] : min * sc[3]; + il = il & 3; + const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales); + const half d = il < 2 ? xb->d : xb->d / 16.h; + const half min = xb->dmin; + const half dl = d * sc[0]; + const half ml = min * sc[1]; - const ushort mask = il<2 ? 0x0F : 0xF0; - const float qh_val = il<2 ? 16.f : 256.f; + const ushort mask = il<2 ? 0x0F : 0xF0; + const half qh_val = il<2 ? 16.h : 256.h; for (int i = 0; i < 16; ++i) { reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml; } @@ -1916,7 +2124,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg template void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) { - const float d_all = (float)(xb->d); + const half d_all = xb->d; device const uint8_t * ql = (device const uint8_t *)xb->ql; device const uint8_t * qh = (device const uint8_t *)xb->qh; device const int8_t * scales = (device const int8_t *)xb->scales; @@ -1924,19 +2132,21 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg #if QK_K == 256 ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1); qh = qh + 32*(il/8) + 16*(il&1); - float sc = scales[(il%2) + 2 * ((il/2))]; - il = (il/2)%4; + half sc = scales[(il%2) + 2 * ((il/2))]; + il = (il/2) & 3; #else ql = ql + 16 * (il&1); - float sc = scales[il]; + half sc = scales[il]; #endif + const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); + const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F; + const half coef = il>1 ? 1.f/16.h : 1.h; + const half ml = d_all * sc * 32.h; + const half dl = d_all * sc * coef; for (int i = 0; i < 16; ++i) { - uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); - uint16_t kmask2 = il>1 ? 0xF0 : 0x0F; - const float coef = il>1 ? 1.f/16.f : 1.f; - float q = il&1 ? ((ql[i]&kmask2)|((qh[i]&kmask1)<<2)) - 32.f/coef : \ - ((ql[i]&kmask2)|((qh[i]&kmask1)<<4)) - 32.f/coef; - reg[i/4][i%4] = d_all * sc * q * coef; + const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2)) + : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4)); + reg[i/4][i%4] = dl * q - ml; } } @@ -1976,22 +2186,25 @@ kernel void kernel_get_rows( // each block_q contains 16*nl weights template kernel void kernel_mul_mm(device const uchar * src0, - device const float * src1, - device float * dst, - constant int64_t & ne00, - constant int64_t & ne02, - constant int64_t & nb01, - constant int64_t & nb02, - constant int64_t & ne12, - constant int64_t & ne0, - constant int64_t & ne1, - constant uint & gqa, - threadgroup uchar * shared_memory [[threadgroup(0)]], - uint3 tgpig[[threadgroup_position_in_grid]], - uint tiitg[[thread_index_in_threadgroup]], - uint sgitg[[simdgroup_index_in_threadgroup]]) { + device const uchar * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne02, + constant int64_t & nb01, + constant int64_t & nb02, + constant int64_t & ne12, + constant int64_t & nb10, + constant int64_t & nb11, + constant int64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + constant uint & gqa, + threadgroup uchar * shared_memory [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiitg[[thread_index_in_threadgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]]) { - threadgroup half * sa = ((threadgroup half *)shared_memory); + threadgroup half * sa = (threadgroup half *)(shared_memory); threadgroup float * sb = (threadgroup float *)(shared_memory + 4096); const uint r0 = tgpig.y; @@ -2004,7 +2217,7 @@ kernel void kernel_mul_mm(device const uchar * src0, short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1; short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1; - simdgroup_half8x8 ma[4]; + simdgroup_half8x8 ma[4]; simdgroup_float8x8 mb[2]; simdgroup_float8x8 c_res[8]; for (int i = 0; i < 8; i++){ @@ -2012,10 +2225,15 @@ kernel void kernel_mul_mm(device const uchar * src0, } short il = (tiitg % THREAD_PER_ROW); - uint offset0 = im/gqa*nb02; ushort offset1 = il/nl; - device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1; - device const float * y = src1 + (r1 * BLOCK_SIZE_N + thread_col) * ne00 \ - + BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL) + im * ne00 * ne1; + + uint offset0 = im/gqa*nb02; + ushort offset1 = il/nl; + + device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1; + device const float * y = (device const float *)(src1 + + nb12 * im + + nb11 * (r1 * BLOCK_SIZE_N + thread_col) + + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL))); for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) { //load data and store to threadgroup memory @@ -2095,6 +2313,7 @@ kernel void kernel_mul_mm(device const uchar * src0, typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \ constant uint64_t &, constant uint64_t &, uint, uint, uint); +template [[host_name("kernel_get_rows_f32")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows; @@ -2105,14 +2324,28 @@ template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows; -typedef void (mat_mm_t)(device const uchar *, device const float *, device float *, constant int64_t &,\ - constant int64_t &, constant int64_t &, constant int64_t &, constant int64_t &, \ - constant int64_t &, constant int64_t &, constant uint &, threadgroup uchar *, uint3, uint, uint); +typedef void (mat_mm_t)( + device const uchar * src0, + device const uchar * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne02, + constant int64_t & nb01, + constant int64_t & nb02, + constant int64_t & ne12, + constant int64_t & nb10, + constant int64_t & nb11, + constant int64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + constant uint & gqa, + threadgroup uchar *, uint3, uint, uint); -template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm; diff --git a/ggml.c b/ggml.c index 8a677ab2a..a0be068d6 100644 --- a/ggml.c +++ b/ggml.c @@ -1,4 +1,3 @@ -#define _GNU_SOURCE // Defines CLOCK_MONOTONIC on Linux #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows #include "ggml.h" @@ -47,6 +46,10 @@ // disable "possible loss of data" to avoid hundreds of casts // we should just be careful :) #pragma warning(disable: 4244 4267) + +// disable POSIX deprecation warnigns +// these functions are never going away, anyway +#pragma warning(disable: 4996) #endif #if defined(_WIN32) @@ -103,6 +106,9 @@ typedef void * thread_ret_t; #include #include +#endif +#ifdef GGML_USE_CPU_HBM +#include #endif // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512 @@ -192,8 +198,14 @@ typedef void * thread_ret_t; #define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr) #else inline static void * ggml_aligned_malloc(size_t size) { + if (size == 0) { + GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n"); + return NULL; + } void * aligned_memory = NULL; -#ifdef GGML_USE_METAL +#ifdef GGML_USE_CPU_HBM + int result = hbw_posix_memalign(&aligned_memory, 16, size); +#elif GGML_USE_METAL int result = posix_memalign(&aligned_memory, sysconf(_SC_PAGESIZE), size); #else int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size); @@ -215,8 +227,12 @@ inline static void * ggml_aligned_malloc(size_t size) { return aligned_memory; } #define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size) +#ifdef GGML_USE_CPU_HBM +#define GGML_ALIGNED_FREE(ptr) if(NULL != ptr) hbw_free(ptr) +#else #define GGML_ALIGNED_FREE(ptr) free(ptr) #endif +#endif #define UNUSED GGML_UNUSED #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0) @@ -267,7 +283,7 @@ typedef double ggml_float; // 16-bit float // on Arm, we use __fp16 // on x86, we use uint16_t -#ifdef __ARM_NEON +#if defined(__ARM_NEON) && !defined(_MSC_VER) // if YCM cannot find , make a symbolic link to it, for example: // @@ -294,12 +310,14 @@ typedef double ggml_float; #if defined(_MSC_VER) || defined(__MINGW32__) #include #else +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) #if !defined(__riscv) #include #endif #endif #endif #endif +#endif #ifdef __riscv_v_intrinsic #include @@ -4285,10 +4303,21 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) { } size_t ggml_nbytes(const struct ggml_tensor * tensor) { - size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type); - for (int i = 1; i < GGML_MAX_DIMS; ++i) { - nbytes += (tensor->ne[i] - 1)*tensor->nb[i]; + size_t nbytes; + size_t blck_size = ggml_blck_size(tensor->type); + if (blck_size == 1) { + nbytes = ggml_type_size(tensor->type); + for (int i = 0; i < GGML_MAX_DIMS; ++i) { + nbytes += (tensor->ne[i] - 1)*tensor->nb[i]; + } } + else { + nbytes = tensor->ne[0]*tensor->nb[0]/blck_size; + for (int i = 1; i < GGML_MAX_DIMS; ++i) { + nbytes += (tensor->ne[i] - 1)*tensor->nb[i]; + } + } + return nbytes; } @@ -4566,6 +4595,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { return NULL; } + // allow to call ggml_init with 0 size + if (params.mem_size == 0) { + params.mem_size = GGML_MEM_ALIGN; + } + const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN); *ctx = (struct ggml_context) { @@ -17260,10 +17294,18 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { } else { // wait for other threads to finish const int last = node_n; - do { - //sched_yield(); + while (true) { + // TODO: this sched_yield can have significant impact on the performance - either positive or negative + // depending on the workload and the operating system. + // since it is not clear what is the best approach, it should potentially become user-configurable + // ref: https://github.com/ggerganov/ggml/issues/291 +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + sched_yield(); +#endif + node_n = atomic_load(&state->shared->node_n); - } while (node_n == last); + if (node_n != last) break; + }; } // check if we should stop @@ -18314,10 +18356,11 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { for (int i = 0; i < cgraph->n_leafs; i++) { struct ggml_tensor * node = cgraph->leafs[i]; - GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n", + GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n", i, node->ne[0], node->ne[1], - ggml_op_name(node->op)); + ggml_op_name(node->op), + ggml_get_name(node)); } for (int i = 0; i < GGML_OP_COUNT; i++) { @@ -18854,7 +18897,6 @@ static enum ggml_opt_result linesearch_backtracking( // strong Wolfe condition (GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) return count; } - return count; } } @@ -20077,27 +20119,27 @@ const char * gguf_type_name(enum gguf_type type) { return GGUF_TYPE_NAME[type]; } -int gguf_get_version(struct gguf_context * ctx) { +int gguf_get_version(const struct gguf_context * ctx) { return ctx->header.version; } -size_t gguf_get_alignment(struct gguf_context * ctx) { +size_t gguf_get_alignment(const struct gguf_context * ctx) { return ctx->alignment; } -size_t gguf_get_data_offset(struct gguf_context * ctx) { +size_t gguf_get_data_offset(const struct gguf_context * ctx) { return ctx->offset; } -void * gguf_get_data(struct gguf_context * ctx) { +void * gguf_get_data(const struct gguf_context * ctx) { return ctx->data; } -int gguf_get_n_kv(struct gguf_context * ctx) { +int gguf_get_n_kv(const struct gguf_context * ctx) { return ctx->header.n_kv; } -int gguf_find_key(struct gguf_context * ctx, const char * key) { +int gguf_find_key(const struct gguf_context * ctx, const char * key) { // return -1 if key not found int keyfound = -1; @@ -20113,85 +20155,85 @@ int gguf_find_key(struct gguf_context * ctx, const char * key) { return keyfound; } -const char * gguf_get_key(struct gguf_context * ctx, int i) { +const char * gguf_get_key(const struct gguf_context * ctx, int i) { return ctx->kv[i].key.data; } -enum gguf_type gguf_get_kv_type(struct gguf_context * ctx, int i) { +enum gguf_type gguf_get_kv_type(const struct gguf_context * ctx, int i) { return ctx->kv[i].type; } -enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i) { +enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.arr.type; } -const void * gguf_get_arr_data(struct gguf_context * ctx, int i) { +const void * gguf_get_arr_data(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.arr.data; } -const char * gguf_get_arr_str(struct gguf_context * ctx, int key_id, int i) { +const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i) { struct gguf_kv * kv = &ctx->kv[key_id]; struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[i]; return str->data; } -int gguf_get_arr_n(struct gguf_context * ctx, int i) { +int gguf_get_arr_n(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.arr.n; } -uint8_t gguf_get_val_u8(struct gguf_context * ctx, int i) { +uint8_t gguf_get_val_u8(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.uint8; } -int8_t gguf_get_val_i8(struct gguf_context * ctx, int i) { +int8_t gguf_get_val_i8(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.int8; } -uint16_t gguf_get_val_u16(struct gguf_context * ctx, int i) { +uint16_t gguf_get_val_u16(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.uint16; } -int16_t gguf_get_val_i16(struct gguf_context * ctx, int i) { +int16_t gguf_get_val_i16(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.int16; } -uint32_t gguf_get_val_u32(struct gguf_context * ctx, int i) { +uint32_t gguf_get_val_u32(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.uint32; } -int32_t gguf_get_val_i32(struct gguf_context * ctx, int i) { +int32_t gguf_get_val_i32(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.int32; } -float gguf_get_val_f32(struct gguf_context * ctx, int i) { +float gguf_get_val_f32(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.float32; } -uint64_t gguf_get_val_u64(struct gguf_context * ctx, int i) { +uint64_t gguf_get_val_u64(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.uint64; } -int64_t gguf_get_val_i64(struct gguf_context * ctx, int i) { +int64_t gguf_get_val_i64(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.int64; } -double gguf_get_val_f64(struct gguf_context * ctx, int i) { +double gguf_get_val_f64(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.float64; } -bool gguf_get_val_bool(struct gguf_context * ctx, int i) { +bool gguf_get_val_bool(const struct gguf_context * ctx, int i) { return ctx->kv[i].value.bool_; } -const char * gguf_get_val_str (struct gguf_context * ctx, int i) { +const char * gguf_get_val_str (const struct gguf_context * ctx, int i) { return ctx->kv[i].value.str.data; } -int gguf_get_n_tensors(struct gguf_context * ctx) { +int gguf_get_n_tensors(const struct gguf_context * ctx) { return ctx->header.n_tensors; } -int gguf_find_tensor(struct gguf_context * ctx, const char * name) { +int gguf_find_tensor(const struct gguf_context * ctx, const char * name) { // return -1 if tensor not found int tensorfound = -1; @@ -20207,11 +20249,11 @@ int gguf_find_tensor(struct gguf_context * ctx, const char * name) { return tensorfound; } -size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i) { +size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i) { return ctx->infos[i].offset; } -char * gguf_get_tensor_name(struct gguf_context * ctx, int i) { +char * gguf_get_tensor_name(const struct gguf_context * ctx, int i) { return ctx->infos[i].name.data; } @@ -20494,7 +20536,7 @@ static void gguf_bwrite_el(struct gguf_buf * buf, const void * val, size_t el_si buf->offset += el_size; } -static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta) { +static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta) { // write header gguf_bwrite_el(buf, &ctx->header.magic, sizeof(ctx->header.magic)); gguf_bwrite_el(buf, &ctx->header.version, sizeof(ctx->header.version)); @@ -20609,7 +20651,7 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf, } } -void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta) { +void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) { FILE * file = fopen(fname, "wb"); if (!file) { GGML_ASSERT(false && "failed to open file for writing"); @@ -20626,7 +20668,7 @@ void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only fclose(file); } -size_t gguf_get_meta_size(struct gguf_context * ctx) { +size_t gguf_get_meta_size(const struct gguf_context * ctx) { // no allocs - only compute size struct gguf_buf buf = gguf_buf_init(0); @@ -20635,7 +20677,7 @@ size_t gguf_get_meta_size(struct gguf_context * ctx) { return buf.offset; } -void gguf_get_meta_data(struct gguf_context * ctx, void * data) { +void gguf_get_meta_data(const struct gguf_context * ctx, void * data) { struct gguf_buf buf = gguf_buf_init(16*1024); gguf_write_to_buf(ctx, &buf, true); @@ -20711,6 +20753,14 @@ int ggml_cpu_has_arm_fma(void) { #endif } +int ggml_cpu_has_metal(void) { +#if defined(GGML_USE_METAL) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_f16c(void) { #if defined(__F16C__) return 1; diff --git a/ggml.h b/ggml.h index c936823d6..f45456876 100644 --- a/ggml.h +++ b/ggml.h @@ -195,6 +195,14 @@ # define GGML_DEPRECATED(func, hint) func #endif +#ifndef __GNUC__ +# define GGML_ATTRIBUTE_FORMAT(...) +#elif defined(__MINGW32__) +# define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +#else +# define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) +#endif + #include #include #include @@ -685,6 +693,7 @@ extern "C" { GGML_API const char * ggml_get_name (const struct ggml_tensor * tensor); GGML_API struct ggml_tensor * ggml_set_name ( struct ggml_tensor * tensor, const char * name); + GGML_ATTRIBUTE_FORMAT(2, 3) GGML_API struct ggml_tensor * ggml_format_name( struct ggml_tensor * tensor, const char * fmt, ...); // @@ -1866,39 +1875,39 @@ extern "C" { GGML_API const char * gguf_type_name(enum gguf_type type); - GGML_API int gguf_get_version (struct gguf_context * ctx); - GGML_API size_t gguf_get_alignment (struct gguf_context * ctx); - GGML_API size_t gguf_get_data_offset(struct gguf_context * ctx); - GGML_API void * gguf_get_data (struct gguf_context * ctx); + GGML_API int gguf_get_version (const struct gguf_context * ctx); + GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx); + GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx); + GGML_API void * gguf_get_data (const struct gguf_context * ctx); - GGML_API int gguf_get_n_kv(struct gguf_context * ctx); - GGML_API int gguf_find_key(struct gguf_context * ctx, const char * key); - GGML_API const char * gguf_get_key (struct gguf_context * ctx, int i); + GGML_API int gguf_get_n_kv(const struct gguf_context * ctx); + GGML_API int gguf_find_key(const struct gguf_context * ctx, const char * key); + GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int i); - GGML_API enum gguf_type gguf_get_kv_type (struct gguf_context * ctx, int i); - GGML_API enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i); + GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int i); + GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int i); // results are undefined if the wrong type is used for the key - GGML_API uint8_t gguf_get_val_u8 (struct gguf_context * ctx, int i); - GGML_API int8_t gguf_get_val_i8 (struct gguf_context * ctx, int i); - GGML_API uint16_t gguf_get_val_u16 (struct gguf_context * ctx, int i); - GGML_API int16_t gguf_get_val_i16 (struct gguf_context * ctx, int i); - GGML_API uint32_t gguf_get_val_u32 (struct gguf_context * ctx, int i); - GGML_API int32_t gguf_get_val_i32 (struct gguf_context * ctx, int i); - GGML_API float gguf_get_val_f32 (struct gguf_context * ctx, int i); - GGML_API uint64_t gguf_get_val_u64 (struct gguf_context * ctx, int i); - GGML_API int64_t gguf_get_val_i64 (struct gguf_context * ctx, int i); - GGML_API double gguf_get_val_f64 (struct gguf_context * ctx, int i); - GGML_API bool gguf_get_val_bool(struct gguf_context * ctx, int i); - GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i); - GGML_API int gguf_get_arr_n (struct gguf_context * ctx, int i); - GGML_API const void * gguf_get_arr_data(struct gguf_context * ctx, int i); - GGML_API const char * gguf_get_arr_str (struct gguf_context * ctx, int key_id, int i); + GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int i); + GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int i); + GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int i); + GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int i); + GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int i); + GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int i); + GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int i); + GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int i); + GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int i); + GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int i); + GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int i); + GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int i); + GGML_API int gguf_get_arr_n (const struct gguf_context * ctx, int i); + GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int i); + GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i); - GGML_API int gguf_get_n_tensors (struct gguf_context * ctx); - GGML_API int gguf_find_tensor (struct gguf_context * ctx, const char * name); - GGML_API size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i); - GGML_API char * gguf_get_tensor_name (struct gguf_context * ctx, int i); + GGML_API int gguf_get_n_tensors (const struct gguf_context * ctx); + GGML_API int gguf_find_tensor (const struct gguf_context * ctx, const char * name); + GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i); + GGML_API char * gguf_get_tensor_name (const struct gguf_context * ctx, int i); // overrides existing values or adds a new one GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val); @@ -1943,11 +1952,11 @@ extern "C" { // // write the entire context to a binary file - GGML_API void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta); + GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta); // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding - GGML_API size_t gguf_get_meta_size(struct gguf_context * ctx); - GGML_API void gguf_get_meta_data(struct gguf_context * ctx, void * data); + GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx); + GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data); // // system info @@ -1961,6 +1970,7 @@ extern "C" { GGML_API int ggml_cpu_has_fma (void); GGML_API int ggml_cpu_has_neon (void); GGML_API int ggml_cpu_has_arm_fma (void); + GGML_API int ggml_cpu_has_metal (void); GGML_API int ggml_cpu_has_f16c (void); GGML_API int ggml_cpu_has_fp16_va (void); GGML_API int ggml_cpu_has_wasm_simd (void); diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index d377cd56d..e0e0dbcbb 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -77,12 +77,14 @@ KEY_TOKENIZER_RWKV = "tokenizer.rwkv.world" class MODEL_ARCH(IntEnum): - LLAMA : int = auto() - FALCON : int = auto() - GPT2 : int = auto() - GPTJ : int = auto() - GPTNEOX: int = auto() - MPT : int = auto() + LLAMA : int = auto() + FALCON : int = auto() + BAICHUAN : int = auto() + GPT2 : int = auto() + GPTJ : int = auto() + GPTNEOX : int = auto() + MPT : int = auto() + STARCODER : int = auto() class MODEL_TENSOR(IntEnum): @@ -106,12 +108,14 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { - MODEL_ARCH.LLAMA: "llama", - MODEL_ARCH.FALCON: "falcon", - MODEL_ARCH.GPT2: "gpt2", - MODEL_ARCH.GPTJ: "gptj", - MODEL_ARCH.GPTNEOX: "gptneox", - MODEL_ARCH.MPT: "mpt", + MODEL_ARCH.LLAMA: "llama", + MODEL_ARCH.FALCON: "falcon", + MODEL_ARCH.BAICHUAN: "baichuan", + MODEL_ARCH.GPT2: "gpt2", + MODEL_ARCH.GPTJ: "gptj", + MODEL_ARCH.GPTNEOX: "gptneox", + MODEL_ARCH.MPT: "mpt", + MODEL_ARCH.STARCODER: "starcoder", } MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = { @@ -153,6 +157,34 @@ MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = { MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", }, + MODEL_ARCH.BAICHUAN: { + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ROPE_FREQS: "rope_freqs", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q", + MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k", + MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd", + MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", + }, + MODEL_ARCH.STARCODER: { + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.POS_EMBD: "position_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", + }, MODEL_ARCH.GPT2: { # TODO }, @@ -165,6 +197,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.BAICHUAN: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } @@ -187,7 +223,7 @@ class TensorNameMap: # Output MODEL_TENSOR.OUTPUT: ( "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf + "lm_head", # gpt2 mpt falcon llama-hf baichuan "output", # llama-pth ), @@ -195,7 +231,7 @@ class TensorNameMap: MODEL_TENSOR.OUTPUT_NORM: ( "gpt_neox.final_layer_norm", # gptneox "transformer.ln_f", # gpt2 falcon - "model.norm", # llama-hf + "model.norm", # llama-hf baichuan "norm", # llama-pth ), @@ -311,6 +347,7 @@ class TensorNameMap: tensor_name = tensor_names.get(tensor) if tensor_name is None: continue + mapping[tensor_name] = (tensor, tensor_name) for key in keys: mapping[key] = (tensor, tensor_name) for bid in range(n_blocks): @@ -319,11 +356,12 @@ class TensorNameMap: if tensor_name is None: continue tensor_name = tensor_name.format(bid = bid) + mapping[tensor_name] = (tensor, tensor_name) for key in keys: key = key.format(bid = bid) mapping[key] = (tensor, tensor_name) - def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> tuple[MODEL_TENSOR, str] | None: + def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None: result = self.mapping.get(key) if result is not None: return result @@ -334,13 +372,13 @@ class TensorNameMap: return (result[0], result[1] + suffix) return None - def get_name(self, key: str, try_suffixes: Sequence[str]) -> str | None: + def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None: result = self.get_type_and_name(key, try_suffixes = try_suffixes) if result is None: return None return result[1] - def get_type(self, key: str, try_suffixes: Sequence[str]) -> MODEL_TENSOR | None: + def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None: result = self.get_type_and_name(key, try_suffixes = try_suffixes) if result is None: return None diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index 8da60de1b..9489ccd6f 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.3.2" +version = "0.3.3" description = "Write ML models in GGUF for GGML" authors = ["GGML "] packages = [ diff --git a/k_quants.c b/k_quants.c index eb702ce86..62085882d 100644 --- a/k_quants.c +++ b/k_quants.c @@ -2609,7 +2609,10 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri memcpy(utmp, x[i].scales, 12); - const uint32x2_t mins8 = {utmp[1] & kmask1, ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4)}; + 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; diff --git a/llama.cpp b/llama.cpp index 208dcef0e..a65026122 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1,8 +1,4 @@ -// Defines fileno on msys: -#ifndef _GNU_SOURCE -#define _GNU_SOURCE -#endif - +#define LLAMA_API_INTERNAL #include "llama.h" #include "ggml.h" @@ -113,7 +109,7 @@ static size_t utf8_len(char src) { return lookup[highbits]; } -void replace_all(std::string & s, const std::string & search, const std::string & replace) { +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { std::string result; for (size_t pos = 0; ; pos += search.length()) { auto new_pos = s.find(search, pos); @@ -126,6 +122,9 @@ void replace_all(std::string & s, const std::string & search, const std::string } s = std::move(result); } +#ifdef GGML_USE_CPU_HBM +#include +#endif static void zeros(std::ofstream & file, size_t n) { char zero = 0; @@ -157,20 +156,24 @@ static std::string format(const char * fmt, ...) { enum llm_arch { LLM_ARCH_LLAMA, LLM_ARCH_FALCON, + LLM_ARCH_BAICHUAN, LLM_ARCH_GPT2, LLM_ARCH_GPTJ, LLM_ARCH_GPTNEOX, LLM_ARCH_MPT, + LLM_ARCH_STARCODER, LLM_ARCH_UNKNOWN, }; static std::map LLM_ARCH_NAMES = { - { LLM_ARCH_LLAMA, "llama" }, - { LLM_ARCH_FALCON, "falcon" }, - { LLM_ARCH_GPT2, "gpt2" }, - { LLM_ARCH_GPTJ, "gptj" }, - { LLM_ARCH_GPTNEOX, "gptneox" }, - { LLM_ARCH_MPT, "mpt" }, + { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GPT2, "gpt2" }, + { LLM_ARCH_GPTJ, "gptj" }, + { LLM_ARCH_GPTNEOX, "gptneox" }, + { LLM_ARCH_MPT, "mpt" }, + { LLM_ARCH_BAICHUAN, "baichuan" }, + { LLM_ARCH_STARCODER, "starcoder" }, }; enum llm_kv { @@ -311,6 +314,25 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_BAICHUAN, + { + { 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_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_FALCON, { @@ -357,6 +379,21 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, }, }, + { + LLM_ARCH_STARCODER, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_POS_EMBD, "position_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { 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_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -450,6 +487,9 @@ static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * #elif GGML_USE_METAL # define llama_host_malloc(n) ggml_metal_host_malloc(n) # define llama_host_free(data) ggml_metal_host_free(data) +#elif GGML_USE_CPU_HBM +# define llama_host_malloc(n) hbw_malloc(n) +# define llama_host_free(data) if (data != NULL) hbw_free(data) #else # define llama_host_malloc(n) malloc(n) # define llama_host_free(data) free(data) @@ -873,9 +913,11 @@ static llama_state g_state; // available llama models enum e_model { MODEL_UNKNOWN, + MODEL_1B, MODEL_3B, MODEL_7B, MODEL_13B, + MODEL_15B, MODEL_30B, MODEL_34B, MODEL_40B, @@ -944,13 +986,22 @@ struct llama_layer { struct ggml_tensor * wo; struct ggml_tensor * wqkv; + // attention bias + struct ggml_tensor * bo; + struct ggml_tensor * bqkv; + // normalization struct ggml_tensor * ffn_norm; + struct ggml_tensor * ffn_norm_b; // ff struct ggml_tensor * w1; // ffn_gate struct ggml_tensor * w2; // ffn_down struct ggml_tensor * w3; // ffn_up + + // ff bias + struct ggml_tensor * b2; // ffn_down + struct ggml_tensor * b3; // ffn_up }; struct llama_kv_cache { @@ -1028,6 +1079,7 @@ struct llama_model { llama_vocab vocab; struct ggml_tensor * tok_embeddings; + struct ggml_tensor * pos_embeddings; struct ggml_tensor * output_norm; struct ggml_tensor * output_norm_b; @@ -1489,7 +1541,11 @@ struct llama_model_loader { // allocate temp buffer if not using mmap if (!use_mmap && cur->data == NULL) { GGML_ASSERT(cur->backend != GGML_BACKEND_CPU); - cur->data = malloc(ggml_nbytes(cur)); + #ifdef GGML_USE_CPU_HBM + cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur)); + #else + cur->data = (uint8_t*)malloc(ggml_nbytes(cur)); + #endif } load_data_for(cur); @@ -1534,7 +1590,7 @@ struct llama_model_loader { // load LLaMA models // -std::string llama_model_ftype_name(enum llama_ftype ftype) { +static std::string llama_model_ftype_name(enum llama_ftype ftype) { if (ftype & LLAMA_FTYPE_GUESSED) { return llama_model_ftype_name((enum llama_ftype) (ftype & ~LLAMA_FTYPE_GUESSED)) + " (guessed)"; } @@ -1567,9 +1623,11 @@ std::string llama_model_ftype_name(enum llama_ftype ftype) { static const char * llama_model_type_name(e_model type) { switch (type) { + case MODEL_1B: return "1B"; case MODEL_3B: return "3B"; case MODEL_7B: return "7B"; case MODEL_13B: return "13B"; + case MODEL_15B: return "15B"; case MODEL_30B: return "30B"; case MODEL_34B: return "34B"; case MODEL_40B: return "40B"; @@ -1678,6 +1736,26 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_BAICHUAN: + { + GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_STARCODER: + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 36: model.type = e_model::MODEL_3B; break; + case 42: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_15B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; }; @@ -1918,7 +1996,6 @@ static void llm_load_tensors( const int64_t n_vocab = hparams.n_vocab; const auto tn = LLM_TN(model.arch); - switch (model.arch) { case LLM_ARCH_LLAMA: { @@ -1961,6 +2038,72 @@ static void llm_load_tensors( model.layers.resize(n_layer); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + } + } + } break; + case LLM_ARCH_BAICHUAN: + { + model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + { + ggml_backend backend_norm; + ggml_backend backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU +#ifndef _WIN32 + backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#else + backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#endif // _WIN32 + + backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + for (uint32_t i = 0; i < n_layer; ++i) { const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT @@ -2066,6 +2209,85 @@ static void llm_load_tensors( } } } break; + case LLM_ARCH_STARCODER: + { + model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.pos_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); + + // output + { + ggml_backend backend_norm; + ggml_backend backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU +#ifndef _WIN32 + backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#else + backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#endif // _WIN32 + + backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); + + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + + layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); + + layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + + ggml_nbytes(layer.w2) + ggml_nbytes(layer.b2) + + ggml_nbytes(layer.w3) + ggml_nbytes(layer.b3); + } + } + } break; default: throw std::runtime_error("unknown architecture"); }; @@ -2537,6 +2759,367 @@ static struct ggml_cgraph * llm_build_llama( return gf; } + +static struct ggml_cgraph * llm_build_baichaun( + llama_context & lctx, + const llama_token * tokens, + const float * embd, + int n_tokens, + int n_past) { + + GGML_ASSERT((!tokens && embd) || (tokens && !embd)); // NOLINT + + const int N = n_tokens; + + const auto & model = lctx.model; + const auto & hparams = model.hparams; + + const auto & kv_self = lctx.kv_self; + + GGML_ASSERT(!!kv_self.ctx); + + const int64_t n_embd = hparams.n_embd; + const int64_t n_layer = hparams.n_layer; + const int64_t n_ctx = hparams.n_ctx; + const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_rot); + + const float freq_base = hparams.rope_freq_base; + const float freq_scale = hparams.rope_freq_scale; + const float norm_rms_eps = hparams.f_norm_rms_eps; + + const int n_gpu_layers = model.n_gpu_layers; + + auto & buf_compute = lctx.buf_compute; + + struct ggml_init_params params = { + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.data, + /*.no_alloc =*/ false, + }; + + params.no_alloc = true; + + struct ggml_context * ctx0 = ggml_init(params); + + ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + if (tokens) { + struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + + ggml_allocr_alloc(lctx.alloc, inp_tokens); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens)); + } + ggml_set_name(inp_tokens, "inp_tokens"); + + inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); + } else { +#ifdef GGML_USE_MPI + GGML_ASSERT(false && "not implemented"); +#endif + + inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N); + + ggml_allocr_alloc(lctx.alloc, inpL); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL)); + } + } + + const int i_gpu_start = n_layer - n_gpu_layers; + (void) i_gpu_start; + + // offload functions set the tensor output backend to GPU + // tensors are GPU-accelerated if any input or the output has been offloaded + // + // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal + // in that case ggml_cuda_assign_buffers has no effect + offload_func_t offload_func_nr = llama_nop; // nr = non-repeating + offload_func_t offload_func_kq = llama_nop; + offload_func_t offload_func_v = llama_nop; + +#ifdef GGML_USE_CUBLAS + if (n_gpu_layers > n_layer) { + offload_func_nr = ggml_cuda_assign_buffers_no_alloc; + } + if (n_gpu_layers > n_layer + 1) { + offload_func_v = ggml_cuda_assign_buffers_no_alloc; + } + if (n_gpu_layers > n_layer + 2) { + offload_func_kq = ggml_cuda_assign_buffers_no_alloc; + } +#endif // GGML_USE_CUBLAS + + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + ggml_allocr_alloc(lctx.alloc, KQ_scale); + if (!ggml_allocr_is_measure(lctx.alloc)) { + ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); + } + ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); + + for (int il = 0; il < n_layer; ++il) { + ggml_format_name(inpL, "layer_inp_%d", il); + + offload_func_t offload_func = llama_nop; + +#ifdef GGML_USE_CUBLAS + if (il >= i_gpu_start) { + offload_func = ggml_cuda_assign_buffers_no_alloc; + } +#endif // GGML_USE_CUBLAS + + struct ggml_tensor * inpSA = inpL; + + // norm + { + cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); + offload_func(cur); + ggml_set_name(cur, "rms_norm_0"); + + // cur = cur*attn_norm(broadcasted) + cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); + offload_func(cur); + ggml_set_name(cur, "attention_norm_0"); + } + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + offload_func_kq(tmpk); + ggml_set_name(tmpk, "tmpk"); + + struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + offload_func_kq(tmpq); + ggml_set_name(tmpq, "tmpq"); + + struct ggml_tensor * Kcur; + struct ggml_tensor * Qcur; + switch (model.type) { + case MODEL_7B: + Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), n_past, n_embd_head, 0, 0, freq_base, freq_scale); + Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), n_past, n_embd_head, 0, 0, freq_base, freq_scale); + break; + case MODEL_13B: + Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N); + Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N); + break; + default: + GGML_ASSERT(false); + } + + offload_func_kq(Kcur); + ggml_set_name(Kcur, "Kcur"); + + offload_func_kq(Qcur); + ggml_set_name(Qcur, "Qcur"); + + // store key and value to memory + { + // compute the transposed [N, n_embd] V matrix + + struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + offload_func_v(tmpv); + ggml_set_name(tmpv, "tmpv"); + + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, N)); + offload_func_v(Vcur); + ggml_set_name(Vcur, "Vcur"); + + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past)); + offload_func_kq(k); + ggml_set_name(k, "k"); + + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v)); + offload_func_v(v); + ggml_set_name(v, "v"); + + // important: storing RoPE-ed version of K in the KV cache! + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); + } + + struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + offload_func_kq(Q); + ggml_set_name(Q, "Q"); + + struct ggml_tensor * K = + ggml_view_3d(ctx0, kv_self.k, + n_embd_head, n_past + N, n_head_kv, + ggml_element_size(kv_self.k)*n_embd_gqa, + ggml_element_size(kv_self.k)*n_embd_head, + ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); + offload_func_kq(K); + ggml_set_name(K, "K"); + + // K * Q + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + offload_func_kq(KQ); + ggml_set_name(KQ, "KQ"); + + // KQ_scaled = KQ / sqrt(n_embd_head) + // KQ_scaled shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); + offload_func_kq(KQ_scaled); + ggml_set_name(KQ_scaled, "KQ_scaled"); + + struct ggml_tensor * KQ_masked; + struct ggml_tensor * KQ_scaled_alibi; + + switch (model.type) { + case MODEL_7B: + KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); + break; + case MODEL_13B: + KQ_scaled_alibi =ggml_alibi(ctx0, KQ_scaled, n_past, n_head, 8); + ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); + KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, n_past); + break; + default: + GGML_ASSERT(false); + } + // KQ_masked = mask_past(KQ_scaled) + // struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); + // struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, n_past); + // offload_func_kq(KQ_masked); + // ggml_set_name(KQ_masked, "KQ_masked"); + + // KQ = soft_max(KQ_masked) + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); + offload_func_v(KQ_soft_max); + ggml_set_name(KQ_soft_max, "KQ_soft_max"); + + // split cached V into n_head heads + struct ggml_tensor * V = + ggml_view_3d(ctx0, kv_self.v, + n_past + N, n_embd_head, n_head_kv, + ggml_element_size(kv_self.v)*n_ctx, + ggml_element_size(kv_self.v)*n_ctx*n_embd_head, + ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); + offload_func_v(V); + ggml_set_name(V, "V"); + +#if 1 + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + offload_func_v(KQV); + ggml_set_name(KQV, "KQV"); +#else + // make V contiguous in memory to speed up the matmul, however we waste time on the copy + // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation + // is there a better way? + struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd_head, n_head)); + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); +#endif + + // KQV_merged = KQV.permute(0, 2, 1, 3) + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + offload_func_v(KQV_merged); + ggml_set_name(KQV_merged, "KQV_merged"); + + // cur = KQV_merged.contiguous().view(n_embd, N) + cur = ggml_cpy(ctx0, + KQV_merged, + ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + offload_func_v(cur); + ggml_set_name(cur, "KQV_merged_contiguous"); + + // projection (no bias) + cur = ggml_mul_mat(ctx0, + model.layers[il].wo, + cur); + offload_func(cur); + ggml_set_name(cur, "result_wo"); + } + + struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); + offload_func(inpFF); + ggml_set_name(inpFF, "inpFF"); + + // feed-forward network + { + // norm + { + cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); + offload_func(cur); + ggml_set_name(cur, "rms_norm_1"); + + // cur = cur*ffn_norm(broadcasted) + cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); + offload_func(cur); + ggml_set_name(cur, "ffn_norm"); + } + + struct ggml_tensor * tmp = ggml_mul_mat(ctx0, + model.layers[il].w3, + cur); + offload_func(tmp); + ggml_set_name(tmp, "result_w3"); + + cur = ggml_mul_mat(ctx0, + model.layers[il].w1, + cur); + offload_func(cur); + ggml_set_name(cur, "result_w1"); + + // SILU activation + cur = ggml_silu(ctx0, cur); + offload_func(cur); + ggml_set_name(cur, "silu"); + + cur = ggml_mul(ctx0, cur, tmp); + offload_func(cur); + ggml_set_name(cur, "silu_x_result_w3"); + + cur = ggml_mul_mat(ctx0, + model.layers[il].w2, + cur); + offload_func(cur); + ggml_set_name(cur, "result_w2"); + } + + cur = ggml_add(ctx0, cur, inpFF); + offload_func(cur); + ggml_set_name(cur, "inpFF_+_result_w2"); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + // norm + { + cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); + offload_func_nr(cur); + ggml_set_name(cur, "rms_norm_2"); + + // cur = cur*norm(broadcasted) + cur = ggml_mul(ctx0, cur, model.output_norm); + // offload_func_nr(cur); // TODO CPU + GPU mirrored backend + ggml_set_name(cur, "result_norm"); + } + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + ggml_set_name(cur, "result_output"); + + ggml_build_forward_expand(gf, cur); + + ggml_free(ctx0); + + return gf; +} + static struct ggml_cgraph * llm_build_falcon( llama_context & lctx, const llama_token * tokens, @@ -2844,6 +3427,235 @@ static struct ggml_cgraph * llm_build_falcon( return gf; } +static struct ggml_cgraph * llm_build_starcoder( + llama_context & lctx, + const llama_token * tokens, + const float * embd, + int n_tokens, + int n_past) { + + GGML_ASSERT((!tokens && embd) || (tokens && !embd)); // NOLINT + + const int N = n_tokens; + + const auto & model = lctx.model; + const auto & hparams = model.hparams; + + const auto & kv_self = lctx.kv_self; + + GGML_ASSERT(!!kv_self.ctx); + + const int64_t n_embd = hparams.n_embd; + const int64_t n_layer = hparams.n_layer; + const int64_t n_ctx = hparams.n_ctx; + const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_rot); + + const float norm_eps = hparams.f_norm_eps; + + auto & buf_compute = lctx.buf_compute; + + struct ggml_init_params params = { + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.data, + /*.no_alloc =*/ false, + }; + + params.no_alloc = true; + + struct ggml_context * ctx0 = ggml_init(params); + + ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * token; + struct ggml_tensor * position; + struct ggml_tensor * inpL; + + if (tokens) { + struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + + ggml_allocr_alloc(lctx.alloc, inp_tokens); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens)); + } + ggml_set_name(inp_tokens, "inp_tokens"); + + token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); + } else { +#ifdef GGML_USE_MPI + GGML_ASSERT(false && "not implemented"); +#endif + + token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N); + + ggml_allocr_alloc(lctx.alloc, token); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(token->data, embd, N * n_embd * ggml_element_size(inpL)); + } + } + + { + // Compute position embeddings. + struct ggml_tensor * inp_positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + ggml_allocr_alloc(lctx.alloc, inp_positions); + if (!ggml_allocr_is_measure(lctx.alloc)) { + for (int i = 0; i < N; ++i) { + ((int32_t *) inp_positions->data)[i] = n_past + i; + } + } + ggml_set_name(inp_positions, "inp_positions"); + + position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions); + } + + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + ggml_allocr_alloc(lctx.alloc, KQ_scale); + if (!ggml_allocr_is_measure(lctx.alloc)) { + ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); + } + ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); + + inpL = ggml_add(ctx0, token, position); + ggml_set_name(inpL, "inpL"); + + for (int il = 0; il < n_layer; ++il) { + { + // Norm + cur = ggml_norm(ctx0, inpL, norm_eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); + } + + { + // Self Attention + cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv); + + struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); + struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, N, cur->nb[1], sizeof(float)*n_embd); + struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, N, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa)); + + struct ggml_tensor * Qcur = tmpq; + struct ggml_tensor * Kcur = tmpk; + + { + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, N)); + ggml_set_name(Vcur, "Vcur"); + + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past)); + ggml_set_name(k, "k"); + + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v)); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); + } + + struct ggml_tensor * Q = + ggml_permute(ctx0, + ggml_cpy(ctx0, + Qcur, + ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, N)), + 0, 2, 1, 3); + ggml_set_name(Q, "Q"); + + struct ggml_tensor * K = + ggml_view_3d(ctx0, kv_self.k, + n_embd_head, n_past + N, n_head_kv, + ggml_element_size(kv_self.k)*n_embd_gqa, + ggml_element_size(kv_self.k)*n_embd_head, + ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); + ggml_set_name(K, "K"); + + // K * Q + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + ggml_set_name(KQ, "KQ"); + + // KQ_scaled = KQ / sqrt(n_embd_head) + // KQ_scaled shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); + ggml_set_name(KQ_scaled, "KQ_scaled"); + + // KQ_masked = mask_past(KQ_scaled) + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); + ggml_set_name(KQ_masked, "KQ_masked"); + + // KQ = soft_max(KQ_masked) + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); + ggml_set_name(KQ_soft_max, "KQ_soft_max"); + + // split cached V into n_head heads + struct ggml_tensor * V = + ggml_view_3d(ctx0, kv_self.v, + n_past + N, n_embd_head, n_head_kv, + ggml_element_size(kv_self.v)*n_ctx, + ggml_element_size(kv_self.v)*n_ctx*n_embd_head, + ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); + ggml_set_name(V, "V"); + + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + ggml_set_name(KQV, "KQV"); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + ggml_set_name(KQV_merged, "KQV_merged"); + + // cur = KQV_merged.contiguous().view(n_embd, N) + cur = ggml_cpy(ctx0, + KQV_merged, + ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + ggml_set_name(cur, "KQV_merged_contiguous"); + } + + // Projection + cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo); + + // Add the input + cur = ggml_add(ctx0, cur, inpL); + + struct ggml_tensor * inpFF = cur; + + // FF + { + // Norm + { + cur = ggml_norm(ctx0, inpFF, norm_eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b); + } + + cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3); + + // GELU activation + cur = ggml_gelu(ctx0, cur); + + // Projection + cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2); + } + + inpL = ggml_add(ctx0, cur, inpFF); + } + + // Output Norm + { + cur = ggml_norm(ctx0, inpL, norm_eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b); + } + ggml_set_name(cur, "result_norm"); + + cur = ggml_mul_mat(ctx0, model.output, cur); + ggml_set_name(cur, "result_output"); + + ggml_build_forward_expand(gf, cur); + ggml_free(ctx0); + + return gf; +} + static struct ggml_cgraph * llama_build_graph( llama_context & lctx, const llama_token * tokens, @@ -2859,10 +3671,18 @@ static struct ggml_cgraph * llama_build_graph( { result = llm_build_llama(lctx, tokens, embd, n_tokens, n_past); } break; + case LLM_ARCH_BAICHUAN: + { + result = llm_build_baichaun(lctx, tokens, embd, n_tokens, n_past); + } break; case LLM_ARCH_FALCON: { result = llm_build_falcon(lctx, tokens, embd, n_tokens, n_past); } break; + case LLM_ARCH_STARCODER: + { + result = llm_build_starcoder(lctx, tokens, embd, n_tokens, n_past); + } break; default: GGML_ASSERT(false); }; @@ -2964,10 +3784,6 @@ static bool llama_eval_internal( if (lctx.ctx_metal) { ggml_metal_set_n_cb (lctx.ctx_metal, n_threads); ggml_metal_graph_compute(lctx.ctx_metal, gf); - ggml_metal_get_tensor (lctx.ctx_metal, res); - if (!lctx.embedding.empty()) { - ggml_metal_get_tensor(lctx.ctx_metal, embeddings); - } } else { ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads); } @@ -3116,10 +3932,9 @@ struct llm_tokenizer_spm { while (offs < text.size()) { llm_symbol sym; size_t len = utf8_len(text[offs]); - GGML_ASSERT(offs + len <= text.size()); sym.text = text.c_str() + offs; - sym.n = len; - offs += len; + sym.n = std::min(len, text.size() - offs); + offs += sym.n; sym.prev = index - 1; sym.next = offs == text.size() ? -1 : index + 1; index++; @@ -3481,7 +4296,7 @@ struct llama_grammar_candidate { // Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as // pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`. -std::pair, llama_partial_utf8> decode_utf8( +static std::pair, llama_partial_utf8> decode_utf8( const char * src, llama_partial_utf8 partial_start) { static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 }; @@ -4635,7 +5450,16 @@ void llama_beam_search(llama_context * ctx, // quantization // -static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vector & output, const size_t nelements, const int nthread) { +template +struct no_init { + T value; + no_init() { /* do nothing */ } +}; + +static void llama_convert_tensor_internal( + struct ggml_tensor * tensor, std::vector> & output, std::vector & workers, + const size_t nelements, const int nthread +) { if (output.size() < nelements) { output.resize(nelements); } @@ -4670,7 +5494,6 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect auto blocks_per_thread = nblocks / nthread; auto spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count - std::vector workers; for (auto tnum = 0, in_buff_offs = 0, out_buff_offs = 0; tnum < nthread; tnum++) { auto thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread auto thr_elems = thr_blocks * block_size; // number of elements for this thread @@ -4683,15 +5506,124 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect qtype.to_float(inbuf, outbuf, nels); } }; - workers.push_back(std::thread(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems)); + workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems); in_buff_offs += thr_block_bytes; out_buff_offs += thr_elems; } - for (auto & worker : workers) { - worker.join(); - } + for (auto & w : workers) { w.join(); } + workers.clear(); } +#ifdef GGML_USE_K_QUANTS +static ggml_type get_k_quant_type( + ggml_type new_type, const ggml_tensor * tensor, const llama_model & model, llama_ftype ftype, int * i_attention_wv, + int n_attention_wv, int * i_feed_forward_w2, int n_feed_forward_w2 +) { + const std::string name = ggml_get_name(tensor); + // TODO: avoid hardcoded tensor names - use the TN_* constants + const auto tn = LLM_TN(model.arch); + + auto use_more_bits = [](int i_layer, int num_layers) -> bool { + return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2; + }; + + if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { + int nx = tensor->ne[0]; + if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) { + new_type = GGML_TYPE_Q8_0; + } + else if (new_type != GGML_TYPE_Q8_0) { + new_type = GGML_TYPE_Q6_K; + } + } else if (name.find("attn_v.weight") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { + new_type = *i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; + else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && + use_more_bits(*i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && *i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; + else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) && + (*i_attention_wv < n_attention_wv/8 || *i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; + if (model.type == MODEL_70B) { + // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is + // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with + // nearly negligible increase in model size by quantizing this tensor with more bits: + if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K; + } + ++*i_attention_wv; + } else if (name.find("ffn_down.weight") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { + new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K + : model.arch != LLM_ARCH_FALCON || use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K + : GGML_TYPE_Q3_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) { + new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { + if (model.arch == LLM_ARCH_FALCON) { + new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K : + use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + } else { + if (use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + } + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && *i_feed_forward_w2 < 4) { + new_type = GGML_TYPE_Q5_K; + } + ++*i_feed_forward_w2; + } else if (name.find("attn_output.weight") != std::string::npos) { + if (model.arch != LLM_ARCH_FALCON) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; + } else { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K; + } + } + else if (name.find("attn_qkv.weight") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; + } + else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + } + // This can be used to reduce the size of the Q5_K_S model. + // The associated PPL increase is fully in line with the size reduction + //else { + // if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K; + //} + bool convert_incompatible_tensor = false; + if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K || + new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) { + int nx = tensor->ne[0]; + int ny = tensor->ne[1]; + if (nx % QK_K != 0) { + LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K); + convert_incompatible_tensor = true; + } + } + if (convert_incompatible_tensor) { + if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { + new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. + LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n"); + } else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) { + new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. + LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n"); + } else { + throw std::runtime_error("Unsupported tensor size encountered\n"); + } + } + + return new_type; +} +#endif + static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) { ggml_type quantized_type; llama_ftype ftype = params->ftype; @@ -4775,18 +5707,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s std::vector hist_all(1 << 4, 0); std::vector workers; + workers.reserve(nthread); std::mutex mutex; -#ifdef GGML_USE_K_QUANTS - auto use_more_bits = [] (int i_layer, int num_layers) -> bool { - return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2; - }; -#endif - int idx = 0; - std::vector read_data; - std::vector work; + std::vector> read_data; + std::vector> work; + std::vector> f32_conv_buf; // populate the original tensors so we get an initial meta data for (int i = 0; i < ml->n_tensors; ++i) { @@ -4808,7 +5736,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const std::string name = ggml_get_name(tensor); - read_data.resize(ggml_nbytes(tensor)); + if (read_data.size() < ggml_nbytes(tensor)) { + read_data.resize(ggml_nbytes(tensor)); + } tensor->data = read_data.data(); ml->load_data_for(tensor); @@ -4833,101 +5763,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (quantize) { new_type = quantized_type; #ifdef GGML_USE_K_QUANTS - // TODO: avoid hardcoded tensor names - use the TN_* constants - const auto tn = LLM_TN(ml->get_arch()); - - if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { - int nx = tensor->ne[0]; - if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) { - new_type = GGML_TYPE_Q8_0; - } - else if (new_type != GGML_TYPE_Q8_0) { - new_type = GGML_TYPE_Q6_K; - } - } else if (name.find("attn_v.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; - else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && - use_more_bits(i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; - else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) && - (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; - if (model.type == MODEL_70B) { - // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is - // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with - // nearly negligible increase in model size by quantizing this tensor with more bits: - if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K; - } - ++i_attention_wv; - } else if (name.find("ffn_down.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K - : model.arch != LLM_ARCH_FALCON || use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K - : GGML_TYPE_Q3_K; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) { - new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { - if (model.arch == LLM_ARCH_FALCON) { - new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K : - use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; - } else { - if (use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; - } - } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && i_feed_forward_w2 < 4) { - new_type = GGML_TYPE_Q5_K; - } - ++i_feed_forward_w2; - } else if (name.find("attn_output.weight") != std::string::npos) { - if (model.arch != LLM_ARCH_FALCON) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; - } else { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K; - } - } - else if (name.find("attn_qkv.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; - } - else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; - } - // This can be used to reduce the size of the Q5_K_S model. - // The associated PPL increase is fully in line with the size reduction - //else { - // if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K; - //} - bool convert_incompatible_tensor = false; - if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K || - new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) { - int nx = tensor->ne[0]; - int ny = tensor->ne[1]; - if (nx % QK_K != 0) { - LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K); - convert_incompatible_tensor = true; - } - } - if (convert_incompatible_tensor) { - if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { - new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. - LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n"); - } else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) { - new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. - LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n"); - } else { - throw std::runtime_error("Unsupported tensor size encountered\n"); - } - } + new_type = get_k_quant_type( + new_type, tensor, model, ftype, &i_attention_wv, n_attention_wv, &i_feed_forward_w2, n_feed_forward_w2 + ); #endif // If we've decided to quantize to the same type the tensor is already // in then there's nothing to do. @@ -4942,23 +5780,24 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const size_t nelements = ggml_nelements(tensor); float * f32_data; - std::vector f32_conv_buf; if (tensor->type == GGML_TYPE_F32) { f32_data = (float *) tensor->data; } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) { throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type))); } else { - llama_convert_tensor_internal(tensor, f32_conv_buf, nelements, nthread); + llama_convert_tensor_internal(tensor, f32_conv_buf, workers, nelements, nthread); f32_data = (float *) f32_conv_buf.data(); } LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type)); fflush(stdout); - work.resize(nelements * 4); // upper bound on size + if (work.size() < nelements * 4) { + work.resize(nelements * 4); // upper bound on size + } new_data = work.data(); - std::vector hist_cur(1 << 4, 0); + std::array hist_cur = {}; static const int chunk_size = 32 * 512; const int nchunk = (nelements + chunk_size - 1)/chunk_size; @@ -4969,13 +5808,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s size_t counter = 0; new_size = 0; auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements]() { - std::vector local_hist; + std::array local_hist = {}; size_t local_size = 0; while (true) { std::unique_lock lock(mutex); size_t first = counter; counter += chunk_size; if (first >= nelements) { - if (!local_hist.empty()) { + if (local_size > 0) { for (int j=0; j %8.2f MB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); @@ -5062,7 +5894,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } // TODO: after the GGUF PR, this likely won't work and needs to be updated -int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) { +static int llama_apply_lora_from_file_internal( + const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads +) { LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora); const int64_t t_start_lora_us = ggml_time_us(); @@ -5609,7 +6443,7 @@ struct llama_context * llama_new_context_with_model( return ctx; } -struct llama_context * llama_init_from_file( +static struct llama_context * llama_init_from_file( const char * path_model, struct llama_context_params params) { struct llama_model * model = llama_load_model_from_file(path_model, params); @@ -5628,15 +6462,19 @@ void llama_free(struct llama_context * ctx) { } int llama_n_vocab(const struct llama_context * ctx) { - return ctx->model.vocab.id_to_token.size(); + return llama_model_n_vocab(&ctx->model); } int llama_n_ctx(const struct llama_context * ctx) { - return ctx->model.hparams.n_ctx; + return llama_model_n_ctx(&ctx->model); +} + +int llama_n_ctx_train(const struct llama_context * ctx) { + return llama_model_n_ctx_train(&ctx->model); } int llama_n_embd(const struct llama_context * ctx) { - return ctx->model.hparams.n_embd; + return llama_model_n_embd(&ctx->model); } enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx) { @@ -5651,6 +6489,10 @@ int llama_model_n_ctx(const struct llama_model * model) { return model->hparams.n_ctx; } +int llama_model_n_ctx_train(const struct llama_model * model) { + return model->hparams.n_ctx_train; +} + int llama_model_n_embd(const struct llama_model * model) { return model->hparams.n_embd; } @@ -5806,7 +6648,7 @@ struct llama_data_file_context : llama_data_context { * llama_copy_state_data(ctx, &data_ctx); * */ -void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) { +static void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) { // copy rng { std::stringstream rng_ss; @@ -6205,7 +7047,7 @@ int llama_tokenize_with_model( auto res = llama_tokenize_internal(model->vocab, text, add_bos); if (n_max_tokens < (int) res.size()) { - LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); + // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); return -((int) res.size()); } @@ -6344,7 +7186,9 @@ void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) { } // For internal test use -const std::vector>& llama_internal_get_tensor_map(struct llama_context * ctx) { +const std::vector> & llama_internal_get_tensor_map( + struct llama_context * ctx +) { return ctx->model.tensors_by_name; } diff --git a/llama.h b/llama.h index 5b95aaa87..c6ee038c7 100644 --- a/llama.h +++ b/llama.h @@ -245,15 +245,17 @@ extern "C" { LLAMA_API bool llama_mmap_supported (void); LLAMA_API bool llama_mlock_supported(void); - LLAMA_API int llama_n_vocab(const struct llama_context * ctx); - LLAMA_API int llama_n_ctx (const struct llama_context * ctx); - LLAMA_API int llama_n_embd (const struct llama_context * ctx); + LLAMA_API int llama_n_vocab (const struct llama_context * ctx); + LLAMA_API int llama_n_ctx (const struct llama_context * ctx); + LLAMA_API int llama_n_ctx_train(const struct llama_context * ctx); + LLAMA_API int llama_n_embd (const struct llama_context * ctx); LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx); - LLAMA_API int llama_model_n_vocab(const struct llama_model * model); - LLAMA_API int llama_model_n_ctx (const struct llama_model * model); - LLAMA_API int llama_model_n_embd (const struct llama_model * model); + LLAMA_API int llama_model_n_vocab (const struct llama_model * model); + LLAMA_API int llama_model_n_ctx (const struct llama_model * model); + LLAMA_API int llama_model_n_ctx_train(const struct llama_model * model); + LLAMA_API int llama_model_n_embd (const struct llama_model * model); // Get a string describing the model type LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); @@ -538,7 +540,9 @@ extern "C" { struct ggml_tensor; -const std::vector>& llama_internal_get_tensor_map(struct llama_context * ctx); +const std::vector> & llama_internal_get_tensor_map( + struct llama_context * ctx +); #endif // LLAMA_API_INTERNAL diff --git a/pocs/vdot/vdot.cpp b/pocs/vdot/vdot.cpp index 48758cda8..e96372c4b 100644 --- a/pocs/vdot/vdot.cpp +++ b/pocs/vdot/vdot.cpp @@ -16,7 +16,7 @@ constexpr int kVecSize = 1 << 18; -float drawFromGaussianPdf(std::mt19937& rndm) { +static float drawFromGaussianPdf(std::mt19937& rndm) { constexpr double kScale = 1./(1. + std::mt19937::max()); constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale; static float lastX; @@ -28,7 +28,8 @@ float drawFromGaussianPdf(std::mt19937& rndm) { haveX = true; return r*cos(phi); } -void fillRandomGaussianFloats(std::vector& values, std::mt19937& rndm, float mean = 0) { + +static void fillRandomGaussianFloats(std::vector& values, std::mt19937& rndm, float mean = 0) { for (auto& v : values) v = mean + drawFromGaussianPdf(rndm); } diff --git a/prompts/chat-with-baichuan.txt b/prompts/chat-with-baichuan.txt new file mode 100644 index 000000000..11626b692 --- /dev/null +++ b/prompts/chat-with-baichuan.txt @@ -0,0 +1,4 @@ +以下内容为人类用户与与一位智能助手的对话。 + +用户:你好！ +助手: diff --git a/run_with_preset.py b/run_with_preset.py index 8f90f52a9..9b4d7ecbe 100755 --- a/run_with_preset.py +++ b/run_with_preset.py @@ -13,7 +13,7 @@ CLI_ARGS_MAIN_PERPLEXITY = [ "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix", "instruct", "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base", "low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock", - "model", "mtest", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q", + "model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q", "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt", "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "random-prompt", "repeat-last-n", "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed", diff --git a/scripts/LlamaConfig.cmake.in b/scripts/LlamaConfig.cmake.in new file mode 100644 index 000000000..e1fadc361 --- /dev/null +++ b/scripts/LlamaConfig.cmake.in @@ -0,0 +1,69 @@ +set(LLAMA_VERSION @LLAMA_INSTALL_VERSION@) +set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@) +set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@) +set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@) +set(LLAMA_BLAS @LLAMA_BLAS@) +set(LLAMA_CUBLAS @LLAMA_CUBLAS@) +set(LLAMA_METAL @LLAMA_METAL@) +set(LLAMA_MPI @LLAMA_MPI@) +set(LLAMA_CLBLAST @LLAMA_CLBLAST@) +set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@) +set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@) + +@PACKAGE_INIT@ + +set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@") +set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@") +set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@") + +# Ensure transient dependencies satisfied + +find_package(Threads REQUIRED) +if (APPLE AND LLAMA_ACCELERATE) + find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED) +endif() + +if (LLAMA_BLAS) + find_package(BLAS REQUIRED) +endif() + +if (LLAMA_CUBLAS) + find_package(CUDAToolkit REQUIRED) +endif() + +if (LLAMA_METAL) + find_library(FOUNDATION_LIBRARY Foundation REQUIRED) + find_library(METAL_FRAMEWORK Metal REQUIRED) + find_library(METALKIT_FRAMEWORK MetalKit REQUIRED) +endif() + +if (LLAMA_MPI) + find_package(MPI REQUIRED) +endif() + +if (LLAMA_CLBLAST) + find_package(CLBlast REQUIRED) +endif() + +if (LLAMA_HIPBLAS) + find_package(hip REQUIRED) + find_package(hipblas REQUIRED) + find_package(rocblas REQUIRED) +endif() + +find_library(llama_LIBRARY llama + REQUIRED + HINTS ${LLAMA_LIB_DIR}) + +set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@") +add_library(llama UNKNOWN IMPORTED) +set_target_properties(llama + PROPERTIES + INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}" + INTERFACE_LINK_LIBRARIES "${_llama_link_deps}" + IMPORTED_LINK_INTERFACE_LANGUAGES "CXX" + IMPORTED_LOCATION "${llama_LIBRARY}" + INTERFACE_COMPILE_FEATURES cxx_std_11 + POSITION_INDEPENDENT_CODE ON ) + +check_required_components(Llama) diff --git a/scripts/build-info.cmake b/scripts/build-info.cmake index 5023b77ab..e33f3349a 100644 --- a/scripts/build-info.cmake +++ b/scripts/build-info.cmake @@ -2,6 +2,8 @@ set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in") set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h") set(BUILD_NUMBER 0) set(BUILD_COMMIT "unknown") +set(BUILD_COMPILER "unknown") +set(BUILD_TARGET "unknown") # Look for git find_package(Git) @@ -41,11 +43,45 @@ if(Git_FOUND) endif() endif() +if(GIT_HEAD_RESULT EQUAL 0 AND GIT_COUNT_RESULT EQUAL 0) + set(BUILD_COMMIT ${HEAD}) + set(BUILD_NUMBER ${COUNT}) +endif() + +execute_process( + COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER} + OUTPUT_VARIABLE OUT + OUTPUT_STRIP_TRAILING_WHITESPACE + RESULT_VARIABLE RES +) +if (RES EQUAL 0) + set(BUILD_COMPILER ${OUT}) +endif() + +execute_process( + COMMAND ${CMAKE_C_COMPILER} -dumpmachine + OUTPUT_VARIABLE OUT + OUTPUT_STRIP_TRAILING_WHITESPACE + RESULT_VARIABLE RES +) +if (RES EQUAL 0) + set(BUILD_TARGET ${OUT}) +endif() + # Only write the header if it's changed to prevent unnecessary recompilation if(EXISTS ${HEADER_FILE}) - file(STRINGS ${HEADER_FILE} CONTENTS REGEX "BUILD_COMMIT \"([^\"]*)\"") - list(GET CONTENTS 0 EXISTING) - if(NOT EXISTING STREQUAL "#define BUILD_COMMIT \"${BUILD_COMMIT}\"") + file(READ ${HEADER_FILE} CONTENTS) + string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS}) + set(OLD_COMMIT ${CMAKE_MATCH_1}) + string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS}) + set(OLD_COMPILER ${CMAKE_MATCH_1}) + string(REGEX MATCH "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} ${HEADER_FILE}) endif() else() diff --git a/scripts/build-info.h.in b/scripts/build-info.h.in index 75d1e16fd..e996faef0 100644 --- a/scripts/build-info.h.in +++ b/scripts/build-info.h.in @@ -3,5 +3,7 @@ #define BUILD_NUMBER @BUILD_NUMBER@ #define BUILD_COMMIT "@BUILD_COMMIT@" +#define BUILD_COMPILER "@BUILD_COMPILER@" +#define BUILD_TARGET "@BUILD_TARGET@" #endif // BUILD_INFO_H diff --git a/scripts/build-info.sh b/scripts/build-info.sh index ed0d6c56a..3c8b1fb85 100755 --- a/scripts/build-info.sh +++ b/scripts/build-info.sh @@ -1,23 +1,35 @@ #!/bin/sh -BUILD_NUMBER="0" -BUILD_COMMIT="unknown" +CC=$1 -REV_LIST=$(git rev-list --count HEAD) -if [ $? -eq 0 ]; then - BUILD_NUMBER=$REV_LIST +build_number="0" +build_commit="unknown" +build_compiler="unknown" +build_target="unknown" + +if out=$(git rev-list --count HEAD); then + # git is broken on WSL so we need to strip extra newlines + build_number=$(printf '%s' "$out" | tr -d '\n') fi -REV_PARSE=$(git rev-parse --short HEAD) -if [ $? -eq 0 ]; then - BUILD_COMMIT=$REV_PARSE +if out=$(git rev-parse --short HEAD); then + build_commit=$(printf '%s' "$out" | tr -d '\n') +fi + +if out=$($CC --version | head -1); then + build_compiler=$out +fi + +if out=$($CC -dumpmachine); then + build_target=$out fi echo "#ifndef BUILD_INFO_H" echo "#define BUILD_INFO_H" -echo "" -echo "#define BUILD_NUMBER $BUILD_NUMBER" | tr -d '\n' -echo "" -echo "#define BUILD_COMMIT \"$BUILD_COMMIT\"" | tr -d '\n' -echo "" +echo +echo "#define BUILD_NUMBER $build_number" +echo "#define BUILD_COMMIT \"$build_commit\"" +echo "#define BUILD_COMPILER \"$build_compiler\"" +echo "#define BUILD_TARGET \"$build_target\"" +echo echo "#endif // BUILD_INFO_H" diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 483210d7b..916dc9d05 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -29,9 +29,8 @@ llama_build_executable(test-tokenizer-0-llama.cpp) llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) llama_build_executable(test-tokenizer-0-falcon.cpp) #llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) -llama_build_executable(test-tokenizer-1.cpp) -# test-tokenizer-1 requires a BPE vocab. re-enable when we have one. -#llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) +llama_build_executable(test-tokenizer-1-llama.cpp) +llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) #llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf) llama_build_and_test_executable(test-grammar-parser.cpp) llama_build_and_test_executable(test-llama-grammar.cpp) diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp index 8ab240202..ce4976858 100644 --- a/tests/test-opt.cpp +++ b/tests/test-opt.cpp @@ -36,15 +36,15 @@ #define GGML_PRINT(...) printf(__VA_ARGS__) -float frand(void) { +static float frand(void) { return (float)rand()/(float)RAND_MAX; } -int irand(int n) { +static int irand(int n) { return rand()%n; } -void get_random_dims(int64_t * dims, int ndims) { +static void get_random_dims(int64_t * dims, int ndims) { dims[0] = dims[1] = dims[2] = dims[3] = 1; for (int i = 0; i < ndims; i++) { @@ -52,7 +52,7 @@ void get_random_dims(int64_t * dims, int ndims) { } } -void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) { +static void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) { dims[0] = dims[1] = dims[2] = dims[3] = 1; for (int i = 0; i < ndims; i++) { @@ -61,12 +61,9 @@ void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) { } -struct ggml_tensor * get_random_tensor( - struct ggml_context * ctx0, - int ndims, - int64_t ne[], - float fmin, - float fmax) { +static struct ggml_tensor * get_random_tensor( + struct ggml_context * ctx0, int ndims, int64_t ne[], float fmin, float fmax +) { struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne); switch (ndims) { @@ -109,11 +106,11 @@ struct ggml_tensor * get_random_tensor( return result; } -float get_element(const struct ggml_tensor * t, int idx) { +static float get_element(const struct ggml_tensor * t, int idx) { return ((float *)t->data)[idx]; } -void set_element(struct ggml_tensor * t, int idx, float value) { +static void set_element(struct ggml_tensor * t, int idx, float value) { ((float *)t->data)[idx] = value; } diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp index 8d3c162d2..884af4054 100644 --- a/tests/test-quantize-fns.cpp +++ b/tests/test-quantize-fns.cpp @@ -13,24 +13,24 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif -const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f; -const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f; -const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f; -const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f; -const float MAX_DOT_PRODUCT_ERROR = 0.02f; +constexpr float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f; +constexpr float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f; +constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f; +constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f; +constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f; -const char* RESULT_STR[] = {"ok", "FAILED"}; +static const char* RESULT_STR[] = {"ok", "FAILED"}; // Generate synthetic data -void generate_data(float offset, size_t n, float * dst) { +static void generate_data(float offset, size_t n, float * dst) { for (size_t i = 0; i < n; i++) { dst[i] = 0.1 + 2*cosf(i + offset); } } // Calculate RMSE between two float arrays -float array_rmse(const float * a1, const float * a2, size_t n) { +static float array_rmse(const float * a1, const float * a2, size_t n) { double sum = 0; for (size_t i = 0; i < n; i++) { double diff = a1[i] - a2[i]; @@ -40,7 +40,7 @@ float array_rmse(const float * a1, const float * a2, size_t n) { } // Total quantization error on test data -float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) { +static float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) { std::vector tmp_q(2*test_size); std::vector tmp_out(test_size); @@ -50,7 +50,7 @@ float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, cons } // Total quantization error on test data -float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) { +static float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) { std::vector tmp_q(2*test_size); std::vector tmp_out(test_size); std::vector tmp_out_ref(test_size); @@ -64,7 +64,7 @@ float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size); } -float dot_product(const float * a1, const float * a2, size_t test_size) { +static float dot_product(const float * a1, const float * a2, size_t test_size) { double sum = 0; for (size_t i = 0; i < test_size; i++) { sum += a1[i] * a2[i]; @@ -73,7 +73,9 @@ float dot_product(const float * a1, const float * a2, size_t test_size) { } // Total dot product error -float dot_product_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) { +static float dot_product_error( + ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2 +) { std::vector tmp_q1(2*test_size); std::vector tmp_q2(2*test_size); diff --git a/tests/test-quantize-perf.cpp b/tests/test-quantize-perf.cpp index cbea7d452..01aa69877 100644 --- a/tests/test-quantize-perf.cpp +++ b/tests/test-quantize-perf.cpp @@ -61,22 +61,22 @@ inline int64_t cpu_cycles() { // Generate synthetic data -void generate_data(float offset, size_t n, float * dst) { +static void generate_data(float offset, size_t n, float * dst) { for (size_t i = 0; i < n; i++) { dst[i] = 0.1 + 2*cosf(i + offset); } } -float gigabytes_per_second(size_t bytes, int64_t usecs) { +static float gigabytes_per_second(size_t bytes, int64_t usecs) { return bytes / (float) usecs * 1000000 / (1024*1024*1024); } -void * align_with_offset(void * ptr, int offset) { +static void * align_with_offset(void * ptr, int offset) { size_t dummy_size = MAX_ALIGNMENT * 4; return (char *) std::align(MAX_ALIGNMENT, MAX_ALIGNMENT, ptr, dummy_size) + offset; } -void benchmark_function(size_t size, size_t q_size, int64_t iterations, const std::function & function) { +static void benchmark_function(size_t size, size_t q_size, int64_t iterations, const std::function & function) { int64_t min_time_us = INT64_MAX; int64_t total_time_us = 0; int64_t min_time_cycles = INT64_MAX; @@ -108,7 +108,7 @@ void benchmark_function(size_t size, size_t q_size, int64_t iterations, const st printf(" quantized throughput : %9.2f GB/s\n", gigabytes_per_second(q_size * iterations, total_time_us)); } -void usage(char * argv[]) { +static void usage(char * argv[]) { printf("Benchmark quantization specific functions on synthetic data\n"); printf("\n"); printf("usage: %s [options]\n", argv[0]); diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index 4437c3948..019c0d462 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -12,7 +12,8 @@ #include #include -void dump(const llama_token_data_array * candidates) { + +static void dump(const llama_token_data_array * candidates) { for (size_t i = 0; i < candidates->size; i++) { printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit); } @@ -21,9 +22,7 @@ void dump(const llama_token_data_array * candidates) { #define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0) -void test_top_k(const std::vector & probs, - const std::vector & expected_probs, - int k) { +static void test_top_k(const std::vector & probs, const std::vector & expected_probs, int k) { size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); @@ -45,10 +44,7 @@ void test_top_k(const std::vector & probs, } -void test_top_p(const std::vector & probs, - const std::vector & expected_probs, - float p) { - +static void test_top_p(const std::vector & probs, const std::vector & expected_probs, float p) { size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); @@ -70,9 +66,7 @@ void test_top_p(const std::vector & probs, } -void test_tfs(const std::vector & probs, - const std::vector & expected_probs, - float z) { +static void test_tfs(const std::vector & probs, const std::vector & expected_probs, float z) { size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); @@ -93,9 +87,7 @@ void test_tfs(const std::vector & probs, } -void test_typical(const std::vector & probs, - const std::vector & expected_probs, - float p) { +static void test_typical(const std::vector & probs, const std::vector & expected_probs, float p) { size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); @@ -116,11 +108,10 @@ void test_typical(const std::vector & probs, } -void test_repetition_penalty( - const std::vector & probs, - const std::vector & last_tokens, - const std::vector & expected_probs, - float penalty) { +static void test_repetition_penalty( + const std::vector & probs, const std::vector & last_tokens, + const std::vector & expected_probs, float penalty +) { assert(probs.size() == expected_probs.size()); size_t n_vocab = probs.size(); @@ -145,11 +136,10 @@ void test_repetition_penalty( } -void test_frequency_presence_penalty( - const std::vector & probs, - const std::vector & last_tokens, - const std::vector & expected_probs, - float alpha_frequency, float alpha_presence) { +static void test_frequency_presence_penalty( + const std::vector & probs, const std::vector & last_tokens, + const std::vector & expected_probs, float alpha_frequency, float alpha_presence +) { assert(probs.size() == expected_probs.size()); size_t n_vocab = probs.size(); diff --git a/tests/test-tokenizer-0-llama.cpp b/tests/test-tokenizer-0-llama.cpp index 8630742c6..edbd86f85 100644 --- a/tests/test-tokenizer-0-llama.cpp +++ b/tests/test-tokenizer-0-llama.cpp @@ -1,5 +1,6 @@ #include "llama.h" #include "common.h" +#include "console.h" #include #include @@ -89,6 +90,12 @@ int main(int argc, char **argv) { return 2; } +#ifdef _WIN32 + // We need this for unicode console support + console::init(false, false); + atexit([]() { console::cleanup(); }); +#endif + bool success = true; for (const auto & test_kv : k_tests()) { diff --git a/tests/test-tokenizer-1-llama.cpp b/tests/test-tokenizer-1-llama.cpp new file mode 100644 index 000000000..804ea2486 --- /dev/null +++ b/tests/test-tokenizer-1-llama.cpp @@ -0,0 +1,127 @@ +#include "llama.h" +#include "common.h" +#include "console.h" + +#include +#include +#include +#include +#include +#include +#include +#include + +typedef int codepoint; + +static std::string codepoint_to_utf8(codepoint cp) { + std::string result; + if (0x00 <= cp && cp <= 0x7f) { + result.push_back(cp); + } else if (0x80 <= cp && cp <= 0x7ff) { + result.push_back(0xc0 | ((cp >> 6) & 0x1f)); + result.push_back(0x80 | (cp & 0x3f)); + } else if (0x800 <= cp && cp <= 0xffff) { + result.push_back(0xe0 | ((cp >> 12) & 0x0f)); + result.push_back(0x80 | ((cp >> 6) & 0x3f)); + result.push_back(0x80 | (cp & 0x3f)); + } else if (0x10000 <= cp && cp <= 0x10ffff) { + result.push_back(0xf0 | ((cp >> 18) & 0x07)); + result.push_back(0x80 | ((cp >> 12) & 0x3f)); + result.push_back(0x80 | ((cp >> 6) & 0x3f)); + result.push_back(0x80 | (cp & 0x3f)); + } else { + throw std::invalid_argument("invalid codepoint"); + } + return result; +} + +int main(int argc, char **argv) { + if (argc < 2) { + fprintf(stderr, "Usage: %s \n", argv[0]); + return 1; + } + + const std::string fname = argv[1]; + + fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str()); + + llama_model * model; + llama_context * ctx; + + llama_backend_init(false); + + // load the vocab + { + auto lparams = llama_context_default_params(); + + lparams.vocab_only = true; + + model = llama_load_model_from_file(fname.c_str(), lparams); + + if (model == NULL) { + fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); + return 1; + } + + ctx = llama_new_context_with_model(model, lparams); + + if (ctx == NULL) { + fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); + llama_free_model(model); + return 1; + } + } + + GGML_ASSERT(llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM); + +#ifdef _WIN32 + // We need this for unicode console support + console::init(false, false); + atexit([]() { console::cleanup(); }); +#endif + + const int n_vocab = llama_n_vocab(ctx); + + for (int i = 0; i < n_vocab; ++i) { + std::string str = llama_detokenize_spm(ctx, std::vector(1, i)); + std::vector tokens = llama_tokenize(ctx, str, false); + std::string check = llama_detokenize_spm(ctx, tokens); + if (check != str) { + fprintf(stderr, "%s : error: token %d detokenizes to >%s<(%llu) but tokenization of this detokenizes to >%s<(%llu)\n", + __func__, i, str.c_str(), str.length(), check.c_str(), check.length()); + if(i != 3) + return 2; + } + } + + for (codepoint cp = 0x0000; cp < 0xffff; ++cp) { + if (cp < 0xd800 || cp > 0xdfff) { + std::string str = codepoint_to_utf8(cp); + std::vector tokens = llama_tokenize(ctx, str, false); + std::string check = llama_detokenize_spm(ctx, tokens); + if (str != check) { + fprintf(stderr, "%s : error: codepoint %d detokenizes to >%s<(%llu) instead of >%s<(%llu)\n", + __func__, cp, check.c_str(), check.length(), str.c_str(), str.length()); + if(cp != 0 && cp != 9601) + return 3; + } + } + } + for (codepoint cp = 0x10000; cp < 0x0010ffff; ++cp) { + std::string str = codepoint_to_utf8(cp); + std::vector tokens = llama_tokenize(ctx, str, false); + std::string check = llama_detokenize_spm(ctx, tokens); + if (str != check) { + fprintf(stderr, "%s : error: codepoint %d detokenizes to >%s<(%llu) instead of >%s<(%llu)\n", + __func__, cp, check.c_str(), check.length(), str.c_str(), str.length()); + return 4; + } + } + + llama_free_model(model); + llama_free(ctx); + + llama_backend_free(); + + return 0; +} diff --git a/tests/test-tokenizer-1.cpp b/tests/test-tokenizer-1.cpp deleted file mode 100644 index ce4f2898c..000000000 --- a/tests/test-tokenizer-1.cpp +++ /dev/null @@ -1,108 +0,0 @@ -#include "llama.h" -#include "common.h" - -#include -#include -#include -#include -#include -#include -#include -#include - -static std::string escape_whitespace(const std::string& text) { - std::string result = "\xe2\x96\x81"; - for (size_t offs = 0; offs < text.length(); ++offs) { - if (text[offs] == ' ') { - result += "\xe2\x96\x81"; - } else { - result += text[offs]; - } - } - return result; -} - -int main(int argc, char **argv) { - if (argc < 2) { - fprintf(stderr, "Usage: %s \n", argv[0]); - return 1; - } - - const std::string fname = argv[1]; - - fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str()); - - llama_model * model; - llama_context * ctx; - - llama_backend_init(false); - - // load the vocab - { - auto lparams = llama_context_default_params(); - - lparams.vocab_only = true; - - model = llama_load_model_from_file(fname.c_str(), lparams); - - if (model == NULL) { - fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); - return 1; - } - - ctx = llama_new_context_with_model(model, lparams); - - if (ctx == NULL) { - fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); - llama_free_model(model); - return 1; - } - } - - GGML_ASSERT(llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_BPE); - - const int n_vocab = llama_n_vocab(ctx); - - for (int i = 0; i < n_vocab; ++i) { - std::string forward = llama_token_to_piece(ctx, i); - std::vector tokens = llama_tokenize(ctx, forward, false); - if (tokens.size() == 1) { - if (i != tokens[0]) { - std::string backward = llama_token_to_piece(ctx, tokens[0]); - fprintf(stderr, "%s : error: token %d is string %s but bpe returns token %d %s\n", - __func__, i, llama_token_to_piece(ctx, i).c_str(), tokens[0], backward.c_str()); - return 2; - } - } - } - -#ifdef _WIN32 - std::wstring_convert, char16_t> u16converter; - for (char16_t ch = 0x0000; ch < 0xffff; ++ch) { - std::u16string u16str(1, ch); - std::string str = u16converter.to_bytes(u16str); - std::vector tokens = llama_tokenize(ctx, escape_whitespace(str).c_str(), false); - if (tokens.size() == 1) { - fprintf(stderr, "%s : info: %s tokenized to %d \n", - __func__, str.c_str(), tokens[0]); - } - } - - std::wstring_convert, char32_t> u32converter; - for (char32_t ch = 0x0000; ch < 0x0010ffff; ++ch) { - std::u32string u32str(1, ch); - std::string str = u32converter.to_bytes(u32str); - std::vector tokens = llama_tokenize(ctx, escape_whitespace(str).c_str(), false); - if (tokens.size() == 1) { - fprintf(stderr, "%s : info: %s tokenized to %d \n", __func__, str.c_str(), tokens[0]); - } - } -#endif - - llama_free_model(model); - llama_free(ctx); - - llama_backend_free(); - - return 0; -}