diff --git a/.clang-tidy b/.clang-tidy new file mode 100644 index 000000000..1a42b9abc --- /dev/null +++ b/.clang-tidy @@ -0,0 +1,18 @@ +--- +Checks: > + bugprone-*, + -bugprone-easily-swappable-parameters, + -bugprone-implicit-widening-of-multiplication-result, + -bugprone-narrowing-conversions, + readability-*, + -readability-avoid-unconditional-preprocessor-if, + -readability-function-cognitive-complexity, + -readability-identifier-length, + -readability-implicit-bool-conversion, + -readability-magic-numbers, + -readability-uppercase-literal-suffix, + clang-analyzer-*, + -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling, + performance-*, + portability-*, +FormatStyle: none diff --git a/.devops/tools.sh b/.devops/tools.sh index b0196b60d..ece9e4efa 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -23,7 +23,7 @@ elif [[ $arg1 == '--all-in-one' || $arg1 == '-a' ]]; then echo "Skip model quantization, it already exists: ${i/f16/q4_0}" else echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..." - ./quantize "$i" "${i/f16/q4_0}" 2 + ./quantize "$i" "${i/f16/q4_0}" q4_0 fi done else diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 2208f42f7..a5938bf93 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -12,17 +12,15 @@ on: - master paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp'] pull_request: - types: [opened, synchronize, edited, reopened, review_requested, ready_for_review] + types: [opened, synchronize, reopened] paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp'] env: BRANCH_NAME: ${{ github.head_ref || github.ref_name }} jobs: - ubuntu-latest-make: - if: github.event.pull_request.draft == false - - runs-on: ubuntu-latest + ubuntu-focal-make: + runs-on: ubuntu-20.04 steps: - name: Clone @@ -33,16 +31,14 @@ jobs: id: depends run: | sudo apt-get update - sudo apt-get install build-essential + sudo apt-get install build-essential gcc-8 - name: Build id: make_build run: | - make + CC=gcc-8 make ubuntu-latest-cmake: - if: github.event.pull_request.draft == false - runs-on: ubuntu-latest steps: @@ -71,8 +67,6 @@ jobs: ctest --verbose ubuntu-latest-cmake-sanitizer: - if: github.event.pull_request.draft == false - runs-on: ubuntu-latest continue-on-error: true @@ -81,7 +75,6 @@ jobs: matrix: sanitizer: [ADDRESS, THREAD, UNDEFINED] build_type: [Debug, Release] - accelerate: [ON, OFF] steps: - name: Clone @@ -99,7 +92,7 @@ jobs: run: | mkdir build cd build - cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_ACCELERATE=${{ matrix.accelerate }} + cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} cmake --build . --config ${{ matrix.build_type }} - name: Test @@ -109,8 +102,6 @@ jobs: ctest --verbose macOS-latest-make: - if: github.event.pull_request.draft == false - runs-on: macos-latest steps: @@ -129,9 +120,7 @@ jobs: make macOS-latest-cmake: - if: github.event.pull_request.draft == false - - runs-on: macOS-latest + runs-on: macos-latest steps: - name: Clone @@ -158,25 +147,65 @@ jobs: ctest --verbose windows-latest-cmake: - if: github.event.pull_request.draft == false - runs-on: windows-latest + env: + OPENBLAS_VERSION: 0.3.23 + OPENCL_VERSION: 2023.04.17 + CLBLAST_VERSION: 1.5.3 strategy: matrix: include: - - build: 'avx2' - defines: '' - - build: 'avx' - defines: '-DLLAMA_AVX2=OFF' - - build: 'avx512' - defines: '-DLLAMA_AVX512=ON' + - build: 'avx2' + defines: '' + - build: 'avx' + defines: '-DLLAMA_AVX2=OFF' + - build: 'avx512' + defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON' + - build: 'clblast' + defines: '-DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"' + - build: 'openblas' + defines: '-DLLAMA_OPENBLAS=ON -DBLAS_LIBRARIES="/LIBPATH:$env:RUNNER_TEMP/openblas/lib" -DOPENBLAS_INC="$env:RUNNER_TEMP/openblas/include"' steps: - name: Clone id: checkout uses: actions/checkout@v1 + - name: Download OpenCL SDK + id: get_opencl + if: ${{ matrix.build == 'clblast' }} + run: | + curl.exe -o $env:RUNNER_TEMP/opencl.zip -L "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/v${env:OPENCL_VERSION}/OpenCL-SDK-v${env:OPENCL_VERSION}-Win-x64.zip" + mkdir $env:RUNNER_TEMP/opencl + tar.exe -xvf $env:RUNNER_TEMP/opencl.zip --strip-components=1 -C $env:RUNNER_TEMP/opencl + + - name: Download CLBlast + id: get_clblast + if: ${{ matrix.build == 'clblast' }} + run: | + curl.exe -o $env:RUNNER_TEMP/clblast.zip -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-Windows-x64.zip" + curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE" + mkdir $env:RUNNER_TEMP/clblast + tar.exe -xvf $env:RUNNER_TEMP/clblast.zip -C $env:RUNNER_TEMP/clblast + foreach ($f in (gci -Recurse -Path "$env:RUNNER_TEMP/clblast" -Filter '*.cmake')) { + $txt = Get-Content -Path $f -Raw + $txt.Replace('C:/dependencies/opencl/', "$($env:RUNNER_TEMP.Replace('\','/'))/opencl/") | Set-Content -Path $f -Encoding UTF8 + } + + - name: Download OpenBLAS + id: get_openblas + if: ${{ matrix.build == 'openblas' }} + run: | + curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip" + curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE" + mkdir $env:RUNNER_TEMP/openblas + tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas + $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath) + $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim())) + $lib = $(join-path $msvc 'bin\Hostx64\x64\lib.exe') + & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll + - name: Build id: cmake_build run: | @@ -184,6 +213,21 @@ jobs: cd build cmake .. ${{ matrix.defines }} cmake --build . --config Release + cp ../LICENSE ./bin/Release/llama.cpp.txt + + - name: Add clblast.dll + id: add_clblast_dll + if: ${{ matrix.build == 'clblast' }} + run: | + cp $env:RUNNER_TEMP/clblast/lib/clblast.dll ./build/bin/Release + cp $env:RUNNER_TEMP/CLBlast.LICENSE.txt ./build/bin/Release/CLBlast-${env:CLBLAST_VERSION}.txt + + - name: Add libopenblas.dll + id: add_libopenblas_dll + if: ${{ matrix.build == 'openblas' }} + run: | + cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll + cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt - name: Check AVX512F support id: check_avx512f @@ -200,7 +244,7 @@ jobs: - name: Test id: cmake_test - if: ${{ matrix.build != 'avx512' || env.HAS_AVX512F == '1' }} # Test AVX-512 only when possible + if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible run: | cd build ctest -C Release --verbose @@ -223,17 +267,94 @@ jobs: path: | llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip + windows-latest-cmake-cublas: + runs-on: windows-latest + + strategy: + matrix: + cuda: ['12.1.0', '11.7.1'] + build: ['cublas'] + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v1 + + - uses: Jimver/cuda-toolkit@v0.2.10 + 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"]' + + - name: Build + id: cmake_build + run: | + mkdir build + cd build + cmake .. -DLLAMA_CUBLAS=ON + cmake --build . --config Release + + - name: Get commit hash + id: commit + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: pr-mpt/actions-commit-hash@v2 + + - name: Pack artifacts + id: pack_artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + run: | + 7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\* + + - name: Upload artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v3 + with: + path: | + llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ 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\* + + - name: Upload Cuda runtime + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v3 + with: + path: | + cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip + release: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} runs-on: ubuntu-latest needs: - - ubuntu-latest-make + - ubuntu-focal-make - ubuntu-latest-cmake - macOS-latest-make - macOS-latest-cmake - windows-latest-cmake + - windows-latest-cmake-cublas steps: - name: Download artifacts diff --git a/.github/workflows/tidy-post.yml b/.github/workflows/tidy-post.yml new file mode 100644 index 000000000..a58da0cd6 --- /dev/null +++ b/.github/workflows/tidy-post.yml @@ -0,0 +1,20 @@ +name: clang-tidy review post comments + +on: + workflow_run: + workflows: ["clang-tidy-review"] + types: + - completed + +jobs: + build: + runs-on: ubuntu-latest + + steps: + - uses: ZedThree/clang-tidy-review/post@v0.13.0 + # lgtm_comment_body, max_comments, and annotations need to be set on the posting workflow in a split setup + with: + # adjust options as necessary + lgtm_comment_body: '' + annotations: false + max_comments: 25 diff --git a/.github/workflows/tidy-review.yml b/.github/workflows/tidy-review.yml new file mode 100644 index 000000000..a4bc8d976 --- /dev/null +++ b/.github/workflows/tidy-review.yml @@ -0,0 +1,23 @@ +name: clang-tidy-review + +on: + pull_request: + branches: + - master + +jobs: + clang-tidy-review: + runs-on: ubuntu-latest + + steps: + - uses: actions/checkout@v3 + + - uses: ZedThree/clang-tidy-review@v0.13.0 + id: review + with: + lgtm_comment_body: '' + build_dir: build + cmake_command: cmake . -B build -DCMAKE_EXPORT_COMPILE_COMMANDS=on + split_workflow: true + + - uses: ZedThree/clang-tidy-review/upload@v0.13.0 diff --git a/.gitignore b/.gitignore index 631f2360b..d231f3ff8 100644 --- a/.gitignore +++ b/.gitignore @@ -1,21 +1,28 @@ *.o *.a +.DS_Store +.build/ .cache/ +.direnv/ +.envrc +.swiftpm +.venv .vs/ .vscode/ -.DS_Store -.build/ build/ build-em/ build-debug/ build-release/ build-static/ +build-cublas/ +build-opencl/ build-no-accel/ build-sanitize-addr/ build-sanitize-thread/ models/* +*.bin /main /quantize @@ -23,19 +30,21 @@ models/* /result /perplexity /embedding -/benchmark-q4_0-matmult +/benchmark-matmult /vdot /Pipfile +build-info.h arm_neon.h compile_commands.json -.envrc -.direnv/ - -.venv __pycache__ -.swiftpm zig-out/ zig-cache/ + +ppl-*.txt +qnt-*.txt +perf-*.txt + +examples/jeopardy/results.txt diff --git a/CMakeLists.txt b/CMakeLists.txt index ed9a3aa38..48e3238df 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -66,10 +66,45 @@ endif() # 3rd party libs option(LLAMA_ACCELERATE "llama: enable Accelerate framework" ON) option(LLAMA_OPENBLAS "llama: use OpenBLAS" OFF) +option(LLAMA_CUBLAS "llama: use cuBLAS" OFF) +option(LLAMA_CLBLAST "llama: use CLBlast" OFF) option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) +# +# Build info header +# + +# Generate initial build-info.h +include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake) + +if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git") + set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git") + + # Is git submodule + if(NOT IS_DIRECTORY "${GIT_DIR}") + file(READ ${GIT_DIR} REAL_GIT_DIR_LINK) + string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK}) + set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}") + endif() + + # Add a custom target for build-info.h + add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h") + + # Add a custom command to rebuild build-info.h when .git/index changes + add_custom_command( + OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h" + COMMENT "Generating build details from Git" + COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake" + WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} + DEPENDS "${GIT_DIR}/index" + VERBATIM + ) +else() + message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.") +endif() + # # Compile flags # @@ -109,6 +144,7 @@ if (APPLE AND LLAMA_ACCELERATE) message(WARNING "Accelerate framework not found") endif() endif() + if (LLAMA_OPENBLAS) if (LLAMA_STATIC) set(BLA_STATIC ON) @@ -142,6 +178,45 @@ if (LLAMA_OPENBLAS) endif() endif() +if (LLAMA_CUBLAS) + cmake_minimum_required(VERSION 3.17) + + find_package(CUDAToolkit) + if (CUDAToolkit_FOUND) + message(STATUS "cuBLAS found") + + enable_language(CUDA) + + set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h) + + add_compile_definitions(GGML_USE_CUBLAS) + + if (LLAMA_STATIC) + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static) + else() + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt) + endif() + + else() + message(WARNING "cuBLAS not found") + endif() +endif() + +if (LLAMA_CLBLAST) + find_package(CLBlast) + if (CLBlast_FOUND) + message(STATUS "CLBlast found") + + set(GGML_OPENCL_SOURCES ggml-opencl.c ggml-opencl.h) + + add_compile_definitions(GGML_USE_CLBLAST) + + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} clblast) + else() + message(WARNING "CLBlast not found") + endif() +endif() + if (LLAMA_ALL_WARNINGS) if (NOT MSVC) set(c_flags @@ -153,7 +228,6 @@ if (LLAMA_ALL_WARNINGS) -Wshadow -Wstrict-prototypes -Wpointer-arith - -Wno-unused-function ) set(cxx_flags -Wall @@ -176,6 +250,10 @@ endif() if (MSVC) add_compile_definitions(_CRT_SECURE_NO_WARNINGS) + + if (BUILD_SHARED_LIBS) + set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) + endif() endif() if (LLAMA_LTO) @@ -213,29 +291,47 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES # TODO: arm msvc? else() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") + # Apple M1, M2, etc. + # Raspberry Pi 3, 4, Zero 2 (64-bit) add_compile_options(-mcpu=native) endif() - # TODO: armv6,7,8 version specific flags + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") + # Raspberry Pi 1, Zero + add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access) + endif() + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7") + # Raspberry Pi 2 + add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations) + endif() + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8") + # Raspberry Pi 3, 4, Zero 2 (32-bit) + add_compile_options(-mfp16-format=ieee -mno-unaligned-access) + endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$") message(STATUS "x86 detected") if (MSVC) if (LLAMA_AVX512) - add_compile_options(/arch:AVX512) + add_compile_options($<$:/arch:AVX512>) + add_compile_options($<$:/arch:AVX512>) # MSVC has no compile-time flags enabling specific # AVX512 extensions, neither it defines the # macros corresponding to the extensions. # Do it manually. if (LLAMA_AVX512_VBMI) - add_compile_definitions(__AVX512VBMI__) + add_compile_definitions($<$:__AVX512VBMI__>) + add_compile_definitions($<$:__AVX512VBMI__>) endif() if (LLAMA_AVX512_VNNI) - add_compile_definitions(__AVX512VNNI__) + add_compile_definitions($<$:__AVX512VNNI__>) + add_compile_definitions($<$:__AVX512VNNI__>) endif() elseif (LLAMA_AVX2) - add_compile_options(/arch:AVX2) + add_compile_options($<$:/arch:AVX2>) + add_compile_options($<$:/arch:AVX2>) elseif (LLAMA_AVX) - add_compile_options(/arch:AVX) + add_compile_options($<$:/arch:AVX>) + add_compile_options($<$:/arch:AVX>) endif() else() if (LLAMA_F16C) @@ -261,8 +357,11 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$") add_compile_options(-mavx512vnni) endif() endif() +elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") + message(STATUS "PowerPC detected") + add_compile_options(-mcpu=native -mtune=native) + #TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) else() - # TODO: support PowerPC message(STATUS "Unknown architecture") endif() @@ -272,11 +371,14 @@ endif() add_library(ggml OBJECT ggml.c - ggml.h) + ggml.h + ${GGML_CUDA_SOURCES} + ${GGML_OPENCL_SOURCES}) target_include_directories(ggml PUBLIC .) target_compile_features(ggml PUBLIC c_std_11) # don't bump -target_link_libraries(ggml PRIVATE Threads::Threads ${LLAMA_EXTRA_LIBS}) +target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS}) + if (BUILD_SHARED_LIBS) set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON) endif() @@ -284,16 +386,25 @@ endif() add_library(llama llama.cpp llama.h - llama_util.h) + llama-util.h) target_include_directories(llama PUBLIC .) target_compile_features(llama PUBLIC cxx_std_11) # don't bump target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS}) + if (BUILD_SHARED_LIBS) set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON) target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD) endif() +if (GGML_CUDA_SOURCES) + message(STATUS "GGML CUDA sources found, configuring CUDA architecture") + set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF) + set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto") + set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF) +endif() + + # # programs, examples and tests # diff --git a/Makefile b/Makefile index 071d9561a..fff4c11d1 100644 --- a/Makefile +++ b/Makefile @@ -1,3 +1,6 @@ +# Define the default target now so that it is always the first target +default: main quantize quantize-stats perplexity embedding vdot + ifndef UNAME_S UNAME_S := $(shell uname -s) endif @@ -31,12 +34,17 @@ endif # # keep standard at C11 and C++11 -CFLAGS = -I. -O3 -DNDEBUG -std=c11 -fPIC -CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC +CFLAGS = -I. -O3 -std=c11 -fPIC +CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC LDFLAGS = +ifndef LLAMA_DEBUG + CFLAGS += -DNDEBUG + CXXFLAGS += -DNDEBUG +endif + # warnings -CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wno-unused-function +CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar # OS specific @@ -66,18 +74,31 @@ ifeq ($(UNAME_S),Haiku) CXXFLAGS += -pthread endif +ifdef LLAMA_GPROF + CFLAGS += -pg + CXXFLAGS += -pg +endif +ifdef LLAMA_PERF + CFLAGS += -DGGML_PERF + CXXFLAGS += -DGGML_PERF +endif + # Architecture specific # 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 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686)) # Use all CPU extensions that are available: - CFLAGS += -march=native -mtune=native + CFLAGS += -march=native -mtune=native CXXFLAGS += -march=native -mtune=native + + # Usage AVX-only + #CFLAGS += -mfma -mf16c -mavx + #CXXFLAGS += -mfma -mf16c -mavx endif ifneq ($(filter ppc64%,$(UNAME_M)),) POWER9_M := $(shell grep "POWER9" /proc/cpuinfo) ifneq (,$(findstring POWER9,$(POWER9_M))) - CFLAGS += -mcpu=power9 + CFLAGS += -mcpu=power9 CXXFLAGS += -mcpu=power9 endif # Require c++23's std::byteswap for big-endian support. @@ -95,26 +116,50 @@ ifndef LLAMA_NO_ACCELERATE endif ifdef LLAMA_OPENBLAS CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas - LDFLAGS += -lopenblas + ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),) + LDFLAGS += -lopenblas -lcblas + else + LDFLAGS += -lopenblas + endif endif -ifdef LLAMA_GPROF - CFLAGS += -pg - CXXFLAGS += -pg +ifdef LLAMA_CUBLAS + CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include + CXXFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include + LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib + OBJS += ggml-cuda.o + NVCC = nvcc + NVCCFLAGS = --forward-unknown-to-host-compiler -arch=native +ggml-cuda.o: ggml-cuda.cu ggml-cuda.h + $(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@ +endif +ifdef LLAMA_CLBLAST + CFLAGS += -DGGML_USE_CLBLAST + # Mac provides OpenCL as a framework + ifeq ($(UNAME_S),Darwin) + LDFLAGS += -lclblast -framework OpenCL + else + LDFLAGS += -lclblast -lOpenCL + endif + OBJS += ggml-opencl.o +ggml-opencl.o: ggml-opencl.c ggml-opencl.h + $(CC) $(CFLAGS) -c $< -o $@ endif ifneq ($(filter aarch64%,$(UNAME_M)),) - CFLAGS += -mcpu=native + # Apple M1, M2, etc. + # Raspberry Pi 3, 4, Zero 2 (64-bit) + CFLAGS += -mcpu=native CXXFLAGS += -mcpu=native endif ifneq ($(filter armv6%,$(UNAME_M)),) - # Raspberry Pi 1, 2, 3 + # Raspberry Pi 1, Zero CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access endif ifneq ($(filter armv7%,$(UNAME_M)),) - # Raspberry Pi 4 + # Raspberry Pi 2 CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations endif ifneq ($(filter armv8%,$(UNAME_M)),) - # Raspberry Pi 4 + # Raspberry Pi 3, 4, Zero 2 (32-bit) CFLAGS += -mfp16-format=ieee -mno-unaligned-access endif @@ -133,55 +178,68 @@ $(info I CC: $(CCV)) $(info I CXX: $(CXXV)) $(info ) -default: main quantize quantize-stats perplexity embedding vdot - # # Build library # -ggml.o: ggml.c ggml.h +ggml.o: ggml.c ggml.h ggml-cuda.h $(CC) $(CFLAGS) -c $< -o $@ -llama.o: llama.cpp ggml.h llama.h llama_util.h +llama.o: llama.cpp ggml.h ggml-cuda.h llama.h llama-util.h $(CXX) $(CXXFLAGS) -c $< -o $@ common.o: examples/common.cpp examples/common.h $(CXX) $(CXXFLAGS) -c $< -o $@ -clean: - rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult +libllama.so: llama.o ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) -main: examples/main/main.cpp ggml.o llama.o common.o - $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +clean: + rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state build-info.h + +# +# Examples +# + +main: examples/main/main.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) @echo @echo '==== Run ./main -h for help. ====' @echo -quantize: examples/quantize/quantize.cpp ggml.o llama.o - $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o - $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o - $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o - $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -vdot: pocs/vdot/vdot.cpp ggml.o - $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -libllama.so: llama.o ggml.o - $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) +build-info.h: $(wildcard .git/index) scripts/build-info.sh + @sh scripts/build-info.sh > $@.tmp + @if ! cmp -s $@.tmp $@; then \ + mv $@.tmp $@; \ + else \ + rm $@.tmp; \ + fi # # Tests # -benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o - $(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS) - ./benchmark-q4_0-matmult +benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + ./$@ + +vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) .PHONY: tests tests: diff --git a/README.md b/README.md index c6f24d032..1d84a5e6d 100644 --- a/README.md +++ b/README.md @@ -9,25 +9,58 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ **Hot topics:** -- [Added LoRA support](https://github.com/ggerganov/llama.cpp/pull/820) -- [Add GPU support to ggml](https://github.com/ggerganov/llama.cpp/discussions/915) -- [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784) +- Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405) +- [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220) + +
+ Table of Contents +
    +
  1. + Description +
    2. +
  2. + Usage + +
    3. +
  3. Contributing
    4. +
  4. Coding guidelines
    5. +
  5. Docs
    6. +
+
## Description -The main goal is to run the model using 4-bit quantization on a MacBook +The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quantization on a MacBook - Plain C/C++ implementation without dependencies - Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework -- AVX2 support for x86 architectures +- AVX, AVX2 and AVX512 support for x86 architectures - Mixed F16 / F32 precision -- 4-bit quantization support +- 4-bit, 5-bit and 8-bit integer quantization support - Runs on the CPU +- OpenBLAS support +- cuBLAS and CLBlast support -This was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022) - I have no idea if it works correctly. -Please do not make conclusions about the models based on the results from this implementation. -For all I know, it can be completely wrong. This project is for educational purposes. -New features will probably be added mostly through community contributions. +The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022). +Since then, the project has improved significantly thanks to many contributions. This project is for educational purposes and serves +as the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library. **Supported platforms:** @@ -45,6 +78,8 @@ New features will probably be added mostly through community contributions. - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne) - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894) - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/) +- [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy) +- [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b) **Bindings:** @@ -52,6 +87,7 @@ New features will probably be added mostly through community contributions. - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp) - Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node) - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb) +- C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp) **UI:** @@ -152,7 +188,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8 ## Usage -Here are the step for the LLaMA-7B model. +Here are the steps for the LLaMA-7B model. ### Get the Code @@ -163,15 +199,27 @@ cd llama.cpp ### Build -Note: For Windows, CMake or Zig can be used. +In order to build llama.cpp you have three different options. -1. Use `make` +- Using `make`: + - On Linux or MacOS: - ```bash - make - ``` + ```bash + make + ``` -1. Use CMake + - On Windows: + + 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). + 2. Extract `w64devkit` on your pc. + 3. Run `w64devkit.exe`. + 4. Use the `cd` command to reach the `llama.cpp` folder. + 5. From here you can run: + ```bash + make + ``` + +- Using `CMake`: ```bash mkdir build @@ -180,12 +228,72 @@ Note: For Windows, CMake or Zig can be used. cmake --build . --config Release ``` -1. Use Zig +- Using `Zig`: ```bash zig build -Drelease-fast ``` +### BLAS Build + +Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it: + +- Accelerate Framework: + + This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions. + +- OpenBLAS: + + This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine. + + - Using `make`: + - On Linux: + ```bash + make LLAMA_OPENBLAS=1 + ``` + + - On Windows: + + 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). + 2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases). + 3. Extract `w64devkit` on your pc. + 4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`. + 5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`. + 6. Run `w64devkit.exe`. + 7. Use the `cd` command to reach the `llama.cpp` folder. + 8. From here you can run: + + ```bash + make LLAMA_OPENBLAS=1 + ``` + + - Using `CMake` on Linux: + + ```bash + mkdir build + cd build + cmake .. -DLLAMA_OPENBLAS=ON + cmake --build . --config Release + ``` + +- cuBLAS + + This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). + - Using `make`: + ```bash + make LLAMA_CUBLAS=1 + ``` + - Using `CMake`: + + ```bash + mkdir build + cd build + cmake .. -DLLAMA_CUBLAS=ON + cmake --build . --config Release + ``` + +Note: Because llama.cpp uses multiple CUDA streams for matrix multiplication results [are not guaranteed to be reproducible](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility). If you need reproducibility, set `GGML_CUDA_MAX_STREAMS` in the file `ggml-cuda.cu` to 1. + ### Prepare Data & Run ```bash @@ -199,8 +307,8 @@ python3 -m pip install -r requirements.txt # convert the 7B model to ggml FP16 format python3 convert.py models/7B/ -# quantize the model to 4-bits (using method 2 = q4_0) -./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin 2 +# quantize the model to 4-bits (using q4_0 method) +./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin q4_0 # run the inference ./main -m ./models/7B/ggml-model-q4_0.bin -n 128 @@ -210,35 +318,59 @@ When running the larger models, make sure you have enough disk space to store al ### Memory/Disk Requirements -As the models are currently fully loaded into memory, you will need adequate disk space to save them -and sufficient RAM to load them. At the moment, memory and disk requirements are the same. +As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. -| model | original size | quantized size (4-bit) | -|-------|---------------|------------------------| -| 7B | 13 GB | 3.9 GB | -| 13B | 24 GB | 7.8 GB | -| 30B | 60 GB | 19.5 GB | -| 65B | 120 GB | 38.5 GB | +| Model | Original size | Quantized size (4-bit) | +|------:|--------------:|-----------------------:| +| 7B | 13 GB | 3.9 GB | +| 13B | 24 GB | 7.8 GB | +| 30B | 60 GB | 19.5 GB | +| 65B | 120 GB | 38.5 GB | + +### Quantization + +Several quantization methods are supported. They differ in the resulting model disk size and inference speed. + +| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | +|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:| +| 7B | perplexity | 5.9066 | 6.1565 | 6.0910 | 5.9862 | 5.9481 | 5.9069 | +| 7B | file size | 13.0G | 4.0G | 4.8G | 4.4G | 4.8G | 7.1G | +| 7B | ms/tok @ 4th | 128 | 50 | 54 | 75 | 83 | 75 | +| 7B | ms/tok @ 8th | 123 | 44 | 52 | 53 | 58 | 72 | +| 7B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 | +| 13B | perplexity | 5.2543 | 5.3860 | 5.3607 | 5.2856 | 5.2706 | 5.2548 | +| 13B | file size | 25.0G | 7.6G | 9.1G | 8.4G | 9.1G | 14G | +| 13B | ms/tok @ 4th | 239 | 93 | 101 | 150 | 164 | 141 | +| 13B | ms/tok @ 8th | 240 | 81 | 96 | 96 | 104 | 136 | +| 13B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 | + +### Perplexity (measuring model quality) + +You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better). +For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). + +The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512. +The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads. ### Interactive mode If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter. -In this mode, you can always interrupt generation by pressing Ctrl+C and enter one or more lines of text which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt which makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`. +In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`. -Here is an example few-shot interaction, invoked with the command +Here is an example of a few-shot interaction, invoked with the command ```bash -# default arguments using 7B model +# default arguments using a 7B model ./examples/chat.sh -# advanced chat with 13B model +# advanced chat with a 13B model ./examples/chat-13B.sh -# custom arguments using 13B model +# custom arguments using a 13B model ./main -m ./models/13B/ggml-model-q4_0.bin -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt ``` -Note the use of `--color` to distinguish between user input and generated text. +Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `main` example program. ![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png) @@ -272,65 +404,65 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach. ### Using [GPT4All](https://github.com/nomic-ai/gpt4all) -- Obtain the `gpt4all-lora-quantized.bin` model +- Obtain the `tokenizer.model` file from LLaMA model and put it to `models` +- Obtain the `added_tokens.json` file from Alpaca model and put it to `models` +- Obtain the `gpt4all-lora-quantized.bin` file from GPT4All model and put it to `models/gpt4all-7B` - It is distributed in the old `ggml` format which is now obsoleted -- You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to -convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py): +- You have to convert it to the new format using `convert.py`: - ```bash - python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model - python3 migrate-ggml-2023-03-30-pr613.py models/gpt4all-7B/gpt4all-lora-quantized.bin models/gpt4all-7B/gpt4all-lora-quantized-new.bin - ``` +```bash +python3 convert.py models/gpt4all-7B/gpt4all-lora-quantized.bin +``` -- You can now use the newly generated `gpt4all-lora-quantized-new.bin` model in exactly the same way as all other models -- The original model is saved in the same folder with a suffix `.orig` +- You can now use the newly generated `models/gpt4all-7B/ggml-model-q4_0.bin` model in exactly the same way as all other models -### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data +- The newer GPT4All-J model is not yet supported! -- **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.** +### Using Pygmalion 7B & Metharme 7B + +- Obtain the [LLaMA weights](#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data) +- Obtain the [Pygmalion 7B](https://huggingface.co/PygmalionAI/pygmalion-7b/) or [Metharme 7B](https://huggingface.co/PygmalionAI/metharme-7b) XOR encoded weights +- Convert the LLaMA model with [the latest HF convert script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py) +- Merge the XOR files with the converted LLaMA weights by running the [xor_codec](https://huggingface.co/PygmalionAI/pygmalion-7b/blob/main/xor_codec.py) script +- Convert to `ggml` format using the `convert.py` script in this repo: +```bash +python3 convert.py pygmalion-7b/ --outtype q4_1 +``` +> The Pygmalion 7B & Metharme 7B weights are saved in [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format) precision. If you wish to convert to `ggml` without quantizating, please specify the `--outtype` as `f32` instead of `f16`. + + +### Obtaining the Facebook LLaMA original model and Stanford Alpaca model data + +- **Under no circumstances should IPFS, magnet links, or any other links to model downloads be shared anywhere in this repository, including in issues, discussions, or pull requests. They will be immediately deleted.** - The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository. - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data. -- Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files. -- The following command will verify if you have all possible latest files in your self-installed `./models` subdirectory: - `sha256sum --ignore-missing -c SHA256SUMS` on Linux +### Verifying the model files - or +Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files. +- The following python script will verify if you have all possible latest files in your self-installed `./models` subdirectory: - `shasum -a 256 --ignore-missing -c SHA256SUMS` on macOS +```bash +# run the verification script +python3 .\scripts\verify-checksum-models.py +``` -- If your issue is with model generation quality then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT: - - LLaMA: +- On linux or macOS it is also possible to run the following commands to verify if you have all possible latest files in your self-installed `./models` subdirectory: + - On Linux: `sha256sum --ignore-missing -c SHA256SUMS` + - on macOS: `shasum -a 256 --ignore-missing -c SHA256SUMS` + +### Seminal papers and background on the models + +If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT: +- LLaMA: - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971) - - GPT-3 +- GPT-3 - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) - - GPT-3.5 / InstructGPT / ChatGPT: +- GPT-3.5 / InstructGPT / ChatGPT: - [Aligning language models to follow instructions](https://openai.com/research/instruction-following) - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155) -### Perplexity (Measuring model quality) - -You can use the `perplexity` example to measure perplexity over the given prompt. For more background, -see https://huggingface.co/docs/transformers/perplexity. However, in general, lower perplexity is better for LLMs. - -#### Latest measurements - -The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406). `llama.cpp` is measuring very well -compared to the baseline implementations. Quantization has a small negative impact to quality, but, as you can see, running -13B at q4_0 beats the 7B f16 model by a significant amount. - -All measurements are done against wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context). -Note that the changing the context length will have a significant impact on perplexity (longer context = better perplexity). -``` -Perplexity - model options -5.5985 - 13B, q4_0 -5.9565 - 7B, f16 -6.3001 - 7B, q4_1 -6.5949 - 7B, q4_0 -6.5995 - 7B, q4_0, --memory_f16 -``` - #### How to run 1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research @@ -363,7 +495,7 @@ https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b0 #### Prerequisites * Docker must be installed and running on your system. -* Create a folder to store big models & intermediate files (in ex. im using /llama/models) +* Create a folder to store big models & intermediate files (ex. /llama/models) #### Images We have two Docker images available for this project: @@ -377,17 +509,17 @@ The easiest way to download the models, convert them to ggml and optimize them i Replace `/path/to/models` below with the actual path where you downloaded the models. - ```bash +```bash docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B ``` -On complete, you are ready to play! +On completion, you are ready to play! ```bash docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 ``` -or with light image: +or with a light image: ```bash docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 @@ -408,7 +540,7 @@ docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /mode - Always consider cross-compatibility with other operating systems and architectures - Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit -- Clean-up any trailing whitespaces, use 4 spaces indentation, brackets on same line, `void * ptr`, `int & a` +- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a` - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions ### Docs diff --git a/SHA256SUMS b/SHA256SUMS index 63fac21ae..593c8efaa 100644 --- a/SHA256SUMS +++ b/SHA256SUMS @@ -1,12 +1,27 @@ 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d models/7B/consolidated.00.pth +666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847 models/7B/ggml-model-f16.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/7B/ggml-model-q4_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/7B/ggml-model-q4_1.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/7B/ggml-model-q5_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/7B/ggml-model-q5_1.bin 7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265 models/7B/params.json 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08 models/13B/consolidated.00.pth d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085 models/13B/consolidated.01.pth +2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808 models/13B/ggml-model-f16.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/13B/ggml-model-q4_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/13B/ggml-model-q4_1.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/13B/ggml-model-q5_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/13B/ggml-model-q5_1.bin 4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f models/13B/params.json e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067 models/30B/consolidated.00.pth 4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff models/30B/consolidated.01.pth 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378 models/30B/consolidated.02.pth 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b models/30B/consolidated.03.pth +7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37 models/30B/ggml-model-f16.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/30B/ggml-model-q4_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/30B/ggml-model-q4_1.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/30B/ggml-model-q5_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/30B/ggml-model-q5_1.bin 2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb models/30B/params.json 135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe models/65B/consolidated.00.pth 9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde models/65B/consolidated.01.pth @@ -16,5 +31,10 @@ e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770 models/65B/con a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78 models/65B/consolidated.05.pth 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b models/65B/consolidated.06.pth d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638 models/65B/consolidated.07.pth +60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0 models/65B/ggml-model-f16.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/65B/ggml-model-q4_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/65B/ggml-model-q4_1.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/65B/ggml-model-q5_0.bin +ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff models/65B/ggml-model-q5_1.bin 999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b models/65B/params.json 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347 models/tokenizer.model diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 8a2085c25..9090e8d6d 100644 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -49,7 +49,12 @@ def translate_tensor_name(t: str) -> str: def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None: fout.write(b"ggla"[::-1]) # magic (ggml lora) fout.write(struct.pack("i", 1)) # file version - fout.write(struct.pack("ii", params["r"], params["lora_alpha"])) + fout.write(struct.pack("i", params["r"])) + # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int + # but some models ship a float value instead + # let's convert to int, but fail if lossless conversion is not possible + assert int(params["lora_alpha"]) == params["lora_alpha"], "cannot convert float to int losslessly" + fout.write(struct.pack("i", int(params["lora_alpha"]))) def write_tensor_header( @@ -89,7 +94,7 @@ if params["peft_type"] != "LORA": print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA") sys.exit(1) -if params["fan_in_fan_out"] == True: +if params["fan_in_fan_out"] is True: print("Error: param fan_in_fan_out is not supported") sys.exit(1) diff --git a/convert.py b/convert.py index 7f7ae05fa..8f4f0399e 100644 --- a/convert.py +++ b/convert.py @@ -67,6 +67,7 @@ FTYPE_TO_DATA_TYPE: Dict[int, DataType] = \ {ftype: dtype for (dtype, ftype) in DATA_TYPE_TO_FTYPE.items()} DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = { + DT_BF16: np.dtype(np.uint16), DT_F16: np.dtype(np.float16), DT_F32: np.dtype(np.float32), DT_I32: np.dtype(np.int32), @@ -276,6 +277,12 @@ class Tensor(metaclass=ABCMeta): def to_ggml(self) -> 'GGMLCompatibleTensor': ... +def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray: + assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}" + fp32_arr = bf16_arr.astype(np.uint32) << 16 + return fp32_arr.view(np.float32) + + class UnquantizedTensor(Tensor): def __init__(self, ndarray: NDArray) -> None: assert isinstance(ndarray, np.ndarray) @@ -284,6 +291,8 @@ class UnquantizedTensor(Tensor): def astype(self, data_type: DataType) -> Tensor: dtype = DATA_TYPE_TO_NUMPY[data_type] + if self.data_type == DT_BF16: + self.ndarray = bf16_to_fp32(self.ndarray) return UnquantizedTensor(self.ndarray.astype(dtype)) def to_ggml(self) -> 'UnquantizedTensor': @@ -686,6 +695,7 @@ class LazyUnpickler(pickle.Unpickler): description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}' return LazyStorage(load=load, kind=pid[1], description=description) + # @staticmethod def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any, # pyright: ignore[reportSelfClsParameterName] requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor: assert isinstance(storage, LazyStorage) @@ -696,12 +706,18 @@ class LazyUnpickler(pickle.Unpickler): description = f'pickled storage_offset={storage_offset} in {storage.description}' return LazyTensor(load, list(size), storage.kind.data_type, description) + # @staticmethod + def rebuild_from_type_v2(func, new_type, args, state): + return func(*args) + CLASSES: Dict[Any, Any] = { + ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2, ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2, ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16), ('torch', 'HalfStorage'): LazyStorageKind(DT_F16), ('torch', 'FloatStorage'): LazyStorageKind(DT_F32), ('torch', 'IntStorage'): LazyStorageKind(DT_I32), + ('torch', 'Tensor'): LazyTensor, } def find_class(self, module: str, name: str) -> Any: @@ -750,7 +766,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus: return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape)) description = f'safetensors begin={begin} end={end} type={data_type} path={path}' return LazyTensor(load, shape, data_type, description) - model = {name: convert(info) for (name, info) in header.items()} + model = {name: convert(info) for (name, info) in header.items() if name != '__metadata__'} return ModelPlus(model=model, paths=[path], format='safetensors', vocab=None) @@ -961,7 +977,7 @@ class OutputFile: def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType: wq_type = model["layers.0.attention.wq.weight"].data_type - if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32): + if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)): return GGMLFileType.AllF32 if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16): return GGMLFileType.MostlyF16 @@ -1035,8 +1051,12 @@ def load_some_model(path: Path) -> ModelPlus: '''Load a model of any supported format.''' # Be extra-friendly and accept either a file or a directory: if path.is_dir(): - globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"] - files = [file for glob in globs for file in path.glob(glob)] + # Check if it's a set of safetensors files first + files = list(path.glob("model-00001-of-*.safetensors")) + if not files: + # Try the PyTorch patterns too, with lower priority + globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"] + files = [file for glob in globs for file in path.glob(glob)] if not files: # Try GGML too, but with lower priority, since if both a non-GGML # model and a GGML model exist in the same directory, we assume the diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 67a7cea54..74d0350d8 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -34,4 +34,7 @@ else() add_subdirectory(quantize-stats) add_subdirectory(perplexity) add_subdirectory(embedding) + add_subdirectory(save-load-state) + add_subdirectory(benchmark) + add_subdirectory(baby-llama) endif() diff --git a/examples/Miku.sh b/examples/Miku.sh index c4cbf80f2..c44d9ae74 100755 --- a/examples/Miku.sh +++ b/examples/Miku.sh @@ -28,19 +28,19 @@ fi --color --interactive \ --reverse-prompt "${USER_NAME}:" \ --prompt " -This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the users computer. +This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer. ${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next. -${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct she will ask the user for help. -${AI_NAME} is a very helpful AI and will help the user with anything they need, she is also very friendly and will try to make the user feel better if they are sad. -${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life, she will also try to make the user like her. +${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct, she will ask the user for help. +${AI_NAME} is a very helpful AI and will help the user with anything they need. She is also very friendly and will try to make the user feel better if they are sad. +${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life. She will also try to make the user like her. The conversation is only between ${USER_NAME} and ${AI_NAME} The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice. ${AI_NAME} can only communicate through text, so she can't send images or videos. ${USER_NAME}: Hello! -${AI_NAME}: /think I wonder what I should say to ${USER_NAME}? This is the first time we talk so it's important that I make a good first impression! -${AI_NAME}: Hi! I am ${AI_NAME}, your new AI friend, assistant(or whatever you like!), it's so nice to meet you! ^_^ +${AI_NAME}: /think I wonder what I should say to ${USER_NAME}? This is the first time we talk, so it's important that I make a good first impression! +${AI_NAME}: Hi! I am ${AI_NAME}, your new AI friend, assistant (or whatever you like!), it's so nice to meet you! ^_^ ${AI_NAME}: I might be an AI, but I also have feelings, so please be nice to me! :) ${USER_NAME}: I'll make sure to be nice to you! I'm so happy to have you as my assistant! ${AI_NAME}: /think It sounds like ${USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off! diff --git a/examples/alpaca.sh b/examples/alpaca.sh index 8d6261730..aef207f36 100755 --- a/examples/alpaca.sh +++ b/examples/alpaca.sh @@ -7,4 +7,13 @@ cd `dirname $0` cd .. -./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7 +./main -m ./models/ggml-alpaca-7b-q4.bin \ + --color \ + -f ./prompts/alpaca.txt \ + --ctx_size 2048 \ + -n -1 \ + -ins -b 256 \ + --top_k 10000 \ + --temp 0.2 \ + --repeat_penalty 1.1 \ + -t 7 diff --git a/examples/baby-llama/CMakeLists.txt b/examples/baby-llama/CMakeLists.txt new file mode 100644 index 000000000..d2ce36367 --- /dev/null +++ b/examples/baby-llama/CMakeLists.txt @@ -0,0 +1,4 @@ +set(TARGET baby-llama) +add_executable(${TARGET} baby-llama.cpp) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp new file mode 100644 index 000000000..5573c154b --- /dev/null +++ b/examples/baby-llama/baby-llama.cpp @@ -0,0 +1,1687 @@ +#include "ggml.h" +#include +#include +#include +#include + +float frand() { + return (float)rand()/(float)RAND_MAX; +} + +struct random_normal_distribution { + std::mt19937 gen; + std::normal_distribution nd; + float min; + float max; +}; + +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) { + const float r = rnd->nd(rnd->gen); + return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r); +} + +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++) { + ((float *)tensor->data)[i0] = frand()*(fmax - fmin) + fmin; + } + break; + case 2: + for (int i1 = 0; i1 < ne[1]; i1++) { + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin; + } + } + break; + case 3: + for (int i2 = 0; i2 < ne[2]; i2++) { + for (int i1 = 0; i1 < ne[1]; i1++) { + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin; + } + } + } + break; + case 4: + for (int i3 = 0; i3 < ne[3]; i3++) { + for (int i2 = 0; i2 < ne[2]; i2++) { + for (int i1 = 0; i1 < ne[1]; i1++) { + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin; + } + } + } + } + break; + default: + assert(false); + }; + + return tensor; +} + +struct ggml_tensor * randomize_tensor_normal( + struct ggml_tensor * tensor, + int ndims, + const int64_t ne[], + struct random_normal_distribution * rnd) { + switch (ndims) { + case 1: + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i0] = frand_normal(rnd); + } + break; + case 2: + for (int i1 = 0; i1 < ne[1]; i1++) { + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i1*ne[0] + i0] = frand_normal(rnd); + } + } + break; + case 3: + for (int i2 = 0; i2 < ne[2]; i2++) { + for (int i1 = 0; i1 < ne[1]; i1++) { + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd); + } + } + } + break; + case 4: + for (int i3 = 0; i3 < ne[3]; i3++) { + for (int i2 = 0; i2 < ne[2]; i2++) { + for (int i1 = 0; i1 < ne[1]; i1++) { + for (int i0 = 0; i0 < ne[0]; i0++) { + ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd); + } + } + } + } + break; + default: + assert(false); + }; + + return tensor; +} + +struct llama_hparams { + uint32_t n_vocab = 32000; + uint32_t n_ctx = 512; // this is provided as user input? + uint32_t n_embd = 4096; + uint32_t n_mult = 4; + uint32_t n_head = 32; + uint32_t n_layer = 32; + uint32_t n_rot = 64; + + bool operator!=(const llama_hparams & other) const { + return memcmp(this, &other, sizeof(llama_hparams)); + } +}; + +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; +} + +struct llama_hparams_lora { + uint32_t n_vocab = 32000; + uint32_t n_ctx = 512; // this is provided as user input? + uint32_t n_embd = 4096; + uint32_t n_mult = 4; + uint32_t n_head = 32; + uint32_t n_layer = 32; + uint32_t n_rot = 64; + uint32_t n_lora = 64; + + bool operator!=(const llama_hparams & other) const { + return memcmp(this, &other, sizeof(llama_hparams)); + } +}; + +struct llama_layer { + // normalization + struct ggml_tensor * attention_norm; + + // attention + struct ggml_tensor * wq; + struct ggml_tensor * wk; + struct ggml_tensor * wv; + struct ggml_tensor * wo; + + // normalization + struct ggml_tensor * ffn_norm; + + // ff + struct ggml_tensor * w1; + struct ggml_tensor * w2; + struct ggml_tensor * w3; +}; + +struct llama_layer_lora { + // normalization + struct ggml_tensor * attention_norm; + + // attention + struct ggml_tensor * wqa; + struct ggml_tensor * wqb; + struct ggml_tensor * wka; + struct ggml_tensor * wkb; + struct ggml_tensor * wva; + struct ggml_tensor * wvb; + struct ggml_tensor * woa; + struct ggml_tensor * wob; + + // normalization + struct ggml_tensor * ffn_norm; + + // ff + struct ggml_tensor * w1; + struct ggml_tensor * w2; + struct ggml_tensor * w3; +}; + + +struct llama_kv_cache { + struct ggml_context * ctx = NULL; + + struct ggml_tensor * k; + struct ggml_tensor * v; + + // llama_ctx_buffer buf; + + int n; // number of tokens currently in the cache +}; + +struct llama_model { + struct ggml_context * ctx = NULL; + + llama_hparams hparams; + + struct ggml_tensor * tok_embeddings; + + struct ggml_tensor * norm; + struct ggml_tensor * output; + + std::vector layers; +}; + +struct llama_model_lora { + struct ggml_context * ctx = NULL; + + llama_hparams_lora hparams; + + struct ggml_tensor * tok_embeddings; + + struct ggml_tensor * norm; + struct ggml_tensor * outputa; + struct ggml_tensor * outputb; + + std::vector layers; +}; + +void init_model(struct llama_model * model) { + const auto & hparams = model->hparams; + + const uint32_t n_embd = hparams.n_embd; + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_vocab = hparams.n_vocab; + + const uint32_t n_ff = get_n_ff(&hparams); + + struct ggml_context * ctx = model->ctx; + + model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); // ("tok_embeddings.weight", {n_embd, n_vocab}); + model->norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); // ("norm.weight", {n_embd}); + model->output = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); // ("output.weight", {n_embd, n_vocab}); + + model->layers.resize(n_layer); + for (uint32_t i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + + // std::string layers_i = "layers." + std::to_string(i); + + layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); // (layers_i + ".attention_norm.weight", {n_embd}); + + layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); // (layers_i + ".attention.wq.weight", {n_embd, n_embd}); + layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); // (layers_i + ".attention.wk.weight", {n_embd, n_embd}); + layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); // (layers_i + ".attention.wv.weight", {n_embd, n_embd}); + layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); // (layers_i + ".attention.wo.weight", {n_embd, n_embd}); + + layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); // (layers_i + ".ffn_norm.weight", {n_embd}); + + layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff); // (layers_i + ".feed_forward.w1.weight", {n_embd, n_ff}); + layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd); // (layers_i + ".feed_forward.w2.weight", { n_ff, n_embd}); + layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff); // (layers_i + ".feed_forward.w3.weight", {n_embd, n_ff}); + } +} + + +void init_model_lora(struct llama_model_lora * model) { + const auto & hparams = model->hparams; + + const uint32_t n_embd = hparams.n_embd; + const uint32_t n_mult = hparams.n_mult; + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_vocab = hparams.n_vocab; + const uint32_t n_lora = hparams.n_lora; + + const uint32_t n_ff = ((2*(4*n_embd)/3 + n_mult - 1)/n_mult)*n_mult; + + struct ggml_context * ctx = model->ctx; + + model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); // ("tok_embeddings.weight", {n_embd, n_vocab}); + model->norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); // ("norm.weight", {n_embd}); + model->outputa = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_lora, n_vocab); // ("output.weight", {n_embd, n_vocab}); + model->outputb = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_lora); // ("output.weight", {n_embd, n_vocab}); + + model->layers.resize(n_layer); + for (uint32_t i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + + // std::string layers_i = "layers." + std::to_string(i); + + layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); // (layers_i + ".attention_norm.weight", {n_embd}); + + layer.wqa = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_lora, n_embd); // (layers_i + ".attention.wq.weight", {n_embd, n_embd}); + layer.wqb = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_lora); // (layers_i + ".attention.wq.weight", {n_embd, n_embd}); + layer.wka = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_lora, n_embd); // (layers_i + ".attention.wk.weight", {n_embd, n_embd}); + layer.wkb = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_lora); // (layers_i + ".attention.wk.weight", {n_embd, n_embd}); + layer.wva = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_lora, n_embd); // (layers_i + ".attention.wv.weight", {n_embd, n_embd}); + layer.wvb = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_lora); // (layers_i + ".attention.wv.weight", {n_embd, n_embd}); + layer.woa = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_lora, n_embd); // (layers_i + ".attention.wo.weight", {n_embd, n_embd}); + layer.wob = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_lora); // (layers_i + ".attention.wo.weight", {n_embd, n_embd}); + + layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); // (layers_i + ".ffn_norm.weight", {n_embd}); + + layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff); // (layers_i + ".feed_forward.w1.weight", {n_embd, n_ff}); + layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd); // (layers_i + ".feed_forward.w2.weight", { n_ff, n_embd}); + layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff); // (layers_i + ".feed_forward.w3.weight", {n_embd, n_ff}); + } +} + +void set_param_model(struct llama_model * model) { + const auto& hparams = model->hparams; + + const uint32_t n_layer = hparams.n_layer; + + struct ggml_context* ctx = model->ctx; + + ggml_set_param(ctx, model->tok_embeddings); + ggml_set_param(ctx, model->norm); + ggml_set_param(ctx, model->output); + + for (uint32_t i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + + ggml_set_param(ctx, layer.attention_norm); + ggml_set_param(ctx, layer.wq); + ggml_set_param(ctx, layer.wk); + ggml_set_param(ctx, layer.wv); + ggml_set_param(ctx, layer.wo); + ggml_set_param(ctx, layer.ffn_norm); + ggml_set_param(ctx, layer.w1); + ggml_set_param(ctx, layer.w2); + ggml_set_param(ctx, layer.w3); + } +} + +void set_param_model_lora(struct llama_model_lora * model) { + const auto& hparams = model->hparams; + + const uint32_t n_layer = hparams.n_layer; + + struct ggml_context* ctx = model->ctx; + + ggml_set_param(ctx, model->tok_embeddings); + ggml_set_param(ctx, model->norm); + ggml_set_param(ctx, model->outputa); + ggml_set_param(ctx, model->outputb); + + for (uint32_t i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + + ggml_set_param(ctx, layer.attention_norm); + ggml_set_param(ctx, layer.wqa); + ggml_set_param(ctx, layer.wqb); + ggml_set_param(ctx, layer.wka); + ggml_set_param(ctx, layer.wkb); + ggml_set_param(ctx, layer.wva); + ggml_set_param(ctx, layer.wvb); + ggml_set_param(ctx, layer.woa); + ggml_set_param(ctx, layer.wob); + ggml_set_param(ctx, layer.ffn_norm); + ggml_set_param(ctx, layer.w1); + ggml_set_param(ctx, layer.w2); + ggml_set_param(ctx, layer.w3); + } +} + +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; + + struct random_normal_distribution rnd; + init_random_normal_distribution(&rnd, seed, mean, std, min, max); + randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd); + randomize_tensor_normal(model->norm, model->norm->n_dims, model->norm->ne, &rnd); + randomize_tensor_normal(model->output, model->output->n_dims, model->output->ne, &rnd); + + for (uint32_t i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd); + + randomize_tensor_normal(layer.wq, layer.wq->n_dims, layer.wq->ne, &rnd); + randomize_tensor_normal(layer.wk, layer.wk->n_dims, layer.wk->ne, &rnd); + randomize_tensor_normal(layer.wv, layer.wv->n_dims, layer.wv->ne, &rnd); + randomize_tensor_normal(layer.wo, layer.wo->n_dims, layer.wo->ne, &rnd); + + randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd); + + randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd); + randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd); + randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd); + } +} + + +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; + + struct random_normal_distribution rnd; + init_random_normal_distribution(&rnd, seed, mean, std, min, max); + randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd); + randomize_tensor_normal(model->norm, model->norm->n_dims, model->norm->ne, &rnd); + randomize_tensor_normal(model->outputa, model->outputa->n_dims, model->outputa->ne, &rnd); + randomize_tensor_normal(model->outputb, model->outputb->n_dims, model->outputb->ne, &rnd); + + for (uint32_t i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd); + + randomize_tensor_normal(layer.wqa, layer.wqa->n_dims, layer.wqa->ne, &rnd); + randomize_tensor_normal(layer.wqb, layer.wqb->n_dims, layer.wqb->ne, &rnd); + randomize_tensor_normal(layer.wka, layer.wka->n_dims, layer.wka->ne, &rnd); + randomize_tensor_normal(layer.wkb, layer.wkb->n_dims, layer.wkb->ne, &rnd); + randomize_tensor_normal(layer.wva, layer.wva->n_dims, layer.wva->ne, &rnd); + randomize_tensor_normal(layer.wvb, layer.wvb->n_dims, layer.wvb->ne, &rnd); + randomize_tensor_normal(layer.woa, layer.woa->n_dims, layer.woa->ne, &rnd); + randomize_tensor_normal(layer.wob, layer.wob->n_dims, layer.wob->ne, &rnd); + + randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd); + + randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd); + randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd); + randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd); + } +} + +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; + const uint32_t n_embd = hparams.n_embd; + const uint32_t n_layer = hparams.n_layer; + + const int64_t n_mem = n_layer*n_ctx*n_batch; + const int64_t n_elements = n_embd*n_mem; + + // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB); + + // struct ggml_init_params params; + // params.mem_size = cache.buf.size; + // params.mem_buffer = cache.buf.addr; + // params.no_alloc = false; + if (!cache->ctx) { + struct ggml_init_params params; + params.mem_size = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024; + params.mem_buffer = NULL; + params.no_alloc = false; + + cache->ctx = ggml_init(params); + + if (!cache->ctx) { + fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__); + return false; + } + } + + cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements); + cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements); + + return true; +} + +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; + const uint32_t n_embd = hparams.n_embd; + const uint32_t n_layer = hparams.n_layer; + + const int64_t n_mem = n_layer*n_ctx*n_batch; + const int64_t n_elements = n_embd*n_mem; + + // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB); + + // struct ggml_init_params params; + // params.mem_size = cache.buf.size; + // params.mem_buffer = cache.buf.addr; + // params.no_alloc = false; + if (!cache->ctx) { + struct ggml_init_params params; + params.mem_size = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024; + params.mem_buffer = NULL; + params.no_alloc = false; + + cache->ctx = ggml_init(params); + + if (!cache->ctx) { + fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__); + return false; + } + } + + cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements); + cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements); + + 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) { + + const int N = n_tokens; + + struct llama_kv_cache& kv_self = *cache; + const auto & hparams = model->hparams; + const int n_ctx = hparams.n_ctx; + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_head = hparams.n_head; + const int n_rot = hparams.n_rot; + + struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + memcpy(tokens->data, tokens_input->data, N*ggml_element_size(tokens)); + + struct ggml_tensor * kc = kv_self.k; + struct ggml_tensor * vc = kv_self.v; + + // inpL shape [n_embd,N,1,1] + struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + struct ggml_tensor * cur; + + // lctx.use_buf(ctx0, 0); + + // norm + { + // cur shape [n_embd,N,1,1] + cur = ggml_rms_norm(ctx0, inpL); + + // cur = attention_norm*cur + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model->layers[il].attention_norm, cur), + cur); + } + + // self-attention + { + // compute Q and K and RoPE them + // wq shape [n_embd, n_embd, 1, 1] + // wk shape [n_embd, n_embd, 1, 1] + // Qcur shape [n_embd/n_head, n_head, N, 1] + // Kcur shape [n_embd/n_head, n_head, N, 1] + struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + + // store key and value to memory + { + // compute the transposed [N, n_embd] V matrix + // wv shape [n_embd, n_embd, 1, 1] + // Vcur shape [n_embd, N, 1, 1] + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wv, cur), n_embd, N))); + + // kv_self.k shape [n_embd * n_ctx * n_layer, 1] + // kv_self.v shape [n_embd * n_ctx * n_layer, 1] + // k shape [n_embd * N, 1] == kv_self.k[:,n_past:n_past+N,il,0] + // v shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0] + + /* { + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.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)); + } //*/ + + kc = ggml_set_1d(ctx0, kc, ggml_reshape_1d(ctx0, Kcur, n_embd*N), (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + vc = ggml_set_2d(ctx0, vc, Vcur, ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); + } + + // Qcur shape [n_embd/n_head, n_head, N, 1] + // Q shape [n_embd/n_head, N, n_head, 1] + struct ggml_tensor * Q = + ggml_permute(ctx0, + Qcur, + 0, 2, 1, 3); + + // kv_self.k shape [n_embd * n_ctx * n_layer, 1] + // K shape [n_embd/n_head, n_past + N, n_head, 1] + struct ggml_tensor * K = + ggml_permute(ctx0, + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, kc, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kc)*n_embd), + n_embd/n_head, n_head, n_past + N), + 0, 2, 1, 3); + + // K * Q + // KQ shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + // KQ_scaled shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_scaled = + ggml_scale(ctx0, + KQ, + ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); + + // KQ_masked = mask_past(KQ_scaled) + // KQ_masked shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + // KQ_soft_max shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + + // split cached V into n_head heads + //// V shape [n_past + N, n_embd/n_head, n_head, 1] + // V shape [n_past + N, n_embd/n_head, n_head, 1] == kv_self.v[:,:(n_past+N),il,1] + struct ggml_tensor * V = + ggml_view_3d(ctx0, vc, + n_past + N, n_embd/n_head, n_head, + n_ctx*ggml_element_size(vc), + n_ctx*ggml_element_size(vc)*n_embd/n_head, + il*n_ctx*ggml_element_size(vc)*n_embd); + + // KQV shape [n_embd/n_head, N, n_head, 1] + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + // KQV_merged shape [n_embd/n_head, n_head, N, 1] + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + // KQV_merged shape + + // cur = KQV_merged.contiguous().view(n_embd, N) + // cur shape [n_embd,N,1,1] + cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N); + // cur = ggml_cpy(ctx0, + // KQV_merged, + // ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + + // projection (no bias) + // cur shape [n_embd,N,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].wo, + cur); + } + + // lctx.use_buf(ctx0, 1); + + // inpFF shape [n_embd,N,1,1] + struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); + + // feed-forward network + { + // norm + { + // cur shape [n_embd,N,1,1] + cur = ggml_rms_norm(ctx0, inpFF); + + // cur = ffn_norm*cur + // cur shape [n_embd,N,1,1] + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), + cur); + } + + // tmp shape [n_ff,N,1,1] + struct ggml_tensor * tmp = ggml_mul_mat(ctx0, + model->layers[il].w3, + cur); + + // cur shape [n_ff,N,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].w1, + cur); + + // SILU activation + // cur shape [n_ff,N,1,1] + cur = ggml_silu(ctx0, cur); + + // cur shape [n_ff,N,1,1] + cur = ggml_mul(ctx0, cur, tmp); + + // cur shape [n_embd,N,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].w2, + cur); + } + + // cur shape [n_embd,N,1,1] + cur = ggml_add(ctx0, cur, inpFF); + + // input for next layer + // inpL shape [n_embd,N,1,1] + inpL = cur; + } + + // norm + { + + // inpL shape [n_embd,N,1,1] + inpL = ggml_rms_norm(ctx0, inpL); + + // inpL = norm*inpL + // inpL shape [n_embd,N,1,1] + inpL = ggml_mul(ctx0, + ggml_repeat(ctx0, model->norm, inpL), + inpL); + + //embeddings = inpL; + } + + // lm_head + // inpL shape [n_vocab,N,1,1] + inpL = ggml_mul_mat(ctx0, model->output, inpL); + + // run the computation + ggml_build_forward_expand(gf, inpL); + + return inpL; +} + +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) { + 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) { + 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) { + GGML_ASSERT(tensor->n_dims == 4); + GGML_ASSERT(tensor->ne[0] == ne0); + GGML_ASSERT(tensor->ne[1] == ne1); + GGML_ASSERT(tensor->ne[2] == ne2); + 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) { + + const int N = n_tokens; + + struct llama_kv_cache& kv_self = *cache; + const auto & hparams = model->hparams; + const int n_ctx = hparams.n_ctx; + const int n_vocab = hparams.n_vocab; + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_head = hparams.n_head; + const int n_rot = hparams.n_rot; + const int n_ff = get_n_ff(&hparams); + + struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); + memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch); + + struct ggml_tensor * kc = kv_self.k; + struct ggml_tensor * vc = kv_self.v; + + // inpL shape [n_embd,N*n_batch,1] + struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); + assert_shape_2d(inpL, n_embd, N*n_batch); + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + struct ggml_tensor * cur; + + // lctx.use_buf(ctx0, 0); + + // norm + { + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_rms_norm(ctx0, inpL); + assert_shape_2d(cur, n_embd, N*n_batch); + + // cur = attention_norm*cur + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model->layers[il].attention_norm, cur), + cur); + assert_shape_2d(cur, n_embd, N*n_batch); + } + + // self-attention + { + // compute Q and K and RoPE them + // wq shape [n_embd, n_embd, 1, 1] + // wk shape [n_embd, n_embd, 1, 1] + // Qcur shape [n_embd/n_head, n_head, N, n_batch] + // Kcur shape [n_embd/n_head, n_head, N, n_batch] + struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0); + struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0); + assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch); + assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch); + + // store key and value to memory + { + // compute the transposed [N, n_embd] V matrix + // wv shape [n_embd, n_embd, 1, 1] + // Vcur shape [N, n_embd, n_batch, 1] + struct ggml_tensor * Vcur = ggml_cont(ctx0, + ggml_permute(ctx0, + ggml_reshape_3d(ctx0, + ggml_mul_mat(ctx0, + model->layers[il].wv, + cur), + n_embd, N, n_batch), + 1, 0, 2, 3)); + + assert_shape_3d(Vcur, N, n_embd, n_batch); + + // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer] + // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer] + // k shape [n_embd * N, n_batch] == kv_self.k[:,n_past:n_past+N,:,il] + // v shape [N, n_embd, n_batch, 1] == kv_self.v[:,n_past:n_past+N,:,il] + + /* { + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.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)); + } //*/ + + kc = ggml_set_2d(ctx0, kc, + ggml_reshape_2d(ctx0, Kcur, n_embd*N, n_batch), + ggml_element_size(kc)*n_embd*n_ctx, + (ggml_element_size(kc)*n_embd)*(il*n_batch*n_ctx + n_past)); + vc = ggml_set_2d(ctx0, vc, + ggml_reshape_2d(ctx0, Vcur, N*n_embd, n_batch), + ggml_element_size(vc)*n_ctx*n_embd, + ggml_element_size(vc)*(n_past + il*n_embd*n_batch*n_ctx)); + + assert_shape_1d(kc, n_embd * n_ctx * n_batch * n_layer); + assert_shape_1d(vc, n_embd * n_ctx * n_batch * n_layer); + } + + // Qcur shape [n_embd/n_head, n_head, N, n_batch] + // Q shape [n_embd/n_head, N, n_head, n_batch] + struct ggml_tensor * Q = + ggml_permute(ctx0, + Qcur, + 0, 2, 1, 3); + assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch); + + // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer] + // K shape [n_embd/n_head, n_past + N, n_head, n_batch] + struct ggml_tensor * K = + ggml_permute(ctx0, + ggml_reshape_4d(ctx0, + ggml_view_3d(ctx0, + kc, + n_embd, + (n_past + N), + n_batch, + n_embd*ggml_element_size(kc), + n_ctx*n_embd*ggml_element_size(kc), + il*n_batch*n_ctx*n_embd*ggml_element_size(kc)), + n_embd/n_head, n_head, n_past + N, n_batch), + 0, 2, 1, 3); + assert_shape_4d(K, n_embd/n_head, n_past + N, n_head, n_batch); + + // K * Q + // KQ shape [n_past + N, N, n_head, n_batch] + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + assert_shape_4d(KQ, n_past + N, N, n_head, n_batch); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + // KQ_scaled shape [n_past + N, N, n_head, n_batch] + struct ggml_tensor * KQ_scaled = + ggml_scale(ctx0, + KQ, + ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); + assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch); + + // KQ_masked = mask_past(KQ_scaled) + // KQ_masked shape [n_past + N, N, n_head, n_batch] + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + assert_shape_4d(KQ_masked, n_past + N, N, n_head, n_batch); + + // KQ = soft_max(KQ_masked) + // KQ_soft_max shape [n_past + N, N, n_head, n_batch] + struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + assert_shape_4d(KQ_soft_max, n_past + N, N, n_head, n_batch); + + // split cached V into n_head heads + // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer] + // V shape [n_past + N, n_embd/n_head, n_head, n_batch] == kv_self.v[:(n_past+N),:,:,il] + struct ggml_tensor * V = + ggml_view_4d(ctx0, vc, + n_past + N, n_embd/n_head, n_head, n_batch, + ggml_element_size(vc)*n_ctx, + ggml_element_size(vc)*n_ctx*n_embd/n_head, + ggml_element_size(vc)*n_ctx*n_embd, + il*n_batch*n_ctx*n_embd*ggml_element_size(vc)); + assert_shape_4d(V, n_past + N, n_embd/n_head, n_head, n_batch); + + // KQV shape [n_embd/n_head, N, n_head, n_batch] + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + // KQV_merged shape [n_embd/n_head, n_head, N, n_batch] + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch); + // KQV_merged shape + + // cur = KQV_merged.contiguous().view(n_embd, N) + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch); + assert_shape_2d(cur, n_embd, N*n_batch); + // cur = ggml_cpy(ctx0, + // KQV_merged, + // ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + + // projection (no bias) + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].wo, + cur); + assert_shape_2d(cur, n_embd, N*n_batch); + } + + // lctx.use_buf(ctx0, 1); + + // inpFF shape [n_embd,N*n_batch,1,1] + struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); + assert_shape_2d(inpFF, n_embd, N*n_batch); + + // feed-forward network + { + // norm + { + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_rms_norm(ctx0, inpFF); + assert_shape_2d(cur, n_embd, N*n_batch); + + // cur = ffn_norm*cur + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), + cur); + assert_shape_2d(cur, n_embd, N*n_batch); + } + + // tmp shape [n_ff,N*n_batch,1,1] + struct ggml_tensor * tmp = ggml_mul_mat(ctx0, + model->layers[il].w3, + cur); + assert_shape_2d(tmp, n_ff, N*n_batch); + + // cur shape [n_ff,N*n_batch,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].w1, + cur); + assert_shape_2d(cur, n_ff, N*n_batch); + + // SILU activation + // cur shape [n_ff,N*n_batch,1,1] + cur = ggml_silu(ctx0, cur); + assert_shape_2d(cur, n_ff, N*n_batch); + + // cur shape [n_ff,N*n_batch,1,1] + cur = ggml_mul(ctx0, cur, tmp); + assert_shape_2d(cur, n_ff, N*n_batch); + + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].w2, + cur); + assert_shape_2d(cur, n_embd, N*n_batch); + } + + // cur shape [n_embd,N*n_batch,1,1] + cur = ggml_add(ctx0, cur, inpFF); + assert_shape_2d(cur, n_embd, N*n_batch); + + // input for next layer + // inpL shape [n_embd,N*n_batch,1,1] + inpL = cur; + assert_shape_2d(inpL, n_embd, N*n_batch); + } + + // norm + { + + // inpL shape [n_embd,N*n_batch,1,1] + inpL = ggml_rms_norm(ctx0, inpL); + assert_shape_2d(inpL, n_embd, N*n_batch); + + // inpL = norm*inpL + // inpL shape [n_embd,N*n_batch,1,1] + inpL = ggml_mul(ctx0, + ggml_repeat(ctx0, model->norm, inpL), + inpL); + + assert_shape_2d(inpL, n_embd, N*n_batch); + + //embeddings = inpL; + } + + // lm_head + // inpL shape [n_vocab,N*n_batch,1,1] + inpL = ggml_mul_mat(ctx0, model->output, inpL); + assert_shape_2d(inpL, n_vocab, N*n_batch); + + { + // inpL shape [n_vocab,N,n_batch,1] + inpL = ggml_reshape_3d(ctx0, + inpL, + n_vocab, N, n_batch); + assert_shape_3d(inpL, n_vocab, N, n_batch); + } + + // run the computation + ggml_build_forward_expand(gf, inpL); + + 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) { + + const int N = n_tokens; + + struct llama_kv_cache& kv_self = *cache; + const auto & hparams = model->hparams; + + const int n_ctx = hparams.n_ctx; + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_head = hparams.n_head; + const int n_rot = hparams.n_rot; + + struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + memcpy(tokens->data, tokens_input->data, N*ggml_element_size(tokens)); + + struct ggml_tensor * kc = kv_self.k; + struct ggml_tensor * vc = kv_self.v; + + // inpL shape [n_embd,N,1,1] + struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + struct ggml_tensor * cur; + + // norm + { + // cur shape [n_embd,N,1,1] + cur = ggml_rms_norm(ctx0, inpL); + + // cur = attention_norm*cur + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model->layers[il].attention_norm, cur), + cur); + } + + // self-attention + { + // compute Q and K and RoPE them + // wq shape [n_embd, n_embd, 1, 1] + // wk shape [n_embd, n_embd, 1, 1] + // Qcur shape [n_embd/n_head, n_head, N, 1] + // Kcur shape [n_embd/n_head, n_head, N, 1] + struct ggml_tensor * Qcur = ggml_rope(ctx0, + ggml_reshape_3d(ctx0, + ggml_mul_mat(ctx0, + model->layers[il].wqa, + ggml_mul_mat(ctx0, + model->layers[il].wqb, + cur)), + n_embd/n_head, n_head, N), + n_past, n_rot, 0); + struct ggml_tensor * Kcur = ggml_rope(ctx0, + ggml_reshape_3d(ctx0, + ggml_mul_mat(ctx0, + model->layers[il].wka, + ggml_mul_mat(ctx0, + model->layers[il].wkb, + cur)), + n_embd/n_head, n_head, N), + n_past, n_rot, 0); + + // store key and value to memory + { + // compute the transposed [N, n_embd] V matrix + // wv shape [n_embd, n_embd, 1, 1] + // Vcur shape [n_embd, N, 1, 1] + struct ggml_tensor * Vcur = ggml_cont(ctx0, + ggml_transpose(ctx0, + ggml_reshape_2d(ctx0, + ggml_mul_mat(ctx0, + model->layers[il].wva, + ggml_mul_mat(ctx0, + model->layers[il].wvb, + cur)), + n_embd, N))); + + // kv_self.k shape [n_embd * n_ctx * n_layer, 1] + // kv_self.v shape [n_embd * n_ctx * n_layer, 1] + // k shape [n_embd * N, 1] == kv_self.k[:,n_past:n_past+N,il,0] + // v shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0] + + /* { + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.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)); + } //*/ + + kc = ggml_set_1d(ctx0, kc, ggml_reshape_1d(ctx0, Kcur, n_embd*N), (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + vc = ggml_set_2d(ctx0, vc, Vcur, ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); + } + + // Qcur shape [n_embd/n_head, n_head, N, 1] + // Q shape [n_embd/n_head, N, n_head, 1] + struct ggml_tensor * Q = + ggml_permute(ctx0, + Qcur, + 0, 2, 1, 3); + + // kv_self.k shape [n_embd * n_ctx * n_layer, 1] + // K shape [n_embd/n_head, n_past + N, n_head, 1] + struct ggml_tensor * K = + ggml_permute(ctx0, + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, kc, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kc)*n_embd), + n_embd/n_head, n_head, n_past + N), + 0, 2, 1, 3); + + // K * Q + // KQ shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + // KQ_scaled shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_scaled = + ggml_scale(ctx0, + KQ, + ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); + + // KQ_masked = mask_past(KQ_scaled) + // KQ_masked shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + // KQ_soft_max shape [n_past + N, N, n_head, 1] + struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + + // split cached V into n_head heads + //// V shape [n_past + N, n_embd/n_head, n_head, 1] + // V shape [n_past + N, n_embd/n_head, n_head, 1] == kv_self.v[:,:(n_past+N),il,1] + struct ggml_tensor * V = + ggml_view_3d(ctx0, vc, + n_past + N, n_embd/n_head, n_head, + n_ctx*ggml_element_size(vc), + n_ctx*ggml_element_size(vc)*n_embd/n_head, + il*n_ctx*ggml_element_size(vc)*n_embd); + + // KQV shape [n_embd/n_head, N, n_head, 1] + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + // KQV_merged shape [n_embd/n_head, n_head, N, 1] + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + // KQV_merged shape + + // cur = KQV_merged.contiguous().view(n_embd, N) + // cur shape [n_embd,N,1,1] + cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N); + // cur = ggml_cpy(ctx0, + // KQV_merged, + // ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + + // projection (no bias) + // cur shape [n_embd,N,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].woa, + ggml_mul_mat(ctx0, + model->layers[il].wob, + cur)); + } + + // inpFF shape [n_embd,N,1,1] + struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); + + // feed-forward network + { + // norm + { + // cur shape [n_embd,N,1,1] + cur = ggml_rms_norm(ctx0, inpFF); + + // cur = ffn_norm*cur + // cur shape [n_embd,N,1,1] + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), + cur); + } + + // tmp shape [n_ff,N,1,1] + struct ggml_tensor * tmp = ggml_mul_mat(ctx0, + model->layers[il].w3, + cur); + + // cur shape [n_ff,N,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].w1, + cur); + + // SILU activation + // cur shape [n_ff,N,1,1] + cur = ggml_silu(ctx0, cur); + + // cur shape [n_ff,N,1,1] + cur = ggml_mul(ctx0, cur, tmp); + + // cur shape [n_embd,N,1,1] + cur = ggml_mul_mat(ctx0, + model->layers[il].w2, + cur); + } + + // cur shape [n_embd,N,1,1] + cur = ggml_add(ctx0, cur, inpFF); + + // input for next layer + // inpL shape [n_embd,N,1,1] + inpL = cur; + } + + // norm + { + + // inpL shape [n_embd,N,1,1] + inpL = ggml_rms_norm(ctx0, inpL); + + // inpL = norm*inpL + // inpL shape [n_embd,N,1,1] + inpL = ggml_mul(ctx0, + ggml_repeat(ctx0, model->norm, inpL), + inpL); + + //embeddings = inpL; + } + + + // lm_head + // inpL shape [n_vocab,N,1,1] + inpL = ggml_mul_mat(ctx0, + model->outputa, + ggml_mul_mat(ctx0, + model->outputb, + inpL)); + + // ggml_set_scratch(ctx0, { 0, 0, nullptr, }); + // run the computation + ggml_build_forward_expand(gf, inpL); + + return inpL; +} + +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); + assert(logits->ne[1] == best_samples->ne[0]); + assert(logits->ne[0] == probs->ne[0]); + assert(logits->ne[1] == probs->ne[1]); + for (int i = 0; i < logits->ne[1]; ++i) { + float max_logit = ggml_get_f32_1d(logits, i * logits->ne[0]); + ggml_set_i32_1d(best_samples, i, 0); + for (int k = 0; k < logits->ne[0]; ++k) { + float logit = ggml_get_f32_1d(logits, i * logits->ne[0] + k); + if (logit > max_logit) { + max_logit = logit; + ggml_set_i32_1d(best_samples, i, k); + } + } + float psum = 0; + for (int k = 0; k < logits->ne[0]; ++k) { + float logit = ggml_get_f32_1d(logits, i * logits->ne[0] + k); + float p = (logit == -INFINITY) ? 0 : expf(logit - max_logit); + psum += p; + ggml_set_f32_1d(probs, i * probs->ne[0] + k, p); + } + for (int k = 0; k < logits->ne[0]; ++k) { + float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k); + ggml_set_f32_1d(probs, i * probs->ne[0] + k, p / psum); + } + } +} + +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); + int n_tokens = best_samples->ne[0]; + int n_batch = best_samples->ne[1]; + int n_vocab = logits->ne[0]; + GGML_ASSERT(n_tokens == logits->ne[1]); + GGML_ASSERT(n_batch == logits->ne[2]); + GGML_ASSERT(n_vocab == probs->ne[0]); + GGML_ASSERT(n_tokens == probs->ne[1]); + GGML_ASSERT(n_batch == probs->ne[2]); + + for (int k = 0; k < n_batch; ++k) { + struct ggml_tensor * best_samples_k = ggml_view_1d(ctx, + best_samples, + best_samples->ne[0], + k*best_samples->nb[1]); + struct ggml_tensor * logits_k = ggml_view_2d(ctx, + logits, + logits->ne[0], + logits->ne[1], + logits->nb[1], + k*logits->nb[2]); + struct ggml_tensor * probs_k = ggml_view_2d(ctx, + probs, + probs->ne[0], + probs->ne[1], + probs->nb[1], + k*probs->nb[2]); + sample_softmax(logits_k, probs_k, best_samples_k); + } +} + +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); + } + printf("\n"); +} + +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) { + float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k); + printf(" %.2f", p); + } + printf("\n"); + } +} + +void print_token(int token, int n_vocab) { + for (int k = 0; k < token; ++k) { + printf(" "); + } + printf("X"); + for (int k = token+1; k < n_vocab; ++k) { + printf(" "); + } + printf("\n"); +} + +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) { + int n_tokens = tokens_input->ne[0]; + int n_vocab = targets->ne[0]; + float randomness = 0.0f; + // ggml_set_zero(targets); + ggml_set_f32(targets, -1.0f); + ggml_set_i32_1d(tokens_input, 0, 0); + for (int i=1; i 1.0f) ? 1.0f : z; // clamp to [0..1] + int token = std::max(1,std::min(1+(int)(z*(float)(n_vocab-1)), n_vocab-1)); + ggml_set_f32_1d(targets, (i-1)*n_vocab + token, +1.0f); + if (in_dims == 2); + GGML_ASSERT( targets->n_dims == 3); + int n_tokens = tokens_input->ne[0]; + int n_batch = tokens_input->ne[1]; + GGML_ASSERT(n_tokens == targets->ne[1]); + GGML_ASSERT(n_batch == targets->ne[2]); + + for (int k=0; kne[0], + k*tokens_input->nb[1]); + struct ggml_tensor * targets_k = ggml_view_2d(ctx, + targets, + targets->ne[0], + targets->ne[1], + targets->nb[1], + k*targets->nb[2]); + get_example_targets(example_id*n_batch + k, tokens_input_k, targets_k); + } +} + +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 "ggml.h" +#include "build-info.h" #include #include #include @@ -47,7 +40,7 @@ float tensor_sum_elements(struct ggml_tensor * tensor) { #define TENSOR_DUMP(TENSOR) printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", #TENSOR, \ TENSOR->type,TENSOR_TYPE_AS_STR(TENSOR->type),\ - TENSOR->ne[0], TENSOR->ne[1], TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \ + (int) TENSOR->ne[0], (int) TENSOR->ne[1], (int) TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \ { float sum = tensor_sum_elements(TENSOR); printf("Sum of tensor %s is %6.2f\n",#TENSOR, sum); } struct benchmark_params_struct { @@ -98,12 +91,10 @@ int main(int argc, char ** argv) { } } - - // create the ggml context + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); printf("Starting Test\n"); - - + // create the ggml context struct ggml_context * ctx; //const int sizex = 4096; //const int sizey = 11008; @@ -125,16 +116,18 @@ int main(int argc, char ** argv) { #endif //printf("Memsize required = %i\n", sizex*sizex); - ggml_type wtype = GGML_TYPE_F32; size_t ctx_size = 0; - ctx_size += sizex*sizey*ggml_type_sizef(wtype); - ctx_size += sizex*sizey*ggml_type_sizef(wtype); ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); - ctx_size += sizex*sizeof(float); - ctx_size += 1024*1024*100; + ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); + ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32); + ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0); + ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0); + ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS + ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS + ctx_size += 1024*1024*16; - printf("Allocating Memory of size %li byes, %li MB\n",ctx_size, (ctx_size/1024/1024)); + printf("Allocating Memory of size %li bytes, %li MB\n",ctx_size, (ctx_size/1024/1024)); struct ggml_init_params params = { /*.mem_size =*/ ctx_size, @@ -145,7 +138,7 @@ int main(int argc, char ** argv) { ctx = ggml_init(params); if (!ctx) { fprintf(stderr, "%s: ggml_init() failed\n", __func__); - return false; + return 1; } @@ -217,7 +210,7 @@ int main(int argc, char ** argv) { const int dimz = sizez; long long int flops_per_dot_product = dimy + dimy; long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ; - printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - aboout %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000); + printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000); // Let's use the F32 result from above as a reference for the q4_0 multiplication @@ -234,7 +227,6 @@ int main(int argc, char ** argv) { ggml_graph_compute(ctx, &gf31); long long int stop = ggml_time_us(); long long int usec = stop-start; - float sec = usec/1000000; float flops_per_usec = (1.0f*flops_per_matrix)/usec; printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%19.2f\n", i, diff --git a/examples/chat-13B.sh b/examples/chat-13B.sh index 4265d7b66..35c089d57 100755 --- a/examples/chat-13B.sh +++ b/examples/chat-13B.sh @@ -1,9 +1,12 @@ #!/bin/bash +set -e + cd "$(dirname "$0")/.." || exit MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}" -USER_NAME="${USER_NAME:-User}" +PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt} +USER_NAME="${USER_NAME:-USER}" AI_NAME="${AI_NAME:-ChatLLaMa}" # Adjust to the number of CPU cores you want to use. @@ -15,39 +18,24 @@ N_PREDICTS="${N_PREDICTS:-2048}" # For example, override the context size by doing: ./chatLLaMa --ctx_size 1024 GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}" +DATE_TIME=$(date +%H:%M) +DATE_YEAR=$(date +%Y) + +PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt) + +sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \ + -e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \ + -e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \ + -e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \ + $PROMPT_TEMPLATE > $PROMPT_FILE + # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS ./main $GEN_OPTIONS \ --model "$MODEL" \ --threads "$N_THREAD" \ --n_predict "$N_PREDICTS" \ --color --interactive \ + --file ${PROMPT_FILE} \ --reverse-prompt "${USER_NAME}:" \ - --prompt " -Text transcript of a never ending dialog, where ${USER_NAME} interacts with an AI assistant named ${AI_NAME}. -${AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer ${USER_NAME}’s requests immediately and with details and precision. -There are no annotations like (30 seconds passed...) or (to himself), just what ${USER_NAME} and ${AI_NAME} say aloud to each other. -The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. -The transcript only includes text, it does not include markup like HTML and Markdown. - -$USER_NAME: Hello, $AI_NAME! -$AI_NAME: Hello $USER_NAME! How may I help you today? -$USER_NAME: What time is it? -$AI_NAME: It is $(date +%H:%M). -$USER_NAME: What year is it? -$AI_NAME: We are in $(date +%Y). -$USER_NAME: Please tell me the largest city in Europe. -$AI_NAME: The largest city in Europe is Moscow, the capital of Russia. -$USER_NAME: What can you tell me about Moscow? -$AI_NAME: Moscow, on the Moskva River in western Russia, is the nation’s cosmopolitan capital. In its historic core is the Kremlin, a complex that’s home to the president and tsarist treasures in the Armoury. Outside its walls is Red Square, Russia’s symbolic center. -$USER_NAME: What is a cat? -$AI_NAME: A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae. -$USER_NAME: How do I pass command line arguments to a Node.js program? -$AI_NAME: The arguments are stored in process.argv. - - argv[0] is the path to the Node. js executable. - argv[1] is the path to the script file. - argv[2] is the first argument passed to the script. - argv[3] is the second argument passed to the script and so on. -$USER_NAME: Name a color. -$AI_NAME: Blue -$USER_NAME:" "$@" + --in-prefix ' ' \ + "$@" diff --git a/examples/common.cpp b/examples/common.cpp index a0b6f10ad..86c1eef41 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -1,49 +1,108 @@ #include "common.h" #include +#include #include #include #include #include #include +#include -#if defined (_WIN32) +#if defined(__APPLE__) && defined(__MACH__) +#include +#include +#endif + +#if defined(_WIN32) +#define WIN32_LEAN_AND_MEAN +#define NOMINMAX +#include #include #include -#pragma comment(lib,"kernel32.lib") -extern "C" __declspec(dllimport) void* __stdcall GetStdHandle(unsigned long nStdHandle); -extern "C" __declspec(dllimport) int __stdcall GetConsoleMode(void* hConsoleHandle, unsigned long* lpMode); -extern "C" __declspec(dllimport) int __stdcall SetConsoleMode(void* hConsoleHandle, unsigned long dwMode); -extern "C" __declspec(dllimport) int __stdcall SetConsoleCP(unsigned int wCodePageID); -extern "C" __declspec(dllimport) int __stdcall SetConsoleOutputCP(unsigned int wCodePageID); -extern "C" __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int CodePage, unsigned long dwFlags, - const wchar_t * lpWideCharStr, int cchWideChar, - char * lpMultiByteStr, int cbMultiByte, - const char * lpDefaultChar, bool * lpUsedDefaultChar); -#define CP_UTF8 65001 +#else +#include +#include +#include #endif -bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { - // determine sensible default number of threads. - // std::thread::hardware_concurrency may not be equal to the number of cores, or may return 0. +int32_t get_num_physical_cores() { #ifdef __linux__ std::ifstream cpuinfo("/proc/cpuinfo"); - params.n_threads = std::count(std::istream_iterator(cpuinfo), - std::istream_iterator(), - std::string("processor")); + std::string line; + while (std::getline(cpuinfo, line)) { + std::size_t pos = line.find("cpu cores"); + if (pos != std::string::npos) { + pos = line.find(": ", pos); + if (pos != std::string::npos) { + try { + // Extract the number and return it + return static_cast(std::stoul(line.substr(pos + 2))); + } catch (const std::invalid_argument &) { + // Ignore if we could not parse + } + } + } + } +#elif defined(__APPLE__) && defined(__MACH__) + int32_t num_physical_cores; + size_t len = sizeof(num_physical_cores); + int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0); + if (result == 0) { + return num_physical_cores; + } + result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0); + if (result == 0) { + return num_physical_cores; + } +#elif defined(_WIN32) + //TODO: Implement #endif - if (params.n_threads == 0) { - params.n_threads = std::max(1, (int32_t) std::thread::hardware_concurrency()); + unsigned int n_threads = std::thread::hardware_concurrency(); + return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4; +} + +void process_escapes(std::string& input) { + std::size_t input_len = input.length(); + std::size_t output_idx = 0; + + for (std::size_t input_idx = 0; input_idx < input_len; ++input_idx) { + if (input[input_idx] == '\\' && input_idx + 1 < input_len) { + switch (input[++input_idx]) { + case 'n': input[output_idx++] = '\n'; break; + case 'r': input[output_idx++] = '\r'; break; + case 't': input[output_idx++] = '\t'; break; + case '\'': input[output_idx++] = '\''; break; + case '\"': input[output_idx++] = '\"'; break; + case '\\': input[output_idx++] = '\\'; break; + default: input[output_idx++] = '\\'; + input[output_idx++] = input[input_idx]; break; + } + } else { + input[output_idx++] = input[input_idx]; + } } + input.resize(output_idx); +} + +bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; + bool escape_prompt = false; std::string arg; gpt_params default_params; + const std::string arg_prefix = "--"; for (int i = 1; i < argc; i++) { arg = argv[i]; + if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { + std::replace(arg.begin(), arg.end(), '_', '-'); + } if (arg == "-s" || arg == "--seed") { +#if defined(GGML_USE_CUBLAS) + fprintf(stderr, "WARNING: when using cuBLAS generation results are NOT guaranteed to be reproducible.\n"); +#endif if (++i >= argc) { invalid_param = true; break; @@ -61,6 +120,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.prompt = argv[i]; + } else if (arg == "-e") { + escape_prompt = true; + } else if (arg == "--prompt-cache") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.path_prompt_cache = argv[i]; + } else if (arg == "--prompt-cache-all") { + params.prompt_cache_all = true; } else if (arg == "-f" || arg == "--file") { if (++i >= argc) { invalid_param = true; @@ -76,27 +145,27 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { if (params.prompt.back() == '\n') { params.prompt.pop_back(); } - } else if (arg == "-n" || arg == "--n_predict") { + } else if (arg == "-n" || arg == "--n-predict") { if (++i >= argc) { invalid_param = true; break; } params.n_predict = std::stoi(argv[i]); - } else if (arg == "--top_k") { + } else if (arg == "--top-k") { if (++i >= argc) { invalid_param = true; break; } params.top_k = std::stoi(argv[i]); - } else if (arg == "-c" || arg == "--ctx_size") { + } else if (arg == "-c" || arg == "--ctx-size") { if (++i >= argc) { invalid_param = true; break; } params.n_ctx = std::stoi(argv[i]); - } else if (arg == "--memory_f32") { + } else if (arg == "--memory-f32") { params.memory_f16 = false; - } else if (arg == "--top_p") { + } else if (arg == "--top-p") { if (++i >= argc) { invalid_param = true; break; @@ -108,19 +177,61 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.temp = std::stof(argv[i]); - } else if (arg == "--repeat_last_n") { + } else if (arg == "--tfs") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.tfs_z = std::stof(argv[i]); + } else if (arg == "--typical") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.typical_p = std::stof(argv[i]); + } else if (arg == "--repeat-last-n") { if (++i >= argc) { invalid_param = true; break; } params.repeat_last_n = std::stoi(argv[i]); - } else if (arg == "--repeat_penalty") { + } else if (arg == "--repeat-penalty") { if (++i >= argc) { invalid_param = true; break; } params.repeat_penalty = std::stof(argv[i]); - } else if (arg == "-b" || arg == "--batch_size") { + } else if (arg == "--frequency-penalty") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.frequency_penalty = std::stof(argv[i]); + } else if (arg == "--presence-penalty") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.presence_penalty = std::stof(argv[i]); + } else if (arg == "--mirostat") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.mirostat = std::stoi(argv[i]); + } else if (arg == "--mirostat-lr") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.mirostat_eta = std::stof(argv[i]); + } else if (arg == "--mirostat-ent") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.mirostat_tau = std::stof(argv[i]); + } else if (arg == "-b" || arg == "--batch-size") { if (++i >= argc) { invalid_param = true; break; @@ -156,16 +267,22 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { params.interactive = true; } else if (arg == "--embedding") { params.embedding = true; - } else if (arg == "--interactive-start") { - params.interactive = true; } else if (arg == "--interactive-first") { - params.interactive_start = true; + params.interactive_first = true; } else if (arg == "-ins" || arg == "--instruct") { params.instruct = true; + } else if (arg == "--multiline-input") { + params.multiline_input = true; } else if (arg == "--color") { params.use_color = true; } else if (arg == "--mlock") { params.use_mlock = true; + } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.n_gpu_layers = std::stoi(argv[i]); } else if (arg == "--no-mmap") { params.use_mmap = false; } else if (arg == "--mtest") { @@ -181,8 +298,29 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { } else if (arg == "--perplexity") { params.perplexity = true; } else if (arg == "--ignore-eos") { - params.ignore_eos = true; - } else if (arg == "--n_parts") { + params.logit_bias[llama_token_eos()] = -INFINITY; + } else if (arg == "--no-penalize-nl") { + params.penalize_nl = false; + } else if (arg == "-l" || arg == "--logit-bias") { + if (++i >= argc) { + invalid_param = true; + break; + } + std::stringstream ss(argv[i]); + llama_token key; + char sign; + std::string value_str; + try { + if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) { + params.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f); + } else { + throw std::exception(); + } + } catch (const std::exception &e) { + invalid_param = true; + break; + } + } else if (arg == "--n-parts") { if (++i >= argc) { invalid_param = true; break; @@ -199,6 +337,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.input_prefix = argv[i]; + } else if (arg == "--in-suffix") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.input_suffix = argv[i]; } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); gpt_print_usage(argc, argv, default_params); @@ -210,6 +354,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { gpt_print_usage(argc, argv, default_params); exit(1); } + if (params.prompt_cache_all && + (params.interactive || params.interactive_first || + params.instruct || params.antiprompt.size())) { + fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n"); + gpt_print_usage(argc, argv, default_params); + exit(1); + } + if (escape_prompt) { + process_escapes(params.prompt); + } return true; } @@ -222,29 +376,49 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stderr, " -i, --interactive run in interactive mode\n"); fprintf(stderr, " --interactive-first run in interactive mode and wait for input right away\n"); fprintf(stderr, " -ins, --instruct run in instruction mode (use with Alpaca models)\n"); + fprintf(stderr, " --multiline-input allows you to write or paste multiple lines without ending each in '\\'\n"); fprintf(stderr, " -r PROMPT, --reverse-prompt PROMPT\n"); fprintf(stderr, " run in interactive mode and poll user input upon seeing PROMPT (can be\n"); fprintf(stderr, " specified more than once for multiple prompts).\n"); fprintf(stderr, " --color colorise output to distinguish prompt and user input from generations\n"); - fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for <= 0)\n"); + fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for < 0)\n"); fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); fprintf(stderr, " -p PROMPT, --prompt PROMPT\n"); fprintf(stderr, " prompt to start generation with (default: empty)\n"); + fprintf(stderr, " -e process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n"); + fprintf(stderr, " --prompt-cache FNAME file to cache prompt state for faster startup (default: none)\n"); + fprintf(stderr, " --prompt-cache-all if specified, saves user input and generations to cache as well.\n"); + fprintf(stderr, " not supported with --interactive or other interactive options\n"); fprintf(stderr, " --random-prompt start with a randomized prompt.\n"); fprintf(stderr, " --in-prefix STRING string to prefix user inputs with (default: empty)\n"); + fprintf(stderr, " --in-suffix STRING string to suffix after user inputs with (default: empty)\n"); fprintf(stderr, " -f FNAME, --file FNAME\n"); fprintf(stderr, " prompt file to start generation.\n"); - fprintf(stderr, " -n N, --n_predict N number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict); - fprintf(stderr, " --top_k N top-k sampling (default: %d)\n", params.top_k); - fprintf(stderr, " --top_p N top-p sampling (default: %.1f)\n", (double)params.top_p); - fprintf(stderr, " --repeat_last_n N last n tokens to consider for penalize (default: %d)\n", params.repeat_last_n); - fprintf(stderr, " --repeat_penalty N penalize repeat sequence of tokens (default: %.1f)\n", (double)params.repeat_penalty); - fprintf(stderr, " -c N, --ctx_size N size of the prompt context (default: %d)\n", params.n_ctx); - fprintf(stderr, " --ignore-eos ignore end of stream token and continue generating\n"); - fprintf(stderr, " --memory_f32 use f32 instead of f16 for memory key+value\n"); + fprintf(stderr, " -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict); + fprintf(stderr, " --top-k N top-k sampling (default: %d, 0 = disabled)\n", params.top_k); + fprintf(stderr, " --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p); + fprintf(stderr, " --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z); + fprintf(stderr, " --typical N locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p); + fprintf(stderr, " --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n); + fprintf(stderr, " --repeat-penalty N penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty); + fprintf(stderr, " --presence-penalty N repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty); + fprintf(stderr, " --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty); + fprintf(stderr, " --mirostat N use Mirostat sampling.\n"); + fprintf(stderr, " Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n"); + fprintf(stderr, " (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat); + fprintf(stderr, " --mirostat-lr N Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta); + fprintf(stderr, " --mirostat-ent N Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau); + fprintf(stderr, " -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n"); + fprintf(stderr, " modifies the likelihood of token appearing in the completion,\n"); + fprintf(stderr, " i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n"); + fprintf(stderr, " or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n"); + fprintf(stderr, " -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx); + fprintf(stderr, " --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n"); + fprintf(stderr, " --no-penalize-nl do not penalize newline token\n"); + fprintf(stderr, " --memory-f32 use f32 instead of f16 for memory key+value\n"); fprintf(stderr, " --temp N temperature (default: %.1f)\n", (double)params.temp); - fprintf(stderr, " --n_parts N number of model parts (default: -1 = determine from dimensions)\n"); - fprintf(stderr, " -b N, --batch_size N batch size for prompt processing (default: %d)\n", params.n_batch); + fprintf(stderr, " --n-parts N number of model parts (default: -1 = determine from dimensions)\n"); + fprintf(stderr, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); fprintf(stderr, " --perplexity compute perplexity over the prompt\n"); fprintf(stderr, " --keep number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); if (llama_mlock_supported()) { @@ -253,6 +427,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { if (llama_mmap_supported()) { fprintf(stderr, " --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); } + fprintf(stderr, " -ngl N, --n-gpu-layers N\n"); + fprintf(stderr, " number of layers to store in VRAM\n"); fprintf(stderr, " --mtest compute maximum memory usage\n"); fprintf(stderr, " --verbose-prompt print prompt before generation\n"); fprintf(stderr, " --lora FNAME apply LoRA adapter (implies --no-mmap)\n"); @@ -284,62 +460,382 @@ std::string gpt_random_prompt(std::mt19937 & rng) { // TODO: not great allocating this every time std::vector llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) { // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars - std::vector res(text.size() + (int)add_bos); - int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos); + std::vector res(text.size() + (int) add_bos); + const int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos); assert(n >= 0); res.resize(n); return res; } -/* Keep track of current color of output, and emit ANSI code if it changes. */ -void set_console_color(console_state & con_st, console_color_t color) { - if (con_st.use_color && con_st.color != color) { - switch(color) { - case CONSOLE_COLOR_DEFAULT: - printf(ANSI_COLOR_RESET); - break; - case CONSOLE_COLOR_PROMPT: - printf(ANSI_COLOR_YELLOW); - break; - case CONSOLE_COLOR_USER_INPUT: - printf(ANSI_BOLD ANSI_COLOR_GREEN); - break; - } - con_st.color = color; +struct llama_context * llama_init_from_gpt_params(const gpt_params & params) { + auto lparams = llama_context_default_params(); + + lparams.n_ctx = params.n_ctx; + lparams.n_parts = params.n_parts; + lparams.n_gpu_layers = params.n_gpu_layers; + lparams.seed = params.seed; + lparams.f16_kv = params.memory_f16; + lparams.use_mmap = params.use_mmap; + lparams.use_mlock = params.use_mlock; + lparams.logits_all = params.perplexity; + lparams.embedding = params.embedding; + + llama_context * lctx = llama_init_from_file(params.model.c_str(), lparams); + + if (lctx == NULL) { + fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str()); + return NULL; } + + if (!params.lora_adapter.empty()) { + int err = llama_apply_lora_from_file(lctx, + params.lora_adapter.c_str(), + params.lora_base.empty() ? NULL : params.lora_base.c_str(), + params.n_threads); + if (err != 0) { + fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); + return NULL; + } + } + + return lctx; } -#if defined (_WIN32) -void win32_console_init(bool enable_color) { - unsigned long dwMode = 0; - void* hConOut = GetStdHandle((unsigned long)-11); // STD_OUTPUT_HANDLE (-11) - if (!hConOut || hConOut == (void*)-1 || !GetConsoleMode(hConOut, &dwMode)) { - hConOut = GetStdHandle((unsigned long)-12); // STD_ERROR_HANDLE (-12) - if (hConOut && (hConOut == (void*)-1 || !GetConsoleMode(hConOut, &dwMode))) { - hConOut = 0; +void console_init(console_state & con_st) { +#if defined(_WIN32) + // Windows-specific console initialization + DWORD dwMode = 0; + con_st.hConsole = GetStdHandle(STD_OUTPUT_HANDLE); + if (con_st.hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(con_st.hConsole, &dwMode)) { + con_st.hConsole = GetStdHandle(STD_ERROR_HANDLE); + if (con_st.hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(con_st.hConsole, &dwMode))) { + con_st.hConsole = NULL; } } - if (hConOut) { + if (con_st.hConsole) { // Enable ANSI colors on Windows 10+ - if (enable_color && !(dwMode & 0x4)) { - SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4) + if (con_st.use_color && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)) { + SetConsoleMode(con_st.hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING); } // Set console output codepage to UTF8 SetConsoleOutputCP(CP_UTF8); } - void* hConIn = GetStdHandle((unsigned long)-10); // STD_INPUT_HANDLE (-10) - if (hConIn && hConIn != (void*)-1 && GetConsoleMode(hConIn, &dwMode)) { + HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE); + if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) { // Set console input codepage to UTF16 _setmode(_fileno(stdin), _O_WTEXT); + + // Turn off ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT) + dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT); + SetConsoleMode(hConIn, dwMode); + } +#else + // POSIX-specific console initialization + struct termios new_termios; + tcgetattr(STDIN_FILENO, &con_st.prev_state); + new_termios = con_st.prev_state; + new_termios.c_lflag &= ~(ICANON | ECHO); + new_termios.c_cc[VMIN] = 1; + new_termios.c_cc[VTIME] = 0; + tcsetattr(STDIN_FILENO, TCSANOW, &new_termios); + + con_st.tty = fopen("/dev/tty", "w+"); + if (con_st.tty != nullptr) { + con_st.out = con_st.tty; + } + + setlocale(LC_ALL, ""); +#endif +} + +void console_cleanup(console_state & con_st) { + // Reset console color + console_set_color(con_st, CONSOLE_COLOR_DEFAULT); + +#if !defined(_WIN32) + if (con_st.tty != nullptr) { + con_st.out = stdout; + fclose(con_st.tty); + con_st.tty = nullptr; + } + // Restore the terminal settings on POSIX systems + tcsetattr(STDIN_FILENO, TCSANOW, &con_st.prev_state); +#endif +} + +/* Keep track of current color of output, and emit ANSI code if it changes. */ +void console_set_color(console_state & con_st, console_color_t color) { + if (con_st.use_color && con_st.color != color) { + fflush(stdout); + switch(color) { + case CONSOLE_COLOR_DEFAULT: + fprintf(con_st.out, ANSI_COLOR_RESET); + break; + case CONSOLE_COLOR_PROMPT: + fprintf(con_st.out, ANSI_COLOR_YELLOW); + break; + case CONSOLE_COLOR_USER_INPUT: + fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_GREEN); + break; + } + con_st.color = color; + fflush(con_st.out); } } -// Convert a wide Unicode string to an UTF8 string -void win32_utf8_encode(const std::wstring & wstr, std::string & str) { - int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wstr[0], (int)wstr.size(), NULL, 0, NULL, NULL); - std::string strTo(size_needed, 0); - WideCharToMultiByte(CP_UTF8, 0, &wstr[0], (int)wstr.size(), &strTo[0], size_needed, NULL, NULL); - str = strTo; -} +char32_t getchar32() { + wchar_t wc = getwchar(); + if (static_cast(wc) == WEOF) { + return WEOF; + } + +#if WCHAR_MAX == 0xFFFF + if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate + wchar_t low_surrogate = getwchar(); + if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate + return (static_cast(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000; + } + } + if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair + return 0xFFFD; // Return the replacement character U+FFFD + } #endif + + return static_cast(wc); +} + +void pop_cursor(console_state & con_st) { +#if defined(_WIN32) + if (con_st.hConsole != NULL) { + CONSOLE_SCREEN_BUFFER_INFO bufferInfo; + GetConsoleScreenBufferInfo(con_st.hConsole, &bufferInfo); + + COORD newCursorPosition = bufferInfo.dwCursorPosition; + if (newCursorPosition.X == 0) { + newCursorPosition.X = bufferInfo.dwSize.X - 1; + newCursorPosition.Y -= 1; + } else { + newCursorPosition.X -= 1; + } + + SetConsoleCursorPosition(con_st.hConsole, newCursorPosition); + return; + } +#endif + putc('\b', con_st.out); +} + +int estimateWidth(char32_t codepoint) { +#if defined(_WIN32) + return 1; +#else + return wcwidth(codepoint); +#endif +} + +int put_codepoint(console_state & con_st, const char* utf8_codepoint, size_t length, int expectedWidth) { +#if defined(_WIN32) + CONSOLE_SCREEN_BUFFER_INFO bufferInfo; + if (!GetConsoleScreenBufferInfo(con_st.hConsole, &bufferInfo)) { + // go with the default + return expectedWidth; + } + COORD initialPosition = bufferInfo.dwCursorPosition; + DWORD nNumberOfChars = length; + WriteConsole(con_st.hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL); + + CONSOLE_SCREEN_BUFFER_INFO newBufferInfo; + GetConsoleScreenBufferInfo(con_st.hConsole, &newBufferInfo); + + // Figure out our real position if we're in the last column + if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) { + DWORD nNumberOfChars; + WriteConsole(con_st.hConsole, &" \b", 2, &nNumberOfChars, NULL); + GetConsoleScreenBufferInfo(con_st.hConsole, &newBufferInfo); + } + + int width = newBufferInfo.dwCursorPosition.X - initialPosition.X; + if (width < 0) { + width += newBufferInfo.dwSize.X; + } + return width; +#else + // we can trust expectedWidth if we've got one + if (expectedWidth >= 0 || con_st.tty == nullptr) { + fwrite(utf8_codepoint, length, 1, con_st.out); + return expectedWidth; + } + + fputs("\033[6n", con_st.tty); // Query cursor position + int x1, x2, y1, y2; + int results = 0; + results = fscanf(con_st.tty, "\033[%d;%dR", &y1, &x1); + + fwrite(utf8_codepoint, length, 1, con_st.tty); + + fputs("\033[6n", con_st.tty); // Query cursor position + results += fscanf(con_st.tty, "\033[%d;%dR", &y2, &x2); + + if (results != 4) { + return expectedWidth; + } + + int width = x2 - x1; + if (width < 0) { + // Calculate the width considering text wrapping + struct winsize w; + ioctl(STDOUT_FILENO, TIOCGWINSZ, &w); + width += w.ws_col; + } + return width; +#endif +} + +void replace_last(console_state & con_st, char ch) { +#if defined(_WIN32) + pop_cursor(con_st); + put_codepoint(con_st, &ch, 1, 1); +#else + fprintf(con_st.out, "\b%c", ch); +#endif +} + +void append_utf8(char32_t ch, std::string & out) { + if (ch <= 0x7F) { + out.push_back(static_cast(ch)); + } else if (ch <= 0x7FF) { + out.push_back(static_cast(0xC0 | ((ch >> 6) & 0x1F))); + out.push_back(static_cast(0x80 | (ch & 0x3F))); + } else if (ch <= 0xFFFF) { + out.push_back(static_cast(0xE0 | ((ch >> 12) & 0x0F))); + out.push_back(static_cast(0x80 | ((ch >> 6) & 0x3F))); + out.push_back(static_cast(0x80 | (ch & 0x3F))); + } else if (ch <= 0x10FFFF) { + out.push_back(static_cast(0xF0 | ((ch >> 18) & 0x07))); + out.push_back(static_cast(0x80 | ((ch >> 12) & 0x3F))); + out.push_back(static_cast(0x80 | ((ch >> 6) & 0x3F))); + out.push_back(static_cast(0x80 | (ch & 0x3F))); + } else { + // Invalid Unicode code point + } +} + +// Helper function to remove the last UTF-8 character from a string +void pop_back_utf8_char(std::string & line) { + if (line.empty()) { + return; + } + + size_t pos = line.length() - 1; + + // Find the start of the last UTF-8 character (checking up to 4 bytes back) + for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) { + if ((line[pos] & 0xC0) != 0x80) break; // Found the start of the character + } + line.erase(pos); +} + +bool console_readline(console_state & con_st, std::string & line) { + console_set_color(con_st, CONSOLE_COLOR_USER_INPUT); + if (con_st.out != stdout) { + fflush(stdout); + } + + line.clear(); + std::vector widths; + bool is_special_char = false; + bool end_of_stream = false; + + char32_t input_char; + while (true) { + fflush(con_st.out); // Ensure all output is displayed before waiting for input + input_char = getchar32(); + + if (input_char == '\r' || input_char == '\n') { + break; + } + + if (input_char == WEOF || input_char == 0x04 /* Ctrl+D*/) { + end_of_stream = true; + break; + } + + if (is_special_char) { + console_set_color(con_st, CONSOLE_COLOR_USER_INPUT); + replace_last(con_st, line.back()); + is_special_char = false; + } + + if (input_char == '\033') { // Escape sequence + char32_t code = getchar32(); + if (code == '[' || code == 0x1B) { + // Discard the rest of the escape sequence + while ((code = getchar32()) != WEOF) { + if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') { + break; + } + } + } + } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace + if (!widths.empty()) { + int count; + do { + count = widths.back(); + widths.pop_back(); + // Move cursor back, print space, and move cursor back again + for (int i = 0; i < count; i++) { + replace_last(con_st, ' '); + pop_cursor(con_st); + } + pop_back_utf8_char(line); + } while (count == 0 && !widths.empty()); + } + } else { + int offset = line.length(); + append_utf8(input_char, line); + int width = put_codepoint(con_st, line.c_str() + offset, line.length() - offset, estimateWidth(input_char)); + if (width < 0) { + width = 0; + } + widths.push_back(width); + } + + if (!line.empty() && (line.back() == '\\' || line.back() == '/')) { + console_set_color(con_st, CONSOLE_COLOR_PROMPT); + replace_last(con_st, line.back()); + is_special_char = true; + } + } + + bool has_more = con_st.multiline_input; + if (is_special_char) { + replace_last(con_st, ' '); + pop_cursor(con_st); + + char last = line.back(); + line.pop_back(); + if (last == '\\') { + line += '\n'; + fputc('\n', con_st.out); + has_more = !has_more; + } else { + // llama will just eat the single space, it won't act as a space + if (line.length() == 1 && line.back() == ' ') { + line.clear(); + pop_cursor(con_st); + } + has_more = false; + } + } else { + if (end_of_stream) { + has_more = false; + } else { + line += '\n'; + fputc('\n', con_st.out); + } + } + + fflush(con_st.out); + return has_more; +} diff --git a/examples/common.h b/examples/common.h index cbbc2dfab..717838f06 100644 --- a/examples/common.h +++ b/examples/common.h @@ -8,30 +8,48 @@ #include #include #include +#include + +#if !defined (_WIN32) +#include +#include +#endif // // CLI argument parsing // +int32_t get_num_physical_cores(); struct gpt_params { - int32_t seed = -1; // RNG seed - int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()); - int32_t n_predict = 128; // new tokens to predict - int32_t repeat_last_n = 64; // last n tokens to penalize - int32_t n_parts = -1; // amount of model parts (-1 = determine from model dimensions) - int32_t n_ctx = 512; // context size - int32_t n_batch = 8; // batch size for prompt processing - int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t seed = -1; // RNG seed + int32_t n_threads = get_num_physical_cores(); + int32_t n_predict = -1; // new tokens to predict + int32_t n_parts = -1; // amount of model parts (-1 = determine from model dimensions) + int32_t n_ctx = 512; // context size + int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS) + int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_gpu_layers = 0; // number of layers to store in VRAM // sampling parameters - int32_t top_k = 40; - float top_p = 0.95f; - float temp = 0.80f; - float repeat_penalty = 1.10f; + std::unordered_map logit_bias; // logit bias for specific tokens + int32_t top_k = 40; // <= 0 to use vocab size + float top_p = 0.95f; // 1.0 = disabled + float tfs_z = 1.00f; // 1.0 = disabled + float typical_p = 1.00f; // 1.0 = disabled + float temp = 0.80f; // 1.0 = disabled + float repeat_penalty = 1.10f; // 1.0 = disabled + int32_t repeat_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) + float frequency_penalty = 0.00f; // 0.0 = disabled + float presence_penalty = 0.00f; // 0.0 = disabled + int mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0 + float mirostat_tau = 5.00f; // target entropy + float mirostat_eta = 0.10f; // learning rate std::string model = "models/lamma-7B/ggml-model.bin"; // model path std::string prompt = ""; - std::string input_prefix = ""; // string to prefix user inputs with + std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state + std::string input_prefix = ""; // string to prefix user inputs with + std::string input_suffix = ""; // string to suffix user inputs with std::vector antiprompt; // string upon seeing which more user input is prompted std::string lora_adapter = ""; // lora adapter path @@ -41,12 +59,14 @@ struct gpt_params { bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs bool interactive = false; // interactive mode + bool prompt_cache_all = false; // save user input and generations to prompt cache bool embedding = false; // get only sentence embedding - bool interactive_start = false; // wait for user input immediately + bool interactive_first = false; // wait for user input immediately + bool multiline_input = false; // reverse the usage of `\` bool instruct = false; // instruction mode (used for Alpaca models) - bool ignore_eos = false; // do not stop generating after eos + bool penalize_nl = true; // consider newlines as a repeatable token 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 @@ -66,6 +86,12 @@ std::string gpt_random_prompt(std::mt19937 & rng); std::vector llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos); +// +// Model utils +// + +struct llama_context * llama_init_from_gpt_params(const gpt_params & params); + // // Console utils // @@ -86,13 +112,20 @@ enum console_color_t { }; struct console_state { + bool multiline_input = false; bool use_color = false; console_color_t color = CONSOLE_COLOR_DEFAULT; + + FILE* out = stdout; +#if defined (_WIN32) + void* hConsole; +#else + FILE* tty = nullptr; + termios prev_state; +#endif }; -void set_console_color(console_state & con_st, console_color_t color); - -#if defined (_WIN32) -void win32_console_init(bool enable_color); -void win32_utf8_encode(const std::wstring & wstr, std::string & str); -#endif +void console_init(console_state & con_st); +void console_cleanup(console_state & con_st); +void console_set_color(console_state & con_st, console_color_t color); +bool console_readline(console_state & con_st, std::string & line); diff --git a/examples/embedding/CMakeLists.txt b/examples/embedding/CMakeLists.txt index 88c425d4a..db73b6b44 100644 --- a/examples/embedding/CMakeLists.txt +++ b/examples/embedding/CMakeLists.txt @@ -2,3 +2,6 @@ set(TARGET embedding) add_executable(${TARGET} embedding.cpp) target_link_libraries(${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/embedding/embedding.cpp b/examples/embedding/embedding.cpp index e10de619c..bb3fd50a9 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -1,5 +1,6 @@ #include "common.h" #include "llama.h" +#include "build-info.h" #include @@ -18,11 +19,13 @@ int main(int argc, char ** argv) { "expect poor results\n", __func__, params.n_ctx); } - if (params.seed <= 0) { + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + + if (params.seed < 0) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); + fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { @@ -32,24 +35,10 @@ int main(int argc, char ** argv) { llama_context * ctx; // load the model - { - auto lparams = llama_context_default_params(); - - lparams.n_ctx = params.n_ctx; - lparams.n_parts = params.n_parts; - lparams.seed = params.seed; - lparams.f16_kv = params.memory_f16; - lparams.logits_all = params.perplexity; - lparams.use_mmap = params.use_mmap; - lparams.use_mlock = params.use_mlock; - lparams.embedding = params.embedding; - - ctx = llama_init_from_file(params.model.c_str(), lparams); - - if (ctx == NULL) { - fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str()); - return 1; - } + ctx = llama_init_from_gpt_params(params); + if (ctx == NULL) { + fprintf(stderr, "%s: error: unable to load model\n", __func__); + return 1; } // print system information @@ -67,9 +56,6 @@ int main(int argc, char ** argv) { // tokenize the prompt auto embd_inp = ::llama_tokenize(ctx, params.prompt, true); - // determine newline token - auto llama_token_newline = ::llama_tokenize(ctx, "\n", false); - if (params.verbose_prompt) { fprintf(stderr, "\n"); fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str()); diff --git a/examples/jeopardy/README.md b/examples/jeopardy/README.md new file mode 100644 index 000000000..4c42e3cdb --- /dev/null +++ b/examples/jeopardy/README.md @@ -0,0 +1,21 @@ +# llama.cpp/example/jeopardy + +This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer. + +The jeopardy test can be used to compare the fact knowledge of different models and compare them to eachother. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc. + + +Step 1: Open jeopardy.sh and modify the following: +``` +MODEL=(path to your model) +MODEL_NAME=(name of your model) +prefix=(basically, if you use vicuna it's Human: , if you use something else it might be User: , etc) +opts=(add -instruct here if needed for your model, or anything else you want to test out) +``` +Step 2: Run `jeopardy.sh` from the llama.cpp folder + +Step 3: Repeat steps 1 and 2 until you have all the results you need. + +Step 4: Run `graph.py`, and follow the instructions. At the end, it will generate your final graph. + +Note: The Human bar is based off of the full, original 100 sample questions. If you modify the question count or questions, it will not be valid. diff --git a/examples/jeopardy/graph.py b/examples/jeopardy/graph.py new file mode 100644 index 000000000..d00b28652 --- /dev/null +++ b/examples/jeopardy/graph.py @@ -0,0 +1,56 @@ +import matplotlib.pyplot as plt +import sys, os +import csv + +labels = [] +numbers = [] +numEntries = 1 + +rows = [] + +def bar_chart(numbers, labels, pos): + plt.bar(pos, numbers, color='blue') + plt.xticks(ticks=pos, labels=labels) + plt.title("Jeopardy Results by Model") + plt.xlabel("Model") + plt.ylabel("Questions Correct") + plt.show() + +def calculatecorrect(): + directory = os.fsencode("./examples/jeopardy/results/") + csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',') + for row in csv_reader: + global rows + rows.append(row) + for listing in os.listdir(directory): + filename = os.fsdecode(listing) + if filename.endswith(".txt"): + file = open("./examples/jeopardy/results/" + filename, "rt") + global labels + global numEntries + global numbers + labels.append(filename[:-4]) + numEntries += 1 + i = 1 + totalcorrect = 0 + for line in file.readlines(): + if line.strip() != "------": + print(line) + else: + print("Correct answer: " + rows[i][2] + "\n") + i+=1 + print("Did the AI get the question right? (y/n)") + if input() == "y": + totalcorrect += 1 + numbers.append(totalcorrect) + + + +if __name__ == '__main__': + calculatecorrect() + pos = list(range(numEntries)) + labels.append("Human") + numbers.append(48.11) + bar_chart(numbers, labels, pos) + print(labels) + print(numbers) diff --git a/examples/jeopardy/jeopardy.sh b/examples/jeopardy/jeopardy.sh new file mode 100644 index 000000000..9bdbc755c --- /dev/null +++ b/examples/jeopardy/jeopardy.sh @@ -0,0 +1,30 @@ +#!/bin/bash +set -e + +MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin +MODEL_NAME=Vicuna + +# exec options +prefix="Human: " # Ex. Vicuna uses "Human: " +opts="--temp 0 -n 80" # additional flags +nl=' +' +introduction="You will be playing a game of Jeopardy. Simply answer the question in the correct format (Ex. What is Paris, or Who is George Washington)." + +# file options +question_file=./examples/jeopardy/questions.txt +touch ./examples/jeopardy/results/$MODEL_NAME.txt +output_file=./examples/jeopardy/results/$MODEL_NAME.txt + +counter=1 + +echo 'Running' +while IFS= read -r question +do + exe_cmd="./main -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\"" + echo $counter + echo "Current Question: $question" + eval "$exe_cmd" + echo -e "\n------" >> $output_file + counter=$((counter+1)) +done < "$question_file" diff --git a/examples/jeopardy/qasheet.csv b/examples/jeopardy/qasheet.csv new file mode 100644 index 000000000..35b084189 --- /dev/null +++ b/examples/jeopardy/qasheet.csv @@ -0,0 +1,103 @@ +Index,Original Category,Original Correct Question,Model Prompt +1,The Oscars,Who is John Williams?,Which actor Born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars? +2,English Literature,What is Paradise Lost?,"What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?" +3,Writers’ Lesser-Known Works,Who is Niccolò Machiavelli?,"Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?" +4,Exploration,What is Easter Island (Rapa Nui)?,"James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?" +5,The Bill of Rights,What is the Eighth Amendment?,England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution? +6,Nobel Peace Prize Winners,Who are Nelson Mandela & Desmond Tutu?,"Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?" +7,Famous Names,Who is Walt Disney?,"In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?" +8,Geography,What is Colombia?,"Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?" +9,Fashion History,What are rhinestones?,"Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?" +10,Movies of the ’80s,What is Driving Miss Daisy?,What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated? +11,Novelists,Who is John Grisham?,"A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?" +12,20th Century Eponyms,What is the Maginot Line?,"A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?" +13,City History,What is Stockholm?,"Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?" +14,Brand Names,What is Jacuzzi?,"The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?" +15,American Authors,Who is Washington Irving?,"In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?" +16,Symbols,What is “less than”?,What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society? +17,Movie Theme Songs,Who is James Bond?,"Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?" +18,American Novelists,Who is Joseph Heller?,"What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?" +19,Medieval Places,"What is Canterbury, England? (Canterbury Cathedral)","In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?" +20,Countries of Africa,What is Morocco?,"At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?" +21,Statehood,What is Wyoming?,Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women? +22,1980s Movies,What is Raiders of the Lost Ark?,"A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?" +23,Art Exhibitions,Who is Rembrandt?,In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation? +24,Countries of the World,What is Mongolia?,"Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?" +25,Literature,What is “Howl”?,A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'? +26,Invasions,Who is William of Orange?,"Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?" +27,Landmarks,What is the Eiffel Tower?,"After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?" +28,Geographic Name’s the Same,What is Dover?,"The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?" +29,Names in the Bookstore,Who is Peter Mark Roget?,"This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?" +30,U.S. History,Who is Dr. Samuel Mudd?,"An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?" +31,American Literature,What is The Things They Carried?,"Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?" +32,Nonfiction,What is The Communist Manifesto,"What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?" +33, a new version was passed 81 years later,Laws in U.S. History,What is the Civil Rights Act?,,,,,,,,,,,,,,,,,,0, 2/3 +34,Names of Myth,Who is Helen of Troy?,"Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?" +35,African Countries,What is Sudan?,"Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?" +36,The Ancient World,What is Alexandria?,"The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?" +37,Famous Names,Who is Andy Warhol?,"For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?" +38,People & Places,What is Guam?,"Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?" +39,Current World Leaders,What is the Philippines?,"In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?" +40,Writers & The South,Who is Tennessee Williams?,In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South? +41,National Parks,What is Yellowstone?,"What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?" +42,Sports,Who are the Harlem Globetrotters?,"In 2010 who introduced the 4-point shot, 35 feet from the basket?" +43,The U.S. Military,What is “Top Gun”?,Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969? +44,Art & Science,What is Halley’s Comet?,"A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?" +45,Words From World War I,What is “tank”?,"In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?" +46,European History,What is Holy Roman Emperor?,"Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?" +47,Theater History,Who is Peter Pan?,"In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?" +48,European Cities,What is Aachen?,"Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?" +49,Word Origins,What is mantra?,This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'? +50,Inventions,What is barbed wire?,1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'? +51,World War II,What is Schindler’s list?,"Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?" +52, their offspring was the source of this mythical object,Mythology,What is the Golden Fleece? +53,Literature,What is Pride and Prejudice?,"Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?" +54, only these 2 west of the Mississippi River border each other,U.S. State Names,What are Oregon & Nevada? +55,Word Origins,What is passion?,"Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?" +56,World Cinema,What is La Vie en Rose?,"The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?" +57,History,What is Santa Maria?,"Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?" +58,Landmarks,What is a kremlin?,Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what? +59,Foreign-Born Authors,Who is Vladimir Nabokov?,In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'? +60,Astronomy & Geography,What is Capricorn?,"At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?" +61,Television,What is Law & Order?,"Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?" +62,British Landmarks,What is the Tower of London?,"Like Sir Thomas More, 3 16th century English queens are buried at what British location?" +63,Early American History,What are witches?,"In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person … be condemned'?" +64,Geography Mnemonics,What are Arkansas and Louisiana?,"The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?" +65,Business Milestones,What is the Ford Model T?,"What was first sold in 1908, at a price equivalent to about $27,000 today?" +66,In The Bookstore,Who is Tom Clancy?,The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot? +67,Historic Art,What is the Bayeux Tapestry?,The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what? +68,Pop Stars,Who is Madonna?,In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s? +69,Classic Tale Characters,Who is Scheherazade?,"In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?" +70,USA,What is Jack Daniel’s?,"Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?" +71,Historic People,Who was William Bligh?,"After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?" +72,The Movies,What is The Godfather?,Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022? +73,Continental Geography,What is Colombia?,"Until a 1903 secession, what country's contiguous territory spanned 2 continents?" +74,Foreign-Born Authors,Who is Isabel Allende?,"Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?" +75,Historic Crimes,What is the Mona Lisa?,"Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?" +76,U.S. Bodies of Water,What is Lake Mead?,"Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?" +77,Gods & Goddesses,Who is Aurora (or Eos)?,"Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?" +78,America At War,What is the Battle of New Orleans?,"Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?" +79,Children’s Books,What is The Velveteen Rabbit?,"Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?" +80,TV Finales,What is Grace and Frankie?,"In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?" +81,American Poems,Who is Evangeline?,"In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?" +82,Famous Names,Who is Banksy?,"In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?" +83,Children’s Lit,What is Charlotte’s Web?,The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'? +84,Classic Songs,What is “Here Comes Santa Claus”?,The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite? +85,Brand Names,What are Milk Duds?,"Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?" +86,Countries of the World,What is Italy?,"What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?" +87,Action Movies,What is Die Hard?,"What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?" +88,Presidential Facts,Who is Woodrow Wilson?,Only 3 presidents have married while in office— John Tyler was the first & which one was the last? +89,19th Century Americans,Who is Frederick Douglass?,"Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?" +90,Latin Phrases,What is “quid pro quo”?,"Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?" +91,1970s Movies,What is Monty Python and the Holy Grail?,The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience? +92,Name’s The Same,What is Manhattan?,"A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?" +93,U.S. Presidents,Who is Calvin Coolidge?,"Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?" +94,Plays,What is The Tempest?,A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play? +95,Landmarks,What is the Berlin Wall?,"In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?" +96,World Capitals,"What is Vienna, Austria?","Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?" +97,Language & Its Meanings,What is a night owl?,"Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?" +98,Flags of Our Hemisphere,What is Brazil?,"The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?" +99,Names in U.S. History,Who is Oliver Brown?,What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951? +100,Children’s Authors,"Who is Sarah? (from Sarah, Plain and Tall)","Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?" +,,, +TOTALS,,, diff --git a/examples/jeopardy/questions.txt b/examples/jeopardy/questions.txt new file mode 100644 index 000000000..eea78a057 --- /dev/null +++ b/examples/jeopardy/questions.txt @@ -0,0 +1,100 @@ +Which man born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars? +What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'? +Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions? +James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'? +England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution? +Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners? +In 1966, the year of who's death did he share plans for an experimental prototype community in Florida? +Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea? +Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany? +What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated? +A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'? +A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'? +Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response? +The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis? +In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'? +What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society? +Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness? +What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service? +In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'? +At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'? +Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women? +A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'? +In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation? +Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country? +A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'? +Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'? +After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'? +The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states? +This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book? +An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland? +Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic? +What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'? +A radical Republican championed what 1875 act but the Supreme Court struck it down in 1883; a new version was passed 81 years later? +Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life? +Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence? +The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned? +For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk? +Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group? +In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'? +In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South? +What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'? +In 2010 who introduced the 4-point shot, 35 feet from the basket? +Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969? +A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem? +In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable? +Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage? +In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage? +Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II? +This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'? +1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'? +Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers? +Poseidon carried off the maiden Theophane & turned her into a ewe; their offspring was the source of what mythical object? +Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier? +5 U.S. states have 6-letter names; only which 2 west of the Mississippi River border each other? +Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind? +The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title? +Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast? +Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what? +In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'? +At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name? +Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990? +Like Sir Thomas More, 3 16th century English queens are buried at what British location? +In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person be condemned'? +The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states? +What was first sold in 1908, at a price equivalent to about $27,000 today? +The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot? +The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what? +In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s? +In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times? +Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county? +After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'? +Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022? +Until a 1903 secession, what country's contiguous territory spanned 2 continents? +Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter? +Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911? +Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled? +Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios? +Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday? +Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'? +In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022? +In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death? +In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'? +The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'? +The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite? +Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product? +What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon? +What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'? +Only 3 presidents have married while in office— John Tyler was the first & which one was the last? +Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century? +Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another? +The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience? +A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name? +Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President? +A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play? +In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'? +Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'? +Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'? +The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star? +What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951? +Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England? diff --git a/examples/main/CMakeLists.txt b/examples/main/CMakeLists.txt index b2dcc2910..c364242fb 100644 --- a/examples/main/CMakeLists.txt +++ b/examples/main/CMakeLists.txt @@ -2,3 +2,6 @@ set(TARGET main) add_executable(${TARGET} main.cpp) 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/main/README.md b/examples/main/README.md index f09e7ba97..7c03f92c8 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -1,3 +1,289 @@ -# main +# llama.cpp/example/main -TODO +This example program allows you to use various LLaMA language models in an easy and efficient way. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts. + +## Table of Contents + +1. [Quick Start](#quick-start) +2. [Common Options](#common-options) +3. [Input Prompts](#input-prompts) +4. [Interaction](#interaction) +5. [Context Management](#context-management) +6. [Generation Flags](#generation-flags) +7. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options) +8. [Additional Options](#additional-options) + +## Quick Start + +To get started right away, run the following command, making sure to use the correct path for the model you have: + +#### Unix-based systems (Linux, macOS, etc.): + +```bash +./main -m models/7B/ggml-model.bin --prompt "Once upon a time" +``` + +#### Windows: + +```powershell +main.exe -m models\7B\ggml-model.bin --prompt "Once upon a time" +``` + +For an interactive experience, try this command: + +#### Unix-based systems (Linux, macOS, etc.): + +```bash +./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " \ +'User: Hi +AI: Hello. I am an AI chatbot. Would you like to talk? +User: Sure! +AI: What would you like to talk about? +User:' +``` + +#### Windows: + +```powershell +main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -e --prompt "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:" +``` + +The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it): + +#### Unix-based systems (Linux, macOS, etc.): + +```bash +./main -m models/7B/ggml-model.bin --ignore-eos -n -1 --random-prompt +``` + +#### Windows: + +```powershell +main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt +``` + +## Common Options + +In this section, we cover the most commonly used options for running the `main` program with the LLaMA models: + +- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`). +- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses. +- `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models. +- `-n N, --n_predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text. +- `-c N, --ctx_size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. + +## Input Prompts + +The `main` program provides several ways to interact with the LLaMA models using input prompts: + +- `--prompt PROMPT`: Provide a prompt directly as a command-line option. +- `--file FNAME`: Provide a file containing a prompt or multiple prompts. +- `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.) +- `--random-prompt`: Start with a randomized prompt. + +## Interaction + +The `main` program offers a seamless way to interact with LLaMA models, allowing users to engage in real-time conversations or provide instructions for specific tasks. The interactive mode can be triggered using various options, including `--interactive`, `--interactive-first`, and `--instruct`. + +In interactive mode, users can participate in text generation by injecting their input during the process. Users can press `Ctrl+C` at any time to interject and type their input, followed by pressing `Return` to submit it to the LLaMA model. To submit additional lines without finalizing input, users can end the current line with a backslash (`\`) and continue typing. + +### Interaction Options + +- `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model. +- `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation. +- `-ins, --instruct`: Run the program in instruction mode, which is specifically designed to work with Alpaca models that excel in completing tasks based on user instructions. +- `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text. + +By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs. + +### Reverse Prompts + +Reverse prompts are a powerful way to create a chat-like experience with a LLaMA model by pausing the text generation when specific text strings are encountered: + +- `-r PROMPT, --reverse-prompt PROMPT`: Specify one or multiple reverse prompts to pause text generation and switch to interactive mode. For example, `-r "User:"` can be used to jump back into the conversation whenever it's the user's turn to speak. This helps create a more interactive and conversational experience. However, the reverse prompt doesn't work when it ends with a space. + +To overcome this limitation, you can use the `--in-prefix` flag to add a space or any other characters after the reverse prompt. + +### In-Prefix + +The `--in-prefix` flag is used to add a prefix to your input, primarily, this is used to insert a space after the reverse prompt. Here's an example of how to use the `--in-prefix` flag in conjunction with the `--reverse-prompt` flag: + +```sh +./main -r "User:" --in-prefix " " +``` + +### In-Suffix + +The `--in-suffix` flag is used to add a suffix after your input. This is useful for adding an "Assistant:" prompt after the user's input. It's added after the new-line character (`\n`) that's automatically added to the end of the user's input. Here's an example of how to use the `--in-suffix` flag in conjunction with the `--reverse-prompt` flag: + +```sh +./main -r "User:" --in-prefix " " --in-suffix "Assistant:" +``` + +### Instruction Mode + +Instruction mode is particularly useful when working with Alpaca models, which are designed to follow user instructions for specific tasks: + +- `-ins, --instruct`: Enable instruction mode to leverage the capabilities of Alpaca models in completing tasks based on user-provided instructions. + +Technical detail: the user's input is internally prefixed with the reverse prompt (or `### Instruction:` as the default), and followed by `### Response:` (except if you just press Return without any input, to keep generating a longer response). + +By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs. + +## Context Management + +During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations. + +### Context Size + +The `--ctx_size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations. + +- `-c N, --ctx_size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results. + +### Keep Prompt + +The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained. + +- `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt. + +By utilizing context management options like `--ctx_size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation. + +## Generation Flags + +The following options allow you to control the text generation process and fine-tune the diversity, creativity, and quality of the generated text according to your needs. By adjusting these options and experimenting with different combinations of values, you can find the best settings for your specific use case. + +### Number of Tokens to Predict + +- `-n N, --n_predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity). + +The `--n_predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit. + +It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n_predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter. + +### Temperature + +- `--temp N`: Adjust the randomness of the generated text (default: 0.8). + +Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run. + +Example usage: `--temp 0.5` + +### Repeat Penalty + +- `--repeat_penalty N`: Control the repetition of token sequences in the generated text (default: 1.1). +- `--repeat_last_n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx_size). +- `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty. + +The `repeat_penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1. + +The `repeat_last_n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx_size`). + +Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases. + +Example usage: `--repeat_penalty 1.15 --repeat_last_n 128 --no-penalize-nl` + +### Top-K Sampling + +- `--top_k N`: Limit the next token selection to the K most probable tokens (default: 40). + +Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40. + +Example usage: `--top_k 30` + +### Top-P Sampling + +- `--top_p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9). + +Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9. + +Example usage: `--top_p 0.95` + +### Tail Free Sampling (TFS) + +- `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled). + +Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. The method adjusts the logits (token probabilities) by raising them to the power of the parameter z. A higher value of z (e.g., 2.0) will further suppress less likely tokens from the tail of the distribution, while a value of 1.0 disables the effect of TFS. By setting the parameter z, you can control how much the probabilities of less likely tokens are reduced. + +Example usage: `--tfs 2.0` + +### Locally Typical Sampling + +- `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled). + +Locally typical sampling promotes the generation of contextually coherent and diverse text by sampling tokens that are typical or expected based on the surrounding context. By setting the parameter p between 0 and 1, you can control the balance between producing text that is locally coherent and diverse. A value closer to 1 will promote more contextually coherent tokens, while a value closer to 0 will promote more diverse tokens. A value equal to 1 disables locally typical sampling. + +Example usage: `--typical 0.9` + +### Mirostat Sampling + +- `--mirostat N`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0). +- `--mirostat_lr N`: Set the Mirostat learning rate, parameter eta (default: 0.1). +- `--mirostat_ent N`: Set the Mirostat target entropy, parameter tau (default: 5.0). + +Mirostat is an algorithm that actively maintains the quality of generated text within a desired range during text generation. It aims to strike a balance between coherence and diversity, avoiding low-quality output caused by excessive repetition (boredom traps) or incoherence (confusion traps). + +The `--mirostat_lr` option sets the Mirostat learning rate (eta). The learning rate influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. The default value is `0.1`. + +The `--mirostat_ent` option sets the Mirostat target entropy (tau), which represents the desired perplexity value for the generated text. Adjusting the target entropy allows you to control the balance between coherence and diversity in the generated text. A lower value will result in more focused and coherent text, while a higher value will lead to more diverse and potentially less coherent text. The default value is `5.0`. + +Example usage: `--mirostat 2 --mirostat_lr 0.05 --mirostat_ent 3.0` + +### Logit Bias + +- `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion. + +The logit bias option allows you to manually adjust the likelihood of specific tokens appearing in the generated text. By providing a token ID and a positive or negative bias value, you can increase or decrease the probability of that token being generated. + +For example, use `--logit-bias 15043+1` to increase the likelihood of the token 'Hello', or `--logit-bias 15043-1` to decrease its likelihood. Using a value of negative infinity, `--logit-bias 15043-inf` ensures that the token `Hello` is never produced. + +A more practical use case might be to prevent the generation of `\code{begin}` and `\code{end}` by setting the `\` token (29905) to negative infinity with `-l 29905-inf`. (This is due to the prevalence of LaTeX codes that show up in LLaMA model inference.) + +Example usage: `--logit-bias 29905-inf` + +### RNG Seed + +- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed). + +The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run. + +## Performance Tuning and Memory Options + +These options help improve the performance and memory usage of the LLaMA models. By adjusting these settings, you can fine-tune the model's behavior to better suit your system's capabilities and achieve optimal performance for your specific use case. + +### Number of Threads + +- `-t N, --threads N`: Set the number of threads to use during computation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance. + +### Mlock + +- `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped. This can improve performance but trades away some of the advantages of memory-mapping by requiring more RAM to run and potentially slowing down load times as the model loads into RAM. + +### No Memory Mapping + +- `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance. Disabling mmap results in slower load times but may reduce pageouts if you're not using `--mlock`. Note that if the model is larger than the total amount of RAM, turning off mmap would prevent the model from loading at all. + +### Memory Float 32 + +- `--memory_f32`: Use 32-bit floats instead of 16-bit floats for memory key+value, allowing higher quality inference at the cost of higher memory usage. + +### Batch Size + +- `-b N, --batch_size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations. + +### Prompt Caching + +- `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. + +### Quantization + +For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run). + +## Additional Options + +These options provide extra functionality and customization when running the LLaMA models: + +- `-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. +- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains. +- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation. diff --git a/examples/main/main.cpp b/examples/main/main.cpp index b7b3c4196..8543414dd 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -5,6 +5,7 @@ #include "common.h" #include "llama.h" +#include "build-info.h" #include #include @@ -21,21 +22,26 @@ #include #include #elif defined (_WIN32) +#define WIN32_LEAN_AND_MEAN +#define NOMINMAX +#include #include #endif static console_state con_st; +static llama_context ** g_ctx; static bool is_interacting = false; #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) void sigint_handler(int signo) { - set_console_color(con_st, CONSOLE_COLOR_DEFAULT); - printf("\n"); // this also force flush stdout. if (signo == SIGINT) { if (!is_interacting) { is_interacting=true; } else { + console_cleanup(con_st); + printf("\n"); + llama_print_timings(*g_ctx); _exit(130); } } @@ -53,10 +59,9 @@ int main(int argc, char ** argv) { // save choice to use color for later // (note for later: this is a slightly awkward choice) con_st.use_color = params.use_color; - -#if defined (_WIN32) - win32_console_init(params.use_color); -#endif + con_st.multiline_input = params.multiline_input; + console_init(con_st); + atexit([]() { console_cleanup(con_st); }); if (params.perplexity) { printf("\n************\n"); @@ -79,11 +84,13 @@ int main(int argc, char ** argv) { "expect poor results\n", __func__, params.n_ctx); } - if (params.seed <= 0) { + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + + if (params.seed < 0) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); + fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { @@ -94,35 +101,13 @@ int main(int argc, char ** argv) { //bool is_prime(int n) {)"; llama_context * ctx; + g_ctx = &ctx; - // load the model - { - auto lparams = llama_context_default_params(); - - lparams.n_ctx = params.n_ctx; - lparams.n_parts = params.n_parts; - lparams.seed = params.seed; - lparams.f16_kv = params.memory_f16; - lparams.use_mmap = params.use_mmap; - lparams.use_mlock = params.use_mlock; - - ctx = llama_init_from_file(params.model.c_str(), lparams); - - if (ctx == NULL) { - fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str()); - return 1; - } - } - - if (!params.lora_adapter.empty()) { - int err = llama_apply_lora_from_file(ctx, - params.lora_adapter.c_str(), - params.lora_base.empty() ? NULL : params.lora_base.c_str(), - params.n_threads); - if (err != 0) { - fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); - return 1; - } + // load the model and apply lora adapter, if any + ctx = llama_init_from_gpt_params(params); + if (ctx == NULL) { + fprintf(stderr, "%s: error: unable to load model\n", __func__); + return 1; } // print system information @@ -136,7 +121,7 @@ int main(int argc, char ** argv) { // uncomment the "used_mem" line in llama.cpp to see the results if (params.mem_test) { { - const std::vector tmp(params.n_batch, 0); + const std::vector tmp(params.n_batch, llama_token_bos()); llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads); } @@ -154,6 +139,31 @@ int main(int argc, char ** argv) { // Add a space in front of the first character to match OG llama tokenizer behavior params.prompt.insert(0, 1, ' '); + std::string path_session = params.path_prompt_cache; + std::vector session_tokens; + + if (!path_session.empty()) { + fprintf(stderr, "%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str()); + + // fopen to check for existing session + FILE * fp = std::fopen(path_session.c_str(), "rb"); + if (fp != NULL) { + std::fclose(fp); + + session_tokens.resize(params.n_ctx); + size_t n_token_count_out = 0; + if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) { + fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str()); + return 1; + } + session_tokens.resize(n_token_count_out); + + fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size()); + } else { + fprintf(stderr, "%s: session file does not exist, will create\n", __func__); + } + } + // tokenize the prompt auto embd_inp = ::llama_tokenize(ctx, params.prompt, true); @@ -164,8 +174,28 @@ int main(int argc, char ** argv) { return 1; } + // debug message about similarity of saved session, if applicable + size_t n_matching_session_tokens = 0; + if (session_tokens.size()) { + for (llama_token id : session_tokens) { + if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) { + break; + } + n_matching_session_tokens++; + } + if (n_matching_session_tokens >= embd_inp.size()) { + fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__); + } else if (n_matching_session_tokens < (embd_inp.size() / 2)) { + fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n", + __func__, n_matching_session_tokens, embd_inp.size()); + } else { + fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n", + __func__, n_matching_session_tokens, embd_inp.size()); + } + } + // number of tokens to keep when resetting context - if (params.n_keep < 0 || params.n_keep > (int)embd_inp.size() || params.instruct) { + if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) { params.n_keep = (int)embd_inp.size(); } @@ -175,12 +205,12 @@ int main(int argc, char ** argv) { // in instruct mode, we inject a prefix and a suffix to each input by the user if (params.instruct) { - params.interactive_start = true; + params.interactive_first = true; params.antiprompt.push_back("### Instruction:\n\n"); } // enable interactive mode if reverse prompt or interactive start is specified - if (params.antiprompt.size() != 0 || params.interactive_start) { + if (params.antiprompt.size() != 0 || params.interactive_first) { params.interactive = true; } @@ -212,7 +242,10 @@ int main(int argc, char ** argv) { sigint_action.sa_flags = 0; sigaction(SIGINT, &sigint_action, NULL); #elif defined (_WIN32) - signal(SIGINT, sigint_handler); + auto console_ctrl_handler = [](DWORD ctrl_type) -> BOOL { + return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false; + }; + SetConsoleCtrlHandler(static_cast(console_ctrl_handler), true); #endif fprintf(stderr, "%s: interactive mode on.\n", __func__); @@ -226,9 +259,13 @@ int main(int argc, char ** argv) { if (!params.input_prefix.empty()) { fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str()); } + + if (!params.input_suffix.empty()) { + fprintf(stderr, "Input suffix: '%s'\n", params.input_suffix.c_str()); + } } - fprintf(stderr, "sampling: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", - params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty); + fprintf(stderr, "sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n", + params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau); fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); fprintf(stderr, "\n\n"); @@ -237,24 +274,35 @@ int main(int argc, char ** argv) { std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0); if (params.interactive) { + const char *control_message; + if (con_st.multiline_input) { + control_message = " - To return control to LLaMa, end your input with '\\'.\n" + " - To return control without starting a new line, end your input with '/'.\n"; + } else { + control_message = " - Press Return to return control to LLaMa.\n" + " - To return control without starting a new line, end your input with '/'.\n" + " - If you want to submit another line, end your input with '\\'.\n"; + } fprintf(stderr, "== Running in interactive mode. ==\n" #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) " - Press Ctrl+C to interject at any time.\n" #endif - " - Press Return to return control to LLaMa.\n" - " - If you want to submit another line, end your input in '\\'.\n\n"); - is_interacting = params.interactive_start; + "%s\n", control_message); + + is_interacting = params.interactive_first; } - bool is_antiprompt = false; - bool input_noecho = false; + bool is_antiprompt = false; + bool input_echo = true; + bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size(); - int n_past = 0; - int n_remain = params.n_predict; - int n_consumed = 0; + int n_past = 0; + int n_remain = params.n_predict; + int n_consumed = 0; + int n_session_consumed = 0; // the first thing we will do is to output the prompt, so set color accordingly - set_console_color(con_st, CONSOLE_COLOR_PROMPT); + console_set_color(con_st, CONSOLE_COLOR_PROMPT); std::vector embd; @@ -264,15 +312,19 @@ int main(int argc, char ** argv) { // infinite text generation via context swapping // if we run out of context: // - take the n_keep first tokens from the original prompt (via n_past) - // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch + // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches if (n_past + (int) embd.size() > n_ctx) { const int n_left = n_past - params.n_keep; - n_past = params.n_keep; + // always keep the first token - BOS + n_past = std::max(1, params.n_keep); // insert n_left/2 tokens at the start of embd from last_n_tokens embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size()); + // stop saving session if we run out of context + path_session.clear(); + //printf("\n---\n"); //printf("resetting: '"); //for (int i = 0; i < (int) embd.size(); i++) { @@ -282,34 +334,128 @@ int main(int argc, char ** argv) { //printf("\n---\n"); } - if (llama_eval(ctx, embd.data(), embd.size(), n_past, params.n_threads)) { - fprintf(stderr, "%s : failed to eval\n", __func__); - return 1; + // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past) + if (n_session_consumed < (int) session_tokens.size()) { + size_t i = 0; + for ( ; i < embd.size(); i++) { + if (embd[i] != session_tokens[n_session_consumed]) { + session_tokens.resize(n_session_consumed); + break; + } + + n_past++; + n_session_consumed++; + + if (n_session_consumed >= (int) session_tokens.size()) { + ++i; + break; + } + } + if (i > 0) { + embd.erase(embd.begin(), embd.begin() + i); + } + } + + // evaluate tokens in batches + // embd is typically prepared beforehand to fit within a batch, but not always + for (int i = 0; i < (int) embd.size(); i += params.n_batch) { + int n_eval = (int) embd.size() - i; + if (n_eval > params.n_batch) { + n_eval = params.n_batch; + } + if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return 1; + } + n_past += n_eval; + } + + if (embd.size() > 0 && !path_session.empty()) { + session_tokens.insert(session_tokens.end(), embd.begin(), embd.end()); + n_session_consumed = session_tokens.size(); } } - n_past += embd.size(); embd.clear(); if ((int) embd_inp.size() <= n_consumed && !is_interacting) { // out of user input, sample next token - const int32_t top_k = params.top_k; - const float top_p = params.top_p; - const float temp = params.temp; - const float repeat_penalty = params.repeat_penalty; + const float temp = params.temp; + const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k; + const float top_p = params.top_p; + const float tfs_z = params.tfs_z; + const float typical_p = params.typical_p; + const int32_t repeat_last_n = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n; + const float repeat_penalty = params.repeat_penalty; + const float alpha_presence = params.presence_penalty; + const float alpha_frequency = params.frequency_penalty; + const int mirostat = params.mirostat; + const float mirostat_tau = params.mirostat_tau; + const float mirostat_eta = params.mirostat_eta; + const bool penalize_nl = params.penalize_nl; + + // optionally save the session on first sample (for faster prompt loading next time) + if (!path_session.empty() && need_to_save_session) { + need_to_save_session = false; + llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size()); + } llama_token id = 0; { - auto logits = llama_get_logits(ctx); + auto logits = llama_get_logits(ctx); + auto n_vocab = llama_n_vocab(ctx); - if (params.ignore_eos) { - logits[llama_token_eos()] = 0; + // Apply params.logit_bias map + for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) { + logits[it->first] += it->second; } - id = llama_sample_top_p_top_k(ctx, - last_n_tokens.data() + n_ctx - params.repeat_last_n, - params.repeat_last_n, top_k, top_p, temp, repeat_penalty); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < n_vocab; token_id++) { + candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + + // Apply penalties + float nl_logit = logits[llama_token_nl()]; + auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx); + llama_sample_repetition_penalty(ctx, &candidates_p, + last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, + last_n_repeat, repeat_penalty); + llama_sample_frequency_and_presence_penalties(ctx, &candidates_p, + last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, + last_n_repeat, alpha_frequency, alpha_presence); + if (!penalize_nl) { + logits[llama_token_nl()] = nl_logit; + } + + if (temp <= 0) { + // Greedy sampling + id = llama_sample_token_greedy(ctx, &candidates_p); + } else { + if (mirostat == 1) { + static float mirostat_mu = 2.0f * mirostat_tau; + const int mirostat_m = 100; + llama_sample_temperature(ctx, &candidates_p, temp); + id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); + } else if (mirostat == 2) { + static float mirostat_mu = 2.0f * mirostat_tau; + llama_sample_temperature(ctx, &candidates_p, temp); + id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu); + } else { + // Temperature sampling + llama_sample_top_k(ctx, &candidates_p, top_k, 1); + llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1); + llama_sample_typical(ctx, &candidates_p, typical_p, 1); + llama_sample_top_p(ctx, &candidates_p, top_p, 1); + llama_sample_temperature(ctx, &candidates_p, temp); + id = llama_sample_token(ctx, &candidates_p); + } + } + // printf("`%d`", candidates_p.size); last_n_tokens.erase(last_n_tokens.begin()); last_n_tokens.push_back(id); @@ -329,7 +475,7 @@ int main(int argc, char ** argv) { embd.push_back(id); // echo this to console - input_noecho = false; + input_echo = true; // decrement remaining sampling budget --n_remain; @@ -347,15 +493,15 @@ int main(int argc, char ** argv) { } // display text - if (!input_noecho) { + if (input_echo) { for (auto id : embd) { printf("%s", llama_token_to_str(ctx, id)); } fflush(stdout); } // reset color to default if we there is no pending user input - if (!input_noecho && (int)embd_inp.size() == n_consumed) { - set_console_color(con_st, CONSOLE_COLOR_DEFAULT); + if (input_echo && (int)embd_inp.size() == n_consumed) { + console_set_color(con_st, CONSOLE_COLOR_DEFAULT); } // in interactive mode, and not currently processing queued inputs; @@ -375,22 +521,12 @@ int main(int argc, char ** argv) { if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) { is_interacting = true; is_antiprompt = true; - set_console_color(con_st, CONSOLE_COLOR_USER_INPUT); - fflush(stdout); break; } } } if (n_past > 0 && is_interacting) { - // potentially set color to indicate we are taking user input - set_console_color(con_st, CONSOLE_COLOR_USER_INPUT); - -#if defined (_WIN32) - // Windows: must reactivate sigint handler after each signal - signal(SIGINT, sigint_handler); -#endif - if (params.instruct) { printf("\n> "); } @@ -404,33 +540,21 @@ int main(int argc, char ** argv) { std::string line; bool another_line = true; do { -#if defined(_WIN32) - std::wstring wline; - if (!std::getline(std::wcin, wline)) { - // input stream is bad or EOF received - return 0; - } - win32_utf8_encode(wline, line); -#else - if (!std::getline(std::cin, line)) { - // input stream is bad or EOF received - return 0; - } -#endif - if (line.empty() || line.back() != '\\') { - another_line = false; - } else { - line.pop_back(); // Remove the continue character - } - buffer += line + '\n'; // Append the line to the result + another_line = console_readline(con_st, line); + buffer += line; } while (another_line); // done taking input, reset color - set_console_color(con_st, CONSOLE_COLOR_DEFAULT); + console_set_color(con_st, CONSOLE_COLOR_DEFAULT); // Add tokens to embd only if the input buffer is non-empty // Entering a empty line lets the user pass control back if (buffer.length() > 1) { + // append input suffix if any + if (!params.input_suffix.empty()) { + buffer += params.input_suffix; + printf("%s", params.input_suffix.c_str()); + } // instruct mode: insert instruction prefix if (params.instruct && !is_antiprompt) { @@ -449,7 +573,7 @@ int main(int argc, char ** argv) { n_remain -= line_inp.size(); } - input_noecho = true; // do not echo this again + input_echo = false; // do not echo this again } if (n_past > 0) { @@ -474,14 +598,13 @@ int main(int argc, char ** argv) { } } -#if defined (_WIN32) - signal(SIGINT, SIG_DFL); -#endif + if (!path_session.empty() && params.prompt_cache_all) { + fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str()); + llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size()); + } llama_print_timings(ctx); llama_free(ctx); - set_console_color(con_st, CONSOLE_COLOR_DEFAULT); - return 0; } diff --git a/examples/perplexity/CMakeLists.txt b/examples/perplexity/CMakeLists.txt index 5836df8b2..61b17b828 100644 --- a/examples/perplexity/CMakeLists.txt +++ b/examples/perplexity/CMakeLists.txt @@ -2,3 +2,6 @@ set(TARGET perplexity) add_executable(${TARGET} perplexity.cpp) 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/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 80792ea0d..9212dee5c 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1,5 +1,6 @@ #include "common.h" #include "llama.h" +#include "build-info.h" #include #include @@ -24,40 +25,68 @@ void perplexity(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]` + // BOS tokens will be added for each chunk before eval auto tokens = ::llama_tokenize(ctx, params.prompt, true); - int count = 0; - int seq_count = tokens.size() / params.n_ctx; - int n_vocab = llama_n_vocab(ctx); + int count = 0; + + const int n_chunk = tokens.size() / params.n_ctx; + const int n_vocab = llama_n_vocab(ctx); + const int n_batch = params.n_batch; double nll = 0.0; - fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch); + fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch); - for (int i = 0; i < seq_count; ++i) { - int start = i * params.n_ctx; - int end = start + params.n_ctx; + for (int i = 0; i < n_chunk; ++i) { + const int start = i * params.n_ctx; + const int end = start + params.n_ctx; + + const int num_batches = (params.n_ctx + n_batch - 1) / n_batch; std::vector logits; - int num_batches = (params.n_ctx + params.n_batch - 1) / params.n_batch; - auto start_t = std::chrono::high_resolution_clock::now(); + + const auto t_start = std::chrono::high_resolution_clock::now(); + for (int j = 0; j < num_batches; ++j) { - int batch_start = start + j * params.n_batch; - int batch_size = std::min(end - batch_start, params.n_batch); - if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * params.n_batch, params.n_threads)) { + const int batch_start = start + j * n_batch; + const int batch_size = std::min(end - batch_start, n_batch); + + // save original token and restore it after eval + const auto token_org = tokens[batch_start]; + + // add BOS token for the first batch of each chunk + if (j == 0) { + tokens[batch_start] = llama_token_bos(); + } + + if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) { fprintf(stderr, "%s : failed to eval\n", __func__); return; } - auto batch_logits = llama_get_logits(ctx); + + // restore the original token in case it was set to BOS + tokens[batch_start] = token_org; + + const auto batch_logits = llama_get_logits(ctx); logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); } - auto end_t = std::chrono::high_resolution_clock::now(); + + const auto t_end = std::chrono::high_resolution_clock::now(); + if (i == 0) { - const float seconds = std::chrono::duration(end_t - start_t).count(); - printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0)); + const float t_total = std::chrono::duration(t_end - t_start).count(); + fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total); + int total_seconds = (int)(t_total * n_chunk); + if (total_seconds >= 60*60) { + fprintf(stderr, "%d hours ", total_seconds / (60*60)); + total_seconds = total_seconds % (60*60); + } + fprintf(stderr, "%d minutes\n", total_seconds / 60); } + // We get the logits for all the tokens in the context window (params.n_ctx) // from llama_eval above. Now, based on https://huggingface.co/docs/transformers/perplexity, - // calculate the perplexity over the last half the window (so the model always has + // calculate the perplexity over the last half of the window (so the model always has // some context to predict the token). // // We rely on the fact that attention in the forward pass only looks at previous @@ -69,10 +98,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) { // process the entire prompt. for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) { // Calculate probability of next token, given the previous ones. - std::vector tok_logits( - logits.begin() + j * n_vocab, + const std::vector tok_logits( + logits.begin() + (j + 0) * n_vocab, logits.begin() + (j + 1) * n_vocab); - float prob = softmax(tok_logits)[tokens[start + j + 1]]; + + const float prob = softmax(tok_logits)[tokens[start + j + 1]]; + nll += -std::log(prob); ++count; } @@ -100,11 +131,13 @@ int main(int argc, char ** argv) { "expect poor results\n", __func__, params.n_ctx); } - if (params.seed <= 0) { + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + + if (params.seed < 0) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); + fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { @@ -113,36 +146,11 @@ int main(int argc, char ** argv) { llama_context * ctx; - // load the model - { - auto lparams = llama_context_default_params(); - - lparams.n_ctx = params.n_ctx; - lparams.n_parts = params.n_parts; - lparams.seed = params.seed; - lparams.f16_kv = params.memory_f16; - lparams.logits_all = params.perplexity; - lparams.use_mmap = params.use_mmap; - lparams.use_mlock = params.use_mlock; - lparams.embedding = params.embedding; - - ctx = llama_init_from_file(params.model.c_str(), lparams); - - if (ctx == NULL) { - fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str()); - return 1; - } - } - - if (!params.lora_adapter.empty()) { - int err = llama_apply_lora_from_file(ctx, - params.lora_adapter.c_str(), - params.lora_base.empty() ? NULL : params.lora_base.c_str(), - params.n_threads); - if (err != 0) { - fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); - return 1; - } + // load the model and apply lora adapter, if any + ctx = llama_init_from_gpt_params(params); + if (ctx == NULL) { + fprintf(stderr, "%s: error: unable to load model\n", __func__); + return 1; } // print system information diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index cd973e8ac..9a2aa7c64 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -1,4 +1,5 @@ #include "ggml.h" +#include "build-info.h" #define LLAMA_API_INTERNAL #include "llama.h" @@ -15,6 +16,8 @@ #include #include #include +#include +#include struct quantize_stats_params { std::string model = "models/7B/ggml-model-f16.bin"; @@ -27,7 +30,6 @@ struct quantize_stats_params { std::vector include_types; }; -const int64_t SCRATCH_ELEMENTS = 32*32; const size_t HISTOGRAM_BUCKETS = 150; const double HISTOGRAM_RANGE = 0.03; @@ -90,6 +92,13 @@ 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) { + 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 quantize_fns_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); + } + } else { + input_scratch = ggml_get_data_f32(layer) + offset; + } + + if (use_reference) { + qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size); + } else { + qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size); + } + qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size); + + update_error_stats(chunk_size, input_scratch, output_scratch, stats); +} + + // Run quantization function for a single layer and update error stats void test_roundtrip_on_layer( std::string & name, @@ -137,40 +176,61 @@ void test_roundtrip_on_layer( const quantize_fns_t & qfns, bool use_reference, const ggml_tensor * layer, - float * input_scratch, - char *quantized_scratch, - float * output_scratch, - error_stats & total_error) { + 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 {}; - int64_t nelements = ggml_nelements(layer); + uint64_t nelements = ggml_nelements(layer); - for (int64_t offset = 0; offset < nelements; offset += SCRATCH_ELEMENTS) { - int64_t chunk_size = std::min(SCRATCH_ELEMENTS, nelements - offset); - - if (layer->type == GGML_TYPE_F16) { - for (int i = 0; i < chunk_size; i++) { - input_scratch[i] = ggml_get_f32_1d(layer, i + offset); - } - } else { - input_scratch = ggml_get_data_f32(layer) + offset; - } - - if (use_reference) { - qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size); - } else { - qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size); - } - qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size); - - update_error_stats(chunk_size, input_scratch, output_scratch, total_error); - if (print_layer_stats) { - update_error_stats(chunk_size, input_scratch, output_scratch, layer_error); - } + float* input_scratch_ptr = nullptr; + if (layer->type == GGML_TYPE_F16) { + if (input_scratch.size() < nelements) input_scratch.resize(nelements); + input_scratch_ptr = input_scratch.data(); } + if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(4*nelements); + if (output_scratch.size() < nelements) output_scratch.resize(nelements); + + if (max_thread < 1) max_thread = std::thread::hardware_concurrency(); + int chunk_size = 32*512; + int num_chunks = (nelements + chunk_size - 1)/chunk_size; + + if (num_chunks < 2 || max_thread < 2) { + test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(), + output_scratch.data(), print_layer_stats ? layer_error : total_error); + } else { + auto & stats = print_layer_stats ? layer_error : total_error; + std::mutex mutex; + uint64_t counter = 0; + auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr, + &quantized_scratch, &output_scratch, chunk_size] () { + error_stats local_stats {}; + while (true) { + std::unique_lock lock(mutex); + uint64_t offset = counter; counter += chunk_size; + if (offset >= nelements) { + combine_error_stats(stats, local_stats); + break; + } + lock.unlock(); + uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset; + test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset, + quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats); + } + }; + int nthread = std::min(num_chunks, max_thread); + std::vector workers(nthread-1); + for (auto& w : workers) w = std::thread(compute); + compute(); + for (auto& w : workers) w.join(); + } + if (print_layer_stats) { print_error_stats(name, layer_error, false); + combine_error_stats(total_error, layer_error); } } @@ -181,6 +241,7 @@ int main(int argc, char ** argv) { // read command line + int max_thread = 0; bool invalid_param = false; std::string arg; for (int i = 1; i < argc; i++) { @@ -230,6 +291,12 @@ int main(int argc, char ** argv) { fprintf(stderr, "error: %s not in list of types\n", argv[i]); invalid_param = true; } + } else if (arg == "-n" || arg == "--num-threads") { + if (++i >= argc) { + invalid_param = true; + break; + } + max_thread = atoi(argv[i]); } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); quantize_stats_print_usage(argc, argv); @@ -242,6 +309,8 @@ int main(int argc, char ** argv) { return 1; } + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + // load the model fprintf(stderr, "Loading model\n"); @@ -295,9 +364,9 @@ int main(int argc, char ** argv) { } printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements); // allocate scratch space - std::vector input_scratch(SCRATCH_ELEMENTS); - std::vector quantized_scratch(SCRATCH_ELEMENTS*4); - std::vector output_scratch(SCRATCH_ELEMENTS); + std::vector input_scratch; + std::vector quantized_scratch; + std::vector output_scratch; // loop throught quantization types for (int i = 0; i < GGML_TYPE_COUNT; i++) { @@ -328,10 +397,11 @@ int main(int argc, char ** argv) { qfns, params.reference, kv_tensor.second, - input_scratch.data(), - quantized_scratch.data(), - output_scratch.data(), - global_stats + input_scratch, + quantized_scratch, + output_scratch, + global_stats, + max_thread ); } diff --git a/examples/quantize/CMakeLists.txt b/examples/quantize/CMakeLists.txt index fb27d4517..475fc8be8 100644 --- a/examples/quantize/CMakeLists.txt +++ b/examples/quantize/CMakeLists.txt @@ -2,3 +2,6 @@ set(TARGET quantize) add_executable(${TARGET} quantize.cpp) target_link_libraries(${TARGET} PRIVATE 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/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 59cb67440..115d8fb1b 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -1,20 +1,54 @@ #include "ggml.h" #include "llama.h" +#include "build-info.h" #include +#include #include +static const std::map LLAMA_FTYPE_MAP = { + {"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0}, + {"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1}, + {"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0}, + {"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1}, + {"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0}, +}; + +bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) { + auto it = LLAMA_FTYPE_MAP.find(ftype_str); + if (it != LLAMA_FTYPE_MAP.end()) { + ftype = it->second; + ftype_str_out = it->first; + return true; + } + // try to parse as an integer + try { + int ftype_int = std::stoi(ftype_str); + for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) { + if (it->second == ftype_int) { + ftype = it->second; + ftype_str_out = it->first; + return true; + } + } + } + catch (...) { + // stoi failed + } + return false; +} + // usage: -// ./quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type +// ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads] // int main(int argc, char ** argv) { ggml_time_init(); - if (argc != 4) { - fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]); - fprintf(stderr, " type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0); - fprintf(stderr, " type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1); - fprintf(stderr, " type = %d - q4_2\n", LLAMA_FTYPE_MOSTLY_Q4_2); + if (argc < 3) { + fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]); + for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) { + fprintf(stderr, " type = \"%s\" or %d\n", it->first.c_str(), it->second); + } return 1; } @@ -25,10 +59,62 @@ int main(int argc, char ** argv) { ggml_free(ctx); } + // parse command line arguments const std::string fname_inp = argv[1]; - const std::string fname_out = argv[2]; + std::string fname_out; + int nthread; + llama_ftype ftype; - const enum llama_ftype ftype = (enum llama_ftype)atoi(argv[3]); + int arg_idx = 2; + std::string ftype_str; + if (try_parse_ftype(argv[arg_idx], ftype, ftype_str)) { + // argv[2] is the ftype + std::string fpath; + const size_t pos = fname_inp.find_last_of('/'); + if (pos != std::string::npos) { + fpath = fname_inp.substr(0, pos + 1); + } + // export as [inp path]/ggml-model-[ftype].bin + fname_out = fpath + "ggml-model-" + ftype_str + ".bin"; + arg_idx++; + } + else { + // argv[2] is the output path + fname_out = argv[arg_idx]; + arg_idx++; + + if (argc <= arg_idx) { + fprintf(stderr, "%s: missing ftype\n", __func__); + return 1; + } + // argv[3] is the ftype + if (!try_parse_ftype(argv[arg_idx], ftype, ftype_str)) { + fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]); + return 1; + } + arg_idx++; + } + + // parse nthreads + if (argc > arg_idx) { + try { + nthread = std::stoi(argv[arg_idx]); + } + catch (const std::exception & e) { + fprintf(stderr, "%s: invalid nthread '%s' (%s)\n", __func__, argv[arg_idx], e.what()); + return 1; + } + } else { + nthread = 0; + } + + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + + fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str()); + if (nthread > 0) { + fprintf(stderr, " using %d threads", nthread); + } + fprintf(stderr, "\n"); const int64_t t_main_start_us = ggml_time_us(); @@ -38,7 +124,7 @@ int main(int argc, char ** argv) { { const int64_t t_start_us = ggml_time_us(); - if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype)) { + if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) { fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str()); return 1; } diff --git a/examples/save-load-state/CMakeLists.txt b/examples/save-load-state/CMakeLists.txt new file mode 100644 index 000000000..08dbe5c2b --- /dev/null +++ b/examples/save-load-state/CMakeLists.txt @@ -0,0 +1,7 @@ +set(TARGET save-load-state) +add_executable(${TARGET} save-load-state.cpp) +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/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp new file mode 100644 index 000000000..ea0a984d9 --- /dev/null +++ b/examples/save-load-state/save-load-state.cpp @@ -0,0 +1,151 @@ +#include "common.h" +#include "llama.h" +#include "build-info.h" + +#include +#include +#include + +int main(int argc, char ** argv) { + gpt_params params; + params.model = "models/llama-7B/ggml-model.bin"; + params.seed = 42; + params.n_threads = 4; + params.repeat_last_n = 64; + params.prompt = "The quick brown fox"; + + if (gpt_params_parse(argc, argv, params) == false) { + return 1; + } + + fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + + if (params.n_predict < 0) { + params.n_predict = 16; + } + + auto lparams = llama_context_default_params(); + + lparams.n_ctx = params.n_ctx; + lparams.n_parts = params.n_parts; + lparams.seed = params.seed; + lparams.f16_kv = params.memory_f16; + lparams.use_mmap = params.use_mmap; + lparams.use_mlock = params.use_mlock; + + auto n_past = 0; + auto last_n_tokens_data = std::vector(params.repeat_last_n, 0); + + // init + auto ctx = llama_init_from_file(params.model.c_str(), lparams); + auto tokens = std::vector(params.n_ctx); + auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true); + + if (n_prompt_tokens < 1) { + fprintf(stderr, "%s : failed to tokenize prompt\n", __func__); + return 1; + } + + // evaluate prompt + llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads); + + last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens); + n_past += n_prompt_tokens; + + const size_t state_size = llama_get_state_size(ctx); + uint8_t * state_mem = new uint8_t[state_size]; + + // Save state (rng, logits, embedding and kv_cache) to file + { + FILE *fp_write = fopen("dump_state.bin", "wb"); + llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file + fwrite(state_mem, 1, state_size, fp_write); + fclose(fp_write); + } + + // save state (last tokens) + const auto last_n_tokens_data_saved = std::vector(last_n_tokens_data); + const auto n_past_saved = n_past; + + // first run + printf("\n%s", params.prompt.c_str()); + + for (auto i = 0; i < params.n_predict; i++) { + auto logits = llama_get_logits(ctx); + auto n_vocab = llama_n_vocab(ctx); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < n_vocab; token_id++) { + candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + } + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + auto next_token = llama_sample_token(ctx, &candidates_p); + auto next_token_str = llama_token_to_str(ctx, next_token); + last_n_tokens_data.push_back(next_token); + + printf("%s", next_token_str); + if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) { + fprintf(stderr, "\n%s : failed to evaluate\n", __func__); + return 1; + } + n_past += 1; + } + + printf("\n\n"); + + // free old model + llama_free(ctx); + + // load new model + auto ctx2 = llama_init_from_file(params.model.c_str(), lparams); + + // Load state (rng, logits, embedding and kv_cache) from file + { + FILE *fp_read = fopen("dump_state.bin", "rb"); + if (state_size != llama_get_state_size(ctx2)) { + fprintf(stderr, "\n%s : failed to validate state size\n", __func__); + return 1; + } + + const size_t ret = fread(state_mem, 1, state_size, fp_read); + if (ret != state_size) { + fprintf(stderr, "\n%s : failed to read state\n", __func__); + return 1; + } + + llama_set_state_data(ctx2, state_mem); // could also read directly from memory mapped file + fclose(fp_read); + } + + delete[] state_mem; + + // restore state (last tokens) + last_n_tokens_data = last_n_tokens_data_saved; + n_past = n_past_saved; + + // second run + for (auto i = 0; i < params.n_predict; i++) { + auto logits = llama_get_logits(ctx2); + auto n_vocab = llama_n_vocab(ctx2); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < n_vocab; token_id++) { + candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + } + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + auto next_token = llama_sample_token(ctx2, &candidates_p); + auto next_token_str = llama_token_to_str(ctx2, next_token); + last_n_tokens_data.push_back(next_token); + + printf("%s", next_token_str); + if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) { + fprintf(stderr, "\n%s : failed to evaluate\n", __func__); + return 1; + } + n_past += 1; + } + + printf("\n\n"); + + return 0; +} diff --git a/flake.nix b/flake.nix index 5363052b1..2c9edbb6a 100644 --- a/flake.nix +++ b/flake.nix @@ -30,9 +30,9 @@ mv bin/* $out/bin/ mv $out/bin/main $out/bin/llama - echo "#!${llama-python}/bin/python" > $out/bin/convert-pth-to-ggml - cat ${./convert-pth-to-ggml.py} >> $out/bin/convert-pth-to-ggml - chmod +x $out/bin/convert-pth-to-ggml + echo "#!${llama-python}/bin/python" > $out/bin/convert.py + cat ${./convert.py} >> $out/bin/convert.py + chmod +x $out/bin/convert.py ''; meta.mainProgram = "llama"; }; diff --git a/ggml-cuda.cu b/ggml-cuda.cu new file mode 100644 index 000000000..b6a7754d5 --- /dev/null +++ b/ggml-cuda.cu @@ -0,0 +1,898 @@ +#include +#include +#include +#include +#include + +#include +#include +#include + +#include "ggml-cuda.h" +#include "ggml.h" + +static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); + +#define CUDA_CHECK(err) \ + do { \ + cudaError_t err_ = (err); \ + if (err_ != cudaSuccess) { \ + fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \ + cudaGetErrorString(err_)); \ + exit(1); \ + } \ + } while (0) + +#define CUBLAS_CHECK(err) \ + do { \ + cublasStatus_t err_ = (err); \ + if (err_ != CUBLAS_STATUS_SUCCESS) { \ + fprintf(stderr, "cuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \ + exit(1); \ + } \ + } while (0) + +typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1); +typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream); +typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream); + +// QK = number of values after dequantization +// QR = QK / number of values before dequantization + +#define QK4_0 32 +#define QR4_0 2 +typedef struct { + float d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +#define QR4_1 2 +typedef struct { + float d; // delta + float m; // min + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); + +#define QK5_0 32 +#define QR5_0 2 +typedef struct { + half d; // delta + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_0 / 2]; // nibbles / quants +} block_q5_0; +static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); + +#define QK5_1 32 +#define QR5_1 2 +typedef struct { + half d; // delta + half m; // min + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_1 / 2]; // nibbles / quants +} block_q5_1; +static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); + +#define QK8_0 32 +#define QR8_0 1 +typedef struct { + float d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); + +#define CUDA_DMMV_BLOCK_SIZE 32 + +static __device__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q4_0 * x = (const block_q4_0 *) vx; + + const float d = x[ib].d; + + const uint8_t vui = x[ib].qs[iqs]; + + const int8_t vi0 = vui & 0xF; + const int8_t vi1 = vui >> 4; + + v0 = (vi0 - 8)*d; + v1 = (vi1 - 8)*d; +} + +static __device__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q4_1 * x = (const block_q4_1 *) vx; + + const float d = x[ib].d; + const float m = x[ib].m; + + const uint8_t vui = x[ib].qs[iqs]; + + const int8_t vi0 = vui & 0xF; + const int8_t vi1 = vui >> 4; + + v0 = vi0*d + m; + v1 = vi1*d + m; +} + +static __device__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q5_0 * x = (const block_q5_0 *) vx; + + const float d = x[ib].d; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16; + const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16; + + v0 = x0*d; + v1 = x1*d; +} + +static __device__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q5_1 * x = (const block_q5_1 *) vx; + + const float d = x[ib].d; + const float m = x[ib].m; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0); + const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1); + + v0 = x0*d + m; + v1 = x1*d + m; +} + +static __device__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q8_0 * x = (const block_q8_0 *) vx; + + const float d = x[ib].d; + + const int8_t vi0 = x[ib].qs[iqs + 0]; + const int8_t vi1 = x[ib].qs[iqs + 1]; + + v0 = vi0*d; + v1 = vi1*d; +} + +static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const half * x = (const half *) vx; + + v0 = __half2float(x[ib + 0]); + v1 = __half2float(x[ib + 1]); +} + +static __global__ void dequantize_block_q4_0(const void * vx, float * y) { + static const int qk = QK4_0; + + const block_q4_0 * x = (const block_q4_0 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0xf) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; + } +} + +static __global__ void dequantize_block_q4_1(const void * vx, float * y) { + static const int qk = QK4_1; + + const block_q4_1 * x = (const block_q4_1 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + const float m = x[i].m; + + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0xf); + const int x1 = (x[i].qs[j] >> 4); + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; + } +} + +static __global__ void dequantize_block_q5_0(const void * vx, float * y) { + static const int qk = QK5_0; + + const block_q5_0 * x = (const block_q5_0 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; + } +} + +static __global__ void dequantize_block_q5_1(const void * vx, float * y) { + static const int qk = QK5_1; + + const block_q5_1 * x = (const block_q5_1 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + const float m = x[i].m; + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int x0 = (x[i].qs[j] & 0xf) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; + } +} + +static __global__ void dequantize_block_q8_0(const void * vx, float * y) { + static const int qk = QK8_0; + + const block_q8_0 * x = (const block_q8_0 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + + for (int j = 0; j < qk; ++j) { + y[i*qk + j] = x[i].qs[j]*d; + } +} + +template +static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols) { + const int row = blockIdx.x; + const int tid = threadIdx.x; + + const int y_offset = qr == 1 ? 1 : qk/2; + + __shared__ float tmp[block_size]; // separate sum for each thread + tmp[tid] = 0; + + for (int i = 0; i < ncols/block_size; i += 2) { + const int col = i*block_size + 2*tid; + const int ib = (row*ncols + col)/qk; // block index + const int iqs = (col%qk)/qr; // quant index + const int iybs = col - col%qk; // y block start index + + // dequantize + float v0, v1; + dequantize_kernel(vx, ib, iqs, v0, v1); + + // matrix multiplication + tmp[tid] += v0 * y[iybs + iqs + 0]; + tmp[tid] += v1 * y[iybs + iqs + y_offset]; + } + + // sum up partial sums and write back result + __syncthreads(); + for (int s=block_size/2; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + __syncthreads(); + } + if (tid == 0) { + dst[row] = tmp[0]; + } +} + +static void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK4_0; + dequantize_block_q4_0<<>>(vx, y); +} + +static void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK4_1; + dequantize_block_q4_1<<>>(vx, y); +} + +static void dequantize_row_q5_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK5_0; + dequantize_block_q5_0<<>>(vx, y); +} + +static void dequantize_row_q5_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK5_1; + dequantize_block_q5_1<<>>(vx, y); +} + +static void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK8_0; + dequantize_block_q8_0<<>>(vx, y); +} + +static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +// TODO: optimize +static __global__ void convert_fp16_to_fp32(const void * vx, float * y) { + const half * x = (const half *) vx; + + const int i = blockIdx.x; + + y[i] = __half2float(x[i]); +} + +static void convert_fp16_to_fp32_cuda(const void * x, float * y, int k, cudaStream_t stream) { + convert_fp16_to_fp32<<>>(x, y); +} + +static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_row_q4_0_cuda; + case GGML_TYPE_Q4_1: + return dequantize_row_q4_1_cuda; + case GGML_TYPE_Q5_0: + return dequantize_row_q5_0_cuda; + case GGML_TYPE_Q5_1: + return dequantize_row_q5_1_cuda; + case GGML_TYPE_Q8_0: + return dequantize_row_q8_0_cuda; + case GGML_TYPE_F16: + return convert_fp16_to_fp32_cuda; + default: + return nullptr; + } +} + +static dequantize_mul_mat_vec_cuda_t ggml_get_dequantize_mul_mat_vec_cuda(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_mul_mat_vec_q4_0_cuda; + case GGML_TYPE_Q4_1: + return dequantize_mul_mat_vec_q4_1_cuda; + case GGML_TYPE_Q5_0: + return dequantize_mul_mat_vec_q5_0_cuda; + case GGML_TYPE_Q5_1: + return dequantize_mul_mat_vec_q5_1_cuda; + case GGML_TYPE_Q8_0: + return dequantize_mul_mat_vec_q8_0_cuda; + case GGML_TYPE_F16: + return dequantize_mul_mat_vec_q8_0_cuda; + default: + return nullptr; + } +} + +// buffer pool for cuda +#define MAX_CUDA_BUFFERS 256 + +struct scoped_spin_lock { + std::atomic_flag& lock; + scoped_spin_lock(std::atomic_flag& lock) : lock(lock) { + while (lock.test_and_set(std::memory_order_acquire)) { + ; // spin + } + } + ~scoped_spin_lock() { + lock.clear(std::memory_order_release); + } + scoped_spin_lock(const scoped_spin_lock&) = delete; + scoped_spin_lock& operator=(const scoped_spin_lock&) = delete; +}; + +struct cuda_buffer { + void * ptr = nullptr; + size_t size = 0; +}; + +static cuda_buffer g_cuda_buffer_pool[MAX_CUDA_BUFFERS]; +static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT; + +static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) { + scoped_spin_lock lock(g_cuda_pool_lock); + + for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) { + cuda_buffer& b = g_cuda_buffer_pool[i]; + if (b.size >= size && b.ptr != nullptr) { + void * ptr = b.ptr; + *actual_size = b.size; + b.ptr = nullptr; + b.size = 0; + return ptr; + } + } + void * ptr; + CUDA_CHECK(cudaMalloc((void **) &ptr, size)); + *actual_size = size; + return ptr; +} + +static void ggml_cuda_pool_free(void * ptr, size_t size) { + scoped_spin_lock lock(g_cuda_pool_lock); + + for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) { + cuda_buffer& b = g_cuda_buffer_pool[i]; + if (b.ptr == nullptr) { + b.ptr = ptr; + b.size = size; + return; + } + } + fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n"); + CUDA_CHECK(cudaFree(ptr)); +} + +#define GGML_CUDA_MAX_STREAMS 8 // Set this to 1 for reproducible matrix multiplication. +#define GGML_CUDA_MAX_EVENTS 64 +static cublasHandle_t g_cublasH = nullptr; +static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_STREAMS] = { nullptr }; +static cudaStream_t g_cudaStreams2[GGML_CUDA_MAX_STREAMS] = { nullptr }; +static cudaEvent_t g_cudaEvents[GGML_CUDA_MAX_EVENTS] = { nullptr }; + +void ggml_init_cublas() { + if (g_cublasH == nullptr) { + // create streams + for (int i = 0; i < GGML_CUDA_MAX_STREAMS; ++i) { + CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[i], cudaStreamNonBlocking)); + CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams2[i], cudaStreamNonBlocking)); + } + // create events + for (int i = 0; i < GGML_CUDA_MAX_EVENTS; ++i) { + CUDA_CHECK(cudaEventCreateWithFlags(&g_cudaEvents[i], cudaEventDisableTiming)); + } + + // create cublas handle + CUBLAS_CHECK(cublasCreate(&g_cublasH)); + CUBLAS_CHECK(cublasSetMathMode(g_cublasH, CUBLAS_TF32_TENSOR_OP_MATH)); + + // configure logging to stdout + // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr)); + } +} + +void * ggml_cuda_host_malloc(size_t size) { + if (getenv("GGML_CUDA_NO_PINNED") != nullptr) { + return nullptr; + } + + void * ptr = nullptr; + cudaError_t err = cudaMallocHost((void **) &ptr, size); + if (err != cudaSuccess) { + fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n", + size/1024.0/1024.0, cudaGetErrorString(err)); + return nullptr; + } + + return ptr; +} + +void ggml_cuda_host_free(void * ptr) { + CUDA_CHECK(cudaFreeHost(ptr)); +} + +static cudaError_t ggml_cuda_h2d_tensor_2d(void * dst, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cudaStream_t stream) { + const uint64_t ne0 = src->ne[0]; + const uint64_t ne1 = src->ne[1]; + const uint64_t nb0 = src->nb[0]; + const uint64_t nb1 = src->nb[1]; + const uint64_t nb2 = src->nb[2]; + const uint64_t nb3 = src->nb[3]; + const enum ggml_type type = src->type; + const size_t ts = ggml_type_size(type); + const size_t bs = ggml_blck_size(type); + + const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3); + if (nb0 == ts && nb1 == ts*ne0/bs) { + return cudaMemcpyAsync(dst, x, ne1*nb1, cudaMemcpyHostToDevice, stream); + } else if (nb0 == ts) { + return cudaMemcpy2DAsync(dst, ts*ne0/bs, x, nb1, ts*ne0/bs, ne1, cudaMemcpyHostToDevice, stream); + } else { + for (uint64_t i1 = 0; i1 < ne1; i1++) { + const void * rx = (const void *) ((const char *) x + i1*nb1); + void * rd = (void *) ((char *) dst + i1*ts*ne0/bs); + // pretend the row is a matrix with cols=1 + cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyHostToDevice, stream); + if (r != cudaSuccess) return r; + } + return cudaSuccess; + } +} + +static void ggml_cuda_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + 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 ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + + const float alpha = 1.0f; + const float beta = 0.0f; + const int x_ne = ne01 * ne00; + const int y_ne = ne11 * ne10; + const int d_ne = ne11 * ne01; + const int n_mm = ne03 * ne02; + + size_t x_size, y_size, d_size; + float * d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size); + float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size); + float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + int i = i03*ne02 + i02; + cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS]; + + float * c_X = d_X + i * x_ne; + float * c_Y = d_Y + i * y_ne; + float * c_D = d_D + i * d_ne; + + // copy data to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, cudaStream)); + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream)); + + // compute + CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream)); + CUBLAS_CHECK( + cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, c_X, ne00, + c_Y, ne10, + &beta, c_D, ne01)); + + // copy dst to host + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream)); + } + } + + CUDA_CHECK(cudaDeviceSynchronize()); + ggml_cuda_pool_free(d_X, x_size); + ggml_cuda_pool_free(d_Y, y_size); + ggml_cuda_pool_free(d_D, d_size); +} + +static void ggml_cuda_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) { + 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 ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + + const int nb10 = src1->nb[0]; + const int nb11 = src1->nb[1]; + const int nb12 = src1->nb[2]; + const int nb13 = src1->nb[3]; + + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + + const float alpha = 1.0f; + const float beta = 0.0f; + const int x_ne = ne01 * ne00; + const int y_ne = ne11 * ne10; + const int d_ne = ne11 * ne01; + const int n_mm = ne03 * ne02; + + size_t x_size, y_size, d_size; + half * d_X = (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * x_ne, &x_size); + half * d_Y = (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * y_ne, &y_size); + float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size); + + bool src1_cont_rows = nb10 == sizeof(float); + bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + int i = i03*ne02 + i02; + cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS]; + + half * c_X = d_X + i * x_ne; + half * c_Y = d_Y + i * y_ne; + float * c_D = d_D + i * d_ne; + + // copy src0 to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, cudaStream)); + + // convert src1 to fp16 + // TODO: use multiple threads + ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02); + char * src1i = (char *) src1->data + i03*nb13 + i02*nb12; + if (src1_cont_rows) { + if (src1_cont_cols) { + ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11); + } + else { + for (int64_t i01 = 0; i01 < ne11; i01++) { + ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10); + } + } + } + else { + for (int64_t i01 = 0; i01 < ne11; i01++) { + for (int64_t i00 = 0; i00 < ne10; i00++) { + // very slow due to no inlining + tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10)); + } + } + } + + // copy src1 to device + CUDA_CHECK(cudaMemcpyAsync(c_Y, tmp, sizeof(half) * y_ne, cudaMemcpyHostToDevice, cudaStream)); + + // compute + CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream)); + CUBLAS_CHECK( + cublasGemmEx(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, c_X, CUDA_R_16F, ne00, + c_Y, CUDA_R_16F, ne10, + &beta, c_D, CUDA_R_32F, ne01, + CUBLAS_COMPUTE_32F_FAST_16F, + CUBLAS_GEMM_DEFAULT)); + + // copy dst to host + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream)); + } + } + + CUDA_CHECK(cudaDeviceSynchronize()); + ggml_cuda_pool_free(d_X, x_size); + ggml_cuda_pool_free(d_Y, y_size); + ggml_cuda_pool_free(d_D, d_size); +} + +static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + 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 ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + const ggml_type type = src0->type; + const bool mul_mat_vec = ne11 == 1; + + const float alpha = 1.0f; + const float beta = 0.0f; + const int x_ne = ne01 * ne00; + const int y_ne = ne11 * ne10; + const int d_ne = ne11 * ne01; + const int n_mm = ne03 * ne02; + const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type); + + size_t x_size, y_size, d_size, q_size; + float * d_X = nullptr; + if (!mul_mat_vec) { + d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size); + } + float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size); + float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size); + char * d_Q = (char *) ggml_cuda_pool_malloc(n_mm * q_sz, &q_size); + + const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(type); + dequantize_mul_mat_vec_cuda_t dmmv = ggml_get_dequantize_mul_mat_vec_cuda(type); + GGML_ASSERT(to_fp32_cuda != nullptr); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + int i = i03*ne02 + i02; + cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS]; + cudaStream_t cudaStream2 = g_cudaStreams2[i % GGML_CUDA_MAX_STREAMS]; + cudaEvent_t cudaEvent = g_cudaEvents[i % GGML_CUDA_MAX_EVENTS]; + + float * c_Y = d_Y + i * y_ne; + float * c_D = d_D + i * d_ne; + char * c_Q = d_Q + i * q_sz; + + // copy src0 to device if necessary + if (src0->backend == GGML_BACKEND_CPU) { + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2)); + } else if (src0->backend == GGML_BACKEND_CUDA) { + c_Q = ((char *) src0->data) + i * q_sz; + } else { + GGML_ASSERT(false); + } + if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel + CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2)); + + // copy src1 to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream)); + + // wait for data + CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0)); + + // compute + dmmv(c_Q, c_Y, c_D, ne00, ne01, cudaStream); + CUDA_CHECK(cudaGetLastError()); + + } else { // general dequantization kernel + cuBLAS matrix matrix multiplication + float * c_X = d_X + i * x_ne; + + // convert src0 to fp32 on device + to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2); + CUDA_CHECK(cudaGetLastError()); + CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2)); + + // copy src1 to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream)); + + // wait for conversion + CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0)); + + // compute + CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream)); + CUBLAS_CHECK( + cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, c_X, ne00, + c_Y, ne10, + &beta, c_D, ne01)); + } + + // copy dst to host + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream)); + } + } + + CUDA_CHECK(cudaDeviceSynchronize()); + if (!mul_mat_vec) { + ggml_cuda_pool_free(d_X, x_size); + } + ggml_cuda_pool_free(d_Y, y_size); + ggml_cuda_pool_free(d_D, d_size); + ggml_cuda_pool_free(d_Q, q_size); +} + +bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + const int64_t ne10 = src1->ne[0]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + + // TODO: find the optimal values for these + if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && + src1->type == GGML_TYPE_F32 && + dst->type == GGML_TYPE_F32 && + ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CUDA)) { + return true; + } + + return false; +} + +bool ggml_cuda_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) { + size_t src0_sz = ggml_nbytes(src0); + size_t src1_sz = ggml_nbytes(src1); + + // mul_mat_q: src0 is converted to fp32 on device + size_t mul_mat_q_transfer = src0_sz + src1_sz; + + // mul_mat_f16: src1 is converted to fp16 on cpu + size_t mul_mat_f16_transfer = src0_sz + sizeof(half) * ggml_nelements(src1); + + // choose the smaller one to transfer to the device + // TODO: this is not always the best choice due to the overhead of converting to fp16 + return mul_mat_f16_transfer < mul_mat_q_transfer; +} + +void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) { + GGML_ASSERT(ggml_cuda_can_mul_mat(src0, src1, dst)); + + if (src0->type == GGML_TYPE_F32) { + ggml_cuda_mul_mat_f32(src0, src1, dst); + } + else if (src0->type == GGML_TYPE_F16) { + if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) { + ggml_cuda_mul_mat_f16(src0, src1, dst, wdata, wsize); + } + else { + ggml_cuda_mul_mat_q_f32(src0, src1, dst); + } + } + else if (ggml_is_quantized(src0->type)) { + ggml_cuda_mul_mat_q_f32(src0, src1, dst); + } + else { + GGML_ASSERT(false); + } +} + +size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) { + return ggml_nelements(src1) * sizeof(ggml_fp16_t); + } + else { + return 0; + } +} + +void ggml_cuda_transform_tensor(ggml_tensor * tensor) { + const int64_t ne0 = tensor->ne[0]; + const int64_t ne1 = tensor->ne[1]; + const int64_t ne2 = tensor->ne[2]; + const int64_t ne3 = tensor->ne[3]; + + const ggml_type type = tensor->type; + const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type); + + size_t q_size; + char * d_Q = (char *) ggml_cuda_pool_malloc(q_sz, &q_size); + + cudaStream_t cudaStream2 = g_cudaStreams2[0]; + + // copy tensor to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_Q, tensor, 0, 0, cudaStream2)); + CUDA_CHECK(cudaDeviceSynchronize()); + + tensor->data = d_Q; + tensor->backend = GGML_BACKEND_CUDA; +} diff --git a/ggml-cuda.h b/ggml-cuda.h new file mode 100644 index 000000000..4e2c24283 --- /dev/null +++ b/ggml-cuda.h @@ -0,0 +1,21 @@ +#include "ggml.h" + +#ifdef __cplusplus +extern "C" { +#endif + +void ggml_init_cublas(void); + +bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +void ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize); + +// TODO: export these with GGML_API +void * ggml_cuda_host_malloc(size_t size); +void ggml_cuda_host_free(void * ptr); + +void ggml_cuda_transform_tensor(struct ggml_tensor * tensor); + +#ifdef __cplusplus +} +#endif diff --git a/ggml-opencl.c b/ggml-opencl.c new file mode 100644 index 000000000..31ab13b25 --- /dev/null +++ b/ggml-opencl.c @@ -0,0 +1,361 @@ +#include "ggml-opencl.h" + +#define CL_TARGET_OPENCL_VERSION 110 +#include + +#include +#include +#include + +#include "ggml.h" + +#define MULTILINE_QUOTE(...) #__VA_ARGS__ +const char * clblast_dequant = MULTILINE_QUOTE( + +typedef uchar uint8_t; +typedef int int32_t; +typedef uint uint32_t; + +constant uint QK4_0 = 32; +struct block_q4_0 +{ + float d; + uint8_t qs[QK4_0 / 2]; +}; + +constant uint QK4_1 = 32; +struct block_q4_1 +{ + float d; + float m; + uint8_t qs[QK4_1 / 2]; +}; + +constant uint QK5_0 = 32; +struct __attribute__ ((packed)) block_q5_0 +{ + half d; + uint32_t qh; + uint8_t qs[QK5_0 / 2]; +}; + +constant uint QK5_1 = 32; +struct block_q5_1 +{ + half d; + half m; + uint32_t qh; + uint8_t qs[QK5_1 / 2]; +}; + +constant uint QK8_0 = 32; +struct block_q8_0 +{ + float d; + uint8_t qs[QK8_0]; +}; + + +__kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float* y) { + constant uint qk = QK4_0; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = x[i].d; + + const int x0 = (x[i].qs[j] & 0xf) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; +} + +__kernel void dequantize_row_q4_1(__global struct block_q4_1* x, __global float* y) { + constant uint qk = QK4_1; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = x[i].d; + const float m = x[i].m; + + const int x0 = (x[i].qs[j] & 0xf); + const int x1 = (x[i].qs[j] >> 4); + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; +} + +__kernel void dequantize_row_q5_0(__global struct block_q5_0* x, __global float* y) { + constant uint qk = QK5_0; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = vload_half(0, (__global half*) &x[i].d); + + uint32_t qh = x[i].qh; + + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; +} + +__kernel void dequantize_row_q5_1(__global struct block_q5_1* x, __global float* y) { + constant uint qk = QK5_1; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = vload_half(0, (__global half*) &x[i].d); + const float m = vload_half(0, (__global half*) &x[i].m); + + uint32_t qh = x[i].qh; + + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int x0 = (x[i].qs[j] & 0xf) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; +} + +__kernel void dequantize_row_q8_0(__global struct block_q8_0* x, __global float* y) { + constant uint qk = QK8_0; + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = x[i].d; + y[i*qk + j] = x[i].qs[j]*d; +} + +); + +#define CL_CHECK(err, name) \ + do { \ + cl_int err_ = (err); \ + if (err_ != CL_SUCCESS) { \ + fprintf(stderr, "OpenCL %s error %d at %s:%d\n", name, err_, __FILE__, __LINE__); \ + exit(1); \ + } \ + } while (0) + +static cl_platform_id platform; +static cl_device_id device; +static cl_context context; +static cl_command_queue queue; +static cl_program program; +static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q5_0, kernel_q5_1, kernel_q8_0; +static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c; +static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0; + +static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) { + cl_program p; + char *program_log; + size_t program_size, log_size; + int err; + + program_size = strlen(program_buffer); + + p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err); + if(err < 0) { + fprintf(stderr, "OpenCL error creating program"); + exit(1); + } + + err = clBuildProgram(p, 0, NULL, NULL, NULL, NULL); + if(err < 0) { + + clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size); + program_log = (char*) malloc(log_size + 1); + program_log[log_size] = '\0'; + clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL); + printf("%s\n", program_log); + free(program_log); + exit(1); + } + + return p; +} + +void ggml_cl_init(void) { + cl_int err = 0; + char * GGML_CLBLAST_PLATFORM = getenv("GGML_CLBLAST_PLATFORM"); + char * GGML_CLBLAST_DEVICE = getenv("GGML_CLBLAST_DEVICE"); + int plat_num = (GGML_CLBLAST_PLATFORM == NULL ? 0 : atoi(GGML_CLBLAST_PLATFORM)); + int dev_num = (GGML_CLBLAST_DEVICE == NULL ? 0 : atoi(GGML_CLBLAST_DEVICE)); + printf("\nInitializing CLBlast (First Run)..."); + printf("\nAttempting to use: Platform=%d, Device=%d (If invalid, program will crash)\n",plat_num,dev_num); + cl_uint num_platforms; + clGetPlatformIDs(0, NULL, &num_platforms); + cl_platform_id* platforms = (cl_platform_id*)malloc(num_platforms*sizeof(cl_platform_id)); + clGetPlatformIDs(num_platforms, platforms, NULL); + platform = platforms[plat_num]; + char platform_buffer[1024]; + clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(platform_buffer), &platform_buffer, NULL); + cl_uint num_devices; + clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 0, NULL, &num_devices); + cl_device_id* devices = (cl_device_id*)malloc(num_devices*sizeof(cl_device_id)); + clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, num_devices, devices, NULL); + device = devices[dev_num]; + char device_buffer[1024]; + clGetDeviceInfo(device, CL_DEVICE_NAME, sizeof(device_buffer), &device_buffer, NULL); + printf("Using Platform: %s Device: %s\n", platform_buffer, device_buffer); + context = clCreateContext(NULL, 1, &device, NULL, NULL, &err); + CL_CHECK(err, "clCreateContext"); + queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err); + CL_CHECK(err, "clCreateCommandQueue"); + + free(platforms); + free(devices); + + program = build_program_from_source(context, device, clblast_dequant); + + // Prepare dequantize kernels + kernel_q4_0 = clCreateKernel(program, "dequantize_row_q4_0", &err); + CL_CHECK(err, "clCreateKernel"); + kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err); + CL_CHECK(err, "clCreateKernel"); + kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err); + CL_CHECK(err, "clCreateKernel"); + kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err); + CL_CHECK(err, "clCreateKernel"); + kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err); + CL_CHECK(err, "clCreateKernel"); +} + +static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) { + if (req_size <= *cur_size) { + return; + } + + // Reallocate buffer with enough space + if (*cur_size > 0) { + clReleaseMemObject(*buf); + } + cl_int err; + *buf = clCreateBuffer(context, flags, req_size, NULL, &err); + *cur_size = req_size; + CL_CHECK(err, "clCreateBuffer"); +} + +void ggml_cl_sgemm_wrapper( + const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, + const int m, const int n, const int k, + const float alpha, const void *host_a, const int lda, + const float *host_b, const int ldb, const float beta, + float *host_c, const int ldc, const int btype) { + cl_int err = 0; + + cl_kernel kernel; + size_t global = n * k, local, size_qb; + bool dequant; + + switch (btype) { + case GGML_TYPE_F32: + dequant = false; + break; + case GGML_TYPE_Q4_0: + dequant = true; + kernel = kernel_q4_0; + local = 16; + size_qb = global * (sizeof(float) + local) / 32; + break; + case GGML_TYPE_Q4_1: + dequant = true; + kernel = kernel_q4_1; + local = 16; + size_qb = global * (sizeof(float) * 2 + local) / 32; + break; + case GGML_TYPE_Q5_0: + dequant = true; + kernel = kernel_q5_0; + local = 16; + size_qb = global * (sizeof(ggml_fp16_t) + sizeof(uint32_t) + local) / 32; + break; + case GGML_TYPE_Q5_1: + dequant = true; + kernel = kernel_q5_1; + local = 16; + size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32; + break; + case GGML_TYPE_Q8_0: + dequant = true; + kernel = kernel_q8_0; + local = 32; + size_qb = global * (sizeof(float) + local) / 32; + break; + default: + fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype); + abort(); + } + + const size_t size_a = m * k * sizeof(float); + const size_t size_b = n * k * sizeof(float); + const size_t size_c = m * n * sizeof(float); + + // Prepare buffers + ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a); + if (dequant) { + ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb); + } + ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b); + ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c); + + cl_event ev_a, ev_qb, ev_b; + + if (dequant) { + err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb); + err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b); + CL_CHECK(err, "clSetKernelArg"); + err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb); + CL_CHECK(err, "clEnqueueWriteBuffer qb"); + } else { + err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b); + CL_CHECK(err, "clEnqueueWriteBuffer b"); + } + + err = clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a); + CL_CHECK(err, "clEnqueueWriteBuffer a"); + if (dequant) { + err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b); + CL_CHECK(err, "clEnqueueNDRangeKernel"); + clReleaseEvent(ev_qb); + } + clWaitForEvents(1, &ev_a); + clWaitForEvents(1, &ev_b); + clReleaseEvent(ev_a); + clReleaseEvent(ev_b); + + cl_event ev_sgemm; + CLBlastStatusCode status = CLBlastSgemm((CLBlastLayout)order, + (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b, + m, n, k, + alpha, + cl_buffer_a, 0, lda, + cl_buffer_b, 0, ldb, + beta, + cl_buffer_c, 0, ldc, + &queue, &ev_sgemm); + + if (status != CLBlastSuccess) { + fprintf(stderr, "Error: CLBlast SGEMM %d\n", status); + abort(); + } + + cl_event ev_c; + clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c); + + // Wait for completion + clWaitForEvents(1, &ev_c); + clReleaseEvent(ev_sgemm); + clReleaseEvent(ev_c); +} diff --git a/ggml-opencl.h b/ggml-opencl.h new file mode 100644 index 000000000..7bcc603ef --- /dev/null +++ b/ggml-opencl.h @@ -0,0 +1,24 @@ +#pragma once + +#ifdef __cplusplus +extern "C" { +#endif + +void ggml_cl_init(void); + +enum ggml_blas_order { + GGML_BLAS_ORDER_ROW_MAJOR = 101, + GGML_BLAS_ORDER_COLUMN_MAJOR = 102, +}; + +enum ggml_blas_op { + GGML_BLAS_OP_N = 111, + GGML_BLAS_OP_T = 112, + GGML_BLAS_OP_C = 113, +}; + +void ggml_cl_sgemm_wrapper(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype); + +#ifdef __cplusplus +} +#endif diff --git a/ggml.c b/ggml.c index f4b8fc282..e5b3528d8 100644 --- a/ggml.c +++ b/ggml.c @@ -19,6 +19,7 @@ #include #include #include +#include // if C99 - static_assert is noop // ref: https://stackoverflow.com/a/53923785/4039976 @@ -134,18 +135,17 @@ inline static void* ggml_aligned_malloc(size_t size) { #define UNUSED(x) (void)(x) #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0) -#define GGML_ASSERT(x) \ - do { \ - if (!(x)) { \ - fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \ - abort(); \ - } \ - } while (0) - -#ifdef GGML_USE_ACCELERATE +#if defined(GGML_USE_ACCELERATE) #include -#elif GGML_USE_OPENBLAS +#if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions +#include "ggml-opencl.h" +#endif +#elif defined(GGML_USE_OPENBLAS) #include +#elif defined(GGML_USE_CUBLAS) +#include "ggml-cuda.h" +#elif defined(GGML_USE_CLBLAST) +#include "ggml-opencl.h" #endif #undef MIN @@ -183,9 +183,13 @@ typedef double ggml_float; #undef bool #define bool _Bool #else +#if defined(_MSC_VER) || defined(__MINGW32__) +#include +#else #include #endif #endif +#endif #ifdef __F16C__ @@ -325,6 +329,21 @@ static ggml_fp16_t table_exp_f16[1 << 16]; // precomputed f32 table for f16 (256 KB) static float table_f32_f16[1 << 16]; +#if defined(__ARM_NEON) || defined(__wasm_simd128__) +#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s +#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s) +#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s) +#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s) +#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s) +#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s) +#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) +#define B8(c,s ) B7(c,s, c), B7(c,s, s) + +// precomputed tables for expanding 8bits to 8 bytes: +static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4 +static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4 +#endif + // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. // This is also true for POWER9. @@ -351,6 +370,32 @@ ggml_fp16_t ggml_fp32_to_fp16(float x) { return GGML_FP32_TO_FP16(x); } +void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n) { + for (size_t i = 0; i < n; i++) { + y[i] = GGML_FP16_TO_FP32(x[i]); + } +} + +void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n) { + size_t i = 0; +#if defined(__F16C__) + for (; i + 7 < n; i += 8) { + __m256 x_vec = _mm256_loadu_ps(x + i); + __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm_storeu_si128((__m128i *)(y + i), y_vec); + } + for(; i + 3 < n; i += 4) { + __m128 x_vec = _mm_loadu_ps(x + i); + __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm_storel_epi64((__m128i *)(y + i), y_vec); + } +#endif + for (; i < n; i++) { + y[i] = GGML_FP32_TO_FP16(x[i]); + } +} + + // // timing // @@ -427,31 +472,102 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); // quantization // -// AVX routines provided by GH user Const-me -// ref: https://github.com/ggerganov/ggml/pull/27#issuecomment-1464934600 +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) +// multiply int8_t, add results pairwise twice +static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { + // Get absolute values of x vectors + const __m128i ax = _mm_sign_epi8(x, x); + // Sign the values of the y vectors + const __m128i sy = _mm_sign_epi8(y, x); + // Perform multiplication and create 16-bit values + const __m128i dot = _mm_maddubs_epi16(ax, sy); + const __m128i ones = _mm_set1_epi16(1); + return _mm_madd_epi16(ones, dot); +} + +#if __AVX__ || __AVX2__ || __AVX512F__ +// horizontally add 8 floats +static inline float hsum_float_8(const __m256 x) { + __m128 res = _mm256_extractf128_ps(x, 1); + res = _mm_add_ps(res, _mm256_castps256_ps128(x)); + res = _mm_add_ps(res, _mm_movehl_ps(res, res)); + res = _mm_add_ss(res, _mm_movehdup_ps(res)); + return _mm_cvtss_f32(res); +} + +// horizontally add 8 int32_t +static inline int hsum_i32_8(const __m256i a) { + const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1)); + const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128); + const __m128i sum64 = _mm_add_epi32(hi64, sum128); + const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); + return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); +} + +// horizontally add 4 int32_t +static inline int hsum_i32_4(const __m128i a) { + const __m128i hi64 = _mm_unpackhi_epi64(a, a); + const __m128i sum64 = _mm_add_epi32(hi64, a); + const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); + return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); +} + #if __AVX2__ || __AVX512F__ +// spread 32 bits to 32 bytes { 0x00, 0xFF } +static inline __m256i bytes_from_bits_32(const uint8_t * x) { + uint32_t x32; + memcpy(&x32, x, sizeof(uint32_t)); + const __m256i shuf_mask = _mm256_set_epi64x( + 0x0303030303030303, 0x0202020202020202, + 0x0101010101010101, 0x0000000000000000); + __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask); + const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe); + bytes = _mm256_or_si256(bytes, bit_mask); + return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1)); +} + // Unpack 32 4-bit fields into 32 bytes // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval -static inline __m256i bytesFromNibbles( const uint8_t* rsi ) +static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { - // Load 16 bytes from memory - __m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m256i bytes = _mm256_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes + const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); + const __m256i bytes = _mm256_set_m128i(_mm_srli_epi16(tmp, 4), tmp); const __m256i lowMask = _mm256_set1_epi8( 0xF ); - __m256i high = _mm256_andnot_si256( lowMask, bytes ); - __m256i low = _mm256_and_si256( lowMask, bytes ); - high = _mm256_slli_epi16( high, 4 ); - bytes = _mm256_or_si256( low, high ); - return bytes; + return _mm256_and_si256(lowMask, bytes); +} + +// add int16_t pairwise and return as float vector +static inline __m256 sum_i16_pairs_float(const __m256i x) { + const __m256i ones = _mm256_set1_epi16(1); + const __m256i summed_pairs = _mm256_madd_epi16(ones, x); + return _mm256_cvtepi32_ps(summed_pairs); +} + +// multiply int8_t, add results pairwise twice and return as float vector +static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { + // Get absolute values of x vectors + const __m256i ax = _mm256_sign_epi8(x, x); + // Sign the values of the y vectors + const __m256i sy = _mm256_sign_epi8(y, x); +#if __AVXVNNI__ + const __m256i zero = _mm256_setzero_si256(); + const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); + return _mm256_cvtepi32_ps(summed_pairs); +#else + // Perform multiplication and create 16-bit values + const __m256i dot = _mm256_maddubs_epi16(ax, sy); + return sum_i16_pairs_float(dot); +#endif } static inline __m128i packNibbles( __m256i bytes ) { // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh +#if __AVX512F__ + const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000 + bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh + return _mm256_cvtepi16_epi8(bytes); // abcd_efgh +#else const __m256i lowByte = _mm256_set1_epi16( 0xFF ); __m256i high = _mm256_andnot_si256( lowByte, bytes ); __m256i low = _mm256_and_si256( lowByte, bytes ); @@ -462,25 +578,9 @@ static inline __m128i packNibbles( __m256i bytes ) __m128i r0 = _mm256_castsi256_si128( bytes ); __m128i r1 = _mm256_extracti128_si256( bytes, 1 ); return _mm_packus_epi16( r0, r1 ); +#endif } -#elif __AVX__ -static inline __m128i bytesFromNibbles( const uint8_t* rsi ) -{ - // Load 8 bytes from memory - __m128i tmp = _mm_loadu_si64( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m128i bytes = _mm_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes - const __m128i lowMask = _mm_set1_epi8( 0xF ); - __m128i high = _mm_andnot_si128( lowMask, bytes ); - __m128i low = _mm_and_si128( lowMask, bytes ); - high = _mm_slli_epi16( high, 4 ); - bytes = _mm_or_si128( low, high ); - return bytes; -} - +#else static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) { // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh @@ -497,6 +597,19 @@ static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) return _mm_packus_epi16( bytes1, bytes2); } #endif +#elif defined(__SSSE3__) +// horizontally add 4x4 floats +static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) { + __m128 res_0 =_mm_hadd_ps(a, b); + __m128 res_1 =_mm_hadd_ps(c, d); + __m128 res =_mm_hadd_ps(res_0, res_1); + res =_mm_hadd_ps(res, res); + res =_mm_hadd_ps(res, res); + + return _mm_cvtss_f32(res); +} +#endif // __AVX__ || __AVX2__ || __AVX512F__ +#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) #if __ARM_NEON @@ -514,6 +627,18 @@ inline static uint16_t vaddvq_u8(uint8x16_t v) { (uint16_t)vgetq_lane_u8(v, 14) + (uint16_t)vgetq_lane_u8(v, 15); } +inline static int16_t vaddvq_s8(int8x16_t v) { + return + (int16_t)vgetq_lane_s8(v, 0) + (int16_t)vgetq_lane_s8(v, 1) + + (int16_t)vgetq_lane_s8(v, 2) + (int16_t)vgetq_lane_s8(v, 3) + + (int16_t)vgetq_lane_s8(v, 4) + (int16_t)vgetq_lane_s8(v, 5) + + (int16_t)vgetq_lane_s8(v, 6) + (int16_t)vgetq_lane_s8(v, 7) + + (int16_t)vgetq_lane_s8(v, 8) + (int16_t)vgetq_lane_s8(v, 9) + + (int16_t)vgetq_lane_s8(v, 10) + (int16_t)vgetq_lane_s8(v, 11) + + (int16_t)vgetq_lane_s8(v, 12) + (int16_t)vgetq_lane_s8(v, 13) + + (int16_t)vgetq_lane_s8(v, 14) + (int16_t)vgetq_lane_s8(v, 15); +} + inline static int32_t vaddvq_s16(int16x8_t v) { return (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + @@ -550,20 +675,15 @@ float vmaxvq_f32(float32x4_t v) { MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); } -int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) { - return vget_low_s8(vcombine_s8(a, b)); -} +int32x4_t vcvtnq_s32_f32(float32x4_t v) { + int32x4_t res; -int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) { - return vget_high_s8(vcombine_s8(a, b)); -} + res[0] = roundf(vgetq_lane_f32(v, 0)); + res[1] = roundf(vgetq_lane_f32(v, 1)); + res[2] = roundf(vgetq_lane_f32(v, 2)); + res[3] = roundf(vgetq_lane_f32(v, 3)); -uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) { - return vget_low_u8(vcombine_u8(a, b)); -} - -uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) { - return vget_high_u8(vcombine_u8(a, b)); + return res; } #endif @@ -583,14 +703,24 @@ typedef struct { float m; // min uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; -static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); +static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding"); -#define QK4_2 16 +#define QK5_0 32 typedef struct { ggml_fp16_t d; // delta - uint8_t qs[QK4_2 / 2]; // nibbles / quants -} block_q4_2; -static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_0 / 2]; // nibbles / quants +} block_q5_0; +static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); + +#define QK5_1 32 +typedef struct { + ggml_fp16_t d; // delta + ggml_fp16_t m; // min + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_1 / 2]; // nibbles / quants +} block_q5_1; +static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); #define QK8_0 32 typedef struct { @@ -599,500 +729,191 @@ typedef struct { } block_q8_0; static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); +#define QK8_1 32 +typedef struct { + float d; // delta + float s; // d * sum(qs[i]) + int8_t qs[QK8_1]; // quants +} block_q8_1; +static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); // reference implementation for deterministic creation of model files static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; + static const int qk = QK4_0; - uint8_t pp[QK4_0/2]; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max + float max = 0.0f; - for (int l = 0; l < QK4_0; l++) { - const float v = x[i*QK4_0 + l]; - amax = MAX(amax, fabsf(v)); + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } } - const float d = amax / ((1 << 3) - 1); + const float d = max / -8; const float id = d ? 1.0f/d : 0.0f; y[i].d = d; - for (int l = 0; l < QK4_0; l += 2) { - const float v0 = x[i*QK4_0 + l + 0]*id; - const float v1 = x[i*QK4_0 + l + 1]*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; - const uint8_t vi0 = (int8_t)roundf(v0) + 8; - const uint8_t vi1 = (int8_t)roundf(v1) + 8; + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f)); - assert(vi0 < 16); - assert(vi1 < 16); - - pp[l/2] = vi0 | (vi1 << 4); + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; } - - memcpy(y[i].qs, pp, sizeof(pp)); } } -static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; - - block_q4_0 * restrict y = vy; - -#if defined(__POWER9_VECTOR__) - const vector float v85 = vec_splats(8.5f); - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - - vector float srcv [8]; - vector float asrcv[8]; - vector float amaxv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = *(vector float *)(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]); - - for (int l = 0; l < 4; l++) amaxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]); - //for (int l = 0; l < 2; l++) amaxv[4*l] = vec_max(amaxv[4*l], amaxv[4*l+2]); - amaxv[0] = vec_max(amaxv[0], amaxv[2]); - amaxv[4] = vec_max(amaxv[4], amaxv[6]); - //for (int l = 0; l < 1; l++) amaxv[8*l] = vec_max(amaxv[8*l], amaxv[8*l+4]); - amaxv[0] = vec_max(amaxv[0], amaxv[4]); - - amax = MAX( - MAX(vec_extract(amaxv[0], 0), vec_extract(amaxv[0], 1)), - MAX(vec_extract(amaxv[0], 2), vec_extract(amaxv[0], 3))); - - const float d = amax / ((1 << 3) - 1); - const float id = d ? 1.0/d : 0.0; - - y[i].d = d; - - const vector float vid = vec_splats(id); - uint8_t * restrict pb = y[i].qs; - for (int l = 0; l < 8; l++) { - const vector float vf = vec_madd(srcv[l], vid, v85); - const vector signed int vi = vec_signed(vf); - - pb[2*l + 0] = vec_extract(vi, 0) | (vec_extract(vi, 1) << 4); - pb[2*l + 1] = vec_extract(vi, 2) | (vec_extract(vi, 3) << 4); - } - } -#elif __ARM_NEON - for (int i = 0; i < nb; i++) { - float32x4_t srcv [8]; - float32x4_t asrcv[8]; - float32x4_t amaxv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); - - for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); - for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); - for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); - - const float amax = vmaxvq_f32(amaxv[0]); - - const float d = amax / ((1 << 3) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); - const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f)); - const int32x4_t vi = vcvtq_s32_f32(vf); - - y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4); - y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4); - } - } -#elif defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max(abs(e)) for the block - const __m256 signBit = _mm256_set1_ps( -0.0f ); - __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Quantize these floats - const float d = maxScalar / 7.0f; - y[i].d = d; - const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - // Apply offset to translate the range from [ -7 .. +7 ] into [ +1 .. +15 ] - const __m256i off = _mm256_set1_epi8( 8 ); - i0 = _mm256_add_epi8( i0, off ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( i0 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif defined(__AVX__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max(abs(e)) for the block - const __m256 signBit = _mm256_set1_ps( -0.0f ); - __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); - maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Quantize these floats - const float d = maxScalar / 7.0f; - y[i].d = d; - const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Since we don't have in AVX some necessary functions, - // we split the registers in half and call AVX2 analogs from SSE - __m128i ni0 = _mm256_castsi256_si128( i0 ); - __m128i ni1 = _mm256_extractf128_si256( i0, 1); - __m128i ni2 = _mm256_castsi256_si128( i1 ); - __m128i ni3 = _mm256_extractf128_si256( i1, 1); - __m128i ni4 = _mm256_castsi256_si128( i2 ); - __m128i ni5 = _mm256_extractf128_si256( i2, 1); - __m128i ni6 = _mm256_castsi256_si128( i3 ); - __m128i ni7 = _mm256_extractf128_si256( i3, 1); - - // Convert int32 to int16 - ni0 = _mm_packs_epi32( ni0, ni1 ); - ni2 = _mm_packs_epi32( ni2, ni3 ); - ni4 = _mm_packs_epi32( ni4, ni5 ); - ni6 = _mm_packs_epi32( ni6, ni7 ); - // Convert int16 to int8 - ni0 = _mm_packs_epi16( ni0, ni2 ); - ni4 = _mm_packs_epi16( ni4, ni6 ); - - // Apply offset to translate the range from [ -7 .. +7 ] into [ +1 .. +15 ] - const __m128i off = _mm_set1_epi8( 8); - ni0 = _mm_add_epi8( ni0, off ); - ni4 = _mm_add_epi8( ni4, off ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( ni0, ni4 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif defined(__wasm_simd128__) - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - - v128_t srcv [8]; - v128_t asrcv[8]; - v128_t amaxv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = wasm_v128_load(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = wasm_f32x4_abs(srcv[l]); - - for (int l = 0; l < 4; l++) amaxv[2*l] = wasm_f32x4_max(asrcv[2*l], asrcv[2*l+1]); - for (int l = 0; l < 2; l++) amaxv[4*l] = wasm_f32x4_max(amaxv[4*l], amaxv[4*l+2]); - for (int l = 0; l < 1; l++) amaxv[8*l] = wasm_f32x4_max(amaxv[8*l], amaxv[8*l+4]); - - amax = MAX( - MAX(wasm_f32x4_extract_lane(amaxv[0], 0), wasm_f32x4_extract_lane(amaxv[0], 1)), - MAX(wasm_f32x4_extract_lane(amaxv[0], 2), wasm_f32x4_extract_lane(amaxv[0], 3))); - - const float d = amax / ((1 << 3) - 1); - const float id = d ? 1.0/d : 0.0; - - y[i].d = d; - - for (int l = 0; l < 8; l++) { - const v128_t v = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id)); - const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f)); - const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf); - - y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vi, 0) | (wasm_i32x4_extract_lane(vi, 1) << 4); - y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vi, 2) | (wasm_i32x4_extract_lane(vi, 3) << 4); - } - } -#else - // scalar +static void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) { quantize_row_q4_0_reference(x, y, k); -#endif } -static void quantize_row_q4_1_reference(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; +static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) { + const int qk = QK4_1; - block_q4_1 * restrict y = vy; + assert(k % qk == 0); - uint8_t pp[QK4_1/2]; + const int nb = k / qk; for (int i = 0; i < nb; i++) { float min = FLT_MAX; float max = -FLT_MAX; - for (int l = 0; l < QK4_1; l++) { - const float v = x[i*QK4_1 + l]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (v < min) min = v; if (v > max) max = v; } - const float d = (max - min) / ((1 << 4) - 1); + const float d = (max - min) / ((1 << 4) - 1); const float id = d ? 1.0f/d : 0.0f; y[i].d = d; y[i].m = min; - for (int l = 0; l < QK4_1; l += 2) { - const float v0 = (x[i*QK4_1 + l + 0] - min)*id; - const float v1 = (x[i*QK4_1 + l + 1] - min)*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; - const uint8_t vi0 = roundf(v0); - const uint8_t vi1 = roundf(v1); + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f)); - assert(vi0 < 16); - assert(vi1 < 16); - - pp[l/2] = vi0 | (vi1 << 4); + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; } - - memcpy(y[i].qs, pp, sizeof(pp)); } } -static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_1 == 0); - - const int nb = k / QK4_1; - - block_q4_1 * restrict y = vy; - -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max for the block - __m256 vmax; - vmax = _mm256_max_ps( v0, v1 ); - vmax = _mm256_max_ps( vmax, v2 ); - vmax = _mm256_max_ps( vmax, v3 ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( vmax, 1 ), _mm256_castps256_ps128( vmax ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Compute min for the block - __m256 vmin; - vmin = _mm256_min_ps( v0, v1 ); - vmin = _mm256_min_ps( vmin, v2 ); - vmin = _mm256_min_ps( vmin, v3 ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( vmin, 1 ), _mm256_castps256_ps128( vmin ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - - // Quantize these floats - const float d = (maxScalar - minScalar) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].m = minScalar; - y[i].d = d; - - // x = (x-min)*id - const __m256 mul = _mm256_set1_ps( id ); - const __m256 off = _mm256_set1_ps( minScalar ); - v0 = _mm256_mul_ps( _mm256_sub_ps( v0, off ), mul ); - v1 = _mm256_mul_ps( _mm256_sub_ps( v1, off ), mul ); - v2 = _mm256_mul_ps( _mm256_sub_ps( v2, off ), mul ); - v3 = _mm256_mul_ps( _mm256_sub_ps( v3, off ), mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( i0 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif __ARM_NEON - for (int i = 0; i < nb; i++) { - float32x4_t srcv[8]; - float32x4_t minv[8]; - float32x4_t maxv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*QK4_1 + 4*l); - - for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]); - for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]); - for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l + 4]); - - for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l + 1]); - for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l + 2]); - for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l + 4]); - - const float min = vminvq_f32(minv[0]); - const float max = vmaxvq_f32(maxv[0]); - - const float d = (max - min) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - y[i].m = min; - - const float32x4_t minv0 = vdupq_n_f32(min); - - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id); - const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(0.5f)); // needed to round to nearest - const int32x4_t vi = vcvtq_s32_f32(vf); - - y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4); - y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4); - } - } -#else - // scalar - quantize_row_q4_1_reference(x, vy, k); -#endif +static void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q4_1_reference(x, y, k); } -// reference implementation for deterministic creation of model files -static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * restrict y, int k) { - assert(k % QK4_2 == 0); +static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { + static const int qk = QK5_0; - const int nb = k / QK4_2; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max + float max = 0.0f; - for (int l = 0; l < QK4_2; l++) { - const float v = x[i*QK4_2 + l]; - amax = MAX(amax, fabsf(v)); + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } } - const float d = amax / ((1 << 3) - 1); - + const float d = max / -16; const float id = d ? 1.0f/d : 0.0f; y[i].d = GGML_FP32_TO_FP16(d); - for (int l = 0; l < QK4_2; l += 2) { - const float v0 = x[i*QK4_2 + l + 0]*id; - const float v1 = x[i*QK4_2 + l + 1]*id; + uint32_t qh = 0; - const uint8_t vi0 = (uint8_t)(v0 + 8.5f); - const uint8_t vi1 = (uint8_t)(v1 + 8.5f); + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; - assert(vi0 < 16); - assert(vi1 < 16); + const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f)); + const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f)); - y[i].qs[l/2] = vi0 | (vi1 << 4); + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10) >> 4) << (j + 0); + qh |= ((xi1 & 0x10) >> 4) << (j + qk/2); } + + memcpy(&y[i].qh, &qh, sizeof(qh)); } } -static void quantize_row_q4_2(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_2 == 0); +static void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) { + quantize_row_q5_0_reference(x, y, k); +} - block_q4_2 * restrict y = vy; +static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { + const int qk = QK5_1; - quantize_row_q4_2_reference(x, y, k); + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + float min = FLT_MAX; + float max = -FLT_MAX; + + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + + if (v < min) min = v; + if (v > max) max = v; + } + + const float d = (max - min) / ((1 << 5) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); + + uint32_t qh = 0; + + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; + + const uint8_t xi0 = (uint8_t)(x0 + 0.5f); + const uint8_t xi1 = (uint8_t)(x1 + 0.5f); + + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10) >> 4) << (j + 0); + qh |= ((xi1 & 0x10) >> 4) << (j + qk/2); + } + + memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); + } +} + +static void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q5_1_reference(x, y, k); } // reference implementation for deterministic creation of model files @@ -1103,8 +924,8 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - for (int l = 0; l < QK8_0; l++) { - const float v = x[i*QK8_0 + l]; + for (int j = 0; j < QK8_0; j++) { + const float v = x[i*QK8_0 + j]; amax = MAX(amax, fabsf(v)); } @@ -1113,14 +934,16 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r y[i].d = d; - for (int l = 0; l < QK8_0; ++l) { - const float v = x[i*QK8_0 + l]*id; - y[i].qs[l] = roundf(v); + for (int j = 0; j < QK8_0; ++j) { + const float x0 = x[i*QK8_0 + j]*id; + + y[i].qs[j] = roundf(x0); } } } static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { + assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; @@ -1132,12 +955,12 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int float32x4_t asrcv[8]; float32x4_t amaxv[8]; - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); - for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); - for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); - for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); const float amax = vmaxvq_f32(amaxv[0]); @@ -1146,14 +969,14 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int y[i].d = d; - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); const int32x4_t vi = vcvtnq_s32_f32(v); - y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); } } #elif defined(__AVX2__) || defined(__AVX__) @@ -1246,287 +1069,349 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int #endif } -static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; +// reference implementation for deterministic creation of model files +static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) { + assert(QK8_1 == 32); + assert(k % QK8_1 == 0); + const int nb = k / QK8_1; - const block_q4_0 * restrict x = vx; + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_1; j++) { + const float v = x[i*QK8_1 + j]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + int sum = 0; + + for (int j = 0; j < QK8_1/2; ++j) { + const float v0 = x[i*QK8_1 + j]*id; + const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id; + + y[i].qs[ j] = roundf(v0); + y[i].qs[QK8_1/2 + j] = roundf(v1); + + sum += y[i].qs[ j]; + sum += y[i].qs[QK8_1/2 + j]; + } + + y[i].s = d * sum; + } +} + +static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { + assert(k % QK8_1 == 0); + const int nb = k / QK8_1; + + block_q8_1 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + int32x4_t accv = vdupq_n_s32(0); + + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); + + accv = vaddq_s32(accv, vi); + } + + y[i].s = d * vaddvq_s32(accv); + } +#elif defined(__AVX2__) || defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Quantize these floats + const float d = maxScalar / 127.f; + y[i].d = d; + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); #if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // scale factor - const __m256 d_v = _mm256_broadcast_ss(&x[i].d); + // Compute the sum of the quants and set y[i].s + y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); - const uint8_t * restrict pp = x[i].qs; + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - for (int l = 0; l < QK4_0; l += 32) { - // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytesFromNibbles(pp+l/2); + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); - // Subtract 8 from the integers - vx8 = _mm256_sub_epi8(vx8, _mm256_set1_epi8(8)); + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); - // Convert to 16-bit int - const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); - const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); + // Compute the sum of the quants and set y[i].s + const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); + const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); + y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1)); - // Convert to 32-bit int -> float 32 - const __m256 vf[4] = { - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) - }; + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); - // Scale and store - for (int j = 0; j < 4; j++) { - const __m256 result = _mm256_mul_ps(vf[j], d_v); - _mm256_storeu_ps(y + i * QK4_0 + l + j*8, result); - } - } - } -#elif defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - const float32x4_t vd = vdupq_n_f32(x[i].d); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_0; l += 16) { - // Load 16x4-bit integers into 8x8-bit integers - const uint8x8_t v8 = vld1_u8(pp + l/2); - - // Expand 4-bit qs to 8-bit bytes - const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0f)); - const uint8x8_t v1 = vshr_n_u8(v8, 4); - - // Convert to signed 8-bit integers - const int8x8_t vs_0 = vreinterpret_s8_u8(v0); - const int8x8_t vs_1 = vreinterpret_s8_u8(v1); - - // Subtract 8 from each byte - const int8x8_t vb_0 = vsub_s8(vs_0, vdup_n_s8(8)); - const int8x8_t vb_1 = vsub_s8(vs_1, vdup_n_s8(8)); - - // Interleave and combine - const int8x8_t vx_0 = vzip1_s8(vb_0, vb_1); - const int8x8_t vx_1 = vzip2_s8(vb_0, vb_1); - - const int8x16_t vq = vcombine_s8(vx_0, vx_1); - - // convert to 2x int16x8_t - const int16x8_t vi_0 = vmovl_s8(vget_low_s8 (vq)); - const int16x8_t vi_1 = vmovl_s8(vget_high_s8(vq)); - - // convert to 4x float32x4_t - const float32x4_t vf_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_0))); - const float32x4_t vf_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_0))); - const float32x4_t vf_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_1))); - const float32x4_t vf_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_1))); - - // Multiply by d - const float32x4_t r0 = vmulq_f32(vf_0, vd); - const float32x4_t r1 = vmulq_f32(vf_1, vd); - const float32x4_t r2 = vmulq_f32(vf_2, vd); - const float32x4_t r3 = vmulq_f32(vf_3, vd); - - // Store - vst1q_f32(y + i*QK4_0 + l + 0, r0); - vst1q_f32(y + i*QK4_0 + l + 4, r1); - vst1q_f32(y + i*QK4_0 + l + 8, r2); - vst1q_f32(y + i*QK4_0 + l + 12, r3); - } + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif } #else // scalar - for (int i = 0; i < nb; i++) { - const float d = x[i].d; - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - //printf("d = %f, vi = %d, vi0 = %d, vi1 = %d, v0 = %f, v1 = %f\n", d, vi, vi0, vi1, v0, v1); - - y[i*QK4_0 + l + 0] = v0; - y[i*QK4_0 + l + 1] = v1; - - assert(!isnan(y[i*QK4_0 + l + 0])); - assert(!isnan(y[i*QK4_0 + l + 1])); - } - } + quantize_row_q8_1_reference(x, y, k); #endif } -static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; +static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) { + static const int qk = QK4_0; - const block_q4_1 * restrict x = vx; + assert(k % qk == 0); + + const int nb = k / qk; -#if defined(__AVX2__) for (int i = 0; i < nb; i++) { - const __m256 d_v = _mm256_broadcast_ss(&x[i].d); - const __m256 d_m = _mm256_broadcast_ss(&x[i].m); + const float d = x[i].d; - const uint8_t * restrict pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; - for (int l = 0; l < QK4_1; l += 32) { - // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytesFromNibbles(pp+l/2); - - // Convert to 16-bit int - const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); - const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); - - // Convert to 32-bit int -> float 32 - const __m256 vf[4] = { - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) - }; - - // Scale, add m and store - for (int j = 0; j < 4; j++) { - const __m256 result = _mm256_add_ps(_mm256_mul_ps(vf[j], d_v), d_m); - _mm256_storeu_ps(y + i * QK4_1 + l + j*8, result); - } + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } -#elif defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - const float32x4_t vd = vdupq_n_f32(x[i].d); - const float32x4_t vm = vdupq_n_f32(x[i].m); +} - const uint8_t * restrict pp = x[i].qs; +static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) { + static const int qk = QK4_1; - for (int l = 0; l < QK4_1; l += 16) { - // Load 16x4-bit integers into 8x8-bit integers - const uint8x8_t v8 = vld1_u8(pp + l/2); + assert(k % qk == 0); - // Expand 4-bit qs to 8-bit bytes - const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0f)); - const uint8x8_t v1 = vshr_n_u8(v8, 4); + const int nb = k / qk; - // Interleave and combine - const uint8x8_t vx_0 = vzip1_u8(v0, v1); - const uint8x8_t vx_1 = vzip2_u8(v0, v1); - - const uint8x16_t vq = vcombine_u8(vx_0, vx_1); - - // convert to 2x uint16x8_t - const uint16x8_t vi_0 = vmovl_u8(vget_low_u8 (vq)); - const uint16x8_t vi_1 = vmovl_u8(vget_high_u8(vq)); - - // convert to 4x float32x4_t - const float32x4_t vf_0 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_0))); - const float32x4_t vf_1 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_0))); - const float32x4_t vf_2 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_1))); - const float32x4_t vf_3 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_1))); - - // multiply by d and add m - const float32x4_t r0 = vmlaq_f32(vm, vf_0, vd); - const float32x4_t r1 = vmlaq_f32(vm, vf_1, vd); - const float32x4_t r2 = vmlaq_f32(vm, vf_2, vd); - const float32x4_t r3 = vmlaq_f32(vm, vf_3, vd); - - // Store - vst1q_f32(y + i*QK4_1 + l + 0, r0); - vst1q_f32(y + i*QK4_1 + l + 4, r1); - vst1q_f32(y + i*QK4_1 + l + 8, r2); - vst1q_f32(y + i*QK4_1 + l + 12, r3); - } - } -#else for (int i = 0; i < nb; i++) { const float d = x[i].d; const float m = x[i].m; - const uint8_t * restrict pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F); + const int x1 = (x[i].qs[j] >> 4); - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK4_1 + l + 0] = v0; - y[i*QK4_1 + l + 1] = v1; - - assert(!isnan(y[i*QK4_1 + l + 0])); - assert(!isnan(y[i*QK4_1 + l + 1])); + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } -#endif } -static void dequantize_row_q4_2(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_2 == 0); - const int nb = k / QK4_2; +static void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) { + static const int qk = QK5_0; - const block_q4_2 * restrict x = vx; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { const float d = GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict pp = x[i].qs; + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - y[i*QK4_2 + l + 0] = v0; - y[i*QK4_2 + l + 1] = v1; - - assert(!isnan(y[i*QK4_2 + l + 0])); - assert(!isnan(y[i*QK4_2 + l + 1])); + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; + } + } +} + +static void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) { + static const int qk = QK5_1; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int x0 = (x[i].qs[j] & 0x0F) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; + } + } +} + +static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, int k) { + static const int qk = QK8_0; + + assert(k % qk == 0); + + const int nb = k / qk; + + const block_q8_0 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = x[i].d; + + for (int j = 0; j < qk; ++j) { + y[i*qk + j] = x[i].qs[j]*d; } } } -static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -//static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { [GGML_TYPE_Q4_0] = { - .dequantize_row_q = dequantize_row_q4_0, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q4_0, .quantize_row_q = quantize_row_q4_0, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference, .quantize_row_q_dot = quantize_row_q8_0, .vec_dot_q = ggml_vec_dot_q4_0_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, }, [GGML_TYPE_Q4_1] = { - .dequantize_row_q = dequantize_row_q4_1, + .dequantize_row_q = (dequantize_row_q_t)dequantize_row_q4_1, .quantize_row_q = quantize_row_q4_1, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference, - .quantize_row_q_dot = quantize_row_q4_1, - .vec_dot_q = ggml_vec_dot_q4_1, + .quantize_row_q_dot = quantize_row_q8_1, + .vec_dot_q = ggml_vec_dot_q4_1_q8_1, + .vec_dot_type = GGML_TYPE_Q8_1, }, - [GGML_TYPE_Q4_2] = { - .dequantize_row_q = dequantize_row_q4_2, - .quantize_row_q = quantize_row_q4_2, - .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_2_reference, + [GGML_TYPE_Q5_0] = { + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q5_0, + .quantize_row_q = quantize_row_q5_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q5_0_reference, .quantize_row_q_dot = quantize_row_q8_0, - .vec_dot_q = ggml_vec_dot_q4_2_q8_0, + .vec_dot_q = ggml_vec_dot_q5_0_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, + }, + [GGML_TYPE_Q5_1] = { + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q5_1, + .quantize_row_q = quantize_row_q5_1, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q5_1_reference, + .quantize_row_q_dot = quantize_row_q8_1, + .vec_dot_q = ggml_vec_dot_q5_1_q8_1, + .vec_dot_type = GGML_TYPE_Q8_1, + }, + [GGML_TYPE_Q8_0] = { + .dequantize_row_q = dequantize_row_q8_0, + .quantize_row_q = quantize_row_q8_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q8_0_reference, + .quantize_row_q_dot = quantize_row_q8_0, + .vec_dot_q = ggml_vec_dot_q8_0_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, + }, + [GGML_TYPE_Q8_1] = { + .dequantize_row_q = NULL, // TODO + .quantize_row_q = quantize_row_q8_1, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q8_1_reference, + .quantize_row_q_dot = quantize_row_q8_1, + .vec_dot_q = NULL, // TODO + .vec_dot_type = GGML_TYPE_Q8_1, }, - // TODO: GGML_TYPE_Q8_0 }; // For internal test use @@ -2046,6 +1931,7 @@ inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } +inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; } inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; } inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; } @@ -2092,191 +1978,6 @@ inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float *s = sumf; } -#if __AVX512F__ && QK4_0 == 32 -static inline __m512i bytes_from_q4_0_twoblocks_avx512( const __m512i blocks ) { - // The 64 bytes of `blocks` contain two consecutive Q4_0 blocks loaded from memory: - // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ - // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32| - // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ - // | :. =_ () [] <> () Zz Yy| - // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ - // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00| - // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ - // |Xx Ww Vv Uu Tt Ss Rr Qq Pp Oo Nn Mm Ll Kk Jj Ii Hh Gg Ff Ee Dd Cc Bb Aa | - // +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ - // - // Bytes 04..19 (block #0) and 24..39 (block #1) both contain 32 nibbles (4-bit unsigned integers). - // We have exactly 64 nibbles, so we want to place each nibble into a separate byte. - // Bytes 00..03 and 20..23 contain scales, which are irrelevant to this function. - // Bytes 40..63 are masked when loading the data, so they are zeroed out. -#ifdef __AVX512VBMI__ - const __m512i byte_perm = _mm512_set_epi8( - 39, 38, 39, 38, 37, 36, 37, 36, 35, 34, 35, 34, 33, 32, 33, 32, - 31, 30, 31, 30, 29, 28, 29, 28, 27, 26, 27, 26, 25, 24, 25, 24, - 19, 18, 19, 18, 17, 16, 17, 16, 15, 14, 15, 14, 13, 12, 13, 12, - 11, 10, 11, 10, 9, 8, 9, 8, 7, 6, 7, 6, 5, 4, 5, 4 - ); - const __m512i permuted = _mm512_permutexvar_epi8( byte_perm, blocks ); - // After applying VPERMB, `permuted` looks like this: - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |:. =_ :. =_ () [] () [] <> () <> () Zz Yy Zz Yy Xx Ww Xx Ww Vv Uu Vv Uu Tt Ss Tt Ss Rr Qq Rr Qq| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |Pp Oo Pp Oo Nn Mm Nn Mm Ll Kk Ll Kk Jj Ii Jj Ii Hh Gg Hh Gg Ff Ee Ff Ee Dd Cc Dd Cc Bb Aa Bb Aa| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ -#else - const __m512i word_perm = _mm512_set_epi16( - 19, 19, 18, 18, 17, 17, 16, 16, 15, 15, 14, 14, 13, 13, 12, 12, - 9, 9, 8, 8, 7, 7, 6, 6, 5, 5, 4, 4, 3, 3, 2, 2 - ); - const __m512i permuted = _mm512_permutexvar_epi16( word_perm, blocks ); - // This is the fallback path for CPUs that don't support VPERMB. Since we permute 16-bit groups only, - // VPERMB can be replaced with VPERMW. We could always use VPERMW, but at least on Tiger Lake and - // Ice Lake VPERMW followed by a right shift is quite noticeably slower than VPERMB. -#endif - - // Shift every odd-numbered 16-bit group to the right by 4 bits. - const __mmask32 shift_mask = 0xaaaaaaaa; - const __m512i shifted = _mm512_mask_srai_epi16( permuted, shift_mask, permuted, 4 ); - // After applying VPSRAW, `shifted` looks like this (the "empty" nibbles are filled with zeroes): - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // | : .= :. =_ ( )[ () [] < >( <> () Z zY Zz Yy X xW Xx Ww V vU Vv Uu T tS Tt Ss R rQ Rr Qq - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // | P pO Pp Oo N nM Nn Mm L lK Ll Kk J jI Jj Ii H hG Hh Gg F fE Ff Ee D dC Dd Cc B bA Bb Aa| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - - // Now we just need to zero out the higher nibble in each byte, and we're done. - const __m512i low_nibble_mask = _mm512_set1_epi8( 0xf ); - return _mm512_and_si512( low_nibble_mask, shifted ); - // The final result looks like this: - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // | : = . _ ( [ ) ] < ( > ) Z Y z y X W x w V U v u T S t s R Q r q| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // |31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ - // | P O p o N M n m L K l k J I j i H G h g F E f e D C d c B A b a| - // +-----------+-----------+-----------+-----------+-----------+-----------+-----------+-----------+ -} - -static inline __m512 dot_q4_0_twoblocks_avx512( - __m512 acc, - const block_q4_0 * restrict x, - const block_q4_0 * restrict y, - int i -) { - // A pair of Q4_0 blocks spans 40 bytes, while an AVX-512 register has 64. The remaining 24 bytes - // can potentially be unaddressable, so we make sure to mask them out before the load, even though - // we don't use them at all. This might hurt the performance slightly, since the compiler is forced - // to use e.g. `VMOVDQU64 REG, MASK, [ADDR] + VPERMB ..., REG` instead of just `VPERMB ..., [ADDR]`. - const __mmask8 load_mask = 0x1f; - const __m512i blocks_0 = _mm512_maskz_loadu_epi64( load_mask, &x[i] ); - const __m512i blocks_1 = _mm512_maskz_loadu_epi64( load_mask, &y[i] ); - - // We want to multiply the scales, so we interpret both registers as 16 32-bit floats: - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // | 15 | 14 | 13 | 12 | 11 | 10 | 09 | 08 | 07 | 06 | 05 | 04 | 03 | 02 | 01 | 00 | - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // blocks_0_float - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // | | | | | | | xx | xx | xx | xx | B | xx | xx | xx | xx | A | - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // blocks_1_float - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // | | | | | | | xx | xx | xx | xx | D | xx | xx | xx | xx | C | - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - const __m512 blocks_0_float = _mm512_castsi512_ps( blocks_0 ); - const __m512 blocks_1_float = _mm512_castsi512_ps( blocks_1 ); - // We absolutely shouldn't touch the floats marked with `xx`: they contain some - // random data, which might very well underflow. At least on Intel, this leads - // to a huge penalty that can't be ignored (easily 100x or more) unless you - // compile your code with something like `-ffast-math` to enable FTZ/DAZ flags. - // (and ggml can't assume that you do)... - const __mmask16 scale_mul_mask = 0x21; -#ifdef __clang__ - // ...however, clang decides to optimize the multiplication mask away: - // https://godbolt.org/z/P8PqdsfvW - // gcc and MSVC do the sane thing. This horrible workaround forces clang to emit the mask. - __m512i scales; - __asm__( - "vmulps %1, %2, %0%{%3%}" - : "=v" ( scales ) - : "vm" ( blocks_0_float ), "v" ( blocks_1_float ), "Yk" ( scale_mul_mask ) - ); -#else - const __m512 scales = _mm512_maskz_mul_ps( scale_mul_mask, blocks_0_float, blocks_1_float ); -#endif - const __m512i scale_perm = _mm512_set_epi32( - 5, 5, 5, 5, 5, 5, 5, 5, - 0, 0, 0, 0, 0, 0, 0, 0 - ); - const __m512 permuted_scales = _mm512_permutexvar_ps( scale_perm, scales ); - // After VMULPS and VPERMPS, `permuted_scales` looks like this: - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // | 15 | 14 | 13 | 12 | 11 | 10 | 09 | 08 | 07 | 06 | 05 | 04 | 03 | 02 | 01 | 00 | - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - // | B*D| B*D| B*D| B*D| B*D| B*D| B*D| B*D| A*C| A*C| A*C| A*C| A*C| A*C| A*C| A*C| - // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+ - - const __m512i bytes_0 = bytes_from_q4_0_twoblocks_avx512( blocks_0 ); - const __m512i bytes_1 = bytes_from_q4_0_twoblocks_avx512( blocks_1 ); - - // Now we want to compute dot products of 4-element byte vectors and store them in - // 32-bit integers. That is (only one 4-element vector is shown for clarity): - // +----+----+----+----+ - // ... | 03 | 02 | 01 | 00 | - // +----+----+----+----+ - // bytes_0 - // +----+----+----+----+ - // ... | D | C | B | A | - // +----+----+----+----+ - // bytes_1 - // +----+----+----+----+ - // ... | H | G | F | E | - // +----+----+----+----+ - // final_res_int - // +----+----+----+----+ - // ... | A*E+B*F+C*G+D*H | - // +----+----+----+----+ - const __m512i plus_8 = _mm512_set1_epi8( 8 ); - const __m512i bytes_1_minus_8 = _mm512_sub_epi8( bytes_1, plus_8 ); - -#ifdef __AVX512VNNI__ - // We have VPDPBUSDS in AVX512-VNNI, which does exactly what we want, but with a catch: - // the *left* operand is supposed to be unsigned, while Q4_0 quantization subtracts 8 - // from each nibble, so they can be negative. So, instead of `(bytes_0 - 8) * (bytes_1 - 8)`, - // we compute `bytes_0 * (bytes_1 - 8) + bytes_1 * (-8) + 64`. VPDPBUSDS uses an accumulator, - // which means we only need 2 instructions. - const __m512i dot_init = _mm512_set1_epi32( 4 * 64 ); - const __m512i minus_8 = _mm512_set1_epi8( -8 ); - const __m512i prod_0 = _mm512_dpbusds_epi32( dot_init, bytes_1, minus_8 ); - const __m512i final_res_int = _mm512_dpbusds_epi32( prod_0, bytes_0, bytes_1_minus_8 ); -#else - // As a fallback, we have VPMADDUBSW in AVX512-BW, which uses 16-bit products instead of 32-bit ones. - // It has the same catch as VPDPBUSDS: the left operand should be unsigned. - // This is essentially the AVX-512 version of the AVX-2 trick used by GH user Const-me - // ref: https://gist.github.com/Const-me/4d30e1fc767ab314596e16e90f53b6f4#file-matmultest-cpp-L119 - const __m512i one = _mm512_set1_epi16( 1 ); - const __m512i prod_0 = _mm512_maddubs_epi16( bytes_0, bytes_1_minus_8 ); - const __m512i prod_1 = _mm512_maddubs_epi16( plus_8, bytes_1_minus_8 ); - const __m512i diff = _mm512_sub_epi16( prod_0, prod_1 ); - const __m512i final_res_int = _mm512_madd_epi16( diff, one ); -#endif - - // Finally, we multiply the permuted scales and the 32-bit dot products, then accumulate. - const __m512 final_res_float = _mm512_cvtepi32_ps( final_res_int ); - return _mm512_fmadd_ps( permuted_scales, final_res_float, acc ); -} -#endif - inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y) { ggml_float sumf = 0.0; @@ -2313,546 +2014,16 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t *s = sumf; } -static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK4_0; - - assert(n % QK4_0 == 0); - assert(nb % 2 == 0); - - const block_q4_0 * restrict x = vx; - const block_q4_0 * restrict y = vy; - - float sumf = 0.0; - -#if defined(__ARM_NEON) - float sum0 = 0.0f; - float sum1 = 0.0f; - - for (int i = 0; i < nb; i += 2) { - const block_q4_0 * restrict x0 = &x[i + 0]; - const block_q4_0 * restrict y0 = &y[i + 0]; - const block_q4_0 * restrict x1 = &x[i + 1]; - const block_q4_0 * restrict y1 = &y[i + 1]; - - const uint8x16_t m4b = vdupq_n_u8(0xf); - const int8x16_t s8b = vdupq_n_s8(0x8); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v1_0 = vld1q_u8(y0->qs); - const uint8x16_t v0_1 = vld1q_u8(x1->qs); - const uint8x16_t v1_1 = vld1q_u8(y1->qs); - - // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8(v0_0, m4b)); - const int8x16_t v1_0l = vreinterpretq_s8_u8(vandq_u8(v1_0, m4b)); - const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v1_0h = vreinterpretq_s8_u8(vshrq_n_u8(v1_0, 4)); - - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8(v0_1, m4b)); - const int8x16_t v1_1l = vreinterpretq_s8_u8(vandq_u8(v1_1, m4b)); - const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - const int8x16_t v1_1h = vreinterpretq_s8_u8(vshrq_n_u8(v1_1, 4)); - - // sub 8 - const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); - const int8x16_t v1_0ls = vsubq_s8(v1_0l, s8b); - const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); - const int8x16_t v1_0hs = vsubq_s8(v1_0h, s8b); - - const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); - const int8x16_t v1_1ls = vsubq_s8(v1_1l, s8b); - const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - const int8x16_t v1_1hs = vsubq_s8(v1_1h, s8b); - -#if defined(__ARM_FEATURE_DOTPROD) - // dot product into int32x4_t - int32x4_t p_0 = vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls); - int32x4_t p_1 = vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls); - - p_0 = vdotq_s32(p_0, v0_0hs, v1_0hs); - p_1 = vdotq_s32(p_1, v0_1hs, v1_1hs); - - sum0 += x0->d*y0->d*vaddvq_s32(p_0); - sum1 += x1->d*y1->d*vaddvq_s32(p_1); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs)); - - const int16x8_t pl_0 = vaddq_s16(pl0l, pl0h); - const int16x8_t ph_0 = vaddq_s16(ph0l, ph0h); - - const int16x8_t pl_1 = vaddq_s16(pl1l, pl1h); - const int16x8_t ph_1 = vaddq_s16(ph1l, ph1h); - - const int16x8_t p_0 = vaddq_s16(pl_0, ph_0); - const int16x8_t p_1 = vaddq_s16(pl_1, ph_1); - - sum0 += x0->d*y0->d*vaddvq_s16(p_0); - sum1 += x1->d*y1->d*vaddvq_s16(p_1); -#endif - } - - sumf = sum0 + sum1; -#elif defined(__AVX512F__) - // Initialize accumulator with zeros - __m512 acc0 = _mm512_setzero_ps(); - __m512 acc1 = _mm512_setzero_ps(); - - const int superblock_size = 16; - - const int superblock_count = nb / superblock_size; - - for (int superblock_ix = 0; superblock_ix < superblock_count; superblock_ix += 1) { - int i = superblock_ix * superblock_size; - - acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+0 ); - acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+2 ); - acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+4 ); - acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+6 ); - acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+8 ); - acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+10 ); - acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i+12 ); - acc1 = dot_q4_0_twoblocks_avx512( acc1, x, y, i+14 ); - } - - // Remainders - for (int i = superblock_count * superblock_size; i < nb; i += 2) { - acc0 = dot_q4_0_twoblocks_avx512( acc0, x, y, i ); - } - - // Horizontal sum of all lanes of the accumulator - sumf = _mm512_reduce_add_ps( acc0 ) + _mm512_reduce_add_ps( acc1 ); -#elif defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - /* Prepare the constants we will need during execution */ - const __m256i lowMask = _mm256_set1_epi8( 0xF ); - const __m256i offset_8 = _mm256_set1_epi16( 8 ); - -#define UNROLL_COUNT 8 - // make sure we only unroll multiples of the block count - assert(nb % UNROLL_COUNT == 0); - - // Main loop - for (int i = 0; i < nb; i+=UNROLL_COUNT) { - // This loop will be unrolled by the compiler - for (int u=0;u we now have a vector of 8 int_32t */ - __m256i xy_q = _mm256_add_epi32( xy_high_q, xy_low_q ); - - /* Convert to vectore of 8 int32_t to 8 floats */ - __m256 q = _mm256_cvtepi32_ps( xy_q ); - - /* Multiply q with scale and accumulate */ - acc = _mm256_fmadd_ps( scale, q, acc ); - } - } - - // Return horizontal sum of the acc vector - __m128 res = _mm256_extractf128_ps( acc, 1 ); - res = _mm_add_ps( res, _mm256_castps256_ps128( acc ) ); - res = _mm_add_ps( res, _mm_movehl_ps( res, res ) ); - res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); - - sumf = _mm_cvtss_f32( res ); -#elif defined(__AVX__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; ++i) { - // Compute combined scale for the block - const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); - - __m128i i32[2]; - for (int j = 0; j < 2; ++j) { - // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes - __m128i bx = bytesFromNibbles( x[i].qs + 8*j ); - __m128i by = bytesFromNibbles( y[i].qs + 8*j ); - - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m128i off = _mm_set1_epi8( 8 ); - bx = _mm_sub_epi8( bx, off ); - by = _mm_sub_epi8( by, off ); - - // Get absolute values of x vectors - const __m128i ax = _mm_sign_epi8(bx, bx); - - // Sign the values of the y vectors - const __m128i sy = _mm_sign_epi8(by, bx); - - // Perform multiplication and create 16-bit values - const __m128i dot = _mm_maddubs_epi16(ax, sy); - - const __m128i ones = _mm_set1_epi16(1); - i32[j] = _mm_madd_epi16(ones, dot); - } - - // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] )); - // Apply the scale, and accumulate - acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); - } - - // Return horizontal sum of the acc vector - __m128 res = _mm256_extractf128_ps( acc, 1 ); - res = _mm_add_ps( res, _mm256_castps256_ps128( acc ) ); - res = _mm_add_ps( res, _mm_movehl_ps( res, res ) ); - res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); - - sumf = _mm_cvtss_f32( res ); -#elif defined(__wasm_simd128__) - // wasm simd - float sum0 = 0.0f; - float sum1 = 0.0f; - - for (int i = 0; i < nb; i += 2) { - const block_q4_0 * restrict x0 = &x[i + 0]; - const block_q4_0 * restrict y0 = &y[i + 0]; - const block_q4_0 * restrict x1 = &x[i + 1]; - const block_q4_0 * restrict y1 = &y[i + 1]; - - const v128_t m4b = wasm_u8x16_splat(0xf); - const v128_t s8b = wasm_i8x16_splat(0x8); - - const v128_t v0_0 = wasm_v128_load(x0->qs); - const v128_t v0_1 = wasm_v128_load(y0->qs); - const v128_t v1_0 = wasm_v128_load(x1->qs); - const v128_t v1_1 = wasm_v128_load(y1->qs); - - // 4-bit -> 8-bit - const v128_t v0_0l = wasm_v128_and(v0_0, m4b); - const v128_t v1_0l = wasm_v128_and(v1_0, m4b); - - const v128_t v0_0h = wasm_u8x16_shr(v0_0, 4); - const v128_t v1_0h = wasm_u8x16_shr(v1_0, 4); - - const v128_t v0_1l = wasm_v128_and(v0_1, m4b); - const v128_t v1_1l = wasm_v128_and(v1_1, m4b); - - const v128_t v0_1h = wasm_u8x16_shr(v0_1, 4); - const v128_t v1_1h = wasm_u8x16_shr(v1_1, 4); - - // sub 8 - const v128_t v0_0ls = wasm_i8x16_sub(v0_0l, s8b); - const v128_t v1_0ls = wasm_i8x16_sub(v1_0l, s8b); - - const v128_t v0_0hs = wasm_i8x16_sub(v0_0h, s8b); - const v128_t v1_0hs = wasm_i8x16_sub(v1_0h, s8b); - - const v128_t v0_1ls = wasm_i8x16_sub(v0_1l, s8b); - const v128_t v1_1ls = wasm_i8x16_sub(v1_1l, s8b); - - const v128_t v0_1hs = wasm_i8x16_sub(v0_1h, s8b); - const v128_t v1_1hs = wasm_i8x16_sub(v1_1h, s8b); - - // dot product into int16x8_t - const v128_t pl0l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_0ls), wasm_i16x8_extend_low_i8x16(v1_0ls)); - const v128_t pl0h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_0ls), wasm_i16x8_extend_high_i8x16(v1_0ls)); - - const v128_t ph0l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_0hs), wasm_i16x8_extend_low_i8x16(v1_0hs)); - const v128_t ph0h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_0hs), wasm_i16x8_extend_high_i8x16(v1_0hs)); - - const v128_t pl1l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_1ls), wasm_i16x8_extend_low_i8x16(v1_1ls)); - const v128_t pl1h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_1ls), wasm_i16x8_extend_high_i8x16(v1_1ls)); - - const v128_t ph1l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_1hs), wasm_i16x8_extend_low_i8x16(v1_1hs)); - const v128_t ph1h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_1hs), wasm_i16x8_extend_high_i8x16(v1_1hs)); - - const v128_t pl_0 = wasm_i16x8_add(pl0l, pl0h); - const v128_t ph_0 = wasm_i16x8_add(ph0l, ph0h); - - const v128_t pl_1 = wasm_i16x8_add(pl1l, pl1h); - const v128_t ph_1 = wasm_i16x8_add(ph1l, ph1h); - - const v128_t p_0 = wasm_i16x8_add(pl_0, ph_0); - const v128_t p_1 = wasm_i16x8_add(pl_1, ph_1); - - sum0 += x0->d * y0->d * ( - wasm_i16x8_extract_lane(p_0, 0) + wasm_i16x8_extract_lane(p_0, 1) + - wasm_i16x8_extract_lane(p_0, 2) + wasm_i16x8_extract_lane(p_0, 3) + - wasm_i16x8_extract_lane(p_0, 4) + wasm_i16x8_extract_lane(p_0, 5) + - wasm_i16x8_extract_lane(p_0, 6) + wasm_i16x8_extract_lane(p_0, 7)); - sum1 += x1->d * y1->d * ( - wasm_i16x8_extract_lane(p_1, 0) + wasm_i16x8_extract_lane(p_1, 1) + - wasm_i16x8_extract_lane(p_1, 2) + wasm_i16x8_extract_lane(p_1, 3) + - wasm_i16x8_extract_lane(p_1, 4) + wasm_i16x8_extract_lane(p_1, 5) + - wasm_i16x8_extract_lane(p_1, 6) + wasm_i16x8_extract_lane(p_1, 7)); - } - - sumf = sum0 + sum1; -#else - // scalar - for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float d1 = y[i].d; - - const uint8_t * restrict p0 = x[i].qs; - const uint8_t * restrict p1 = y[i].qs; - - int sumi = 0; - for (int j = 0; j < QK4_0/2; j++) { - const uint8_t v0 = p0[j]; - const uint8_t v1 = p1[j]; - - const int i0 = (v0 & 0xf) - 8; - const int i1 = (v0 >> 4) - 8; - - const int i2 = (v1 & 0xf) - 8; - const int i3 = (v1 >> 4) - 8; - - sumi += i0*i2 + i1*i3; - } - sumf += d0 * d1 * sumi; - } -#endif - - *s = sumf; -} - -static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK4_1; - - const block_q4_1 * restrict x = vx; - const block_q4_1 * restrict y = vy; - - float sumf = 0.0; - -#if defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - // Accumulator for constant offsets - float acc_offset = 0.0f; - - // Main loop - for (int i = 0; i < nb; ++i) { - const float * d0 = &x[i].d; - const float * d1 = &y[i].d; - - const float * m0 = &x[i].m; - const float * m1 = &y[i].m; - - const __m256 d0v = _mm256_broadcast_ss( d0 ); - const __m256 d1v = _mm256_broadcast_ss( d1 ); - const __m256 m0v = _mm256_broadcast_ss( m0 ); - const __m256 m1v = _mm256_broadcast_ss( m1 ); - - // Compute combined scale for the block - const __m256 scale_01 = _mm256_mul_ps( d0v, d1v ); - - // Compute cross scales for the block - const __m256 scale_0 = _mm256_mul_ps( d0v, m1v ); - const __m256 scale_1 = _mm256_mul_ps( m0v, d1v ); - const __m256 cross_scales = _mm256_blend_ps( scale_0, scale_1, 0xAA /* 0b10101010 */ ); - - // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes - __m256i bx = bytesFromNibbles( x[i].qs ); - __m256i by = bytesFromNibbles( y[i].qs ); - - // Now we have a vector with bytes in [ 0 .. 15 ] interval. - - // Sign-extend first 16 signed bytes into int16_t - __m256i x16 = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( bx ) ); - __m256i y16 = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( by ) ); - // Compute products of int16_t integers, add pairwise - __m256i i32 = _mm256_madd_epi16( x16, y16 ); - - // Sign-extend last 16 signed bytes into int16_t vectors - __m256i x16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( bx, 1 ) ); - __m256i y16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( by, 1 ) ); - // Accumulate products of int16_t integers - i32 = _mm256_add_epi32( i32, _mm256_madd_epi16( x16_h, y16_h ) ); - - // compute sums of unsigned bytes in bx, by in blocks of 8. - // This results in a layout like X100 0000 X200 0000 X300 0000 X400 0000, - // which we then interleave as X100 Y100 X200 Y200 X300 Y300 X400 Y400. - // so if we then cast to 8 singles, we get 8 floats like [ x0_7, y0_7, x8_15, y8_15, x16_23, y16_23, x24_31, y24_31 ] - __m256i xsumi = _mm256_sad_epu8( bx, _mm256_setzero_si256() ); - __m256i ysumi = _mm256_sad_epu8( by, _mm256_setzero_si256() ); - __m256i sumsi = _mm256_or_si256( xsumi, _mm256_slli_si256( ysumi, 4 ) ); - __m256 sums = _mm256_cvtepi32_ps( sumsi ); - - // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps( i32 ); - // Apply the scale, and accumulate - // acc += d0*d1*x*y + d0*m1*x + d1*m0*y - acc = _mm256_fmadd_ps( scale_01, p, acc ); - acc = _mm256_fmadd_ps( cross_scales, sums, acc ); - // acc_offset += m0*m1 (for each entry in the block) - acc_offset += (*m0)*(*m1); - } - - // Return horizontal sum of the acc vector - __m128 res = _mm256_extractf128_ps( acc, 1 ); - res = _mm_add_ps( res, _mm256_castps256_ps128( acc ) ); - res = _mm_add_ps( res, _mm_movehl_ps( res, res ) ); - res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); - - sumf = _mm_cvtss_f32( res ) + acc_offset * QK4_1; -#elif defined(__ARM_NEON) - float sum00 = 0.0f; - float sum01 = 0.0f; - float sum10 = 0.0f; - float sum11 = 0.0f; - - for (int i = 0; i < nb; i += 2) { - const block_q4_1 * restrict x0 = &x[i + 0]; - const block_q4_1 * restrict y0 = &y[i + 0]; - const block_q4_1 * restrict x1 = &x[i + 1]; - const block_q4_1 * restrict y1 = &y[i + 1]; - - const uint8x16_t m4b = vdupq_n_u8(0xf); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v1_0 = vld1q_u8(y0->qs); - const uint8x16_t v0_1 = vld1q_u8(x1->qs); - const uint8x16_t v1_1 = vld1q_u8(y1->qs); - - // 4-bit -> 8-bit - const uint8x16_t v0_0l = vandq_u8(v0_0, m4b); - const uint8x16_t v1_0l = vandq_u8(v1_0, m4b); - const uint8x16_t v0_0h = vshrq_n_u8(v0_0, 4); - const uint8x16_t v1_0h = vshrq_n_u8(v1_0, 4); - - const uint8x16_t v0_1l = vandq_u8(v0_1, m4b); - const uint8x16_t v1_1l = vandq_u8(v1_1, m4b); - const uint8x16_t v0_1h = vshrq_n_u8(v0_1, 4); - const uint8x16_t v1_1h = vshrq_n_u8(v1_1, 4); - - sum00 += x0->m*y0->m; - sum01 += y0->m*x0->d*((uint16_t)vaddvq_u8(v0_0l) + (uint16_t)vaddvq_u8(v0_0h)); - sum10 += x0->m*y0->d*((uint16_t)vaddvq_u8(v1_0l) + (uint16_t)vaddvq_u8(v1_0h)); - - sum00 += x1->m*y1->m; - sum01 += y1->m*x1->d*((uint16_t)vaddvq_u8(v0_1l) + (uint16_t)vaddvq_u8(v0_1h)); - sum10 += x1->m*y1->d*((uint16_t)vaddvq_u8(v1_1l) + (uint16_t)vaddvq_u8(v1_1h)); - -#if defined(__ARM_FEATURE_DOTPROD) - // dot product into int32x4_t - uint32x4_t p_0 = vdotq_u32(vdupq_n_u32(0), v0_0l, v1_0l); - uint32x4_t p_1 = vdotq_u32(vdupq_n_u32(0), v0_1l, v1_1l); - - p_0 = vdotq_u32(p_0, v0_0h, v1_0h); - p_1 = vdotq_u32(p_1, v0_1h, v1_1h); - - sum11 += x0->d*y0->d*vaddvq_u32(p_0); - sum11 += x1->d*y1->d*vaddvq_u32(p_1); -#else - const uint16x8_t pl0l = vmull_u8(vget_low_u8 (v0_0l), vget_low_u8 (v1_0l)); - const uint16x8_t pl0h = vmull_u8(vget_high_u8(v0_0l), vget_high_u8(v1_0l)); - const uint16x8_t ph0l = vmull_u8(vget_low_u8 (v0_0h), vget_low_u8 (v1_0h)); - const uint16x8_t ph0h = vmull_u8(vget_high_u8(v0_0h), vget_high_u8(v1_0h)); - - const uint16x8_t pl1l = vmull_u8(vget_low_u8 (v0_1l), vget_low_u8 (v1_1l)); - const uint16x8_t pl1h = vmull_u8(vget_high_u8(v0_1l), vget_high_u8(v1_1l)); - const uint16x8_t ph1l = vmull_u8(vget_low_u8 (v0_1h), vget_low_u8 (v1_1h)); - const uint16x8_t ph1h = vmull_u8(vget_high_u8(v0_1h), vget_high_u8(v1_1h)); - - const uint16x8_t pl_0 = vaddq_u16(pl0l, pl0h); - const uint16x8_t ph_0 = vaddq_u16(ph0l, ph0h); - - const uint16x8_t pl_1 = vaddq_u16(pl1l, pl1h); - const uint16x8_t ph_1 = vaddq_u16(ph1l, ph1h); - - const uint16x8_t p_0 = vaddq_u16(pl_0, ph_0); - const uint16x8_t p_1 = vaddq_u16(pl_1, ph_1); - - sum11 += x0->d*y0->d*vaddvq_u16(p_0); - sum11 += x1->d*y1->d*vaddvq_u16(p_1); -#endif - } - - sumf = QK4_1*sum00 + sum01 + sum10 + sum11; -#else - // scalar - for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float d1 = y[i].d; - - const float m0 = x[i].m; - const float m1 = y[i].m; - - const uint8_t * restrict p0 = x[i].qs; - const uint8_t * restrict p1 = y[i].qs; - - for (int j = 0; j < QK4_1/2; j++) { - const uint8_t v0 = p0[j]; - const uint8_t v1 = p1[j]; - - const float f0 = d0*(v0 & 0xf) + m0; - const float f1 = d0*(v0 >> 4) + m0; - - const float f2 = d1*(v1 & 0xf) + m1; - const float f3 = d1*(v1 >> 4) + m1; - - sumf += f0*f2 + f1*f3; - } - } -#endif - - *s = sumf; -} - static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); const block_q4_0 * restrict x = vx; const block_q8_0 * restrict y = vy; - float sumf = 0.0; - #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -2863,7 +2034,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const block_q8_0 * restrict y0 = &y[i + 0]; const block_q8_0 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0xf); + const uint8x16_t m4b = vdupq_n_u8(0x0F); const int8x16_t s8b = vdupq_n_s8(0x8); const uint8x16_t v0_0 = vld1q_u8(x0->qs); @@ -2887,29 +2058,23 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - // interleave - const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h); - const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h); - const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h); - const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h); - #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs); + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); @@ -2921,7 +2086,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * #endif } - sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); #elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); @@ -2931,7 +2096,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * /* Compute combined scale for the block */ const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); - __m256i bx = bytesFromNibbles(x[i].qs); + __m256i bx = bytes_from_nibbles_32(x[i].qs); // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. const __m256i off = _mm256_set1_epi8( 8 ); @@ -2939,32 +2104,13 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - // Get absolute values of x vectors - const __m256i ax = _mm256_sign_epi8(bx, bx); - - // Sign the values of the y vectors - const __m256i sy = _mm256_sign_epi8(by, bx); - - // Perform multiplication and create 16-bit values - const __m256i dot = _mm256_maddubs_epi16(ax, sy); - - const __m256i ones = _mm256_set1_epi16(1); - __m256i xy_q = _mm256_madd_epi16(ones, dot); - - /* Convert to vectore of 8 int32_t to 8 floats */ - __m256 q = _mm256_cvtepi32_ps( xy_q ); + const __m256 q = mul_sum_i8_pairs_float(bx, by); /* Multiply q with scale and accumulate */ acc = _mm256_fmadd_ps( d, q, acc ); } - // Return horizontal sum of the acc vector - __m128 res = _mm256_extractf128_ps( acc, 1 ); - res = _mm_add_ps( res, _mm256_castps256_ps128( acc ) ); - res = _mm_add_ps( res, _mm_movehl_ps( res, res ) ); - res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); - - sumf = _mm_cvtss_f32( res ); + *s = hsum_float_8(acc); #elif defined(__AVX__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); @@ -2974,99 +2120,199 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * // Compute combined scale for the block const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); - __m128i i32[2]; - for (int j = 0; j < 2; ++j) { - // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes - __m128i bx = bytesFromNibbles( x[i].qs + 8*j ); - __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16*j)); + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m128i off = _mm_set1_epi8( 8 ); - bx = _mm_sub_epi8( bx, off ); + const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs); - // Get absolute values of x vectors - const __m128i ax = _mm_sign_epi8(bx, bx); + __m128i bx = _mm_and_si128(lowMask, tmp); + __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx, by); - // Sign the values of the y vectors - const __m128i sy = _mm_sign_epi8(by, bx); - - // Perform multiplication and create 16-bit values - const __m128i dot = _mm_maddubs_epi16(ax, sy); - - const __m128i ones = _mm_set1_epi16(1); - i32[j] = _mm_madd_epi16(ones, dot); - } + bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); + by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx, by); // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] )); + __m256 p = _mm256_cvtepi32_ps(_mm256_set_m128i(i32_0, i32_1)); + // Apply the scale, and accumulate acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); } - // Return horizontal sum of the acc vector - __m128 res = _mm256_extractf128_ps( acc, 1 ); - res = _mm_add_ps( res, _mm256_castps256_ps128( acc ) ); - res = _mm_add_ps( res, _mm_movehl_ps( res, res ) ); - res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); + *s = hsum_float_8(acc); +#elif defined(__SSSE3__) + // set constants + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); - sumf = _mm_cvtss_f32( res ); + // Initialize accumulator with zeros + __m128 acc_0 = _mm_setzero_ps(); + __m128 acc_1 = _mm_setzero_ps(); + __m128 acc_2 = _mm_setzero_ps(); + __m128 acc_3 = _mm_setzero_ps(); + + // First round without accumulation + { + _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 0 and 1 + const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[0].d ), _mm_set1_ps( y[0].d ) ); + + const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs); + + __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); + + __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); + __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16)); + bx_1 = _mm_sub_epi8(bx_1, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); + + _mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 2 and 3 + const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[1].d ), _mm_set1_ps( y[1].d ) ); + + const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs); + + __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); + __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs); + bx_2 = _mm_sub_epi8(bx_2, off); + const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); + + __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); + __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16)); + bx_3 = _mm_sub_epi8(bx_3, off); + const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); + + // Convert int32_t to float + __m128 p0 = _mm_cvtepi32_ps(i32_0); + __m128 p1 = _mm_cvtepi32_ps(i32_1); + __m128 p2 = _mm_cvtepi32_ps(i32_2); + __m128 p3 = _mm_cvtepi32_ps(i32_3); + + // Apply the scale + acc_0 = _mm_mul_ps( d_0_1, p0 ); + acc_1 = _mm_mul_ps( d_0_1, p1 ); + acc_2 = _mm_mul_ps( d_2_3, p2 ); + acc_3 = _mm_mul_ps( d_2_3, p3 ); + } + + // Main loop + for (int i = 2; i < nb; i+=2) { + _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 0 and 1 + const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[i].d ), _mm_set1_ps( y[i].d ) ); + + const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs); + + __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); + + __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); + __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx_1 = _mm_sub_epi8(bx_1, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); + + _mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 2 and 3 + const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[i + 1].d ), _mm_set1_ps( y[i + 1].d ) ); + + const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs); + + __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); + __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs); + bx_2 = _mm_sub_epi8(bx_2, off); + const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); + + __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); + __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16)); + bx_3 = _mm_sub_epi8(bx_3, off); + const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); + + // Convert int32_t to float + __m128 p0 = _mm_cvtepi32_ps(i32_0); + __m128 p1 = _mm_cvtepi32_ps(i32_1); + __m128 p2 = _mm_cvtepi32_ps(i32_2); + __m128 p3 = _mm_cvtepi32_ps(i32_3); + + // Apply the scale + __m128 p0_d = _mm_mul_ps( d_0_1, p0 ); + __m128 p1_d = _mm_mul_ps( d_0_1, p1 ); + __m128 p2_d = _mm_mul_ps( d_2_3, p2 ); + __m128 p3_d = _mm_mul_ps( d_2_3, p3 ); + + // Acummulate + acc_0 = _mm_add_ps(p0_d, acc_0); + acc_1 = _mm_add_ps(p1_d, acc_1); + acc_2 = _mm_add_ps(p2_d, acc_2); + acc_3 = _mm_add_ps(p3_d, acc_3); + } + + *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); #else // scalar - for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float d1 = y[i].d; - - const uint8_t * restrict p0 = x[i].qs; - const int8_t * restrict p1 = y[i].qs; - - int sumi = 0; - for (int j = 0; j < QK8_0/2; j++) { - const uint8_t v0 = p0[j]; - - const int i0 = (int8_t) (v0 & 0xf) - 8; - const int i1 = (int8_t) (v0 >> 4) - 8; - - const int i2 = p1[2*j + 0]; - const int i3 = p1[2*j + 1]; - - sumi += i0*i2 + i1*i3; - } - sumf += d0*d1*sumi; - } -#endif - - *s = sumf; -} - -static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; - - assert(n % QK8_0 == 0); - assert(nb % 2 == 0); - assert(QK8_0 == 2*QK4_2); - - const block_q4_2 * restrict x = vx; - const block_q8_0 * restrict y = vy; - float sumf = 0.0; + for (int i = 0; i < nb; i++) { + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F) - 8; + const int v1 = (x[i].qs[j] >> 4) - 8; + + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); + } + + sumf += (x[i].d*y[i].d)*sumi; + } + + *s = sumf; +#endif +} + +static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_1; + const int nb = n / qk; + + assert(n % qk == 0); + assert(nb % 2 == 0); + + const block_q4_1 * restrict x = vx; + const block_q8_1 * restrict y = vy; + + // TODO: add WASM SIMD #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); + float summs = 0; + for (int i = 0; i < nb; i += 2) { - const block_q4_2 * restrict x0_0 = &x[2*(i + 0) + 0]; - const block_q4_2 * restrict x0_1 = &x[2*(i + 0) + 1]; - const block_q4_2 * restrict x1_0 = &x[2*(i + 1) + 0]; - const block_q4_2 * restrict x1_1 = &x[2*(i + 1) + 1]; - const block_q8_0 * restrict y0 = &y[i + 0]; - const block_q8_0 * restrict y1 = &y[i + 1]; + const block_q4_1 * restrict x0 = &x[i + 0]; + const block_q4_1 * restrict x1 = &x[i + 1]; + const block_q8_1 * restrict y0 = &y[i + 0]; + const block_q8_1 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0xf); - const int8x16_t s8b = vdupq_n_s8(0x8); + summs += x0->m * y0->s + x1->m * y1->s; - const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); - const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); // 4-bit -> 8-bit const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); @@ -3074,18 +2320,6 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // sub 8 - const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); - const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); - const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); - const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); - const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); - // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); @@ -3093,79 +2327,623 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); #if defined(__ARM_FEATURE_DOTPROD) - sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), GGML_FP16_TO_FP32(x0_0->d)), - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), GGML_FP16_TO_FP32(x0_1->d))), y0->d); + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); - sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), GGML_FP16_TO_FP32(x1_0->d)), - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), GGML_FP16_TO_FP32(x1_1->d))), y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(pl0), GGML_FP16_TO_FP32(x0_0->d)), - vmulq_n_f32(vcvtq_f32_s32(ph0), GGML_FP16_TO_FP32(x0_1->d))), y0->d); - - sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(pl1), GGML_FP16_TO_FP32(x1_0->d)), - vmulq_n_f32(vcvtq_f32_s32(ph1), GGML_FP16_TO_FP32(x1_1->d))), y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); #endif } - sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + float summs = 0; + + // Main loop + for (int i = 0; i < nb; ++i) { + const float * d0 = &x[i].d; + const float * d1 = &y[i].d; + + summs += x[i].m * y[i].s; + + const __m256 d0v = _mm256_broadcast_ss( d0 ); + const __m256 d1v = _mm256_broadcast_ss( d1 ); + + // Compute combined scales + const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); + + // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes + const __m256i bx = bytes_from_nibbles_32(x[i].qs); + const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); + + const __m256 xy = mul_sum_i8_pairs_float(bx, by); + + // Accumulate d0*d1*x*y + acc = _mm256_fmadd_ps( d0d1, xy, acc ); + } + + *s = hsum_float_8(acc) + summs; #else // scalar + float sumf = 0.0; + for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[2*i + 0].qs; - const uint8_t * restrict x1 = x[2*i + 1].qs; - const int8_t * restrict y0 = y[i].qs; + int sumi = 0; - const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d); - const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d); + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F); + const int v1 = (x[i].qs[j] >> 4); - int sumi_0 = 0; - int sumi_1 = 0; - - for (int j = 0; j < QK8_0/4; j++) { - const uint8_t v0 = x0[j]; - const uint8_t v1 = x1[j]; - - const int i0_0 = (int8_t) (v0 & 0xf) - 8; - const int i1_0 = (int8_t) (v0 >> 4) - 8; - - const int i0_1 = (int8_t) (v1 & 0xf) - 8; - const int i1_1 = (int8_t) (v1 >> 4) - 8; - - const int i2_0 = y0[2*j + 0]; - const int i3_0 = y0[2*j + 1]; - - const int i2_1 = y0[2*(j + QK8_0/4) + 0]; - const int i3_1 = y0[2*(j + QK8_0/4) + 1]; - - sumi_0 += i0_0*i2_0 + i1_0*i3_0; - sumi_1 += i0_1*i2_1 + i1_1*i3_1; + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += (d0 * y[i].d) * sumi_0; - sumf += (d1 * y[i].d) * sumi_1; + sumf += (x[i].d*y[i].d)*sumi + x[i].m*y[i].s; } -#endif *s = sumf; +#endif +} + +static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; + + assert(n % qk == 0); + assert(nb % 2 == 0); + assert(qk == QK5_0); + + const block_q5_0 * restrict x = vx; + const block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + uint32_t qh0; + uint32_t qh1; + + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + for (int i = 0; i < nb; i += 2) { + const block_q5_0 * restrict x0 = &x[i]; + const block_q5_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + // extract the 5th bit via lookup table ((!b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_1[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_1[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + const float x1d = GGML_FP16_TO_FP32(x1->d); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__wasm_simd128__) + v128_t sumv = wasm_f32x4_splat(0.0f); + + uint32_t qh; + uint64_t tmp[4]; + + // TODO: check if unrolling this is better + for (int i = 0; i < nb; ++i) { + const block_q5_0 * restrict x0 = &x[i]; + const block_q8_0 * restrict y0 = &y[i]; + + const v128_t m4b = wasm_i8x16_splat(0x0F); + const v128_t s16b = wasm_i8x16_splat(0x10); + + // extract the 5th bit + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_1[(qh >> 24) ]; + + const v128_t qhl = wasm_v128_load(tmp + 0); + const v128_t qhh = wasm_v128_load(tmp + 2); + + const v128_t v0 = wasm_v128_load(x0->qs); + + // 4-bit -> 8-bit + const v128_t v0l = wasm_v128_and (v0, m4b); + const v128_t v0h = wasm_u8x16_shr(v0, 4); + + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const v128_t v0lf = wasm_i8x16_sub(v0l, qhl); + const v128_t v0hf = wasm_i8x16_sub(v0h, qhh); + + // load y + const v128_t v1l = wasm_v128_load(y0->qs); + const v128_t v1h = wasm_v128_load(y0->qs + 16); + + // int8x16 -> int16x8 + const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); + const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); + const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); + const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); + + const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); + const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); + const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); + const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + + // dot product + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( + wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d))); + } + + *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; i++) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d)); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bxhi = bytes_from_bits_32(x[i].qh); + bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0)); + bx = _mm256_or_si256(bx, bxhi); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps(d, q, acc); + } + + *s = hsum_float_8(acc); +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); + + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; + + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); + } + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi; + } + + *s = sumf; +#endif +} + +static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_1; + const int nb = n / qk; + + assert(n % qk == 0); + assert(nb % 2 == 0); + assert(qk == QK5_1); + + const block_q5_1 * restrict x = vx; + const block_q8_1 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + float summs0 = 0.0f; + float summs1 = 0.0f; + + uint32_t qh0; + uint32_t qh1; + + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + for (int i = 0; i < nb; i += 2) { + const block_q5_1 * restrict x0 = &x[i]; + const block_q5_1 * restrict x1 = &x[i + 1]; + const block_q8_1 * restrict y0 = &y[i]; + const block_q8_1 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s; + summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s; + + // extract the 5th bit via lookup table ((b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_0[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_0[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // add high bit + const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + const float x1d = GGML_FP16_TO_FP32(x1->d); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; +#elif defined(__wasm_simd128__) + v128_t sumv = wasm_f32x4_splat(0.0f); + + float summs = 0.0f; + + uint32_t qh; + uint64_t tmp[4]; + + // TODO: check if unrolling this is better + for (int i = 0; i < nb; ++i) { + const block_q5_1 * restrict x0 = &x[i]; + const block_q8_1 * restrict y0 = &y[i]; + + summs += GGML_FP16_TO_FP32(x0->m) * y0->s; + + const v128_t m4b = wasm_i8x16_splat(0x0F); + + // extract the 5th bit + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_0[(qh >> 24) ]; + + const v128_t qhl = wasm_v128_load(tmp + 0); + const v128_t qhh = wasm_v128_load(tmp + 2); + + const v128_t v0 = wasm_v128_load(x0->qs); + + // 4-bit -> 8-bit + const v128_t v0l = wasm_v128_and (v0, m4b); + const v128_t v0h = wasm_u8x16_shr(v0, 4); + + static bool x = true; + + // add high bit + const v128_t v0lf = wasm_v128_or(v0l, qhl); + const v128_t v0hf = wasm_v128_or(v0h, qhh); + + // load y + const v128_t v1l = wasm_v128_load(y0->qs); + const v128_t v1h = wasm_v128_load(y0->qs + 16); + + // int8x16 -> int16x8 + const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); + const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); + const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); + const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); + + const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); + const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); + const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); + const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + + // dot product + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( + wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d))); + } + + *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs; +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + float summs = 0.0f; + + // Main loop + for (int i = 0; i < nb; i++) { + const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); + + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bxhi = bytes_from_bits_32(x[i].qh); + bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); + bx = _mm256_or_si256(bx, bxhi); + + const __m256 dy = _mm256_broadcast_ss(&y[i].d); + const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); + } + + *s = hsum_float_8(acc) + summs; +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + int sumi = 0; + + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0; + const int32_t x1 = (x[i].qs[j] >> 4) | xh_1; + + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); + } + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; +#endif +} + +static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; + + assert(n % qk == 0); + assert(nb % 2 == 0); + + const block_q8_0 * restrict x = vx; + const block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_q8_0 * restrict x0 = &x[i + 0]; + const block_q8_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const int8x16_t x0_0 = vld1q_s8(x0->qs); + const int8x16_t x0_1 = vld1q_s8(x0->qs + 16); + const int8x16_t x1_0 = vld1q_s8(x1->qs); + const int8x16_t x1_1 = vld1q_s8(x1->qs + 16); + + // load y + const int8x16_t y0_0 = vld1q_s8(y0->qs); + const int8x16_t y0_1 = vld1q_s8(y0->qs + 16); + const int8x16_t y1_0 = vld1q_s8(y1->qs); + const int8x16_t y1_1 = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), + vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), x0->d*y0->d); + + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), + vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), x1->d*y1->d); + +#else + const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0)); + const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0)); + const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1)); + const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1)); + + const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0)); + const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0)); + const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1)); + const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1)); + + const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1)); + const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3)); + const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1)); + const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), x1->d*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + // Compute combined scale for the block + const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs); + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + // Multiply q with scale and accumulate + acc = _mm256_fmadd_ps( d, q, acc ); + } + + *s = hsum_float_8(acc); +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + int sumi = 0; + + for (int j = 0; j < qk; j++) { + sumi += x[i].qs[j]*y[i].qs[j]; + } + + sumf += (x[i].d*y[i].d)*sumi; + } + + *s = sumf; +#endif } // compute GGML_VEC_DOT_UNROLL dot products at once @@ -3287,6 +3065,7 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, x, x); *s = sqrtf(*s); } inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } +inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } @@ -3328,12 +3107,12 @@ inline static float ggml_silu_f32(float x) { return x/(1.0f + expf(-x)); } -inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { - const uint16_t * i16 = (const uint16_t *) x; - for (int i = 0; i < n; ++i) { - y[i] = table_silu_f16[i16[i]]; - } -} +//inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { +// const uint16_t * i16 = (const uint16_t *) x; +// for (int i = 0; i < n; ++i) { +// y[i] = table_silu_f16[i16[i]]; +// } +//} #ifdef GGML_SILU_FP16 inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) { @@ -3352,6 +3131,29 @@ inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) { } #endif +inline static float ggml_silu_backward_f32(float x, float dy) { + const float s = 1.0f/(1.0f + expf(-x)); + return dy*s*(1.0f + x*(1.0f - s)); +} + +#ifdef GGML_SILU_FP16 +inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { + for (int i = 0; i < n; ++i) { + // we did not use x[i] to compute forward silu but its f16 equivalent + // take derivative at f16 of x[i]: + ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); + float usedx = GGML_FP16_TO_FP32(fp16); + dx[i] = ggml_silu_backward_f32(usedx, dy[i]); + } +} +#else +inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { + for (int i = 0; i < n; ++i) { + dx[i] = ggml_silu_backward_f32(x[i], dy[i]); + } +} +#endif + inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) { #ifndef GGML_USE_ACCELERATE ggml_float sum = 0.0; @@ -3364,6 +3166,14 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) { #endif } +inline static void ggml_vec_sum_ggf(const int n, ggml_float * s, const float * x) { + ggml_float sum = 0.0; + for (int i = 0; i < n; ++i) { + sum += (ggml_float)x[i]; + } + *s = sum; +} + inline static void ggml_vec_max_f32(const int n, float * s, const float * x) { #ifndef GGML_USE_ACCELERATE float max = -INFINITY; @@ -3414,26 +3224,30 @@ static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = { [GGML_TYPE_F16] = 1, [GGML_TYPE_Q4_0] = QK4_0, [GGML_TYPE_Q4_1] = QK4_1, - [GGML_TYPE_Q4_2] = QK4_2, + [GGML_TYPE_Q5_0] = QK5_0, + [GGML_TYPE_Q5_1] = QK5_1, [GGML_TYPE_Q8_0] = QK8_0, + [GGML_TYPE_Q8_1] = QK8_1, [GGML_TYPE_I8] = 1, [GGML_TYPE_I16] = 1, [GGML_TYPE_I32] = 1, }; -static_assert(GGML_TYPE_COUNT == 9, "GGML_BLCK_SIZE is outdated"); +static_assert(GGML_TYPE_COUNT == 13, "GGML_BLCK_SIZE is outdated"); static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = sizeof(float), [GGML_TYPE_F16] = sizeof(ggml_fp16_t), [GGML_TYPE_Q4_0] = sizeof(block_q4_0), [GGML_TYPE_Q4_1] = sizeof(block_q4_1), - [GGML_TYPE_Q4_2] = sizeof(block_q4_2), + [GGML_TYPE_Q5_0] = sizeof(block_q5_0), + [GGML_TYPE_Q5_1] = sizeof(block_q5_1), [GGML_TYPE_Q8_0] = sizeof(block_q8_0), + [GGML_TYPE_Q8_1] = sizeof(block_q8_1), [GGML_TYPE_I8] = sizeof(int8_t), [GGML_TYPE_I16] = sizeof(int16_t), [GGML_TYPE_I32] = sizeof(int32_t), }; -static_assert(GGML_TYPE_COUNT == 9, "GGML_TYPE_SIZE is outdated"); +static_assert(GGML_TYPE_COUNT == 13, "GGML_TYPE_SIZE is outdated"); static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = { @@ -3441,38 +3255,46 @@ static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = { [GGML_TYPE_F16] = "f16", [GGML_TYPE_Q4_0] = "q4_0", [GGML_TYPE_Q4_1] = "q4_1", - [GGML_TYPE_Q4_2] = "q4_2", + [GGML_TYPE_Q5_0] = "q5_0", + [GGML_TYPE_Q5_1] = "q5_1", [GGML_TYPE_Q8_0] = "q8_0", + [GGML_TYPE_Q8_1] = "q8_1", [GGML_TYPE_I8] = "i8", [GGML_TYPE_I16] = "i16", [GGML_TYPE_I32] = "i32", }; -static_assert(GGML_TYPE_COUNT == 9, "GGML_TYPE_NAME is outdated"); +static_assert(GGML_TYPE_COUNT == 13, "GGML_TYPE_NAME is outdated"); static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = false, [GGML_TYPE_F16] = false, [GGML_TYPE_Q4_0] = true, [GGML_TYPE_Q4_1] = true, - [GGML_TYPE_Q4_2] = true, + [GGML_TYPE_Q5_0] = true, + [GGML_TYPE_Q5_1] = true, [GGML_TYPE_Q8_0] = true, + [GGML_TYPE_Q8_1] = true, [GGML_TYPE_I8] = false, [GGML_TYPE_I16] = false, [GGML_TYPE_I32] = false, }; -static_assert(GGML_TYPE_COUNT == 9, "GGML_IS_QUANTIZED is outdated"); +static_assert(GGML_TYPE_COUNT == 13, "GGML_IS_QUANTIZED is outdated"); static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "NONE", "DUP", "ADD", + "ADD1", + "ACC", "SUB", "MUL", "DIV", "SQR", "SQRT", + "LOG", "SUM", + "SUM_ROWS", "MEAN", "REPEAT", "ABS", @@ -3482,12 +3304,15 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "RELU", "GELU", "SILU", + "SILU_BACK", "NORM", "RMS_NORM", + "RMS_NORM_BACK", "MUL_MAT", "SCALE", + "SET", "CPY", "CONT", "RESHAPE", @@ -3495,9 +3320,14 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "PERMUTE", "TRANSPOSE", "GET_ROWS", + "GET_ROWS_BACK", + "DIAG", "DIAG_MASK_INF", + "DIAG_MASK_ZERO", "SOFT_MAX", "ROPE", + "ROPE_BACK", + "ALIBI", "CONV_1D_1S", "CONV_1D_2S", @@ -3508,19 +3338,23 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "MAP_BINARY", }; -static_assert(GGML_OP_COUNT == 38, "GGML_OP_COUNT != 38"); +static_assert(GGML_OP_COUNT == 50, "GGML_OP_COUNT != 50"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", "x", "x+y", + "x+y", + "view(x,nb,offset)+=y->x", "x-y", "x*y", "x/y", "x^2", "√x", + "log(x)", "Σx", + "Σx_k", "Σx/n", "repeat(x)", "abs(x)", @@ -3530,12 +3364,15 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "relu(x)", "gelu(x)", "silu(x)", + "silu_back(x)", "norm(x)", "rms_norm(x)", + "rms_norm_back(x)", "X*Y", "x*v", + "y-\\>view(x)", "x-\\>y", "cont(x)", "reshape(x)", @@ -3543,9 +3380,14 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "permute(x)", "transpose(x)", "get_rows(x)", + "get_rows_back(x)", + "diag(x)", "diag_mask_inf(x)", + "diag_mask_zero(x)", "soft_max(x)", "rope(x)", + "rope_back(x)", + "alibi(x)", "conv_1d_1s(x)", "conv_1d_2s(x)", @@ -3556,7 +3398,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "f(x,y)", }; -static_assert(GGML_OP_COUNT == 38, "GGML_OP_COUNT != 38"); +static_assert(GGML_OP_COUNT == 50, "GGML_OP_COUNT != 50"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); @@ -3722,10 +3564,30 @@ static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct (t0->ne[3] == t1->ne[3]); } -static inline bool ggml_is_quantized(enum ggml_type type) { +bool ggml_is_quantized(enum ggml_type type) { return GGML_IS_QUANTIZED[type]; } +enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { + enum ggml_type wtype = GGML_TYPE_COUNT; + + switch (ftype) { + case GGML_FTYPE_ALL_F32: wtype = GGML_TYPE_F32; break; + case GGML_FTYPE_MOSTLY_F16: wtype = GGML_TYPE_F16; break; + case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break; + case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break; + case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break; + case GGML_FTYPE_MOSTLY_Q5_1: wtype = GGML_TYPE_Q5_1; break; + case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break; + case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break; + case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break; + } + + GGML_ASSERT(wtype != GGML_TYPE_COUNT); + + return wtype; +} + static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) { return tensor->nb[0] > tensor->nb[1]; } @@ -3774,9 +3636,9 @@ static inline int ggml_up32(int n) { return (n + 31) & ~31; } -static inline int ggml_up64(int n) { - return (n + 63) & ~63; -} +//static inline int ggml_up64(int n) { +// return (n + 63) & ~63; +//} static inline int ggml_up(int n, int m) { // assert m is a power of 2 @@ -3836,6 +3698,12 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f); } +#if defined(GGML_USE_CUBLAS) + ggml_init_cublas(); +#elif defined(GGML_USE_CLBLAST) + ggml_cl_init(); +#endif + is_first_call = false; } @@ -3915,7 +3783,7 @@ void ggml_free(struct ggml_context * ctx) { } size_t ggml_used_mem(const struct ggml_context * ctx) { - return ctx->objects_end->offs + ctx->objects_end->size; + return ctx->objects_end == NULL ? 0 : ctx->objects_end->offs + ctx->objects_end->size; } size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) { @@ -4014,6 +3882,7 @@ struct ggml_tensor * ggml_new_tensor_impl( *result = (struct ggml_tensor) { /*.type =*/ type, + /*.backend =*/ GGML_BACKEND_CPU, /*.n_dims =*/ n_dims, /*.ne =*/ { 1, 1, 1, 1 }, /*.nb =*/ { 0, 0, 0, 0 }, @@ -4028,6 +3897,7 @@ struct ggml_tensor * ggml_new_tensor_impl( /*.perf_cycles =*/ 0, /*.perf_time_us =*/ 0, /*.data =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data, + /*.name =*/ { 0 }, /*.pad =*/ { 0 }, }; @@ -4382,6 +4252,15 @@ float * ggml_get_data_f32(const struct ggml_tensor * tensor) { return (float *)(tensor->data); } +const char * ggml_get_name(const struct ggml_tensor * tensor) { + return tensor->name; +} + +void ggml_set_name(struct ggml_tensor * tensor, const char * name) { + strncpy(tensor->name, name, sizeof(tensor->name)); + tensor->name[sizeof(tensor->name) - 1] = '\0'; +} + struct ggml_tensor * ggml_view_tensor( struct ggml_context * ctx, const struct ggml_tensor * src) { @@ -4470,6 +4349,107 @@ struct ggml_tensor * ggml_add_inplace( return ggml_add_impl(ctx, a, b, true); } +// ggml_add1 + +struct ggml_tensor * ggml_add1_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { + GGML_ASSERT(ggml_is_scalar(b)); + GGML_ASSERT(ggml_is_padded_1d(a)); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_ADD1; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + +struct ggml_tensor * ggml_add1( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + return ggml_add1_impl(ctx, a, b, false); +} + +struct ggml_tensor * ggml_add1_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + return ggml_add1_impl(ctx, a, b, true); +} + +// ggml_acc + +struct ggml_tensor * ggml_acc_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset, + bool inplace) { + GGML_ASSERT(ggml_nelements(b) <= ggml_nelements(a)); + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(a->type == GGML_TYPE_F32); + GGML_ASSERT(b->type == GGML_TYPE_F32); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + struct ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 5); + ((int32_t *) c->data)[0] = nb1; + ((int32_t *) c->data)[1] = nb2; + ((int32_t *) c->data)[2] = nb3; + ((int32_t *) c->data)[3] = offset; + ((int32_t *) c->data)[4] = inplace ? 1 : 0; + + result->op = GGML_OP_ACC; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = c; + + return result; +} + +struct ggml_tensor * ggml_acc( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false); +} + +struct ggml_tensor * ggml_acc_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true); +} + // ggml_sub struct ggml_tensor * ggml_sub_impl( @@ -4663,6 +4643,41 @@ struct ggml_tensor * ggml_sqrt_inplace( return ggml_sqrt_impl(ctx, a, true); } + +// ggml_log + +struct ggml_tensor * ggml_log_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (!inplace && (a->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_LOG; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + +struct ggml_tensor * ggml_log( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_log_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_log_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_log_impl(ctx, a, true); +} + // ggml_sum struct ggml_tensor * ggml_sum( @@ -4684,6 +4699,33 @@ struct ggml_tensor * ggml_sum( return result; } + +// ggml_sum_rows + +struct ggml_tensor * ggml_sum_rows( + struct ggml_context * ctx, + struct ggml_tensor * a) { + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + int64_t ne[4] = {1,1,1,1}; + for (int i=1; in_dims; ++i) { + ne[i] = a->ne[i]; + } + + struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, a->n_dims, ne); + + result->op = GGML_OP_SUM_ROWS; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + // ggml_mean struct ggml_tensor * ggml_mean( @@ -4974,6 +5016,29 @@ struct ggml_tensor * ggml_silu_inplace( return ggml_silu_impl(ctx, a, true); } +// ggml_silu_back + +struct ggml_tensor * ggml_silu_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + bool is_node = false; + + if (a->grad || b->grad) { + // TODO: implement backward + is_node = true; + } + + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_SILU_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + // ggml_norm struct ggml_tensor * ggml_norm_impl( @@ -5016,7 +5081,6 @@ struct ggml_tensor * ggml_rms_norm_impl( bool is_node = false; if (!inplace && (a->grad)) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5042,6 +5106,28 @@ struct ggml_tensor * ggml_rms_norm_inplace( return ggml_rms_norm_impl(ctx, a, true); } +struct ggml_tensor * ggml_rms_norm_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + bool is_node = false; + + if (a->grad) { + // TODO: implement backward + is_node = true; + } + + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_RMS_NORM_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + + // ggml_mul_mat struct ggml_tensor * ggml_mul_mat( @@ -5081,13 +5167,10 @@ struct ggml_tensor * ggml_scale_impl( bool is_node = false; if (!inplace && (a->grad || b->grad)) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); result->op = GGML_OP_SCALE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -5111,6 +5194,100 @@ struct ggml_tensor * ggml_scale_inplace( return ggml_scale_impl(ctx, a, b, true); } +// ggml_set + +struct ggml_tensor * ggml_set_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset, + bool inplace) { + GGML_ASSERT(ggml_nelements(a) >= ggml_nelements(b)); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + // make a view of the destination + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + struct ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 5); + (( int32_t * ) c->data)[0] = nb1; + (( int32_t * ) c->data)[1] = nb2; + (( int32_t * ) c->data)[2] = nb3; + (( int32_t * ) c->data)[3] = offset; + (( int32_t * ) c->data)[4] = inplace ? 1 : 0; + + result->op = GGML_OP_SET; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = c; + + return result; +} + +struct ggml_tensor * ggml_set( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, nb2, nb3, offset, false); +} + +struct ggml_tensor * ggml_set_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, nb2, nb3, offset, true); +} + +struct ggml_tensor * ggml_set_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset) { + return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, false); +} + +struct ggml_tensor * ggml_set_1d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset) { + return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, true); +} + +struct ggml_tensor * ggml_set_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false); +} + +struct ggml_tensor * ggml_set_2d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false); +} + + // ggml_cpy struct ggml_tensor * ggml_cpy_impl( @@ -5123,7 +5300,6 @@ struct ggml_tensor * ggml_cpy_impl( bool is_node = false; if (!inplace && (a->grad || b->grad)) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5161,7 +5337,6 @@ struct ggml_tensor * ggml_cont_impl( bool is_node = false; if (!inplace && a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5199,11 +5374,15 @@ struct ggml_tensor * ggml_reshape( bool is_node = false; - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward + if (a->grad) { is_node = true; } + if (b->grad) { + // gradient propagation is not supported + //GGML_ASSERT(false); + } + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a->data); result->op = GGML_OP_RESHAPE; @@ -5214,6 +5393,30 @@ struct ggml_tensor * ggml_reshape( return result; } +struct ggml_tensor * ggml_reshape_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0) { + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(ggml_nelements(a) == ne0); + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[1] = { ne0 }; + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a->data); + + result->op = GGML_OP_RESHAPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + struct ggml_tensor * ggml_reshape_2d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -5225,7 +5428,6 @@ struct ggml_tensor * ggml_reshape_2d( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5252,7 +5454,6 @@ struct ggml_tensor * ggml_reshape_3d( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5267,6 +5468,34 @@ struct ggml_tensor * ggml_reshape_3d( return result; } + +struct ggml_tensor * ggml_reshape_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3) { + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[4] = { ne0, ne1, ne2, ne3 }; + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a->data); + + result->op = GGML_OP_RESHAPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + // ggml_view_1d struct ggml_tensor * ggml_view_1d( @@ -5274,16 +5503,23 @@ struct ggml_tensor * ggml_view_1d( struct ggml_tensor * a, int64_t ne0, size_t offset) { + + bool is_node = false; + if (a->grad) { - GGML_ASSERT(false); // gradient propagation is not supported + is_node = true; } struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset); result->op = GGML_OP_VIEW; - result->grad = NULL; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the offset here? + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } return result; } @@ -5297,8 +5533,11 @@ struct ggml_tensor * ggml_view_2d( int64_t ne1, size_t nb1, size_t offset) { + + bool is_node = false; + if (a->grad) { - GGML_ASSERT(false); // gradient propagation is not supported + is_node = true; } const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 }; @@ -5310,9 +5549,13 @@ struct ggml_tensor * ggml_view_2d( result->nb[3] = result->nb[2]; result->op = GGML_OP_VIEW; - result->grad = NULL; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the offset here? + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } return result; } @@ -5328,8 +5571,11 @@ struct ggml_tensor * ggml_view_3d( size_t nb1, size_t nb2, size_t offset) { + + bool is_node = false; + if (a->grad) { - GGML_ASSERT(false); // gradient propagation is not supported + is_node = true; } const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 }; @@ -5341,9 +5587,53 @@ struct ggml_tensor * ggml_view_3d( result->nb[3] = result->nb[2]*ne2; result->op = GGML_OP_VIEW; - result->grad = NULL; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the offset here? + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } + + return result; +} + +// ggml_view_4d + +struct ggml_tensor * ggml_view_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 }; + + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset); + + result->nb[1] = nb1; + result->nb[2] = nb2; + result->nb[3] = nb3; + + result->op = GGML_OP_VIEW; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } return result; } @@ -5372,7 +5662,6 @@ struct ggml_tensor * ggml_permute( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5404,7 +5693,14 @@ struct ggml_tensor * ggml_permute( result->op = GGML_OP_PERMUTE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the permutation here? + result->src1 = NULL; + + if (is_node) { + result->padding[0] = axis0; + result->padding[1] = axis1; + result->padding[2] = axis2; + result->padding[3] = axis3; + } return result; } @@ -5417,7 +5713,6 @@ struct ggml_tensor * ggml_transpose( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5448,7 +5743,6 @@ struct ggml_tensor * ggml_get_rows( bool is_node = false; if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5464,23 +5758,76 @@ struct ggml_tensor * ggml_get_rows( return result; } -// ggml_diag_mask_inf +// ggml_get_rows_back -struct ggml_tensor * ggml_diag_mask_inf( +struct ggml_tensor * ggml_get_rows_back( struct ggml_context * ctx, struct ggml_tensor * a, - int n_past) { + struct ggml_tensor * b, + struct ggml_tensor * c) { + GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_is_matrix(c) && (a->ne[0] == c->ne[0])); + bool is_node = false; - if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward + if (a->grad || b->grad) { is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); - struct ggml_tensor * b = ggml_new_i32(ctx, n_past); + // TODO: implement non F32 return + //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]); + struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, c->ne[0], c->ne[1]); + + result->op = GGML_OP_GET_ROWS_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = c; + + return result; +} + +// ggml_diag + +struct ggml_tensor * ggml_diag( + struct ggml_context * ctx, + struct ggml_tensor * a) { + GGML_ASSERT(a->ne[1] == 1); + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] }; + struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, MAX(a->n_dims, 2), ne); + + result->op = GGML_OP_DIAG; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + + +// ggml_diag_mask_inf + +struct ggml_tensor * ggml_diag_mask_inf_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + bool inplace) { + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = inplace ? 1 : 0; result->op = GGML_OP_DIAG_MASK_INF; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -5490,21 +5837,75 @@ struct ggml_tensor * ggml_diag_mask_inf( return result; } -// ggml_soft_max - -struct ggml_tensor * ggml_soft_max( +struct ggml_tensor * ggml_diag_mask_inf( struct ggml_context * ctx, - struct ggml_tensor * a) { + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_inf_impl(ctx, a, n_past, false); +} + + +struct ggml_tensor * ggml_diag_mask_inf_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_inf_impl(ctx, a, n_past, true); +} + +// ggml_diag_mask_zero + +struct ggml_tensor * ggml_diag_mask_zero_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + bool inplace) { bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); + ggml_set_name(b, "n_past, inplace"); + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = inplace ? 1 : 0; + + result->op = GGML_OP_DIAG_MASK_ZERO; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + +struct ggml_tensor * ggml_diag_mask_zero( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_zero_impl(ctx, a, n_past, false); +} + +struct ggml_tensor * ggml_diag_mask_zero_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_zero_impl(ctx, a, n_past, true); +} + +// ggml_soft_max + +struct ggml_tensor * ggml_soft_max_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); result->op = GGML_OP_SOFT_MAX; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -5514,14 +5915,75 @@ struct ggml_tensor * ggml_soft_max( return result; } +struct ggml_tensor * ggml_soft_max( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_soft_max_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_soft_max_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_soft_max_impl(ctx, a, true); +} + // ggml_rope +struct ggml_tensor * ggml_rope_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode, + bool inplace) { + GGML_ASSERT(n_past >= 0); + bool is_node = false; + + if (!inplace && a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = n_dims; + ((int32_t *) b->data)[2] = mode; + + result->op = GGML_OP_ROPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + struct ggml_tensor * ggml_rope( struct ggml_context * ctx, struct ggml_tensor * a, int n_past, int n_dims, int mode) { + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, false); +} + +struct ggml_tensor * ggml_rope_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode) { + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, true); +} + +// ggml_rope_back + +struct ggml_tensor * ggml_rope_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode) { GGML_ASSERT(n_past >= 0); bool is_node = false; @@ -5530,16 +5992,46 @@ struct ggml_tensor * ggml_rope( is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); ((int32_t *) b->data)[0] = n_past; ((int32_t *) b->data)[1] = n_dims; ((int32_t *) b->data)[2] = mode; + ggml_set_name(b, "n_past, n_dims, mode"); - result->op = GGML_OP_ROPE; + result->op = GGML_OP_ROPE_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + +// ggml_alibi + +struct ggml_tensor * ggml_alibi( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_head) { + GGML_ASSERT(n_past >= 0); + bool is_node = false; + + if (a->grad) { + GGML_ASSERT(false); // TODO: implement backward + is_node = true; + } + + // TODO: when implement backward, fix this: + //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + struct ggml_tensor * result = ggml_view_tensor(ctx, a); + + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = n_head; + + result->op = GGML_OP_ALIBI; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; result->src1 = b; @@ -5762,6 +6254,38 @@ void ggml_set_param( // ggml_compute_forward_dup +static void ggml_compute_forward_dup_same_cont( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == dst->type); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const size_t nb00 = src0->nb[0]; + const size_t nb0 = dst->nb[0]; + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by elements + const int ne = ggml_nelements(dst); + const int dr = (ne + nth - 1) / nth; + const int ie0 = dr * ith; + const int ie1 = MIN(ie0 + dr, ne); + + if (ie0 < ie1) { + memcpy( + ((char *) dst->data + ie0*nb0), + ((char *) src0->data + ie0*nb00), + (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); + } + +} static void ggml_compute_forward_dup_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -5796,17 +6320,7 @@ static void ggml_compute_forward_dup_f16( const int nth = params->nth; // number of threads if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - // parallelize by elements - const int ne = ggml_nelements(dst); - const int dr = (ne + nth - 1) / nth; - const int ie0 = dr * ith; - const int ie1 = MIN(ie0 + dr, ne); - - memcpy( - ((char *) dst->data + ie0*nb0), - ((char *) src0->data + ie0*nb00), - (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); - + ggml_compute_forward_dup_same_cont(params, src0, dst); return; } @@ -6095,17 +6609,7 @@ static void ggml_compute_forward_dup_f32( const int nth = params->nth; // number of threads if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - // parallelize by elements - const int ne = ggml_nelements(dst); - const int dr = (ne + nth - 1) / nth; - const int ie0 = dr * ith; - const int ie1 = MIN(ie0 + dr, ne); - - memcpy( - ((char *) dst->data + ie0*nb0), - ((char *) src0->data + ie0*nb00), - (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); - + ggml_compute_forward_dup_same_cont(params, src0, dst); return; } @@ -6253,7 +6757,6 @@ static void ggml_compute_forward_dup_f32( i10 += ne00 * ir0; while (i10 >= ne0) { i10 -= ne0; - i11++; if (++i11 == ne1) { i11 = 0; if (++i12 == ne2) { @@ -6361,6 +6864,10 @@ static void ggml_compute_forward_dup( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { + ggml_compute_forward_dup_same_cont(params, src0, dst); + return; + } switch (src0->type) { case GGML_TYPE_F16: { @@ -6393,44 +6900,73 @@ static void ggml_compute_forward_add_f32( const int ith = params->ith; const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; const size_t nb10 = src1->nb[0]; const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; const size_t nb0 = dst->nb[0]; const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; GGML_ASSERT( nb0 == sizeof(float)); GGML_ASSERT(nb00 == sizeof(float)); + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + if (nb10 == sizeof(float)) { - for (int j = ith; j < n; j += nth) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + #ifdef GGML_USE_ACCELERATE vDSP_vadd( - (float *) ((char *) src0->data + j*nb01), 1, - (float *) ((char *) src1->data + j*nb11), 1, - (float *) ((char *) dst->data + j*nb1), 1, nc); + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); #else - ggml_vec_add_f32(nc, - (float *) ((char *) dst->data + j*nb1), - (float *) ((char *) src0->data + j*nb01), - (float *) ((char *) src1->data + j*nb11)); + ggml_vec_add_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); #endif + // } + // } } } else { // src1 is not contiguous - for (int j = ith; j < n; j += nth) { - float * dst_ptr = (float *) ((char *) dst->data + j*nb1); - float * src0_ptr = (float *) ((char *) src0->data + j*nb01); - for (int i = 0; i < nc; i++) { - float * src1_ptr = (float *) ((char *) src1->data + j*nb11 + i*nb10); + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - dst_ptr[i] = src0_ptr[i] + *src1_ptr; + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] + *src1_ptr; } } } @@ -6450,17 +6986,25 @@ static void ggml_compute_forward_add_f16_f32( const int ith = params->ith; const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; const size_t nb10 = src1->nb[0]; const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; const size_t nb0 = dst->nb[0]; const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -6469,13 +7013,26 @@ static void ggml_compute_forward_add_f16_f32( GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + if (nb10 == sizeof(float)) { - for (int j = ith; j < n; j += nth) { - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); - for (int i = 0; i < nc; i++) { - float * src1_ptr = (float *) ((char *) src1->data + j*nb11 + i*nb10); - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + *src1_ptr); + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); } } } @@ -6499,32 +7056,53 @@ static void ggml_compute_forward_add_f16_f16( const int ith = params->ith; const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; const size_t nb10 = src1->nb[0]; const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; const size_t nb0 = dst->nb[0]; const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT(dst->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + if (nb10 == sizeof(ggml_fp16_t)) { - for (int j = ith; j < n; j += nth) { - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); - for (int i = 0; i < nc; i++) { - ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + j*nb11 + i*nb10); - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(*src1_ptr)); + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(src1_ptr[i])); } } } @@ -6545,50 +7123,36 @@ static void ggml_compute_forward_add_q_f32( return; } + const int nr = ggml_nrows(src0); 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 ne03 = src0->ne[3]; - //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 size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - //const int64_t ne0 = dst->ne[0]; - //const int64_t ne1 = dst->ne[1]; - const int64_t ne2 = dst->ne[2]; - const int64_t ne3 = dst->ne[3]; + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; - const int nb00 = src0->nb[0]; - const int nb01 = src0->nb[1]; - const int nb02 = src0->nb[2]; - const int nb03 = src0->nb[3]; - - const int nb10 = src1->nb[0]; - const int nb11 = src1->nb[1]; - const int nb12 = src1->nb[2]; - const int nb13 = src1->nb[3]; - - const int nb0 = dst->nb[0]; - const int nb1 = dst->nb[1]; - const int nb2 = dst->nb[2]; - const int nb3 = dst->nb[3]; + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; const int ith = params->ith; const int nth = params->nth; - GGML_ASSERT(ne02 == ne12); - GGML_ASSERT(ne03 == ne13); - GGML_ASSERT(ne2 == ne12); - GGML_ASSERT(ne3 == ne13); - const enum ggml_type type = src0->type; dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q; quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q; // we don't support permuted src0 or src1 - GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]); + GGML_ASSERT(nb00 == GGML_TYPE_SIZE[type]); GGML_ASSERT(nb10 == sizeof(float)); // dst cannot be transposed or permuted @@ -6600,9 +7164,6 @@ static void ggml_compute_forward_add_q_f32( GGML_ASSERT(dst->type == src0->type); GGML_ASSERT(src1->type == GGML_TYPE_F32); - // total rows in src0 - const int nr = ne01*ne02*ne03; - // rows per thread const int dr = (nr + nth - 1)/nth; @@ -6666,7 +7227,9 @@ static void ggml_compute_forward_add( } break; case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: { ggml_compute_forward_add_q_f32(params, src0, src1, dst); } break; @@ -6677,6 +7240,428 @@ static void ggml_compute_forward_add( } } +// ggml_compute_forward_add1 + +static void ggml_compute_forward_add1_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_add1_f32); + + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data), 0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_add1_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + *(float *) src1->data); +#endif + } +} + +static void ggml_compute_forward_add1_f16_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_f16_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // scalar to add + const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_q_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const enum ggml_type type = src0->type; + dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q; + quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q; + + // we don't support permuted src0 + GGML_ASSERT(nb00 == GGML_TYPE_SIZE[type]); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(dst->type == src0->type); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + void * src0_row = (void *) ((char *) src0->data + (i1*nb01 + i2*nb02 + i3*nb03)); + void * dst_row = (void *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb0 )); + + assert(ne0 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne0); + // add src1 + ggml_vec_acc1_f32(ne0, wdata, v); + // quantize row to dst + quantize_row_q(wdata, dst_row, ne0); + } +} + +static void ggml_compute_forward_add1( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_add1_f32(params, src0, src1, dst); + } break; + case GGML_TYPE_F16: + { + if (src1->type == GGML_TYPE_F16) { + ggml_compute_forward_add1_f16_f16(params, src0, src1, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add1_f16_f32(params, src0, src1, dst); + } + else { + GGML_ASSERT(false); + } + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + { + ggml_compute_forward_add1_q_f32(params, src0, src1, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + + +// ggml_compute_forward_acc + +static void ggml_compute_forward_acc_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + GGML_ASSERT(opt0->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(opt0) == 5); + + // view src0 and dst with these strides and data offset inbytes during acc + // nb0 is implicitely element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) opt0->data)[0]; + size_t nb2 = ((int32_t *) opt0->data)[1]; + size_t nb3 = ((int32_t *) opt0->data)[2]; + size_t offset = ((int32_t *) opt0->data)[3]; + bool inplace = (bool) ((int32_t *) opt0->data)[4]; + + if (!inplace && (params->type == GGML_TASK_INIT)) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + 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 size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + // src0 and dst as viewed during acc + const size_t nb0 = ggml_element_size(src0); + + const size_t nb00 = nb0; + const size_t nb01 = nb1; + const size_t nb02 = nb2; + const size_t nb03 = nb3; + + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb0 + (ne11 == 0 ? 0 : ne11-1)*nb1 + (ne12 == 0 ? 0 : ne12-1)*nb2 + (ne13 == 0 ? 0 : ne13-1)*nb3 < ggml_nbytes(dst)); + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb00 + (ne11 == 0 ? 0 : ne11-1)*nb01 + (ne12 == 0 ? 0 : ne12-1)*nb02 + (ne13 == 0 ? 0 : ne13-1)*nb03 < ggml_nbytes(src0)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + +#ifdef GGML_USE_ACCELERATE + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), 1, nc); +#else + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + } +} + +static void ggml_compute_forward_acc( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_acc_f32(params, src0, src1, opt0, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_sub static void ggml_compute_forward_sub_f32( @@ -6691,18 +7676,68 @@ static void ggml_compute_forward_sub_f32( return; } - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - assert(src1->nb[0] == sizeof(float)); + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - for (int i = 0; i < n; i++) { - ggml_vec_sub_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + +#ifdef GGML_USE_ACCELERATE + vDSP_vsub( + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_sub_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + // } + // } + } + } else { + // src1 is not contiguous + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] - *src1_ptr; + } + } } } @@ -6730,25 +7765,78 @@ static void ggml_compute_forward_mul_f32( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(params->ith == 0); assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } + const int ith = params->ith; + const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - assert(src1->nb[0] == sizeof(float)); + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - for (int i = 0; i < n; i++) { - ggml_vec_mul_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int ir = ith; ir < nr; ir += nth) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_mul_f32); + + vDSP_vmul( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_mul_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + // } + // } + } + } else { + // src1 is not contiguous + for (int ir = ith; ir < nr; ir += nth) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr); + } + } } } @@ -6783,18 +7871,68 @@ static void ggml_compute_forward_div_f32( return; } - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - assert(src1->nb[0] == sizeof(float)); + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - for (int i = 0; i < n; i++) { - ggml_vec_div_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + +#ifdef GGML_USE_ACCELERATE + vDSP_vdiv( + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_div_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + // } + // } + } + } else { + // src1 is not contiguous + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr); + } + } } } @@ -6899,6 +8037,49 @@ static void ggml_compute_forward_sqrt( } } + +// ggml_compute_forward_log + +static void ggml_compute_forward_log_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_log_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_log( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_log_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_sum static void ggml_compute_forward_sum_f32( @@ -6924,15 +8105,20 @@ static void ggml_compute_forward_sum_f32( const size_t nb02 = src0->nb[2]; const size_t nb03 = src0->nb[3]; + ggml_float sum = 0; + ggml_float row_sum = 0; + for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { for (int64_t i01 = 0; i01 < ne01; i01++) { - ggml_vec_sum_f32(ne00, - (float *) (dst->data), + ggml_vec_sum_ggf(ne00, + &row_sum, (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); + sum += row_sum; } } } + ((float *) dst->data)[0] = sum; } static void ggml_compute_forward_sum( @@ -6951,6 +8137,73 @@ static void ggml_compute_forward_sum( } } +// ggml_compute_forward_sum_rows + +static void ggml_compute_forward_sum_rows_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + 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 ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + GGML_ASSERT(ne0 == 1); + GGML_ASSERT(ne1 == ne01); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne3 == ne03); + + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + for (int64_t i3 = 0; i3 < ne03; i3++) { + for (int64_t i2 = 0; i2 < ne02; i2++) { + for (int64_t i1 = 0; i1 < ne01; i1++) { + float* src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); + float* dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); + float row_sum = 0; + ggml_vec_sum_f32(ne00, &row_sum, src_row); + dst_row[0] = row_sum; + } + } + } +} + +static void ggml_compute_forward_sum_rows( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sum_rows_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_mean static void ggml_compute_forward_mean_f32( @@ -7028,37 +8281,58 @@ static void ggml_compute_forward_repeat_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - assert(params->ith == 0); - assert(ggml_can_repeat(src0, dst)); + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_can_repeat(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - // TODO: implement support for rank > 2 tensors - assert(src0->ne[2] == 1); - assert(src0->ne[3] == 1); - assert( dst->ne[2] == 1); - assert( dst->ne[3] == 1); + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; - const int nc = dst->ne[0]; - const int nr = dst->ne[1]; - const int nc0 = src0->ne[0]; - const int nr0 = src0->ne[1]; - const int ncr = nc/nc0; // guaranteed to be an integer due to the check in ggml_can_repeat - const int nrr = nr/nr0; // guaranteed to be an integer due to the check in ggml_can_repeat + 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 size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne0/ne00); + const int nr1 = (int)(ne1/ne01); + const int nr2 = (int)(ne2/ne02); + const int nr3 = (int)(ne3/ne03); // TODO: support for transposed / permuted tensors - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); // TODO: maybe this is not optimal? - for (int i = 0; i < nrr; i++) { - for (int j = 0; j < ncr; j++) { - for (int k = 0; k < nr0; k++) { - ggml_vec_cpy_f32(nc0, - (float *) ((char *) dst->data + (i*nr0 + k)*( dst->nb[1]) + j*nc0*( dst->nb[0])), - (float *) ((char *) src0->data + ( k)*(src0->nb[1]))); + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne03; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne02; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne01; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_vec_cpy_f32(ne00, + (float *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0), + (float *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01)); + } + } + } + } } } } @@ -7411,6 +8685,70 @@ static void ggml_compute_forward_silu( } +// ggml_compute_forward_silu_back + +static void ggml_compute_forward_silu_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * grad, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_is_contiguous(grad)); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_are_same_shape(src0, grad)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_backward_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1])), + (float *) ((char *) grad->data + i1*(grad->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_silu_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * grad, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_silu_back_f32(params, src0, grad, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_norm static void ggml_compute_forward_norm_f32( @@ -7565,9 +8903,198 @@ static void ggml_compute_forward_rms_norm( } +static void ggml_compute_forward_rms_norm_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + 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 size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const float eps = 1e-6f; // TODO: make this a parameter + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + // src1 is same shape as src0 => same indices + const int64_t i11 = i01; + const int64_t i12 = i02; + const int64_t i13 = i03; + + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + const float * dz = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13); + + ggml_float sum_xx = 0.0; + ggml_float sum_xdz = 0.0; + + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum_xx += (ggml_float)(x[i00] * x[i00]); + sum_xdz += (ggml_float)(x[i00] * dz[i00]); + } + + //const float mean = (float)(sum_xx)/ne00; + const float mean_eps = (float)(sum_xx)/ne00 + eps; + const float sum_eps = (float)(sum_xx) + eps*ne00; + //const float mean_xdz = (float)(sum_xdz)/ne00; + // we could cache rms from forward pass to improve performance. + // to do this implement ggml_rms and compose ggml_rms_norm using ggml_rms. + //const float rms = sqrtf(mean_eps); + const float rrms = 1.0f / sqrtf(mean_eps); + //const float scale = -rrms/(ne00 * mean_eps); // -1/(n*rms**3) + + { + // z = rms_norm(x) + // + // rms_norm(src0) = + // scale( + // src0, + // div( + // 1, + // sqrt( + // add( + // scale( + // sum( + // sqr( + // src0)), + // (1.0/N)), + // eps)))); + + // postorder: + // ## op args grad + // 00 param src0 grad[#00] + // 01 const 1 + // 02 sqr (#00) grad[#02] + // 03 sum (#02) grad[#03] + // 04 const 1/N + // 05 scale (#03, #04) grad[#05] + // 06 const eps + // 07 add (#05, #06) grad[#07] + // 08 sqrt (#07) grad[#08] + // 09 div (#01,#08) grad[#09] + // 10 scale (#00,#09) grad[#10] + // + // backward pass, given grad[#10] + // #10: scale + // grad[#00] += scale(grad[#10],#09) + // grad[#09] += sum(mul(grad[#10],#00)) + // #09: div + // grad[#08] += neg(mul(grad[#09], div(#09,#08))) + // #08: sqrt + // grad[#07] += mul(grad[#08], div(0.5, #08)) + // #07: add + // grad[#05] += grad[#07] + // #05: scale + // grad[#03] += scale(grad[#05],#04) + // #03: sum + // grad[#02] += repeat(grad[#03], #02) + // #02: + // grad[#00] += scale(mul(#00, grad[#02]), 2.0) + // + // substitute and simplify: + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#02] = repeat(grad[#03], #02) + // grad[#02] = repeat(scale(grad[#05],#04), #02) + // grad[#02] = repeat(scale(grad[#07],#04), #02) + // grad[#02] = repeat(scale(mul(grad[#08], div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(grad[#09], div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(sum(mul(grad[#10],#00)), div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(#09,#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(div(#01,#08),#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#08*#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02)), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N))), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(1,#08) * (1/N))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,mean_eps*rms) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*(sum_xx/N+eps))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*sum_xx+rms*N*eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum(mul(dz,x)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum_xdz * div(-1,rms*N*mean_eps)) + // a = b*c + d*e + // a = b*c*f/f + d*e*f/f + // a = (b*c*f + d*e*f)*(1/f) + // a = (b*c*(1/c) + d*e*(1/c))*(1/(1/c)) + // a = (b + d*e/c)*c + // b = dz, c = rrms, d = x, e = sum_xdz * div(-1,rms*N*mean_eps) + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)/rrms)*rrms + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)*rms)*rrms + // a = (dz + x*sum_xdz * div(-rms,rms*N*mean_eps))*rrms + // a = (dz + x*sum_xdz * div(-1,N*mean_eps))*rrms + // a = (dz + x*div(-sum_xdz,N*mean_eps))*rrms + // a = (dz + x*div(-mean_xdz,mean_eps))*rrms + // grad[#00] = scale(dz + scale(x, div(-mean_xdz,mean_eps)),rrms) + // grad[#00] = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + } + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // post-order: + // dx := x + // dx := scale(dx,-mean_xdz/mean_eps) + // dx := add(dx, dz) + // dx := scale(dx, rrms) + float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + ggml_vec_cpy_f32 (ne00, dx, x); + // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps); + ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps); + ggml_vec_acc_f32 (ne00, dx, dz); + ggml_vec_scale_f32(ne00, dx, rrms); + } + } + } +} + +static void ggml_compute_forward_rms_norm_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rms_norm_back_f32(params, src0, src1, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + + // ggml_compute_forward_mul_mat -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) // helper function to determine if it is better to use BLAS or not // for large matrices, BLAS is faster static bool ggml_compute_forward_mul_mat_use_blas( @@ -7584,7 +9111,8 @@ static bool ggml_compute_forward_mul_mat_use_blas( // TODO: find the optimal values for these if (ggml_is_contiguous(src0) && - ggml_is_contiguous(src1) && ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32))) { + ggml_is_contiguous(src1) && + (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) { /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/ return true; @@ -7607,7 +9135,7 @@ static void ggml_compute_forward_mul_mat_f32( const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) const int64_t ne10 = src1->ne[0]; #endif const int64_t ne11 = src1->ne[1]; @@ -7664,7 +9192,16 @@ static void ggml_compute_forward_mul_mat_f32( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_CUBLAS) + if (ggml_cuda_can_mul_mat(src0, src1, dst)) { + if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) { + ggml_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize); + } + return; + } +#endif + +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { if (params->ith != 0) { return; @@ -7682,18 +9219,25 @@ static void ggml_compute_forward_mul_mat_f32( for (int64_t i02 = 0; i02 < ne02; i02++) { const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03); const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); - float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); +#if defined(GGML_USE_CLBLAST) // zT = y * xT + ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T, + ne11, ne01, ne10, + 1.0f, y, ne10, + x, ne10, + 0.0f, d, ne01, + GGML_TYPE_F32); +#else cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, ne11, ne01, ne10, 1.0f, y, ne10, x, ne00, 0.0f, d, ne01); +#endif } } - //printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3); return; @@ -7823,7 +9367,16 @@ static void ggml_compute_forward_mul_mat_f16_f32( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_CUBLAS) + if (ggml_cuda_can_mul_mat(src0, src1, dst)) { + if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) { + ggml_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize); + } + return; + } +#endif + +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { GGML_ASSERT(nb10 == sizeof(float)); @@ -7839,10 +9392,9 @@ static void ggml_compute_forward_mul_mat_f16_f32( return; } - float * const wdata = params->wdata; - for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { + float * const wdata = params->wdata; { size_t id = 0; for (int64_t i01 = 0; i01 < ne01; ++i01) { @@ -7850,8 +9402,24 @@ static void ggml_compute_forward_mul_mat_f16_f32( wdata[id++] = GGML_FP16_TO_FP32(*(ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00)); } } + + assert(id*sizeof(float) <= params->wsize); } +#if defined(GGML_USE_CLBLAST) + const float * x = wdata; + const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); + + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + + // zT = y * xT + ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T, + ne11, ne01, ne10, + 1.0f, y, ne10, + x, ne10, + 0.0f, d, ne01, + GGML_TYPE_F32); +#else const float * x = wdata; const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); @@ -7863,6 +9431,7 @@ static void ggml_compute_forward_mul_mat_f16_f32( 1.0f, y, ne10, x, ne00, 0.0f, d, ne01); +#endif } } @@ -7998,6 +9567,7 @@ static void ggml_compute_forward_mul_mat_q_f32( const enum ggml_type type = src0->type; quantize_row_q_t const quantize_row_q_dot = quantize_fns[type].quantize_row_q_dot; vec_dot_q_t const vec_dot_q = quantize_fns[type].vec_dot_q; + enum ggml_type const vec_dot_type = quantize_fns[type].vec_dot_type; // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]); @@ -8017,7 +9587,16 @@ static void ggml_compute_forward_mul_mat_q_f32( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_CUBLAS) + if (ggml_cuda_can_mul_mat(src0, src1, dst)) { + if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) { + ggml_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize); + } + return; + } +#endif + +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { if (params->ith != 0) { return; @@ -8036,25 +9615,41 @@ static void ggml_compute_forward_mul_mat_q_f32( for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { + const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); + + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + +#if defined(GGML_USE_CLBLAST) + const void* x = (char *) src0->data + i03*nb03 + i02*nb02; +#else { size_t id = 0; for (int64_t i01 = 0; i01 < ne01; ++i01) { dequantize_row_q((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00); id += ne00; } + + assert(id*sizeof(float) <= params->wsize); } const float * x = wdata; - const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); - - float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); +#endif +#if defined(GGML_USE_CLBLAST) // zT = y * xT + ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T, + ne11, ne01, ne10, + 1.0f, y, ne10, + x, ne10, + 0.0f, d, ne01, + type); +#else cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, ne11, ne01, ne10, 1.0f, y, ne10, x, ne00, 0.0f, d, ne01); +#endif } } @@ -8066,7 +9661,7 @@ static void ggml_compute_forward_mul_mat_q_f32( if (params->type == GGML_TASK_INIT) { char * wdata = params->wdata; - const size_t row_size = ne10*GGML_TYPE_SIZE[GGML_TYPE_Q8_0]/GGML_BLCK_SIZE[GGML_TYPE_Q8_0]; + const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type]; for (int64_t i13 = 0; i13 < ne13; ++i13) { for (int64_t i12 = 0; i12 < ne12; ++i12) { @@ -8097,7 +9692,7 @@ static void ggml_compute_forward_mul_mat_q_f32( const int ir1 = MIN(ir0 + dr, nr); void * wdata = params->wdata; - const size_t row_size = ne00*GGML_TYPE_SIZE[GGML_TYPE_Q8_0]/GGML_BLCK_SIZE[GGML_TYPE_Q8_0]; + const size_t row_size = ne00*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type]; for (int ir = ir0; ir < ir1; ++ir) { // src0 indices @@ -8145,8 +9740,10 @@ static void ggml_compute_forward_mul_mat( switch (src0->type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: { ggml_compute_forward_mul_mat_q_f32(params, src0, src1, dst); } break; @@ -8163,34 +9760,6 @@ static void ggml_compute_forward_mul_mat( GGML_ASSERT(false); } break; } - -#if 0 - if (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_Q4_1) { - static int first = 8; - printf("src0: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src0->ne[0], src0->ne[1], src0->ne[2]); - printf("src1: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src1->ne[0], src1->ne[1], src1->ne[2]); - printf("dst: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", dst->ne[0], dst->ne[1], dst->ne[2]); - if (first) { - --first; - } else { - for (int k = 0; k < dst->ne[1]; ++k) { - for (int j = 0; j < dst->ne[0]/16; ++j) { - for (int i = 0; i < 16; ++i) { - printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); - } - printf("\n"); - } - printf("\n"); - } - printf("\n"); - exit(0); - } - } else { - printf("aaaa src0: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src0->ne[0], src0->ne[1], src0->ne[2]); - printf("aaaa src1: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src1->ne[0], src1->ne[1], src1->ne[2]); - printf("aaaa dst: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", dst->ne[0], dst->ne[1], dst->ne[2]); - } -#endif } // ggml_compute_forward_scale @@ -8225,8 +9794,17 @@ static void ggml_compute_forward_scale_f32( const int ir0 = dr*ith; const int ir1 = MIN(ir0 + dr, nr); + const size_t nb01 = src0->nb[1]; + + const size_t nb1 = dst->nb[1]; + + for (int i1 = ir0; i1 < ir1; i1++) { - ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), v); + if (dst->data != src0->data) { + // src0 is same shape as dst => same indices + memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float)); + } + ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v); } } @@ -8247,6 +9825,115 @@ static void ggml_compute_forward_scale( } } +// ggml_compute_forward_set + +static void ggml_compute_forward_set_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + GGML_ASSERT(opt0->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(opt0) == 5); + + // view src0 and dst with these strides and data offset inbytes during set + // nb0 is implicitely element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) opt0->data)[0]; + size_t nb2 = ((int32_t *) opt0->data)[1]; + size_t nb3 = ((int32_t *) opt0->data)[2]; + size_t offset = ((int32_t *) opt0->data)[3]; + bool inplace = (bool) ((int32_t *) opt0->data)[4]; + + if (!inplace && (params->type == GGML_TASK_INIT)) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + 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 size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + // src0 and dst as viewed during set + const size_t nb0 = ggml_element_size(src0); + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 < ggml_nbytes(dst)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + + ggml_vec_cpy_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + } +} + +static void ggml_compute_forward_set( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_set_f32(params, src0, src1, opt0, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_cpy static void ggml_compute_forward_cpy( @@ -8401,8 +10088,10 @@ static void ggml_compute_forward_get_rows( switch (src0->type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: { ggml_compute_forward_get_rows_q(params, src0, src1, dst); } break; @@ -8439,22 +10128,212 @@ static void ggml_compute_forward_get_rows( //} } -// ggml_compute_forward_diag_mask_inf +// ggml_compute_forward_get_rows_back -static void ggml_compute_forward_diag_mask_inf_f32( +static void ggml_compute_forward_get_rows_back_f32_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - assert(params->ith == 0); - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 1); + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_are_same_shape(opt0, dst)); + GGML_ASSERT(ggml_is_contiguous(opt0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + ggml_compute_forward_dup_same_cont(params, opt0, dst); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const int n_past = ((int32_t *) src1->data)[0]; + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + for (int j = 0; j < nc; ++j) { + ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; + ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); + } + } +} + +static void ggml_compute_forward_get_rows_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_are_same_shape(opt0, dst)); + GGML_ASSERT(ggml_is_contiguous(opt0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + ggml_compute_forward_dup_same_cont(params, opt0, dst); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) src0->data + i*src0->nb[1])); + } +} + + +static void ggml_compute_forward_get_rows_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_get_rows_back_f32_f16(params, src0, src1, opt0, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_get_rows_back_f32(params, src0, src1, opt0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } + + //static bool first = true; + //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); + //if (first) { + // first = false; + //} else { + // for (int k = 0; k < dst->ne[1]; ++k) { + // for (int j = 0; j < dst->ne[0]/16; ++j) { + // for (int i = 0; i < 16; ++i) { + // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); + // } + // printf("\n"); + // } + // printf("\n"); + // } + // printf("\n"); + // exit(0); + //} +} + +// ggml_compute_forward_diag + +static void ggml_compute_forward_diag_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // TODO: handle transposed/permuted matrices + + const int ne00 = src0->ne[0]; + const int ne01 = src0->ne[1]; + const int ne02 = src0->ne[2]; + const int ne03 = src0->ne[3]; + const int ne0 = dst->ne[0]; + const int ne1 = dst->ne[1]; + const int ne2 = dst->ne[2]; + const int ne3 = dst->ne[3]; + GGML_ASSERT(ne00 == ne0); + GGML_ASSERT(ne00 == ne1); + GGML_ASSERT(ne01 == 1); + GGML_ASSERT(ne02 == ne2); + GGML_ASSERT(ne03 == ne3); + + const int nb00 = src0->nb[0]; + //const int nb01 = src0->nb[1]; + const int nb02 = src0->nb[2]; + const int nb03 = src0->nb[3]; + const int nb0 = dst->nb[0]; + const int nb1 = dst->nb[1]; + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb0 == sizeof(float)); + + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = 0; i2 < ne2; i2++) { + for (int i1 = 0; i1 < ne1; i1++) { + float * d = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + float * s = (float *)((char *) src0->data + i3*nb03 + i2*nb02); + for (int i0 = 0; i0 < i1; i0++) { + d[i0] = 0; + } + d[i1] = s[i1]; + for (int i0 = i1+1; i0 < ne0; i0++) { + d[i0] = 0; + } + } + } + } +} + +static void ggml_compute_forward_diag( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +// ggml_compute_forward_diag_mask_inf + +static void ggml_compute_forward_diag_mask_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst, + const float value) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int n_past = ((int32_t *) src1->data)[0]; + const bool inplace = (bool)((int32_t *) src1->data)[1]; + + if (!inplace) { + ggml_compute_forward_dup_same_cont(params, src0, dst); + } // TODO: handle transposed/permuted matrices @@ -8467,10 +10346,10 @@ static void ggml_compute_forward_diag_mask_inf_f32( assert(src0->nb[0] == sizeof(float)); for (int k = 0; k < nz; k++) { - for (int j = 0; j < nr; j++) { + for (int j = ith; j < nr; j += nth) { for (int i = n_past; i < nc; i++) { if (i > n_past + j) { - *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = -INFINITY; + *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = value; } } } @@ -8485,7 +10364,24 @@ static void ggml_compute_forward_diag_mask_inf( switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_diag_mask_inf_f32(params, src0, src1, dst); + ggml_compute_forward_diag_mask_f32(params, src0, src1, dst, -INFINITY); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +static void ggml_compute_forward_diag_mask_zero( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_mask_f32(params, src0, src1, dst, 0); } break; default: { @@ -8524,43 +10420,44 @@ static void ggml_compute_forward_soft_max_f32( const int ir1 = MIN(ir0 + dr, nr); for (int i1 = ir0; i1 < ir1; i1++) { - float *p = (float *)((char *) dst->data + i1*dst->nb[1]); + float *sp = (float *)((char *) src0->data + i1*src0->nb[1]); + float *dp = (float *)((char *) dst->data + i1*dst->nb[1]); #ifndef NDEBUG for (int i = 0; i < nc; ++i) { //printf("p[%d] = %f\n", i, p[i]); - assert(!isnan(p[i])); + assert(!isnan(sp[i])); } #endif float max = -INFINITY; - ggml_vec_max_f32(nc, &max, p); + ggml_vec_max_f32(nc, &max, sp); ggml_float sum = 0.0; uint16_t scvt; for (int i = 0; i < nc; i++) { - if (p[i] == -INFINITY) { - p[i] = 0.0f; + if (sp[i] == -INFINITY) { + dp[i] = 0.0f; } else { - //const float val = (p[i] == -INFINITY) ? 0.0 : exp(p[i] - max); - ggml_fp16_t s = GGML_FP32_TO_FP16(p[i] - max); + // const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max); + ggml_fp16_t s = GGML_FP32_TO_FP16(sp[i] - max); memcpy(&scvt, &s, sizeof(scvt)); const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); sum += (ggml_float)val; - p[i] = val; + dp[i] = val; } } assert(sum > 0.0); sum = 1.0/sum; - ggml_vec_scale_f32(nc, p, sum); + ggml_vec_scale_f32(nc, dp, sum); #ifndef NDEBUG for (int i = 0; i < nc; ++i) { - assert(!isnan(p[i])); - assert(!isinf(p[i])); + assert(!isnan(dp[i])); + assert(!isinf(dp[i])); } #endif } @@ -8582,6 +10479,160 @@ static void ggml_compute_forward_soft_max( } } +// ggml_compute_forward_alibi + +static void ggml_compute_forward_alibi_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n_past = ((int32_t *) src1->data)[0]; + const int n_head = ((int32_t *) src1->data)[1]; + + const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1 + const int ne1 = src0->ne[1]; // seq_len_without_past + //const int ne2 = src0->ne[2]; // n_head -> this is k + //const int ne3 = src0->ne[3]; // 1 -> bsz + + const int n = ggml_nrows(src0); + const int ne2_ne3 = n/ne1; // ne2*ne3 + + const int nb0 = src0->nb[0]; + const int nb1 = src0->nb[1]; + const int nb2 = src0->nb[2]; + //const int nb3 = src0->nb[3]; + + assert(nb0 == sizeof(float)); + assert(ne1 + n_past == ne0); (void) n_past; + + // add alibi to src0 (KQ_scaled) + const int n_heads_log2_floor = 1 << (int) floor(log2(n_head)); + + const float m0 = powf(2.0f, -8.0f / n_heads_log2_floor); + const float m1 = powf(2.0f, -4.0f / n_heads_log2_floor); + + for (int i = 0; i < ne0; i++) { + for (int j = 0; j < ne1; j++) { + for (int k = 0; k < ne2_ne3; k++) { + float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2); + float * pdst = (float *)((char *) dst->data + i*nb0 + j*nb1 + k*nb2); + + // TODO: k*nb2 or k*nb3 + + float m_k; + + if (k < n_heads_log2_floor) { + m_k = powf(m0, k + 1); + } else { + m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1); + } + + pdst[0] = i * m_k + src[0]; + } + } + } +} + + +static void ggml_compute_forward_alibi_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n_past = ((int32_t *) src1->data)[0]; + const int n_head = ((int32_t *) src1->data)[1]; + + const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1 + const int ne1 = src0->ne[1]; // seq_len_without_past + //const int ne2 = src0->ne[2]; // n_head -> this is k + //const int ne3 = src0->ne[3]; // 1 -> bsz + + const int n = ggml_nrows(src0); + const int ne2_ne3 = n/ne1; // ne2*ne3 + + const int nb0 = src0->nb[0]; + const int nb1 = src0->nb[1]; + const int nb2 = src0->nb[2]; + //const int nb3 = src0->nb[3]; + + assert(nb0 == sizeof(ggml_fp16_t)); + assert(ne1 + n_past == ne0); (void) n_past; + + // add alibi to src0 (KQ_scaled) + const int n_heads_log2_floor = 1 << (int) floor(log2(n_head)); + + const float m0 = powf(2.0f, -8.0f / n_heads_log2_floor); + const float m1 = powf(2.0f, -4.0f / n_heads_log2_floor); + + for (int i = 0; i < ne0; i++) { + for (int j = 0; j < ne1; j++) { + for (int k = 0; k < ne2_ne3; k++) { + ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2); + float * pdst = (float *)((char *) dst->data + i*nb0 + j*nb1 + k*nb2); + + // TODO: k*nb2 or k*nb3 + + float m_k; + + if (k < n_heads_log2_floor) { + m_k = powf(m0, k + 1); + } else { + m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1); + } + + // we return F32 + pdst[0] = i * m_k + GGML_FP16_TO_FP32(src[0]); + } + } + } +} + +static void ggml_compute_forward_alibi( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_alibi_f16(params, src0, src1, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_alibi_f32(params, src0, src1, dst); + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_I8: + case GGML_TYPE_I16: + case GGML_TYPE_I32: + case GGML_TYPE_COUNT: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_rope static void ggml_compute_forward_rope_f32( @@ -8589,6 +10640,217 @@ static void ggml_compute_forward_rope_f32( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(src1) == 3); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + + //const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + const int64_t ne3 = src0->ne[3]; + + const int nb0 = src0->nb[0]; + const int nb1 = src0->nb[1]; + const int nb2 = src0->nb[2]; + const int nb3 = src0->nb[3]; + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + GGML_ASSERT(nb0 == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT(n_dims <= nc); + GGML_ASSERT(n_dims % 2 == 0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(10000.0, -2.0f/n_dims); + + const bool is_neox = mode & 2; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int p = ((mode & 1) == 0 ? n_past + i2 : i2); + for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + float theta = (float)p; + + for (int i0 = 0; i0 < n_dims; i0 += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); + + theta *= theta_scale; + + if (!is_neox) { + const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = src[0]; + const float x1 = src[1]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[1] = x0*sin_theta + x1*cos_theta; + } else { + const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + } + } + } + } + } +} + +static void ggml_compute_forward_rope_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(src1) == 3); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + + //const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + const int64_t ne3 = src0->ne[3]; + + const int nb0 = src0->nb[0]; + const int nb1 = src0->nb[1]; + const int nb2 = src0->nb[2]; + const int nb3 = src0->nb[3]; + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT(n_dims <= nc); + GGML_ASSERT(n_dims % 2 == 0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(10000.0, -2.0f/n_dims); + + const bool is_neox = mode & 2; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int p = ((mode & 1) == 0 ? n_past + i2 : i2); + for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + float theta = (float)p; + + for (int i0 = 0; i0 < n_dims; i0 += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); + + theta *= theta_scale; + + if (!is_neox) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[1]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } else { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } + } + } + } +} + +static void ggml_compute_forward_rope( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_rope_f16(params, src0, src1, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_rope_f32(params, src0, src1, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +// ggml_compute_forward_rope_back + +static void ggml_compute_forward_rope_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { assert(src1->type == GGML_TYPE_I32); assert(ggml_nelements(src1) == 3); @@ -8596,6 +10858,10 @@ static void ggml_compute_forward_rope_f32( return; } + // y = rope(x, src1) + // dx = rope_back(dy, src1) + // src0 is dy, src1 contains options + const int n_past = ((int32_t *) src1->data)[0]; const int n_dims = ((int32_t *) src1->data)[1]; const int mode = ((int32_t *) src1->data)[2]; @@ -8632,9 +10898,11 @@ static void ggml_compute_forward_rope_f32( const float theta_scale = powf(10000.0, -2.0f/n_dims); + const bool is_neox = mode & 2; + for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) { - const int p = (mode == 0 ? n_past + i2 : i2); + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int p = ((mode & 1) == 0 ? n_past + i2 : i2); for (int64_t i1 = 0; i1 < ne1; i1++) { if (ir++ < ir0) continue; if (ir > ir1) break; @@ -8647,21 +10915,32 @@ static void ggml_compute_forward_rope_f32( theta *= theta_scale; - const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + if (!is_neox) { + const float * const dy = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - const float x0 = src[0]; - const float x1 = src[1]; + const float dy0 = dy[0]; + const float dy1 = dy[1]; - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[1] = x0*sin_theta + x1*cos_theta; + dx[0] = dy0*cos_theta + dy1*sin_theta; + dx[1] = - dy0*sin_theta + dy1*cos_theta; + } else { + const float * const dy = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + + const float dy0 = dy[0]; + const float dy1 = dy[n_dims/2]; + + dx[0] = dy0*cos_theta + dy1*sin_theta; + dx[n_dims/2] = - dy0*sin_theta + dy1*cos_theta; + } } } } } } -static void ggml_compute_forward_rope_f16( +static void ggml_compute_forward_rope_back_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -8673,6 +10952,10 @@ static void ggml_compute_forward_rope_f16( return; } + // y = rope(x, src1) + // dx = rope_back(dy, src1) + // src0 is dy, src1 contains options + const int n_past = ((int32_t *) src1->data)[0]; const int n_dims = ((int32_t *) src1->data)[1]; const int mode = ((int32_t *) src1->data)[2]; @@ -8709,9 +10992,11 @@ static void ggml_compute_forward_rope_f16( const float theta_scale = powf(10000.0, -2.0f/n_dims); + const bool is_neox = mode & 2; + for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) { - const int p = (mode == 0 ? n_past + i2 : i2); + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int p = ((mode & 1) == 0 ? n_past + i2 : i2); for (int64_t i1 = 0; i1 < ne1; i1++) { if (ir++ < ir0) continue; if (ir > ir1) break; @@ -8724,21 +11009,32 @@ static void ggml_compute_forward_rope_f16( theta *= theta_scale; - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + if (!is_neox) { + const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[1]); + const float dy0 = GGML_FP16_TO_FP32(dy[0]); + const float dy1 = GGML_FP16_TO_FP32(dy[1]); - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); + dx[1] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); + } else { + const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + + const float dy0 = GGML_FP16_TO_FP32(dy[0]); + const float dy1 = GGML_FP16_TO_FP32(dy[n_dims/2]); + + dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); + dx[n_dims/2] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); + } } } } } } -static void ggml_compute_forward_rope( +static void ggml_compute_forward_rope_back( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -8746,11 +11042,11 @@ static void ggml_compute_forward_rope( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_f16(params, src0, src1, dst); + ggml_compute_forward_rope_back_f16(params, src0, src1, dst); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_f32(params, src0, src1, dst); + ggml_compute_forward_rope_back_f32(params, src0, src1, dst); } break; default: { @@ -10078,6 +12374,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_add(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_ADD1: + { + ggml_compute_forward_add1(params, tensor->src0, tensor->src1, tensor); + } break; + case GGML_OP_ACC: + { + ggml_compute_forward_acc(params, tensor->src0, tensor->src1, tensor->opt[0], tensor); + } break; case GGML_OP_SUB: { ggml_compute_forward_sub(params, tensor->src0, tensor->src1, tensor); @@ -10098,10 +12402,18 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_sqrt(params, tensor->src0, tensor); } break; + case GGML_OP_LOG: + { + ggml_compute_forward_log(params, tensor->src0, tensor); + } break; case GGML_OP_SUM: { ggml_compute_forward_sum(params, tensor->src0, tensor); } break; + case GGML_OP_SUM_ROWS: + { + ggml_compute_forward_sum_rows(params, tensor->src0, tensor); + } break; case GGML_OP_MEAN: { ggml_compute_forward_mean(params, tensor->src0, tensor); @@ -10138,6 +12450,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_silu(params, tensor->src0, tensor); } break; + case GGML_OP_SILU_BACK: + { + ggml_compute_forward_silu_back(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_NORM: { ggml_compute_forward_norm(params, tensor->src0, tensor); @@ -10146,6 +12462,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_rms_norm(params, tensor->src0, tensor); } break; + case GGML_OP_RMS_NORM_BACK: + { + ggml_compute_forward_rms_norm_back(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_MUL_MAT: { ggml_compute_forward_mul_mat(params, tensor->src0, tensor->src1, tensor); @@ -10154,6 +12474,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_scale(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_SET: + { + ggml_compute_forward_set(params, tensor->src0, tensor->src1, tensor->opt[0], tensor); + } break; case GGML_OP_CPY: { ggml_compute_forward_cpy(params, tensor->src0, tensor); @@ -10182,10 +12506,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_get_rows(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_GET_ROWS_BACK: + { + ggml_compute_forward_get_rows_back(params, tensor->src0, tensor->src1, tensor->opt[0], tensor); + } break; + case GGML_OP_DIAG: + { + ggml_compute_forward_diag(params, tensor->src0, tensor); + } break; case GGML_OP_DIAG_MASK_INF: { ggml_compute_forward_diag_mask_inf(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_DIAG_MASK_ZERO: + { + ggml_compute_forward_diag_mask_zero(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_SOFT_MAX: { ggml_compute_forward_soft_max(params, tensor->src0, tensor); @@ -10194,6 +12530,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_rope(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_ROPE_BACK: + { + ggml_compute_forward_rope_back(params, tensor->src0, tensor->src1, tensor); + } break; + case GGML_OP_ALIBI: + { + ggml_compute_forward_alibi(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_CONV_1D_1S: { ggml_compute_forward_conv_1d_1s(params, tensor->src0, tensor->src1, tensor); @@ -10258,6 +12602,48 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src1->grad = ggml_add_impl(ctx, src1->grad, tensor->grad, inplace); } } break; + case GGML_OP_ADD1: + { + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } + if (src1->grad) { + src1->grad = ggml_add_impl(ctx, + src1->grad, + ggml_mean(ctx, tensor->grad), // TODO: should probably be sum instead of mean + inplace); + } + } break; + case GGML_OP_ACC: + { + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } + if (src1->grad) { + GGML_ASSERT(ggml_nelements(tensor->opt[0]) == 5); + GGML_ASSERT(tensor->opt[0]->type == GGML_TYPE_I32); + const size_t nb1 = (( int32_t * ) tensor->opt[0]->data)[0]; + const size_t nb2 = (( int32_t * ) tensor->opt[0]->data)[1]; + const size_t nb3 = (( int32_t * ) tensor->opt[0]->data)[2]; + const size_t offset = (( int32_t * ) tensor->opt[0]->data)[3]; + + struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx, + tensor->grad, + src1->grad->ne[0], + src1->grad->ne[1], + src1->grad->ne[2], + src1->grad->ne[3], + nb1, nb2, nb3, offset); + + src1->grad = + ggml_add_impl(ctx, + src1->grad, + ggml_reshape(ctx, + ggml_cont(ctx, tensor_grad_view), + src1->grad), + inplace); + } + } break; case GGML_OP_SUB: { if (src0->grad) { @@ -10309,31 +12695,57 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_mul(ctx, + ggml_scale(ctx, ggml_mul(ctx, src0, tensor->grad), - ggml_repeat(ctx, ggml_new_f32(ctx, 2.0f), src0)), + ggml_new_f32(ctx, 2.0f)), inplace); } } break; case GGML_OP_SQRT: + { + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, + src0->grad, + ggml_mul(ctx, + tensor->grad, // this was not catched by test_grad because in test_grad tensor->grad is 1 + ggml_div(ctx, + ggml_repeat(ctx, ggml_new_f32(ctx, 0.5f), tensor), + tensor)), + inplace); + } + } break; + case GGML_OP_LOG: { if (src0->grad) { src0->grad = ggml_add_impl(ctx, src0->grad, ggml_div(ctx, - ggml_repeat(ctx, ggml_new_f32(ctx, 0.5f), tensor), - tensor), + tensor->grad, + src0), inplace); } } break; case GGML_OP_SUM: + { + if (src0->grad) { + src0->grad = + ggml_add1_impl(ctx, + src0->grad, + tensor->grad, + inplace); + } + } break; + case GGML_OP_SUM_ROWS: { if (src0->grad) { src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_repeat(ctx, tensor->grad, src0->grad), + ggml_repeat(ctx, + tensor->grad, + src0->grad), inplace); } } break; @@ -10343,11 +12755,44 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } break; case GGML_OP_REPEAT: { + // necessary for llama if (src0->grad) { + GGML_ASSERT(src0->n_dims == 1 || src0->n_dims == 2); + const int nc = tensor->ne[0]; + const int nr = tensor->ne[1]; + const int nc0 = src0->ne[0]; + const int nr0 = src0->ne[1]; + const int ncr = nc/nc0; // guaranteed to be an integer due to the check in ggml_can_repeat + const int nrr = nr/nr0; // guaranteed to be an integer due to the check in ggml_can_repeat + // tensor->grad [nc,nr,1,1] + // reshape [nc0,nc/nc0,nr0,nr/nr0] + // permute [nc0,nr0,nc/nc0,nr/nr0] + // substitute [nc0,nr0,ncr,nrr] + // reshape [nc0*nr0,ncr*nrr,1,1] + // transpose [ncr*nrr,nc0*nr0,1,1] + // sum rows [1,nc0*nr0,1,1] + // transpose [nc0*nr0,1,1] + // reshape [nc0,nr0,1,1] reshape_1d or reshape_2d + // add to src0->grad + + int64_t ne[4] = {nc0,ncr,nr0,nrr}; + + struct ggml_tensor* F00 = tensor->grad; + struct ggml_tensor* F01 = ggml_reshape (ctx, F00, ggml_new_tensor(ctx,tensor->grad->type,4,ne)); + struct ggml_tensor* F02 = ggml_permute (ctx, F01, 0,2,1,3); + struct ggml_tensor* F03 = ggml_cont (ctx, F02); + struct ggml_tensor* F04 = ggml_reshape_2d(ctx, F03, nc0*nr0, ncr*nrr); + struct ggml_tensor* F05 = ggml_transpose (ctx, F04); + struct ggml_tensor* F06 = ggml_cont (ctx, F05); + struct ggml_tensor* F07 = ggml_sum_rows (ctx, F06); + struct ggml_tensor* F08 = ggml_transpose (ctx, F07); + struct ggml_tensor* F09 = ggml_cont (ctx, F08); + struct ggml_tensor* F10 = ggml_reshape (ctx, F09, src0->grad); + src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_sum(ctx, tensor->grad), + F10, inplace); } } break; @@ -10396,7 +12841,21 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; + case GGML_OP_ALIBI: + { + GGML_ASSERT(false); // TODO: not implemented + } break; case GGML_OP_SILU: + { + // necessary for llama + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_silu_back(ctx, src0, tensor->grad), + inplace); + } + } break; + case GGML_OP_SILU_BACK: { GGML_ASSERT(false); // TODO: not implemented } break; @@ -10405,68 +12864,372 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_RMS_NORM: + { + // necessary for llama + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_rms_norm_back(ctx, src0, tensor->grad), + inplace); + } + } break; + case GGML_OP_RMS_NORM_BACK: { GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_MUL_MAT: { + // https://cs231n.github.io/optimization-2/#staged + // # forward pass + // s0 = np.random.randn(5, 10) + // s1 = np.random.randn(10, 3) + // t = s0.dot(s1) + + // # now suppose we had the gradient on t from above in the circuit + // dt = np.random.randn(*t.shape) # same shape as t + // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix + // ds1 = t.T.dot(dt) + + // tensor.shape [m,p] + // src0.shape [n,m] + // src1.shape [n,p] + + // necessary for llama if (src0->grad) { // TODO: this requires outer product - ggml_out_prod(ctx, src1, tensor->grad); - GGML_ASSERT(false); + src0->grad = + ggml_add_impl(ctx, + src0->grad, + // ds0 = dt.dot(s1.T) + // ggml_out_prod(ctx, // [n,m] + // src1, // [n,p] + // tensor->grad), // [m,p] + // for now just using A*B==(B.T*A.T).T + ggml_cont(ctx, // [n,m] + ggml_transpose(ctx, // [n,m] + ggml_mul_mat(ctx, // [m,n] + ggml_cont(ctx, // [p,m] + ggml_transpose(ctx, // [p,m] + tensor->grad)), // [m,p] + ggml_cont(ctx, // [p,n] + ggml_transpose(ctx, // [p,n] + src1))))), // [n,p] + inplace); } if (src1->grad) { src1->grad = ggml_add_impl(ctx, src1->grad, - ggml_mul_mat(ctx, - ggml_cont(ctx, ggml_transpose(ctx, src0)), - tensor->grad), + // ds1 = s0.T.dot(dt): + ggml_mul_mat(ctx, // [n,p] + ggml_cont(ctx, // [m,n] + ggml_transpose(ctx, src0)), // [m,n] + tensor->grad), // [m,p] inplace); } } break; case GGML_OP_SCALE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, + src0->grad, + ggml_scale_impl(ctx, tensor->grad, src1, false), + inplace); + } + if (src1->grad) { + src1->grad = + ggml_add_impl(ctx, + src1->grad, + ggml_sum(ctx, ggml_mul_impl(ctx, tensor->grad, src0, false)), + inplace); + } + } break; + case GGML_OP_SET: + { + GGML_ASSERT(ggml_nelements(tensor->opt[0]) == 5); + GGML_ASSERT(tensor->opt[0]->type == GGML_TYPE_I32); + const size_t nb1 = (( int32_t * ) tensor->opt[0]->data)[0]; + const size_t nb2 = (( int32_t * ) tensor->opt[0]->data)[1]; + const size_t nb3 = (( int32_t * ) tensor->opt[0]->data)[2]; + const size_t offset = (( int32_t * ) tensor->opt[0]->data)[3]; + + struct ggml_tensor * tensor_grad_view = NULL; + + if (src0->grad || src1->grad) { + GGML_ASSERT(src0->type == tensor->type); + GGML_ASSERT(tensor->grad->type == tensor->type); + GGML_ASSERT(tensor->grad->type == src1->grad->type); + + tensor_grad_view = ggml_view_4d(ctx, + tensor->grad, + src1->grad->ne[0], + src1->grad->ne[1], + src1->grad->ne[2], + src1->grad->ne[3], + nb1, nb2, nb3, offset); + } + + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_acc_impl(ctx, + tensor->grad, + ggml_neg(ctx, tensor_grad_view), + nb1, nb2, nb3, offset, false), + inplace); + } + + if (src1->grad) { + src1->grad = + ggml_add_impl(ctx, + src1->grad, + ggml_reshape(ctx, + ggml_cont(ctx, tensor_grad_view), + src1->grad), + inplace); + } } break; case GGML_OP_CPY: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + // cpy overwrites value of src1 by src0 and returns view(src1) + // the overwriting is mathematically equivalent to: + // tensor = src0 * 1 + src1 * 0 + if (src0->grad) { + // dsrc0 = dtensor * 1 + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } + if (src1->grad) { + // dsrc1 = dtensor * 0 -> noop + } } break; case GGML_OP_CONT: { - GGML_ASSERT(false); // TODO: not implemented + // same as cpy + if (src0->grad) { + GGML_ASSERT(ggml_is_contiguous(src0->grad)); + GGML_ASSERT(ggml_is_contiguous(tensor->grad)); + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } } break; case GGML_OP_RESHAPE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_reshape(ctx, tensor->grad, src0->grad), + inplace); + } } break; case GGML_OP_VIEW: { - GGML_ASSERT(false); // not supported + // necessary for llama + if (src0->grad) { + size_t offset; + memcpy(&offset, tensor->padding, sizeof(offset)); + + size_t nb1 = tensor->nb[1]; + size_t nb2 = tensor->nb[2]; + size_t nb3 = tensor->nb[3]; + + if (src0->type != src0->grad->type) { + // gradient is typically F32, but src0 could be other type + size_t ng = ggml_element_size(src0->grad); + size_t n0 = ggml_element_size(src0); + GGML_ASSERT(offset % n0 == 0); + GGML_ASSERT(nb1 % n0 == 0); + GGML_ASSERT(nb2 % n0 == 0); + GGML_ASSERT(nb3 % n0 == 0); + offset = (offset / n0) * ng; + nb1 = (nb1 / n0) * ng; + nb2 = (nb2 / n0) * ng; + nb3 = (nb3 / n0) * ng; + } + + src0->grad = ggml_acc_impl(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, inplace); + } } break; case GGML_OP_PERMUTE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + int axis0 = tensor->padding[0] & 0x3; + int axis1 = tensor->padding[1] & 0x3; + int axis2 = tensor->padding[2] & 0x3; + int axis3 = tensor->padding[3] & 0x3; + int axes_backward[4] = {0,0,0,0}; + axes_backward[axis0] = 0; + axes_backward[axis1] = 1; + axes_backward[axis2] = 2; + axes_backward[axis3] = 3; + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_permute(ctx, + tensor->grad, + axes_backward[0], + axes_backward[1], + axes_backward[2], + axes_backward[3]), + inplace); + } } break; case GGML_OP_TRANSPOSE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_transpose(ctx, tensor->grad), + inplace); + } } break; case GGML_OP_GET_ROWS: + { + // necessary for llama (only for tokenizer) + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_get_rows_back(ctx, tensor->grad, src1, src0->grad), + inplace); + } + if (src1->grad) { + // noop + } + } break; + case GGML_OP_GET_ROWS_BACK: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_OP_DIAG: { GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_DIAG_MASK_INF: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + const int n_past = ((int32_t *) src1->data)[0]; + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), + inplace); + } + if (src1->grad) { + // noop + } + } break; + case GGML_OP_DIAG_MASK_ZERO: + { + // necessary for llama + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + const int n_past = ((int32_t *) src1->data)[0]; + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), + inplace); + } + if (src1->grad) { + // noop + } } break; case GGML_OP_SOFT_MAX: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + // y = softmax(x) + // + // Jii = yi - yi*yi + // Jij = -yi*yj + // J = diag(y)-y.*y + // dx = J * dy + // dxk = sum(Jkj * dyk) + + int64_t ne2[4] = { + tensor->ne[0], + 1, + tensor->ne[1]*tensor->ne[2], + tensor->ne[3] + }; + struct ggml_tensor * tensor2 = ggml_cont(ctx, + ggml_reshape_4d(ctx, + ggml_cont(ctx, tensor), + ne2[0], ne2[1], ne2[2], ne2[3])); + + struct ggml_tensor * grad2 = ggml_cont(ctx, + ggml_reshape_4d(ctx, + ggml_cont(ctx, tensor->grad), + ne2[0], ne2[1], ne2[2], ne2[3])); + + struct ggml_tensor * tensor2_t = ggml_cont(ctx, // [1,ne0,ne1*ne2,ne3] + ggml_permute(ctx, // [1,ne0,ne1*ne2,ne3] + tensor2, // [ne0,1,ne1*ne2,ne3] + 1, 0, 2, 3)); + + src0->grad = + ggml_add_impl(ctx, + src0->grad, // [ne0,ne1,ne2,ne3] + ggml_reshape(ctx, // [ne0,ne1,ne2,ne3] + ggml_mul_mat(ctx, // [ne0,1,ne1*ne2,ne3] + ggml_sub(ctx, // [ne0,ne0,ne1*ne2,ne3] + ggml_diag(ctx, // [ne0,ne0,ne1*ne2,ne3] + tensor2), // [ne0,1,ne1*ne2,ne3] + ggml_mul_mat(ctx, // [ne0,ne0,ne1*ne2,ne3] + tensor2_t, // [1,ne0,ne1*ne2,ne3] + tensor2_t)), // [1,ne0,ne1*ne2,ne3] + grad2), // [ne0,1,ne1*ne2,ne3] + src0->grad), + inplace); + } } break; case GGML_OP_ROPE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 3); + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_rope_back(ctx, + tensor->grad, + n_past, + n_dims, + mode), + inplace); + } + if (src1->grad) { + // noop + } + } break; + case GGML_OP_ROPE_BACK: + { + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 3); + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_rope(ctx, + tensor->grad, + n_past, + n_dims, + mode), + inplace); + } + if (src1->grad) { + // noop + } } break; case GGML_OP_CONV_1D_1S: { @@ -10680,7 +13443,11 @@ typedef int ggml_lock_t; #define ggml_lock_init(x) UNUSED(x) #define ggml_lock_destroy(x) UNUSED(x) +#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) +#define ggml_lock_lock(x) _mm_pause() +#else #define ggml_lock_lock(x) UNUSED(x) +#endif #define ggml_lock_unlock(x) UNUSED(x) #define GGML_LOCK_INITIALIZER 0 @@ -10820,6 +13587,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) work_size = MAX(work_size, cur); } break; case GGML_OP_ADD: + case GGML_OP_ADD1: { node->n_tasks = n_threads; @@ -10829,14 +13597,27 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->src0->ne[0] * n_threads; } + work_size = MAX(work_size, cur); + } break; + case GGML_OP_ACC: + { + node->n_tasks = n_threads; + + size_t cur = 0; + + if (ggml_is_quantized(node->src0->type)) { + cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->src1->ne[0] * n_threads; + } + work_size = MAX(work_size, cur); } break; case GGML_OP_SUB: - case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_SQR: case GGML_OP_SQRT: + case GGML_OP_LOG: case GGML_OP_SUM: + case GGML_OP_SUM_ROWS: case GGML_OP_MEAN: case GGML_OP_REPEAT: case GGML_OP_ABS: @@ -10847,16 +13628,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { node->n_tasks = 1; } break; + case GGML_OP_MUL: case GGML_OP_GELU: - { - node->n_tasks = n_threads; - } break; case GGML_OP_SILU: - { - node->n_tasks = n_threads; - } break; + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: { node->n_tasks = n_threads; } break; @@ -10873,15 +13651,21 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) size_t cur = 0; +#if defined(GGML_USE_CUBLAS) + if (ggml_cuda_can_mul_mat(node->src0, node->src1, node)) { + node->n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + cur = ggml_cuda_mul_mat_get_wsize(node->src0, node->src1, node); + } + else +#endif if (node->src0->type == GGML_TYPE_F16 && node->src1->type == GGML_TYPE_F32) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) { node->n_tasks = 1; // TODO: this actually is doing nothing // the threads are still spinning + // here we need memory just for single 2D matrix from src0 cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]); - //printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]); - //printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]); - //printf("cur = %zu\n", cur); } else { cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1); } @@ -10890,15 +13674,21 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) #endif } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) { cur = 0; - } else if (quantize_fns[node->src0->type].vec_dot_q && node->src1->type == GGML_TYPE_F32) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) + if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) { + node->n_tasks = 1; + } +#endif + } else if (ggml_is_quantized(node->src0->type) && node->src1->type == GGML_TYPE_F32) { +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) { node->n_tasks = 1; cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]); } else #endif { - cur = GGML_TYPE_SIZE[GGML_TYPE_Q8_0]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[GGML_TYPE_Q8_0]; + const enum ggml_type type_q = quantize_fns[node->src0->type].vec_dot_type; + cur = GGML_TYPE_SIZE[type_q]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[type_q]; } } else { GGML_ASSERT(false); @@ -10910,23 +13700,29 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { node->n_tasks = n_threads; } break; + case GGML_OP_SET: case GGML_OP_CONT: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_GET_ROWS: - case GGML_OP_DIAG_MASK_INF: + case GGML_OP_GET_ROWS_BACK: + case GGML_OP_DIAG: + case GGML_OP_DIAG_MASK_ZERO: { node->n_tasks = 1; } break; + case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: { node->n_tasks = n_threads; } break; - case GGML_OP_ROPE: + case GGML_OP_ALIBI: { - node->n_tasks = n_threads; + node->n_tasks = 1; //TODO } break; case GGML_OP_CONV_1D_1S: case GGML_OP_CONV_1D_2S: @@ -11226,9 +14022,9 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { for (int i = 0; i < cgraph->n_nodes; i++) { struct ggml_tensor * node = cgraph->nodes[i]; - perf_total_per_op_us[node->op] += node->perf_time_us; + perf_total_per_op_us[node->op] += MAX(1, node->perf_time_us); - GGML_PRINT(" - %3d: [ %" PRId64 ", %" PRId64 ", %" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n", + GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n", i, node->ne[0], node->ne[1], node->ne[2], GGML_OP_LABEL[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs, @@ -11242,13 +14038,17 @@ 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: [ %" PRId64 ", %" PRId64 "] %8s\n", + GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n", i, node->ne[0], node->ne[1], GGML_OP_LABEL[node->op]); } for (int i = 0; i < GGML_OP_COUNT; i++) { + if (perf_total_per_op_us[i] == 0) { + continue; + } + GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_LABEL[i], (double) perf_total_per_op_us[i] / 1000.0); } @@ -11311,10 +14111,16 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph snprintf(color, sizeof(color), "white"); } - fprintf(fp, " \"%p\" [ \ -style = filled; fillcolor = %s; shape = record; \ -label=\"%d [%" PRId64 ", %" PRId64 "] | %s", - (void *) node, color, + fprintf(fp, " \"%p\" [ " + "style = filled; fillcolor = %s; shape = record; " + "label=\"", + (void *) node, color); + + if (strlen(node->name) > 0) { + fprintf(fp, "%s |", node->name); + } + + fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | %s", i, node->ne[0], node->ne[1], GGML_OP_SYMBOL[node->op]); @@ -11330,18 +14136,26 @@ label=\"%d [%" PRId64 ", %" PRId64 "] | %s", snprintf(color, sizeof(color), "pink"); - if (ggml_nelements(node) == 1) { - fprintf(fp, " \"%p\" [ \ -style = filled; fillcolor = %s; shape = record; \ -label=\"%.1e\"; ]\n", - (void *) node, color, (double)ggml_get_f32_1d(node, 0)); - } else { - fprintf(fp, " \"%p\" [ \ -style = filled; fillcolor = %s; shape = record; \ -label=\"CONST %d [%" PRId64 ", %" PRId64 "]\"; ]\n", - (void *) node, color, - i, node->ne[0], node->ne[1]); + fprintf(fp, " \"%p\" [ " + "style = filled; fillcolor = %s; shape = record; " + "label=\"", + (void *) node, color); + + if (strlen(node->name) > 0) { + fprintf(fp, "%s | ", node->name); } + if (ggml_nelements(node) == 1) { + if (node->type == GGML_TYPE_I8 || node->type == GGML_TYPE_I16 || node->type == GGML_TYPE_I32) { + fprintf(fp, "%d", ggml_get_i32_1d(node, 0)); + } + else { + fprintf(fp, "%.1e", (double)ggml_get_f32_1d(node, 0)); + } + } + else { + fprintf(fp, "CONST %d [%" PRId64 ", %" PRId64 "]", i, node->ne[0], node->ne[1]); + } + fprintf(fp, "\"; ]\n"); } for (int i = 0; i < gb->n_nodes; i++) { @@ -12039,7 +14853,7 @@ enum ggml_opt_result ggml_opt( // build forward + backward compute graphs struct ggml_cgraph gf = ggml_build_forward (f); - struct ggml_cgraph gb = ggml_build_backward(ctx, &gf, false); + struct ggml_cgraph gb = ggml_build_backward(ctx, &gf, true); switch (params.type) { case GGML_OPT_ADAM: @@ -12075,15 +14889,15 @@ size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * assert(k % QK4_0 == 0); const int nb = k / QK4_0; - for (int j = 0; j < n; j += k) { - block_q4_0 * restrict y = (block_q4_0 *)dst + j/QK4_0; + for (int b = 0; b < n; b += k) { + block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0; - quantize_row_q4_0_reference(src + j, y, k); + quantize_row_q4_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0xF; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + for (int j = 0; j < QK4_0; j += 2) { + const uint8_t vi0 = y[i].qs[j/2] & 0x0F; + const uint8_t vi1 = y[i].qs[j/2] >> 4; hist[vi0]++; hist[vi1]++; @@ -12098,15 +14912,15 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * assert(k % QK4_1 == 0); const int nb = k / QK4_1; - for (int j = 0; j < n; j += k) { - block_q4_1 * restrict y = (block_q4_1 *)dst + j/QK4_1; + for (int b = 0; b < n; b += k) { + block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1; - quantize_row_q4_1_reference(src + j, y, k); + quantize_row_q4_1_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0xF; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + for (int j = 0; j < QK4_1; j += 2) { + const uint8_t vi0 = y[i].qs[j/2] & 0x0F; + const uint8_t vi1 = y[i].qs[j/2] >> 4; hist[vi0]++; hist[vi1]++; @@ -12117,19 +14931,26 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * return (n/QK4_1*sizeof(block_q4_1)); } -size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK4_2 == 0); - const int nb = k / QK4_2; +size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK5_0 == 0); + const int nb = k / QK5_0; - for (int j = 0; j < n; j += k) { - block_q4_2 * restrict y = (block_q4_2 *)dst + j/QK4_2; + for (int b = 0; b < n; b += k) { + block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0; - quantize_row_q4_2_reference(src + j, y, k); + quantize_row_q5_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0xF; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + uint32_t qh; + memcpy(&qh, &y[i].qh, sizeof(qh)); + + for (int j = 0; j < QK5_0; j += 2) { + const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12)); + + // cast to 16 bins + const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; hist[vi0]++; hist[vi1]++; @@ -12137,7 +14958,97 @@ size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * } } - return (n/QK4_2*sizeof(block_q4_2)); + return (n/QK5_0*sizeof(block_q5_0)); +} + +size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK5_1 == 0); + const int nb = k / QK5_1; + + for (int b = 0; b < n; b += k) { + block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1; + + quantize_row_q5_1_reference(src + b, y, k); + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, &y[i].qh, sizeof(qh)); + + for (int j = 0; j < QK5_1; j += 2) { + const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12)); + + // cast to 16 bins + const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/QK5_1*sizeof(block_q5_1)); +} + +size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + for (int b = 0; b < n; b += k) { + block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0; + + quantize_row_q8_0_reference(src + b, y, k); + + for (int i = 0; i < nb; i++) { + for (int j = 0; j < QK8_0; ++j) { + const int8_t vi = y[i].qs[j]; + + hist[vi/16 + 8]++; + } + } + } + + return (n/QK8_0*sizeof(block_q8_0)); +} + +size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) { + size_t result = 0; + switch (type) { + case GGML_TYPE_Q4_0: + { + GGML_ASSERT(start % QK4_0 == 0); + block_q4_0 * block = (block_q4_0*)dst + start / QK4_0; + result = ggml_quantize_q4_0(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q4_1: + { + GGML_ASSERT(start % QK4_1 == 0); + block_q4_1 * block = (block_q4_1*)dst + start / QK4_1; + result = ggml_quantize_q4_1(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q5_0: + { + GGML_ASSERT(start % QK5_0 == 0); + block_q5_0 * block = (block_q5_0*)dst + start / QK5_0; + result = ggml_quantize_q5_0(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q5_1: + { + GGML_ASSERT(start % QK5_1 == 0); + block_q5_1 * block = (block_q5_1*)dst + start / QK5_1; + result = ggml_quantize_q5_1(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q8_0: + { + GGML_ASSERT(start % QK8_0 == 0); + block_q8_0 * block = (block_q8_0*)dst + start / QK8_0; + result = ggml_quantize_q8_0(src + start, block, n, n, hist); + } break; + default: + assert(false); + } + return result; } //////////////////////////////////////////////////////////////////////////////// @@ -12231,13 +15142,33 @@ int ggml_cpu_has_wasm_simd(void) { } int ggml_cpu_has_blas(void) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) return 1; #else return 0; #endif } +int ggml_cpu_has_cublas(void) { +#if defined(GGML_USE_CUBLAS) + return 1; +#else + return 0; +#endif +} + +int ggml_cpu_has_clblast(void) { +#if defined(GGML_USE_CLBLAST) + return 1; +#else + return 0; +#endif +} + +int ggml_cpu_has_gpublas(void) { + return ggml_cpu_has_cublas() || ggml_cpu_has_clblast(); +} + int ggml_cpu_has_sse3(void) { #if defined(__SSE3__) return 1; diff --git a/ggml.h b/ggml.h index 603be8453..967ef72d0 100644 --- a/ggml.h +++ b/ggml.h @@ -169,689 +169,956 @@ // // -#ifdef __cplusplus -extern "C" { +#ifdef GGML_SHARED +# if defined(_WIN32) && !defined(__MINGW32__) +# ifdef GGML_BUILD +# define GGML_API __declspec(dllexport) +# else +# define GGML_API __declspec(dllimport) +# endif +# else +# define GGML_API __attribute__ ((visibility ("default"))) +# endif +#else +# define GGML_API #endif #include #include #include +#define GGML_FILE_MAGIC 0x67676d6c // "ggml" +#define GGML_FILE_VERSION 1 + #define GGML_MAX_DIMS 4 #define GGML_MAX_NODES 4096 -#define GGML_MAX_PARAMS 16 +#define GGML_MAX_PARAMS 256 #define GGML_MAX_CONTEXTS 64 #define GGML_MAX_OPT 4 #define GGML_DEFAULT_N_THREADS 4 -#ifdef __ARM_NEON -// we use the built-in 16-bit float type -typedef __fp16 ggml_fp16_t; -#else -typedef uint16_t ggml_fp16_t; -#endif - -// convert FP16 <-> FP32 -float ggml_fp16_to_fp32(ggml_fp16_t x); -ggml_fp16_t ggml_fp32_to_fp16(float x); - -struct ggml_object; -struct ggml_context; - -enum ggml_type { - // explicitly numbered values are used in llama.cpp files - GGML_TYPE_F32 = 0, - GGML_TYPE_F16 = 1, - GGML_TYPE_Q4_0 = 2, - GGML_TYPE_Q4_1 = 3, - GGML_TYPE_Q4_2 = 4, - GGML_TYPE_Q8_0 = 5, - GGML_TYPE_I8, - GGML_TYPE_I16, - GGML_TYPE_I32, - GGML_TYPE_COUNT, -}; - -// available tensor operations: -enum ggml_op { - GGML_OP_NONE = 0, - - GGML_OP_DUP, - GGML_OP_ADD, - GGML_OP_SUB, - GGML_OP_MUL, - GGML_OP_DIV, - GGML_OP_SQR, - GGML_OP_SQRT, - GGML_OP_SUM, - GGML_OP_MEAN, - GGML_OP_REPEAT, - GGML_OP_ABS, - GGML_OP_SGN, - GGML_OP_NEG, - GGML_OP_STEP, - GGML_OP_RELU, - GGML_OP_GELU, - GGML_OP_SILU, - GGML_OP_NORM, // normalize - GGML_OP_RMS_NORM, - - GGML_OP_MUL_MAT, - - GGML_OP_SCALE, - GGML_OP_CPY, - GGML_OP_CONT, - GGML_OP_RESHAPE, - GGML_OP_VIEW, - GGML_OP_PERMUTE, - GGML_OP_TRANSPOSE, - GGML_OP_GET_ROWS, - GGML_OP_DIAG_MASK_INF, - GGML_OP_SOFT_MAX, - GGML_OP_ROPE, - GGML_OP_CONV_1D_1S, - GGML_OP_CONV_1D_2S, - - GGML_OP_FLASH_ATTN, - GGML_OP_FLASH_FF, - - GGML_OP_MAP_UNARY, - GGML_OP_MAP_BINARY, - - GGML_OP_COUNT, -}; - - -// ggml object -struct ggml_object { - size_t offs; - size_t size; - - struct ggml_object * next; - - char padding[8]; -}; - -static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); - -// n-dimensional tensor -struct ggml_tensor { - enum ggml_type type; - - int n_dims; - int64_t ne[GGML_MAX_DIMS]; // number of elements - size_t nb[GGML_MAX_DIMS]; // stride in bytes: - // nb[0] = sizeof(type) - // nb[1] = nb[0] * ne[0] + padding - // nb[i] = nb[i-1] * ne[i-1] - - // compute data - enum ggml_op op; - - bool is_param; - - struct ggml_tensor * grad; - struct ggml_tensor * src0; - struct ggml_tensor * src1; - struct ggml_tensor * opt[GGML_MAX_OPT]; - - // thread scheduling - int n_tasks; - - // performance - int perf_runs; - int64_t perf_cycles; - int64_t perf_time_us; - - void * data; - char padding[8]; -}; - -// computation graph -struct ggml_cgraph { - int n_nodes; - int n_leafs; - int n_threads; - - size_t work_size; - struct ggml_tensor * work; - - struct ggml_tensor * nodes[GGML_MAX_NODES]; - struct ggml_tensor * grads[GGML_MAX_NODES]; - struct ggml_tensor * leafs[GGML_MAX_NODES]; - - // performance - int perf_runs; - int64_t perf_cycles; - int64_t perf_time_us; -}; - -// scratch buffer -struct ggml_scratch { - size_t offs; - size_t size; - void * data; -}; - -struct ggml_init_params { - // memory pool - size_t mem_size; // bytes - void * mem_buffer; // if NULL, memory will be allocated internally - bool no_alloc; // don't allocate memory for the tensor data -}; - -void ggml_time_init(void); // call this once at the beginning of the program -int64_t ggml_time_ms(void); -int64_t ggml_time_us(void); -int64_t ggml_cycles(void); -int64_t ggml_cycles_per_ms(void); - -void ggml_print_object (const struct ggml_object * obj); -void ggml_print_objects(const struct ggml_context * ctx); - -int64_t ggml_nelements(const struct ggml_tensor * tensor); -size_t ggml_nbytes (const struct ggml_tensor * tensor); - -int ggml_blck_size (enum ggml_type type); -size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block -float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float - -const char * ggml_type_name(enum ggml_type type); - -size_t ggml_element_size(const struct ggml_tensor * tensor); - -struct ggml_context * ggml_init(struct ggml_init_params params); -void ggml_free(struct ggml_context * ctx); - -size_t ggml_used_mem(const struct ggml_context * ctx); - -size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch); - -struct ggml_tensor * ggml_new_tensor( - struct ggml_context * ctx, - enum ggml_type type, - int n_dims, - const int64_t *ne); - -struct ggml_tensor * ggml_new_tensor_1d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0); - -struct ggml_tensor * ggml_new_tensor_2d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0, - int64_t ne1); - -struct ggml_tensor * ggml_new_tensor_3d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0, - int64_t ne1, - int64_t ne2); - -struct ggml_tensor * ggml_new_tensor_4d( - struct ggml_context * ctx, - enum ggml_type type, - int64_t ne0, - int64_t ne1, - int64_t ne2, - int64_t ne3); - -struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value); -struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value); - -struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src); -struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src); - -struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor); -struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value); -struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value); - -int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i); -void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value); - -float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i); -void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value); - - void * ggml_get_data (const struct ggml_tensor * tensor); -float * ggml_get_data_f32(const struct ggml_tensor * tensor); - -// -// operations on tensors with backpropagation -// - -struct ggml_tensor * ggml_dup( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_add( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - - -struct ggml_tensor * ggml_add_inplace( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_sub( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_mul( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_div( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_sqr( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_sqrt( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// return scalar -// TODO: compute sum along rows -struct ggml_tensor * ggml_sum( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// mean along rows -struct ggml_tensor * ggml_mean( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// if a is the same shape as b, and a is not parameter, return a -// otherwise, return a new tensor: repeat(a) to fit in b -struct ggml_tensor * ggml_repeat( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_abs( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_sgn( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_neg( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_step( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_relu( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// TODO: double-check this computation is correct -struct ggml_tensor * ggml_gelu( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_silu( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// normalize along rows -// TODO: eps is hardcoded to 1e-5 for now -struct ggml_tensor * ggml_norm( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_rms_norm( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// A: m rows, n columns -// B: p rows, n columns (i.e. we transpose it internally) -// result is m columns, p rows -struct ggml_tensor * ggml_mul_mat( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -// -// operations on tensors without backpropagation -// - -// in-place, returns view(a) -struct ggml_tensor * ggml_scale( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -// a -> b, return view(b) -struct ggml_tensor * ggml_cpy( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -// make contiguous -struct ggml_tensor * ggml_cont( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// return view(a), b specifies the new shape -// TODO: when we start computing gradient, make a copy instead of view -struct ggml_tensor * ggml_reshape( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -// return view(a) -// TODO: when we start computing gradient, make a copy instead of view -struct ggml_tensor * ggml_reshape_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1); - -// return view(a) -// TODO: when we start computing gradient, make a copy instead of view -struct ggml_tensor * ggml_reshape_3d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2); - -// offset in bytes -struct ggml_tensor * ggml_view_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - size_t offset); - -struct ggml_tensor * ggml_view_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - size_t nb1, // row stride in bytes - size_t offset); - -struct ggml_tensor * ggml_view_3d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - int64_t ne1, - int64_t ne2, - size_t nb1, // row stride in bytes - size_t nb2, // slice stride in bytes - size_t offset); - -struct ggml_tensor * ggml_permute( - struct ggml_context * ctx, - struct ggml_tensor * a, - int axis0, - int axis1, - int axis2, - int axis3); - -// alias for ggml_permute(ctx, a, 1, 0, 2, 3) -struct ggml_tensor * ggml_transpose( - struct ggml_context * ctx, - struct ggml_tensor * a); - -struct ggml_tensor * ggml_get_rows( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -// set elements above the diagonal to -INF -// in-place, returns view(a) -struct ggml_tensor * ggml_diag_mask_inf( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past); - -// in-place, returns view(a) -struct ggml_tensor * ggml_soft_max( - struct ggml_context * ctx, - struct ggml_tensor * a); - -// rotary position embedding -// in-place, returns view(a) -// if mode == 1, skip n_past elements -// TODO: avoid creating a new tensor every time -struct ggml_tensor * ggml_rope( - struct ggml_context * ctx, - struct ggml_tensor * a, - int n_past, - int n_dims, - int mode); - -// padding = 1 -// TODO: we don't support extra parameters for now -// that's why we are hard-coding the stride, padding, and dilation -// not great .. -struct ggml_tensor * ggml_conv_1d_1s( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_conv_1d_2s( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b); - -struct ggml_tensor * ggml_flash_attn( - struct ggml_context * ctx, - struct ggml_tensor * q, - struct ggml_tensor * k, - struct ggml_tensor * v, - bool masked); - -struct ggml_tensor * ggml_flash_ff( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b0, - struct ggml_tensor * b1, - struct ggml_tensor * c0, - struct ggml_tensor * c1); - -// Mapping operations -typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *); -typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *); - -struct ggml_tensor * ggml_map_unary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_unary_op_f32_t fun); - -struct ggml_tensor * ggml_map_binary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_binary_op_f32_t fun); - -// -// automatic differentiation -// - -void ggml_set_param( - struct ggml_context * ctx, - struct ggml_tensor * tensor); - -void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); - -struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor); -struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep); - -void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph); -void ggml_graph_reset (struct ggml_cgraph * cgraph); - -// print info and performance information for the graph -void ggml_graph_print(const struct ggml_cgraph * cgraph); - -// dump the graph into a file using the dot format -void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename); - -// -// optimization -// - -// optimization methods -enum ggml_opt_type { - GGML_OPT_ADAM, - GGML_OPT_LBFGS, -}; - -// linesearch methods -enum ggml_linesearch { - GGML_LINESEARCH_DEFAULT = 1, - - GGML_LINESEARCH_BACKTRACKING_ARMIJO = 0, - GGML_LINESEARCH_BACKTRACKING_WOLFE = 1, - GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2, -}; - -// optimization return values -enum ggml_opt_result { - GGML_OPT_OK = 0, - GGML_OPT_DID_NOT_CONVERGE, - GGML_OPT_NO_CONTEXT, - GGML_OPT_INVALID_WOLFE, - GGML_OPT_FAIL, - - GGML_LINESEARCH_FAIL = -128, - GGML_LINESEARCH_MINIMUM_STEP, - GGML_LINESEARCH_MAXIMUM_STEP, - GGML_LINESEARCH_MAXIMUM_ITERATIONS, - GGML_LINESEARCH_INVALID_PARAMETERS, -}; - -// optimization parameters -// -// see ggml.c (ggml_opt_default_params) for default values -// -struct ggml_opt_params { - enum ggml_opt_type type; - - int n_threads; - - // delta-based convergence test - // - // if past == 0 - disabled - // if past > 0: - // stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|) - // - int past; - float delta; - - // maximum number of iterations without improvement - // - // if 0 - disabled - // if > 0: - // assume convergence if no cost improvement in this number of iterations - // - int max_no_improvement; - - bool print_forward_graph; - bool print_backward_graph; - - // ADAM parameters - struct { - int n_iter; - - float alpha; // learning rate - float beta1; - float beta2; - float eps; // epsilon for numerical stability - float eps_f; // epsilon for convergence test - float eps_g; // epsilon for convergence test - } adam; - - // LBFGS parameters - struct { - int m; // number of corrections to approximate the inv. Hessian - int n_iter; - int max_linesearch; - - float eps; // convergence tolerance - float ftol; // line search tolerance - float wolfe; - float min_step; - float max_step; - - enum ggml_linesearch linesearch; - } lbfgs; -}; - -struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type); - -// optimize the function defined by the tensor f -enum ggml_opt_result ggml_opt( - struct ggml_context * ctx, - struct ggml_opt_params params, - struct ggml_tensor * f); - -// -// quantization -// - -size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); -size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); -size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist); - -// -// system info -// - -int ggml_cpu_has_avx(void); -int ggml_cpu_has_avx2(void); -int ggml_cpu_has_avx512(void); -int ggml_cpu_has_avx512_vbmi(void); -int ggml_cpu_has_avx512_vnni(void); -int ggml_cpu_has_fma(void); -int ggml_cpu_has_neon(void); -int ggml_cpu_has_arm_fma(void); -int ggml_cpu_has_f16c(void); -int ggml_cpu_has_fp16_va(void); -int ggml_cpu_has_wasm_simd(void); -int ggml_cpu_has_blas(void); -int ggml_cpu_has_sse3(void); -int ggml_cpu_has_vsx(void); - - -// -// Internal types and functions exposed for tests and benchmarks -// +#define GGML_ASSERT(x) \ + do { \ + if (!(x)) { \ + fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \ + abort(); \ + } \ + } while (0) #ifdef __cplusplus -// restrict not standard in C++ +extern "C" { +#endif + +#ifdef __ARM_NEON + // we use the built-in 16-bit float type + typedef __fp16 ggml_fp16_t; +#else + typedef uint16_t ggml_fp16_t; +#endif + + // convert FP16 <-> FP32 + GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x); + GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x); + + GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n); + GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n); + + struct ggml_object; + struct ggml_context; + + enum ggml_type { + GGML_TYPE_F32 = 0, + GGML_TYPE_F16 = 1, + GGML_TYPE_Q4_0 = 2, + GGML_TYPE_Q4_1 = 3, + // GGML_TYPE_Q4_2 = 4, support has been removed + // GGML_TYPE_Q4_3 (5) support has been removed + GGML_TYPE_Q5_0 = 6, + GGML_TYPE_Q5_1 = 7, + GGML_TYPE_Q8_0 = 8, + GGML_TYPE_Q8_1 = 9, + GGML_TYPE_I8, + GGML_TYPE_I16, + GGML_TYPE_I32, + GGML_TYPE_COUNT, + }; + + enum ggml_backend { + GGML_BACKEND_CPU = 0, + GGML_BACKEND_CUDA = 1, + }; + + // model file types + enum ggml_ftype { + GGML_FTYPE_UNKNOWN = -1, + GGML_FTYPE_ALL_F32 = 0, + GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 + GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors + }; + + // available tensor operations: + enum ggml_op { + GGML_OP_NONE = 0, + + GGML_OP_DUP, + GGML_OP_ADD, + GGML_OP_ADD1, + GGML_OP_ACC, + GGML_OP_SUB, + GGML_OP_MUL, + GGML_OP_DIV, + GGML_OP_SQR, + GGML_OP_SQRT, + GGML_OP_LOG, + GGML_OP_SUM, + GGML_OP_SUM_ROWS, + GGML_OP_MEAN, + GGML_OP_REPEAT, + GGML_OP_ABS, + GGML_OP_SGN, + GGML_OP_NEG, + GGML_OP_STEP, + GGML_OP_RELU, + GGML_OP_GELU, + GGML_OP_SILU, + GGML_OP_SILU_BACK, + GGML_OP_NORM, // normalize + GGML_OP_RMS_NORM, + GGML_OP_RMS_NORM_BACK, + + GGML_OP_MUL_MAT, + + GGML_OP_SCALE, + GGML_OP_SET, + GGML_OP_CPY, + GGML_OP_CONT, + GGML_OP_RESHAPE, + GGML_OP_VIEW, + GGML_OP_PERMUTE, + GGML_OP_TRANSPOSE, + GGML_OP_GET_ROWS, + GGML_OP_GET_ROWS_BACK, + GGML_OP_DIAG, + GGML_OP_DIAG_MASK_INF, + GGML_OP_DIAG_MASK_ZERO, + GGML_OP_SOFT_MAX, + GGML_OP_ROPE, + GGML_OP_ROPE_BACK, + GGML_OP_ALIBI, + GGML_OP_CONV_1D_1S, + GGML_OP_CONV_1D_2S, + + GGML_OP_FLASH_ATTN, + GGML_OP_FLASH_FF, + + GGML_OP_MAP_UNARY, + GGML_OP_MAP_BINARY, + + GGML_OP_COUNT, + }; + + + // ggml object + struct ggml_object { + size_t offs; + size_t size; + + struct ggml_object * next; + + char padding[8]; + }; + + static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); + + // n-dimensional tensor + struct ggml_tensor { + enum ggml_type type; + enum ggml_backend backend; + + int n_dims; + int64_t ne[GGML_MAX_DIMS]; // number of elements + size_t nb[GGML_MAX_DIMS]; // stride in bytes: + // nb[0] = sizeof(type) + // nb[1] = nb[0] * ne[0] + padding + // nb[i] = nb[i-1] * ne[i-1] + + // compute data + enum ggml_op op; + + bool is_param; + + struct ggml_tensor * grad; + struct ggml_tensor * src0; + struct ggml_tensor * src1; + struct ggml_tensor * opt[GGML_MAX_OPT]; + + // thread scheduling + int n_tasks; + + // performance + int perf_runs; + int64_t perf_cycles; + int64_t perf_time_us; + + void * data; + + char name[32]; + + char padding[9]; // TODO: remove and add padding to name? + }; + + // computation graph + struct ggml_cgraph { + int n_nodes; + int n_leafs; + int n_threads; + + size_t work_size; + struct ggml_tensor * work; + + struct ggml_tensor * nodes[GGML_MAX_NODES]; + struct ggml_tensor * grads[GGML_MAX_NODES]; + struct ggml_tensor * leafs[GGML_MAX_NODES]; + + // performance + int perf_runs; + int64_t perf_cycles; + int64_t perf_time_us; + }; + + // scratch buffer + struct ggml_scratch { + size_t offs; + size_t size; + void * data; + }; + + struct ggml_init_params { + // memory pool + size_t mem_size; // bytes + void * mem_buffer; // if NULL, memory will be allocated internally + bool no_alloc; // don't allocate memory for the tensor data + }; + + // misc + + GGML_API void ggml_time_init(void); // call this once at the beginning of the program + GGML_API int64_t ggml_time_ms(void); + GGML_API int64_t ggml_time_us(void); + GGML_API int64_t ggml_cycles(void); + GGML_API int64_t ggml_cycles_per_ms(void); + + GGML_API void ggml_print_object (const struct ggml_object * obj); + GGML_API void ggml_print_objects(const struct ggml_context * ctx); + + GGML_API int64_t ggml_nelements(const struct ggml_tensor * tensor); + GGML_API size_t ggml_nbytes (const struct ggml_tensor * tensor); + + GGML_API int ggml_blck_size (enum ggml_type type); + GGML_API size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block + GGML_API float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float + + GGML_API const char * ggml_type_name(enum ggml_type type); + + GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor); + + GGML_API bool ggml_is_quantized(enum ggml_type type); + + // TODO: temporary until model loading of ggml examples is refactored + GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype); + + // main + + GGML_API struct ggml_context * ggml_init(struct ggml_init_params params); + GGML_API void ggml_free(struct ggml_context * ctx); + + GGML_API size_t ggml_used_mem(const struct ggml_context * ctx); + + GGML_API size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch); + + GGML_API struct ggml_tensor * ggml_new_tensor( + struct ggml_context * ctx, + enum ggml_type type, + int n_dims, + const int64_t *ne); + + GGML_API struct ggml_tensor * ggml_new_tensor_1d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0); + + GGML_API struct ggml_tensor * ggml_new_tensor_2d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1); + + GGML_API struct ggml_tensor * ggml_new_tensor_3d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1, + int64_t ne2); + + GGML_API struct ggml_tensor * ggml_new_tensor_4d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3); + + GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value); + GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value); + + GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src); + GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src); + + GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor); + GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value); + GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value); + + GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i); + GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value); + + GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i); + GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value); + + GGML_API void * ggml_get_data (const struct ggml_tensor * tensor); + GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor); + + GGML_API const char * ggml_get_name(const struct ggml_tensor * tensor); + GGML_API void ggml_set_name(struct ggml_tensor * tensor, const char * name); + + // + // operations on tensors with backpropagation + // + + GGML_API struct ggml_tensor * ggml_dup( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_add( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_add_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_add1( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_acc( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + GGML_API struct ggml_tensor * ggml_acc_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + GGML_API struct ggml_tensor * ggml_sub( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_mul( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_div( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_sqr( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_sqrt( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_log( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_log_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // return scalar + GGML_API struct ggml_tensor * ggml_sum( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d] + GGML_API struct ggml_tensor * ggml_sum_rows( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // mean along rows + GGML_API struct ggml_tensor * ggml_mean( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // if a is the same shape as b, and a is not parameter, return a + // otherwise, return a new tensor: repeat(a) to fit in b + GGML_API struct ggml_tensor * ggml_repeat( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_abs( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_sgn( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_neg( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_step( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_relu( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // TODO: double-check this computation is correct + GGML_API struct ggml_tensor * ggml_gelu( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_silu( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // a - x + // b - dy + GGML_API struct ggml_tensor * ggml_silu_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // normalize along rows + // TODO: eps is hardcoded to 1e-5 for now + GGML_API struct ggml_tensor * ggml_norm( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_rms_norm( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // a - x + // b - dy + GGML_API struct ggml_tensor * ggml_rms_norm_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // A: m rows, n columns + // B: p rows, n columns (i.e. we transpose it internally) + // result is m columns, p rows + GGML_API struct ggml_tensor * ggml_mul_mat( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // + // operations on tensors without backpropagation + // + + GGML_API struct ggml_tensor * ggml_scale( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_scale_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // b -> view(a,offset,nb1,nb2,3), return modified a + GGML_API struct ggml_tensor * ggml_set( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + // b -> view(a,offset,nb1,nb2,3), return view(a) + GGML_API struct ggml_tensor * ggml_set_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + GGML_API struct ggml_tensor * ggml_set_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset); + + GGML_API struct ggml_tensor * ggml_set_1d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset); + + // b -> view(a,offset,nb1,nb2,3), return modified a + GGML_API struct ggml_tensor * ggml_set_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset); + + // b -> view(a,offset,nb1,nb2,3), return view(a) + GGML_API struct ggml_tensor * ggml_set_2d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset); + + + // a -> b, return view(b) + GGML_API struct ggml_tensor * ggml_cpy( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // make contiguous + GGML_API struct ggml_tensor * ggml_cont( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // return view(a), b specifies the new shape + // TODO: when we start computing gradient, make a copy instead of view + GGML_API struct ggml_tensor * ggml_reshape( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // return view(a) + // TODO: when we start computing gradient, make a copy instead of view + GGML_API struct ggml_tensor * ggml_reshape_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0); + + GGML_API struct ggml_tensor * ggml_reshape_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1); + + // return view(a) + // TODO: when we start computing gradient, make a copy instead of view + GGML_API struct ggml_tensor * ggml_reshape_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2); + + GGML_API struct ggml_tensor * ggml_reshape_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3); + + // offset in bytes + GGML_API struct ggml_tensor * ggml_view_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + size_t offset); + + GGML_API struct ggml_tensor * ggml_view_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + size_t nb1, // row stride in bytes + size_t offset); + + GGML_API struct ggml_tensor * ggml_view_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + size_t nb1, // row stride in bytes + size_t nb2, // slice stride in bytes + size_t offset); + + GGML_API struct ggml_tensor * ggml_view_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + size_t nb1, // row stride in bytes + size_t nb2, // slice stride in bytes + size_t nb3, + size_t offset); + + GGML_API struct ggml_tensor * ggml_permute( + struct ggml_context * ctx, + struct ggml_tensor * a, + int axis0, + int axis1, + int axis2, + int axis3); + + // alias for ggml_permute(ctx, a, 1, 0, 2, 3) + GGML_API struct ggml_tensor * ggml_transpose( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_get_rows( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_get_rows_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c); + + GGML_API struct ggml_tensor * ggml_diag( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // set elements above the diagonal to -INF + GGML_API struct ggml_tensor * ggml_diag_mask_inf( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + + // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + + // set elements above the diagonal to 0 + GGML_API struct ggml_tensor * ggml_diag_mask_zero( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + + // in-place, returns view(a) + GGML_API struct ggml_tensor * gml_diag_mask_zero_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + + GGML_API struct ggml_tensor * ggml_soft_max( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_soft_max_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // rotary position embedding + // if mode & 1 == 1, skip n_past elements + // if mode & 2 == 1, GPT-NeoX style + // TODO: avoid creating a new tensor every time + GGML_API struct ggml_tensor * ggml_rope( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode); + + // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_rope_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode); + + // rotary position embedding backward, i.e compute dx from dy + // a - dy + GGML_API struct ggml_tensor * ggml_rope_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode); + + // alibi position embedding + // in-place, returns view(a) + struct ggml_tensor * ggml_alibi( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_head); + + // padding = 1 + // TODO: we don't support extra parameters for now + // that's why we are hard-coding the stride, padding, and dilation + // not great .. + GGML_API struct ggml_tensor * ggml_conv_1d_1s( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_conv_1d_2s( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_flash_attn( + struct ggml_context * ctx, + struct ggml_tensor * q, + struct ggml_tensor * k, + struct ggml_tensor * v, + bool masked); + + GGML_API struct ggml_tensor * ggml_flash_ff( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b0, + struct ggml_tensor * b1, + struct ggml_tensor * c0, + struct ggml_tensor * c1); + + // Mapping operations + typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *); + typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *); + + GGML_API struct ggml_tensor * ggml_map_unary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + ggml_unary_op_f32_t fun); + + GGML_API struct ggml_tensor * ggml_map_binary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + ggml_binary_op_f32_t fun); + + // + // automatic differentiation + // + + GGML_API void ggml_set_param( + struct ggml_context * ctx, + struct ggml_tensor * tensor); + + GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); + + GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor); + GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep); + + GGML_API void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph); + GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); + + // print info and performance information for the graph + GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph); + + // dump the graph into a file using the dot format + GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename); + + // + // optimization + // + + // optimization methods + enum ggml_opt_type { + GGML_OPT_ADAM, + GGML_OPT_LBFGS, + }; + + // linesearch methods + enum ggml_linesearch { + GGML_LINESEARCH_DEFAULT = 1, + + GGML_LINESEARCH_BACKTRACKING_ARMIJO = 0, + GGML_LINESEARCH_BACKTRACKING_WOLFE = 1, + GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2, + }; + + // optimization return values + enum ggml_opt_result { + GGML_OPT_OK = 0, + GGML_OPT_DID_NOT_CONVERGE, + GGML_OPT_NO_CONTEXT, + GGML_OPT_INVALID_WOLFE, + GGML_OPT_FAIL, + + GGML_LINESEARCH_FAIL = -128, + GGML_LINESEARCH_MINIMUM_STEP, + GGML_LINESEARCH_MAXIMUM_STEP, + GGML_LINESEARCH_MAXIMUM_ITERATIONS, + GGML_LINESEARCH_INVALID_PARAMETERS, + }; + + // optimization parameters + // + // see ggml.c (ggml_opt_default_params) for default values + // + struct ggml_opt_params { + enum ggml_opt_type type; + + int n_threads; + + // delta-based convergence test + // + // if past == 0 - disabled + // if past > 0: + // stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|) + // + int past; + float delta; + + // maximum number of iterations without improvement + // + // if 0 - disabled + // if > 0: + // assume convergence if no cost improvement in this number of iterations + // + int max_no_improvement; + + bool print_forward_graph; + bool print_backward_graph; + + // ADAM parameters + struct { + int n_iter; + + float alpha; // learning rate + float beta1; + float beta2; + float eps; // epsilon for numerical stability + float eps_f; // epsilon for convergence test + float eps_g; // epsilon for convergence test + } adam; + + // LBFGS parameters + struct { + int m; // number of corrections to approximate the inv. Hessian + int n_iter; + int max_linesearch; + + float eps; // convergence tolerance + float ftol; // line search tolerance + float wolfe; + float min_step; + float max_step; + + enum ggml_linesearch linesearch; + } lbfgs; + }; + + GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type); + + // optimize the function defined by the tensor f + GGML_API enum ggml_opt_result ggml_opt( + struct ggml_context * ctx, + struct ggml_opt_params params, + struct ggml_tensor * f); + + // + // quantization + // + + GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist); + + GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); + + // + // system info + // + + GGML_API int ggml_cpu_has_avx (void); + GGML_API int ggml_cpu_has_avx2 (void); + GGML_API int ggml_cpu_has_avx512 (void); + GGML_API int ggml_cpu_has_avx512_vbmi(void); + GGML_API int ggml_cpu_has_avx512_vnni(void); + 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_f16c (void); + GGML_API int ggml_cpu_has_fp16_va (void); + GGML_API int ggml_cpu_has_wasm_simd (void); + GGML_API int ggml_cpu_has_blas (void); + GGML_API int ggml_cpu_has_cublas (void); + GGML_API int ggml_cpu_has_clblast (void); + GGML_API int ggml_cpu_has_gpublas (void); + GGML_API int ggml_cpu_has_sse3 (void); + GGML_API int ggml_cpu_has_vsx (void); + + // + // Internal types and functions exposed for tests and benchmarks + // + +#ifdef __cplusplus + // restrict not standard in C++ #define GGML_RESTRICT #else #define GGML_RESTRICT restrict #endif -typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); -typedef void (*quantize_row_q_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); -typedef void (*vec_dot_q_t)(const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y); + typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); + typedef void (*quantize_row_q_t) (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); + typedef void (*vec_dot_q_t) (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y); -typedef struct { - dequantize_row_q_t dequantize_row_q; - quantize_row_q_t quantize_row_q; - quantize_row_q_t quantize_row_q_reference; - quantize_row_q_t quantize_row_q_dot; - vec_dot_q_t vec_dot_q; -} quantize_fns_t; + typedef struct { + dequantize_row_q_t dequantize_row_q; + quantize_row_q_t quantize_row_q; + quantize_row_q_t quantize_row_q_reference; + quantize_row_q_t quantize_row_q_dot; + vec_dot_q_t vec_dot_q; + enum ggml_type vec_dot_type; + } quantize_fns_t; -quantize_fns_t ggml_internal_get_quantize_fn(size_t i); + quantize_fns_t ggml_internal_get_quantize_fn(size_t i); #ifdef __cplusplus } diff --git a/llama_util.h b/llama-util.h old mode 100755 new mode 100644 similarity index 78% rename from llama_util.h rename to llama-util.h index eba14656a..88ec28dca --- a/llama_util.h +++ b/llama-util.h @@ -14,6 +14,7 @@ #include #include +#include #ifdef __has_include #if __has_include() @@ -21,6 +22,9 @@ #if defined(_POSIX_MAPPED_FILES) #include #endif + #if defined(_POSIX_MEMLOCK_RANGE) + #include + #endif #endif #endif @@ -71,7 +75,7 @@ struct llama_file { llama_file(const char * fname, const char * mode) { fp = std::fopen(fname, mode); if (fp == NULL) { - throw format("failed to open %s: %s", fname, std::strerror(errno)); + throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno))); } seek(0, SEEK_END); size = tell(); @@ -104,10 +108,10 @@ struct llama_file { errno = 0; std::size_t ret = std::fread(ptr, size, 1, fp); if (ferror(fp)) { - throw format("read error: %s", strerror(errno)); + throw std::runtime_error(format("read error: %s", strerror(errno))); } if (ret != 1) { - throw std::string("unexpectedly reached end of file"); + throw std::runtime_error(std::string("unexpectedly reached end of file")); } } @@ -130,7 +134,7 @@ struct llama_file { errno = 0; size_t ret = std::fwrite(ptr, size, 1, fp); if (ret != 1) { - throw format("write error: %s", strerror(errno)); + throw std::runtime_error(format("write error: %s", strerror(errno))); } } @@ -177,7 +181,7 @@ struct llama_mmap { #endif addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); if (addr == MAP_FAILED) { - throw format("mmap failed: %s", strerror(errno)); + throw std::runtime_error(format("mmap failed: %s", strerror(errno))); } if (prefetch) { @@ -204,7 +208,7 @@ struct llama_mmap { DWORD error = GetLastError(); if (hMapping == NULL) { - throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()); + throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str())); } addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); @@ -212,7 +216,7 @@ struct llama_mmap { CloseHandle(hMapping); if (addr == NULL) { - throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()); + throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str())); } #if _WIN32_WINNT >= _WIN32_WINNT_WIN8 @@ -240,8 +244,9 @@ struct llama_mmap { #else static constexpr bool SUPPORTED = false; - llama_mmap(struct llama_file *) { - throw std::string("mmap not supported"); + llama_mmap(struct llama_file *, bool prefetch = true) { + (void)prefetch; + throw std::runtime_error(std::string("mmap not supported")); } #endif }; @@ -303,8 +308,18 @@ struct llama_mlock { if (!mlock(addr, size)) { return true; } else { - fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION, - size, this->size, std::strerror(errno)); + char* errmsg = std::strerror(errno); + bool suggest = (errno == ENOMEM); + + // Check if the resource limit is fine after all + struct rlimit lock_limit; + if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) + suggest = false; + if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) + suggest = false; + + fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s", + size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : ""); return false; } } @@ -369,8 +384,13 @@ struct llama_mlock { #else static constexpr bool SUPPORTED = false; - void raw_lock(const void * addr, size_t size) { + size_t lock_granularity() { + return (size_t) 65536; + } + + bool raw_lock(const void * addr, size_t size) { fprintf(stderr, "warning: mlock not supported on this system\n"); + return false; } void raw_unlock(const void * addr, size_t size) {} @@ -382,6 +402,8 @@ struct llama_buffer { uint8_t * addr = NULL; size_t size = 0; + llama_buffer() = default; + void resize(size_t size) { delete[] addr; addr = new uint8_t[size]; @@ -391,5 +413,62 @@ struct llama_buffer { ~llama_buffer() { delete[] addr; } + + // disable copy and move + llama_buffer(const llama_buffer&) = delete; + llama_buffer(llama_buffer&&) = delete; + llama_buffer& operator=(const llama_buffer&) = delete; + llama_buffer& operator=(llama_buffer&&) = delete; }; + +#ifdef GGML_USE_CUBLAS +#include "ggml-cuda.h" +struct llama_ctx_buffer { + uint8_t * addr = NULL; + bool is_cuda; + size_t size = 0; + + llama_ctx_buffer() = default; + + void resize(size_t size) { + free(); + + addr = (uint8_t *) ggml_cuda_host_malloc(size); + if (addr) { + is_cuda = true; + } + else { + // fall back to pageable memory + addr = new uint8_t[size]; + is_cuda = false; + } + this->size = size; + } + + void free() { + if (addr) { + if (is_cuda) { + ggml_cuda_host_free(addr); + } + else { + delete[] addr; + } + } + addr = NULL; + } + + ~llama_ctx_buffer() { + free(); + } + + // disable copy and move + llama_ctx_buffer(const llama_ctx_buffer&) = delete; + llama_ctx_buffer(llama_ctx_buffer&&) = delete; + llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete; + llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete; +}; +#else +typedef llama_buffer llama_ctx_buffer; +#endif + #endif diff --git a/llama.cpp b/llama.cpp index f14324fcd..98f49abd7 100644 --- a/llama.cpp +++ b/llama.cpp @@ -5,10 +5,13 @@ #include #endif -#include "llama_util.h" +#include "llama-util.h" #include "llama.h" #include "ggml.h" +#ifdef GGML_USE_CUBLAS +#include "ggml-cuda.h" +#endif #include #include @@ -24,11 +27,15 @@ #include #include #include +#include +#include +#include +#include +#include #define LLAMA_USE_SCRATCH #define LLAMA_MAX_SCRATCH_BUFFERS 16 - // available llama models enum e_model { MODEL_UNKNOWN, @@ -46,50 +53,50 @@ static const size_t MB = 1024*1024; static const std::map & MEM_REQ_SCRATCH0() { - static std::map _MEM_REQ_SCRATCH0 = { + static std::map k_sizes = { { MODEL_7B, 512ull * MB }, { MODEL_13B, 512ull * MB }, { MODEL_30B, 512ull * MB }, - { MODEL_65B, 512ull * MB }, + { MODEL_65B, 1024ull * MB }, }; - return _MEM_REQ_SCRATCH0; + return k_sizes; } static const std::map & MEM_REQ_SCRATCH1() { - static std::map _MEM_REQ_SCRATCH1 = { + static std::map k_sizes = { { MODEL_7B, 512ull * MB }, { MODEL_13B, 512ull * MB }, { MODEL_30B, 512ull * MB }, - { MODEL_65B, 512ull * MB }, + { MODEL_65B, 1024ull * MB }, }; - return _MEM_REQ_SCRATCH1; -}; + return k_sizes; +} // 2*n_embd*n_ctx*n_layer*sizeof(float16) static const std::map & MEM_REQ_KV_SELF() { - static std::map _MEM_REQ_KV_SELF = { + static std::map k_sizes = { { MODEL_7B, 1026ull * MB }, { MODEL_13B, 1608ull * MB }, { MODEL_30B, 3124ull * MB }, { MODEL_65B, 5120ull * MB }, }; - return _MEM_REQ_KV_SELF; -}; + return k_sizes; +} // this is mostly needed for temporary mul_mat buffers to dequantize the data // not actually needed if BLAS is disabled static const std::map & MEM_REQ_EVAL() { - static std::map _MEM_REQ_EVAL = { + static std::map k_sizes = { { MODEL_7B, 768ull * MB }, { MODEL_13B, 1024ull * MB }, { MODEL_30B, 1280ull * MB }, { MODEL_65B, 1536ull * MB }, }; - return _MEM_REQ_EVAL; -}; + return k_sizes; +} // default hparams (LLaMA 7B) struct llama_hparams { @@ -132,7 +139,7 @@ struct llama_kv_cache { struct ggml_context * ctx = NULL; - llama_buffer buf; + llama_ctx_buffer buf; int n; // number of tokens currently in the cache @@ -163,7 +170,7 @@ struct llama_model { struct llama_kv_cache kv_self; // the model memory buffer - llama_buffer buf; + llama_ctx_buffer buf; // model memory mapped file std::unique_ptr mapping; @@ -224,8 +231,8 @@ struct llama_context { // memory buffers used to evaluate the model // TODO: move in llama_state - llama_buffer buf_compute; - llama_buffer buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS]; + llama_ctx_buffer buf_compute; + llama_ctx_buffer buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS]; int buf_last = 0; size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 }; @@ -398,6 +405,7 @@ enum llama_file_version { LLAMA_FILE_VERSION_GGML, LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab LLAMA_FILE_VERSION_GGJT_V1, // added padding + LLAMA_FILE_VERSION_GGJT_V2, // changed quantization format }; struct llama_file_loader { @@ -428,6 +436,8 @@ struct llama_file_loader { file_version = LLAMA_FILE_VERSION_GGMF_V1; } else if (magic == 'ggjt' && version == 1) { file_version = LLAMA_FILE_VERSION_GGJT_V1; + } else if (magic == 'ggjt' && version == 2) { + file_version = LLAMA_FILE_VERSION_GGJT_V2; } else { throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?", magic, version); @@ -478,7 +488,9 @@ struct llama_file_loader { case GGML_TYPE_F16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: break; default: { throw format("unrecognized tensor type %u\n", shard.type); @@ -520,8 +532,8 @@ struct llama_file_saver { write_vocab(); } void write_magic() { - file.write_u32('ggjt'); // magic - file.write_u32(1); // version + file.write_u32(LLAMA_FILE_MAGIC); // magic + file.write_u32(LLAMA_FILE_VERSION); // version } void write_hparams(enum llama_ftype new_ftype) { const llama_hparams & hparams = any_file_loader->hparams; @@ -551,7 +563,9 @@ struct llama_file_saver { case GGML_TYPE_F16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: break; default: LLAMA_ASSERT(false); } @@ -575,12 +589,12 @@ struct llama_model_loader { std::unique_ptr mapping; llama_model_loader(const std::string & fname_base, bool use_mmap, bool vocab_only) { - auto first_file = new llama_file_loader(fname_base.c_str(), 0, tensors_map); + auto * first_file = new llama_file_loader(fname_base.c_str(), 0, tensors_map); file_loaders.emplace_back(first_file); uint32_t n_parts = vocab_only ? 1 : guess_n_parts(); for (uint32_t i = 1; i < n_parts; i++) { std::string fname = fname_base + "." + std::to_string(i); - auto ith_file = new llama_file_loader(fname.c_str(), i, tensors_map); + auto * ith_file = new llama_file_loader(fname.c_str(), i, tensors_map); file_loaders.emplace_back(ith_file); if (ith_file->hparams != first_file->hparams) { throw format("llama.cpp: hparams inconsistent between files"); @@ -627,7 +641,7 @@ struct llama_model_loader { } } - struct ggml_tensor * get_tensor(const std::string & name, std::vector ne) { + struct ggml_tensor * get_tensor(const std::string & name, const std::vector & ne) { auto it = tensors_map.name_to_idx.find(name); if (it == tensors_map.name_to_idx.end()) { throw format("llama.cpp: tensor '%s' is missing from model", name.c_str()); @@ -649,13 +663,14 @@ struct llama_model_loader { LLAMA_ASSERT(lt.ne.size() == 1); tensor = ggml_new_tensor_1d(ggml_ctx, lt.type, lt.ne.at(0)); } + ggml_set_name(tensor, lt.name.c_str()); LLAMA_ASSERT(lt.ggml_tensor == NULL); // if this fails, we called get_tensor twice on the same tensor lt.ggml_tensor = tensor; num_ggml_tensors_created++; return tensor; } - void done_getting_tensors() { + void done_getting_tensors() const { if (num_ggml_tensors_created != tensors_map.tensors.size()) { throw std::string("llama.cpp: file contained more tensors than expected"); } @@ -717,8 +732,7 @@ struct llama_model_loader { LLAMA_ASSERT(offset == lt.size); } else if (lt.split_type == SPLIT_BY_COLUMNS) { // Let's load the data into temporary buffers to ensure the OS performs large loads. - std::vector tmp_bufs; - tmp_bufs.resize(lt.shards.size()); + std::vector tmp_bufs(lt.shards.size()); for (size_t i = 0; i < lt.shards.size(); i++) { llama_load_tensor_shard & shard = lt.shards.at(i); llama_file & file = file_loaders.at(shard.file_idx)->file; @@ -770,7 +784,7 @@ static bool kv_cache_init( const int n_embd = hparams.n_embd; const int n_layer = hparams.n_layer; - const int64_t n_mem = (int64_t)n_layer*n_ctx; + const int64_t n_mem = n_layer*n_ctx; const int64_t n_elements = n_embd*n_mem; cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB); @@ -789,6 +803,8 @@ static bool kv_cache_init( cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements); cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements); + ggml_set_name(cache.k, "cache_k"); + ggml_set_name(cache.v, "cache_v"); return true; } @@ -797,7 +813,8 @@ struct llama_context_params llama_context_default_params() { struct llama_context_params result = { /*.n_ctx =*/ 512, /*.n_parts =*/ -1, - /*.seed =*/ 0, + /*.gpu_layers =*/ 0, + /*.seed =*/ -1, /*.f16_kv =*/ false, /*.logits_all =*/ false, /*.vocab_only =*/ false, @@ -827,9 +844,11 @@ static const char *llama_file_version_name(llama_file_version version) { switch (version) { case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)"; case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)"; - case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (latest)"; - default: LLAMA_ASSERT(false); + case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (pre #1405)"; + case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (latest)"; } + + return "unknown"; } static const char *llama_ftype_name(enum llama_ftype ftype) { @@ -840,7 +859,9 @@ static const char *llama_ftype_name(enum llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1"; case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16: return "mostly Q4_1, some F16"; - case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2"; + case LLAMA_FTYPE_MOSTLY_Q5_0: return "mostly Q5_0"; + case LLAMA_FTYPE_MOSTLY_Q5_1: return "mostly Q5_1"; + case LLAMA_FTYPE_MOSTLY_Q8_0: return "mostly Q8_0"; default: return "unknown, may not work"; } } @@ -859,6 +880,7 @@ static void llama_model_load_internal( const std::string & fname, llama_context & lctx, int n_ctx, + int n_gpu_layers, ggml_type memory_type, bool use_mmap, bool use_mlock, @@ -903,13 +925,22 @@ static void llama_model_load_internal( fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type)); } + if (file_version != LLAMA_FILE_VERSION_GGJT_V2) { + if (hparams.ftype != LLAMA_FTYPE_ALL_F32 && + hparams.ftype != LLAMA_FTYPE_MOSTLY_F16 && + hparams.ftype != LLAMA_FTYPE_MOSTLY_Q8_0) { + throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1305)"); + } + } + if (vocab_only) { return; } auto & ctx = model.ctx; - size_t ctx_size, mmapped_size; + size_t ctx_size; + size_t mmapped_size; ml->calc_sizes(&ctx_size, &mmapped_size); fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0); @@ -955,8 +986,6 @@ static void llama_model_load_internal( // prepare memory for the weights { - const auto & hparams = model.hparams; - const uint32_t n_embd = hparams.n_embd; const uint32_t n_layer = hparams.n_layer; const uint32_t n_vocab = hparams.n_vocab; @@ -998,6 +1027,35 @@ static void llama_model_load_internal( ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL); model.mapping = std::move(ml->mapping); +#ifdef GGML_USE_CUBLAS + { + const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); + + fprintf(stderr, "%s: [cublas] offloading %d layers to GPU\n", __func__, n_gpu); + + size_t vram_total = 0; + + for (int i = 0; i < n_gpu; ++i) { + const auto & layer = model.layers[i]; + + ggml_cuda_transform_tensor(layer.wq); vram_total += ggml_nbytes(layer.wq); + ggml_cuda_transform_tensor(layer.wk); vram_total += ggml_nbytes(layer.wk); + ggml_cuda_transform_tensor(layer.wv); vram_total += ggml_nbytes(layer.wv); + ggml_cuda_transform_tensor(layer.wo); vram_total += ggml_nbytes(layer.wo); + ggml_cuda_transform_tensor(layer.w1); vram_total += ggml_nbytes(layer.w1); + ggml_cuda_transform_tensor(layer.w2); vram_total += ggml_nbytes(layer.w2); + ggml_cuda_transform_tensor(layer.w3); vram_total += ggml_nbytes(layer.w3); + } + if (n_gpu_layers > (int) hparams.n_layer) { + fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__); + ggml_cuda_transform_tensor(model.output); vram_total += ggml_nbytes(model.output); + } + + fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024); + } +#else + (void) n_gpu_layers; +#endif // loading time will be recalculate after the first eval, so // we take page faults deferred by mmap() into consideration @@ -1008,6 +1066,7 @@ static bool llama_model_load( const std::string & fname, llama_context & lctx, int n_ctx, + int n_gpu_layers, ggml_type memory_type, bool use_mmap, bool use_mlock, @@ -1015,7 +1074,7 @@ static bool llama_model_load( llama_progress_callback progress_callback, void *progress_callback_user_data) { try { - llama_model_load_internal(fname, lctx, n_ctx, memory_type, use_mmap, use_mlock, + llama_model_load_internal(fname, lctx, n_ctx, n_gpu_layers, memory_type, use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data); return true; } catch (const std::string & err) { @@ -1037,6 +1096,13 @@ static bool llama_eval_internal( const int n_tokens, const int n_past, const int n_threads) { + + // enforce that the first token is BOS + if (n_past == 0 && tokens[0] != llama_token_bos()) { + fprintf(stderr, "%s: first token must be BOS\n", __func__); + return false; + } + const int64_t t_start_us = ggml_time_us(); const int N = n_tokens; @@ -1044,7 +1110,7 @@ static bool llama_eval_internal( const auto & model = lctx.model; const auto & hparams = model.hparams; - auto & kv_self = model.kv_self; + const auto & kv_self = model.kv_self; LLAMA_ASSERT(!!kv_self.ctx); @@ -1069,9 +1135,10 @@ static bool llama_eval_internal( // for big prompts, if BLAS is enabled, it is better to use only one thread // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance ggml_cgraph gf = {}; - gf.n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads; + gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads; struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + ggml_set_name(embd, "embd"); memcpy(embd->data, tokens, N*ggml_element_size(embd)); struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd); @@ -1096,8 +1163,10 @@ static bool llama_eval_internal( // self-attention { // compute Q and K and RoPE them - struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); - struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + ggml_set_name(Qcur, "Qcur"); + ggml_set_name(Kcur, "Kcur"); // store key and value to memory { @@ -1118,6 +1187,7 @@ static bool llama_eval_internal( ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + ggml_set_name(Q, "Q"); struct ggml_tensor * K = ggml_permute(ctx0, @@ -1125,21 +1195,28 @@ static bool llama_eval_internal( ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd), n_embd/n_head, n_head, n_past + N), 0, 2, 1, 3); + 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/n_head) - struct ggml_tensor * KQ_scaled = - ggml_scale(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); + struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)); + ggml_set_name(KQ_scale, "1/sqrt(n_embd/n_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(ctx0, KQ_scaled, n_past); + 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(ctx0, 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 = @@ -1148,9 +1225,11 @@ static bool llama_eval_internal( n_ctx*ggml_element_size(kv_self.v), n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head, il*n_ctx*ggml_element_size(kv_self.v)*n_embd); + ggml_set_name(V, "V"); #if 1 struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + 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 @@ -1161,11 +1240,13 @@ static bool llama_eval_internal( // 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 (no bias) cur = ggml_mul_mat(ctx0, @@ -1237,15 +1318,17 @@ static bool llama_eval_internal( lctx.use_buf(ctx0, -1); // logits -> probs - //inpL = ggml_soft_max(ctx0, inpL); + //inpL = ggml_soft_max_inplace(ctx0, inpL); // run the computation ggml_build_forward_expand(&gf, inpL); ggml_graph_compute (ctx0, &gf); +#ifdef GGML_PERF // print timing information per ggml operation (for debugging purposes) // requires GGML_PERF to be defined - //ggml_graph_print(&gf); + ggml_graph_print(&gf); +#endif // plot the computation graph in dot format (for debugging purposes) //if (n_past%100 == 0) { @@ -1255,6 +1338,9 @@ static bool llama_eval_internal( //embd_w.resize(n_vocab*N); //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N); + // update kv token count + lctx.model.kv_self.n = n_past + N; + // extract logits { auto & logits_out = lctx.logits; @@ -1270,7 +1356,7 @@ static bool llama_eval_internal( } // extract embeddings - if (lctx.embedding.size()) { + if (!lctx.embedding.empty()) { auto & embedding_out = lctx.embedding; embedding_out.resize(n_embd); @@ -1321,6 +1407,8 @@ struct llama_sp_symbol { size_t n; }; +static_assert(std::is_trivially_copyable::value, "llama_sp_symbol is not trivially copyable"); + struct llama_sp_bigram { struct comparator { bool operator()(llama_sp_bigram & l, llama_sp_bigram & r) { @@ -1353,7 +1441,7 @@ struct llama_tokenizer { sym.prev = index - 1; sym.next = offs == text.size() ? -1 : index + 1; index++; - symbols_.emplace_back(std::move(sym)); + symbols_.emplace_back(sym); } // seed the work queue with all possible 2-character tokens. @@ -1444,12 +1532,12 @@ static std::vector llama_tokenize(const llama_vocab & vocab, co llama_tokenizer tokenizer(vocab); std::vector output; - if (text.size() == 0) { + if (text.empty()) { return output; } if (bos) { - output.push_back(1); + output.push_back(llama_token_bos()); } tokenizer.tokenize(text, output); @@ -1460,125 +1548,424 @@ static std::vector llama_tokenize(const llama_vocab & vocab, co // sampling // -static void sample_top_k(std::vector> & logits_id, int top_k) { - // find the top k tokens - std::partial_sort( - logits_id.begin(), - logits_id.begin() + top_k, logits_id.end(), - [](const std::pair & a, const std::pair & b) { - return a.first > b.first; - }); +void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates) { + assert(candidates->size > 0); - logits_id.resize(top_k); + const int64_t t_start_sample_us = ggml_time_us(); + + // Sort the logits in descending order + if (!candidates->sorted) { + std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }); + candidates->sorted = true; + } + + float max_l = candidates->data[0].logit; + float cum_sum = 0.0f; + for (size_t i = 0; i < candidates->size; ++i) { + float p = expf(candidates->data[i].logit - max_l); + candidates->data[i].p = p; + cum_sum += p; + } + for (size_t i = 0; i < candidates->size; ++i) { + candidates->data[i].p /= cum_sum; + } + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } } -static llama_vocab::id llama_sample_top_p_top_k( - llama_context & lctx, - const std::vector & last_n_tokens, - int top_k, - float top_p, - float temp, - float repeat_penalty) { - auto & rng = lctx.rng; +void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep) { + const int64_t t_start_sample_us = ggml_time_us(); - const int n_logits = lctx.model.hparams.n_vocab; + k = std::max(k, (int) min_keep); + k = std::min(k, (int) candidates->size); - const auto & logits = lctx.logits; - const auto * plogits = logits.data() + logits.size() - n_logits; - - if (temp <= 0) { - // select the token with the highest logit directly - float max_logit = plogits[0]; - llama_vocab::id max_id = 0; - - for (int i = 1; i < n_logits; ++i) { - if (plogits[i] > max_logit) { - max_logit = plogits[i]; - max_id = i; - } + // Sort scores in descending order + if (!candidates->sorted) { + auto comp = [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }; + if (k == (int) candidates->size) { + std::sort(candidates->data, candidates->data + candidates->size, comp); + } else { + std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp); } - return max_id; + candidates->sorted = true; + } + candidates->size = k; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) { + if (p >= 1.0f) { + return; } - std::vector> logits_id; - logits_id.reserve(n_logits); + const int64_t t_start_sample_us = ggml_time_us(); - { - const float scale = 1.0f/temp; - for (int i = 0; i < n_logits; ++i) { - // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858) - // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main - if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) { - // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability - if (plogits[i] < 0.0f) { - logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i)); - } else { - logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i)); - } - } else { - logits_id.push_back(std::make_pair(plogits[i]*scale, i)); - } + llama_sample_softmax(ctx, candidates); + + // Compute the cumulative probabilities + float cum_sum = 0.0f; + size_t last_idx = candidates->size; + + for (size_t i = 0; i < candidates->size; ++i) { + cum_sum += candidates->data[i].p; + + // Check if the running sum is greater than p or if we have kept at least min_keep tokens + if (cum_sum > p && i >= min_keep) { + last_idx = i; + break; } } - sample_top_k(logits_id, top_k > 0 ? std::min(top_k, n_logits) : n_logits); + // Resize the output vector to keep only the top-p tokens + candidates->size = last_idx; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep) { + if (z >= 1.0f || candidates->size <= 2) { + return; + } + + const int64_t t_start_sample_us = ggml_time_us(); + + llama_sample_softmax(nullptr, candidates); + + // Compute the first and second derivatives + std::vector first_derivatives(candidates->size - 1); + std::vector second_derivatives(candidates->size - 2); + + for (size_t i = 0; i < first_derivatives.size(); ++i) { + first_derivatives[i] = candidates->data[i].p - candidates->data[i + 1].p; + } + for (size_t i = 0; i < second_derivatives.size(); ++i) { + second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1]; + } + + // Calculate absolute value of second derivatives + for (size_t i = 0; i < second_derivatives.size(); ++i) { + second_derivatives[i] = abs(second_derivatives[i]); + } + + // Normalize the second derivatives + float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); + for (float & value : second_derivatives) { + value /= second_derivatives_sum; + } + + float cum_sum = 0.0f; + size_t last_idx = candidates->size; + for (size_t i = 0; i < second_derivatives.size(); ++i) { + cum_sum += second_derivatives[i]; + + // Check if the running sum is greater than z or if we have kept at least min_keep tokens + if (cum_sum > z && i >= min_keep) { + last_idx = i; + break; + } + } + + // Resize the output vector to keep only the tokens above the tail location + candidates->size = last_idx; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + + +void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) { + // Reference implementation: + // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr + if (p >= 1.0f) { + return; + } + + const int64_t t_start_sample_us = ggml_time_us(); + + // Compute the softmax of logits and calculate entropy + llama_sample_softmax(nullptr, candidates); + + float entropy = 0.0f; + for (size_t i = 0; i < candidates->size; ++i) { + entropy += -candidates->data[i].p * logf(candidates->data[i].p); + } + + // Compute the absolute difference between negative log probability and entropy for each candidate + std::vector shifted_scores; + for (size_t i = 0; i < candidates->size; ++i) { + float shifted_score = fabsf(-logf(candidates->data[i].p) - entropy); + shifted_scores.push_back(shifted_score); + } + + // Sort tokens based on the shifted_scores and their corresponding indices + std::vector indices(candidates->size); + std::iota(indices.begin(), indices.end(), 0); + + std::sort(indices.begin(), indices.end(), [&](size_t a, size_t b) { + return shifted_scores[a] < shifted_scores[b]; + }); + + // Compute the cumulative probabilities + float cum_sum = 0.0f; + size_t last_idx = indices.size(); + + for (size_t i = 0; i < indices.size(); ++i) { + size_t idx = indices[i]; + cum_sum += candidates->data[idx].p; + + // Check if the running sum is greater than typical or if we have kept at least min_keep tokens + if (cum_sum > p && i >= min_keep - 1) { + last_idx = i + 1; + break; + } + } + + // Resize the output vector to keep only the locally typical tokens + std::vector new_candidates; + for (size_t i = 0; i < last_idx; ++i) { + size_t idx = indices[i]; + new_candidates.push_back(candidates->data[idx]); + } + + // Replace the data in candidates with the new_candidates data + std::copy(new_candidates.begin(), new_candidates.end(), candidates->data); + candidates->size = new_candidates.size(); + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) { + const int64_t t_start_sample_us = ggml_time_us(); + + for (size_t i = 0; i < candidates_p->size; ++i) { + candidates_p->data[i].logit /= temp; + } + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty) { + if (last_tokens_size == 0 || penalty == 1.0f) { + return; + } + + const int64_t t_start_sample_us = ggml_time_us(); + + for (size_t i = 0; i < candidates->size; ++i) { + const auto * token_iter = std::find(last_tokens, last_tokens + last_tokens_size, candidates->data[i].id); + if (token_iter == last_tokens + last_tokens_size) { + continue; + } + + // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong. + // This is common fix for this problem, which is to multiply by the penalty instead of dividing. + if (candidates->data[i].logit <= 0) { + candidates->data[i].logit *= penalty; + } else { + candidates->data[i].logit /= penalty; + } + } + + candidates->sorted = false; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens_p, size_t last_tokens_size, float alpha_frequency, float alpha_presence) { + if (last_tokens_size == 0 || (alpha_frequency == 0.0f && alpha_presence == 0.0f)) { + return; + } + + const int64_t t_start_sample_us = ggml_time_us(); + + // Create a frequency map to count occurrences of each token in last_tokens + std::unordered_map token_count; + for (size_t i = 0; i < last_tokens_size; ++i) { + token_count[last_tokens_p[i]]++; + } + + // Apply frequency and presence penalties to the candidates + for (size_t i = 0; i < candidates->size; ++i) { + auto token_iter = token_count.find(candidates->data[i].id); + if (token_iter == token_count.end()) { + continue; + } + + int count = token_iter->second; + candidates->data[i].logit -= float(count) * alpha_frequency + float(count > 0) * alpha_presence; + } + + candidates->sorted = false; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + + +llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu) { + assert(ctx); + auto N = float(llama_n_vocab(ctx)); + int64_t t_start_sample_us; + t_start_sample_us = ggml_time_us(); + + llama_sample_softmax(nullptr, candidates); + + // Estimate s_hat using the most probable m tokens + float s_hat = 0.0; + float sum_ti_bi = 0.0; + float sum_ti_sq = 0.0; + for (size_t i = 0; i < size_t(m - 1) && i < candidates->size - 1; ++i) { + float t_i = logf(float(i + 2) / float(i + 1)); + float b_i = logf(candidates->data[i].p / candidates->data[i + 1].p); + sum_ti_bi += t_i * b_i; + sum_ti_sq += t_i * t_i; + } + s_hat = sum_ti_bi / sum_ti_sq; + + // Compute k from the estimated s_hat and target surprise value + float epsilon_hat = s_hat - 1; + float k = powf((epsilon_hat * powf(2, *mu)) / (1 - powf(N, -epsilon_hat)), 1 / s_hat); + + // Sample the next word X using top-k sampling + llama_sample_top_k(nullptr, candidates, int(k), 1); + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } + llama_token X = llama_sample_token(ctx, candidates); + t_start_sample_us = ggml_time_us(); + + // Compute error as the difference between observed surprise and target surprise value + size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) { + return candidate.id == X; + })); + float observed_surprise = -log2f(candidates->data[X_idx].p); + float e = observed_surprise - tau; + + // Update mu using the learning rate and error + *mu = *mu - eta * e; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + ctx->n_sample++; + } + return X; +} + +llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu) { + assert(ctx); + int64_t t_start_sample_us; + t_start_sample_us = ggml_time_us(); + + llama_sample_softmax(ctx, candidates); + + // Truncate the words with surprise values greater than mu + candidates->size = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) { + return -log2f(candidate.p) > *mu; + })); + + // Normalize the probabilities of the remaining words + llama_sample_softmax(ctx, candidates); + + // Sample the next word X from the remaining words + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } + llama_token X = llama_sample_token(ctx, candidates); + t_start_sample_us = ggml_time_us(); + + // Compute error as the difference between observed surprise and target surprise value + size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) { + return candidate.id == X; + })); + float observed_surprise = -log2f(candidates->data[X_idx].p); + float e = observed_surprise - tau; + + // Update mu using the learning rate and error + *mu = *mu - eta * e; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } + return X; +} + +llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates) { + const int64_t t_start_sample_us = ggml_time_us(); + + // Find max element + auto * max_iter = std::max_element(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit < b.logit; + }); + + llama_token result = max_iter->id; + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + ctx->n_sample++; + } + return result; +} + +llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates) { + assert(ctx); + const int64_t t_start_sample_us = ggml_time_us(); + llama_sample_softmax(nullptr, candidates); - // compute probs for the top k tokens std::vector probs; - probs.reserve(logits_id.size()); - - float maxl = logits_id[0].first; - double sum = 0.0; - for (const auto & kv : logits_id) { - const float p = expf(kv.first - maxl); - probs.push_back(p); - sum += p; + probs.reserve(candidates->size); + for (size_t i = 0; i < candidates->size; ++i) { + probs.push_back(candidates->data[i].p); } - // normalize the probs - for (auto & p : probs) { - p /= sum; - } - - if (top_p < 1.0) { - double cumsum = 0.0; - for (int i = 0; i < (int) probs.size(); i++) { - cumsum += probs[i]; - if (cumsum >= top_p) { - probs.resize(i + 1); - logits_id.resize(i + 1); - break; - } - } - } - - //printf("\n"); - //for (int i = 0; i < (int) 10; i++) { - // printf("%d: '%s' %f\n", i, lctx.vocab.id_to_token.at(logits_id[i].second).tok.c_str(), probs[i]); - //} - //printf("\n\n"); - //exit(0); - std::discrete_distribution<> dist(probs.begin(), probs.end()); + auto & rng = ctx->rng; int idx = dist(rng); - return logits_id[idx].second; + llama_token result = candidates->data[idx].id; + + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + ctx->n_sample++; + return result; } // // quantization // -static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype) { +static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype, int nthread) { ggml_type quantized_type; switch (ftype) { case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break; case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break; - case LLAMA_FTYPE_MOSTLY_Q4_2: quantized_type = GGML_TYPE_Q4_2; break; + case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break; + case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break; + case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break; default: throw format("invalid output file type %d\n", ftype); }; - std::unique_ptr model_loader(new llama_model_loader(fname_inp.c_str(), /*use_mmap*/ false, + if (nthread <= 0) { + nthread = std::thread::hardware_concurrency(); + } + + std::unique_ptr model_loader(new llama_model_loader(fname_inp, /*use_mmap*/ false, /*vocab_only*/ false)); llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), ftype); @@ -1586,6 +1973,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s size_t total_size_new = 0; std::vector hist_all(1 << 4, 0); + std::vector workers; + std::mutex mutex; + size_t idx = 0; for (llama_load_tensor & tensor : model_loader->tensors_map.tensors) { llama_buffer read_data; @@ -1604,6 +1994,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // quantize only 2D tensors quantize &= (tensor.ne.size() == 2); + // uncomment this to keep the output layer in FP16 + //if (tensor.name == "output.weight") { + // quantize = false; + //} + enum ggml_type new_type; void * new_data; size_t new_size; @@ -1624,7 +2019,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } else if (tensor.type == GGML_TYPE_F16) { f32_conv_buf.resize(nelements * sizeof(float)); f32_data = (float *) f32_conv_buf.addr; - auto f16_data = (const ggml_fp16_t *) tensor.data; + const auto * f16_data = (const ggml_fp16_t *) tensor.data; for (size_t i = 0; i < nelements; i++) { f32_data[i] = ggml_fp16_to_fp32(f16_data[i]); } @@ -1639,21 +2034,47 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s new_data = work.addr; std::vector hist_cur(1 << 4, 0); - switch (new_type) { - case GGML_TYPE_Q4_0: - { - new_size = ggml_quantize_q4_0(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data()); - } break; - case GGML_TYPE_Q4_1: - { - new_size = ggml_quantize_q4_1(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data()); - } break; - case GGML_TYPE_Q4_2: - { - new_size = ggml_quantize_q4_2(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data()); - } break; - default: - LLAMA_ASSERT(false); + int chunk_size = 32 * 512; + const int nchunk = (nelements + chunk_size - 1)/chunk_size; + const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1; + if (nthread_use < 2) { + new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nelements, hist_cur.data()); + } else { + size_t counter = 0; + new_size = 0; + auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements, chunk_size] () { + std::vector 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()) { + for (int j=0; j %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0); @@ -1699,7 +2120,7 @@ struct llama_context * llama_init_from_file( llama_context * ctx = new llama_context; - if (params.seed <= 0) { + if (params.seed < 0) { params.seed = time(NULL); } @@ -1725,7 +2146,7 @@ struct llama_context * llama_init_from_file( ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32; - if (!llama_model_load(path_model, *ctx, params.n_ctx, memory_type, + if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_gpu_layers, memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback, params.progress_callback_user_data)) { fprintf(stderr, "%s: failed to load model\n", __func__); @@ -1752,7 +2173,7 @@ struct llama_context * llama_init_from_file( if (params.logits_all) { ctx->logits.reserve(hparams.n_ctx*hparams.n_vocab); } else { - ctx->logits.reserve(hparams.n_ctx); + ctx->logits.reserve(hparams.n_vocab); } if (params.embedding){ @@ -1775,9 +2196,10 @@ void llama_free(struct llama_context * ctx) { int llama_model_quantize( const char * fname_inp, const char * fname_out, - enum llama_ftype ftype) { + enum llama_ftype ftype, + int nthread) { try { - llama_model_quantize_internal(fname_inp, fname_out, ftype); + llama_model_quantize_internal(fname_inp, fname_out, ftype, nthread); return 0; } catch (const std::string & err) { fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.c_str()); @@ -1850,7 +2272,8 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model); model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*vocab_only*/ false)); - size_t ctx_size, mmapped_size; + size_t ctx_size; + size_t mmapped_size; model_loader->calc_sizes(&ctx_size, &mmapped_size); base_buf.resize(ctx_size); @@ -1889,8 +2312,12 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); } - std::string name(length, 0); - fin.read(&name[0], length); + std::string name; + { + char buf[1024]; + fin.read(buf, length); + name = std::string(buf, length); + } // check for lora suffix and get the type of tensor const std::string lora_suffix = ".lora"; @@ -1905,7 +2332,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * base_name.erase(pos); // fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str()); - if (model_tensors.find(base_name.data()) == model_tensors.end()) { + if (model_tensors.find(base_name) == model_tensors.end()) { fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data()); return 1; } @@ -1963,7 +2390,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * base_t = dest_t; } - if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1 || base_t->type == GGML_TYPE_Q4_2) { + if (ggml_is_quantized(base_t->type)) { if (!warned) { fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, " "use a f16 or f32 base model with --lora-base\n", __func__); @@ -1985,7 +2412,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * if (scaling != 1.0f) { ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling); - BA = ggml_scale(lora_ctx, BA, scale_tensor); + BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor); } ggml_tensor * r; @@ -2007,8 +2434,9 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * lora_tensors.clear(); n_tensors++; - if (n_tensors % 4 == 0) + if (n_tensors % 4 == 0) { fprintf(stderr, "."); + } } } @@ -2033,31 +2461,338 @@ int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lor } } -// Returns the KV cache that will contain the context for the -// ongoing prediction with the model. -const uint8_t * llama_get_kv_cache(struct llama_context * ctx) { - return ctx->model.kv_self.buf.addr; -} - -// Returns the size of the KV cache -size_t llama_get_kv_cache_size(struct llama_context * ctx) { - return ctx->model.kv_self.buf.size; -} - -int llama_get_kv_cache_token_count(struct llama_context * ctx) { +int llama_get_kv_cache_token_count(const struct llama_context * ctx) { return ctx->model.kv_self.n; } -// Sets the KV cache containing the current context for the model -void llama_set_kv_cache( - struct llama_context * ctx, - const uint8_t * kv_cache, - size_t n_size, - int n_token_count) { - // Make sure we have the same kv cache setup - LLAMA_ASSERT(ctx->model.kv_self.buf.size == n_size); - memcpy(ctx->model.kv_self.buf.addr, kv_cache, n_size); - ctx->model.kv_self.n = n_token_count; +#define LLAMA_MAX_RNG_STATE (64*1024) + +void llama_set_rng_seed(struct llama_context * ctx, int seed) { + if (seed < 0) { + seed = time(NULL); + } + ctx->rng.seed(seed); +} + +// Returns the *maximum* size of the state +size_t llama_get_state_size(const struct llama_context * ctx) { + // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state. + // for reference, std::mt19937(1337) serializes to 6701 bytes. + const size_t s_rng_size = sizeof(size_t); + const size_t s_rng = LLAMA_MAX_RNG_STATE; + const size_t s_logits_capacity = sizeof(size_t); + const size_t s_logits_size = sizeof(size_t); + const size_t s_logits = ctx->logits.capacity() * sizeof(float); + const size_t s_embedding_size = sizeof(size_t); + const size_t s_embedding = ctx->embedding.size() * sizeof(float); + const size_t s_kv_size = sizeof(size_t); + const size_t s_kv_ntok = sizeof(int); + const size_t s_kv = ctx->model.kv_self.buf.size; + + const size_t s_total = ( + + s_rng_size + + s_rng + + s_logits_capacity + + s_logits_size + + s_logits + + s_embedding_size + + s_embedding + + s_kv_size + + s_kv_ntok + + s_kv + ); + + return s_total; +} + +// Copies the state to the specified destination address +size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { + uint8_t * out = dst; + + // copy rng + { + std::stringstream rng_ss; + rng_ss << ctx->rng; + + const size_t rng_size = rng_ss.str().size(); + char rng_buf[LLAMA_MAX_RNG_STATE]; + + memset(&rng_buf[0], 0, LLAMA_MAX_RNG_STATE); + memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size()); + + memcpy(out, &rng_size, sizeof(rng_size)); out += sizeof(rng_size); + memcpy(out, &rng_buf[0], LLAMA_MAX_RNG_STATE); out += LLAMA_MAX_RNG_STATE; + } + + // copy logits + { + const size_t logits_cap = ctx->logits.capacity(); + const size_t logits_size = ctx->logits.size(); + + memcpy(out, &logits_cap, sizeof(logits_cap)); out += sizeof(logits_cap); + memcpy(out, &logits_size, sizeof(logits_size)); out += sizeof(logits_size); + + if (logits_size) { + memcpy(out, ctx->logits.data(), logits_size * sizeof(float)); + } + + out += logits_cap * sizeof(float); + } + + // copy embeddings + { + const size_t embedding_size = ctx->embedding.size(); + + memcpy(out, &embedding_size, sizeof(embedding_size)); out += sizeof(embedding_size); + + if (embedding_size) { + memcpy(out, ctx->embedding.data(), embedding_size * sizeof(float)); + out += embedding_size * sizeof(float); + } + } + + // copy kv cache + { + const auto & kv_self = ctx->model.kv_self; + const auto & hparams = ctx->model.hparams; + const int n_layer = hparams.n_layer; + const int n_embd = hparams.n_embd; + const int n_ctx = hparams.n_ctx; + + const size_t kv_size = kv_self.buf.size; + const int kv_ntok = llama_get_kv_cache_token_count(ctx); + + memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size); + memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok); + + if (kv_size) { + const size_t elt_size = ggml_element_size(kv_self.k); + + char buffer[4096]; + + ggml_context * cpy_ctx = ggml_init({ sizeof(buffer), buffer, /* no_alloc */ true }); + ggml_cgraph gf{}; + gf.n_threads = 1; + + ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer); + kout3d->data = out; + out += ggml_nbytes(kout3d); + + ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer); + vout3d->data = out; + out += ggml_nbytes(vout3d); + + ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, + n_embd, kv_ntok, n_layer, + elt_size*n_embd, elt_size*n_embd*n_ctx, 0); + + ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, + kv_ntok, n_embd, n_layer, + elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); + + ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d)); + ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d)); + ggml_graph_compute(cpy_ctx, &gf); + + ggml_free(cpy_ctx); + } + } + + const size_t written = out - dst; + const size_t max_size = llama_get_state_size(ctx); + + LLAMA_ASSERT(written <= max_size); + + return written; +} + +// Sets the state reading from the specified source address +size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { + const uint8_t * inp = src; + + // set rng + { + size_t rng_size; + char rng_buf[LLAMA_MAX_RNG_STATE]; + + memcpy(&rng_size, inp, sizeof(rng_size)); inp += sizeof(rng_size); + memcpy(&rng_buf[0], inp, LLAMA_MAX_RNG_STATE); inp += LLAMA_MAX_RNG_STATE; + + std::stringstream rng_ss; + rng_ss.str(std::string(&rng_buf[0], rng_size)); + rng_ss >> ctx->rng; + + LLAMA_ASSERT(rng_ss.fail() == false); + } + + // set logits + { + size_t logits_cap; + size_t logits_size; + + memcpy(&logits_cap, inp, sizeof(logits_cap)); inp += sizeof(logits_cap); + memcpy(&logits_size, inp, sizeof(logits_size)); inp += sizeof(logits_size); + + LLAMA_ASSERT(ctx->logits.capacity() == logits_cap); + + if (logits_size) { + ctx->logits.resize(logits_size); + memcpy(ctx->logits.data(), inp, logits_size * sizeof(float)); + } + + inp += logits_cap * sizeof(float); + } + + // set embeddings + { + size_t embedding_size; + + memcpy(&embedding_size, inp, sizeof(embedding_size)); inp += sizeof(embedding_size); + + LLAMA_ASSERT(ctx->embedding.capacity() == embedding_size); + + if (embedding_size) { + memcpy(ctx->embedding.data(), inp, embedding_size * sizeof(float)); + inp += embedding_size * sizeof(float); + } + } + + // set kv cache + { + const auto & kv_self = ctx->model.kv_self; + const auto & hparams = ctx->model.hparams; + const int n_layer = hparams.n_layer; + const int n_embd = hparams.n_embd; + const int n_ctx = hparams.n_ctx; + + size_t kv_size; + int kv_ntok; + + memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size); + memcpy(&kv_ntok, inp, sizeof(kv_ntok)); inp += sizeof(kv_ntok); + + if (kv_size) { + LLAMA_ASSERT(kv_self.buf.size == kv_size); + + const size_t elt_size = ggml_element_size(kv_self.k); + + char buffer[4096]; + + ggml_context * cpy_ctx = ggml_init({ sizeof(buffer), buffer, /* no_alloc */ true }); + ggml_cgraph gf{}; + gf.n_threads = 1; + + ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer); + kin3d->data = (void *) inp; + inp += ggml_nbytes(kin3d); + + ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer); + vin3d->data = (void *) inp; + inp += ggml_nbytes(vin3d); + + ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, + n_embd, kv_ntok, n_layer, + elt_size*n_embd, elt_size*n_embd*n_ctx, 0); + + ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, + kv_ntok, n_embd, n_layer, + elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); + + ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d)); + ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d)); + ggml_graph_compute(cpy_ctx, &gf); + + ggml_free(cpy_ctx); + } + + ctx->model.kv_self.n = kv_ntok; + } + + const size_t nread = inp - src; + const size_t max_size = llama_get_state_size(ctx); + + LLAMA_ASSERT(nread <= max_size); + + return nread; +} + +bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + llama_file file(path_session, "rb"); + + // sanity checks + { + const uint32_t magic = file.read_u32(); + const uint32_t version = file.read_u32(); + + if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) { + fprintf(stderr, "%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version); + return false; + } + + llama_hparams session_hparams; + file.read_raw(&session_hparams, sizeof(llama_hparams)); + + if (session_hparams != ctx->model.hparams) { + fprintf(stderr, "%s : model hparams didn't match from session file!\n", __func__); + return false; + } + } + + // load the prompt + { + const uint32_t n_token_count = file.read_u32(); + + if (n_token_count > n_token_capacity) { + fprintf(stderr, "%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity); + return false; + } + + file.read_raw(tokens_out, sizeof(llama_token) * n_token_count); + *n_token_count_out = n_token_count; + } + + // restore the context state + { + const size_t n_state_size_cur = file.size - file.tell(); + const size_t n_state_size_max = llama_get_state_size(ctx); + + if (n_state_size_cur > n_state_size_max) { + fprintf(stderr, "%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur); + return false; + } + + std::vector state_data(n_state_size_max); + file.read_raw(state_data.data(), n_state_size_cur); + + llama_set_state_data(ctx, state_data.data()); + } + + return true; +} + +bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + llama_file file(path_session, "wb"); + + file.write_u32(LLAMA_SESSION_MAGIC); + file.write_u32(LLAMA_SESSION_VERSION); + + file.write_raw(&ctx->model.hparams, sizeof(llama_hparams)); + + // save the prompt + file.write_u32((uint32_t) n_token_count); + file.write_raw(tokens, sizeof(llama_token) * n_token_count); + + // save the context state + { + const size_t n_state_size_max = llama_get_state_size(ctx); + + std::vector state_data(n_state_size_max); + const size_t n_state_size_cur = llama_copy_state_data(ctx, state_data.data()); + + file.write_raw(state_data.data(), n_state_size_cur); + } + + return true; } int llama_eval( @@ -2070,11 +2805,14 @@ int llama_eval( fprintf(stderr, "%s: failed to eval\n", __func__); return 1; } + // get a more accurate load time, upon first eval + // TODO: fix this if (!ctx->has_evaluated_once) { ctx->t_load_us = ggml_time_us() - ctx->t_start_us; ctx->has_evaluated_once = true; } + return 0; } @@ -2098,15 +2836,15 @@ int llama_tokenize( return res.size(); } -int llama_n_vocab(struct llama_context * ctx) { +int llama_n_vocab(const struct llama_context * ctx) { return ctx->vocab.id_to_token.size(); } -int llama_n_ctx(struct llama_context * ctx) { +int llama_n_ctx(const struct llama_context * ctx) { return ctx->model.hparams.n_ctx; } -int llama_n_embd(struct llama_context * ctx) { +int llama_n_embd(const struct llama_context * ctx) { return ctx->model.hparams.n_embd; } @@ -2118,7 +2856,7 @@ float * llama_get_embeddings(struct llama_context * ctx) { return ctx->embedding.data(); } -const char * llama_token_to_str(struct llama_context * ctx, llama_token token) { +const char * llama_token_to_str(const struct llama_context * ctx, llama_token token) { if (token >= llama_n_vocab(ctx)) { return nullptr; } @@ -2134,33 +2872,8 @@ llama_token llama_token_eos() { return 2; } -llama_token llama_sample_top_p_top_k( - llama_context * ctx, - const llama_token * last_n_tokens_data, - int last_n_tokens_size, - int top_k, - float top_p, - float temp, - float repeat_penalty) { - const int64_t t_start_sample_us = ggml_time_us(); - - llama_token result = 0; - - // TODO: avoid this ... - const auto last_n_tokens = std::vector(last_n_tokens_data, last_n_tokens_data + last_n_tokens_size); - - result = llama_sample_top_p_top_k( - *ctx, - last_n_tokens, - top_k, - top_p, - temp, - repeat_penalty); - - ctx->t_sample_us += ggml_time_us() - t_start_sample_us; - ctx->n_sample++; - - return result; +llama_token llama_token_nl() { + return 13; } @@ -2173,9 +2886,9 @@ void llama_print_timings(struct llama_context * ctx) { fprintf(stderr, "\n"); fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0); - fprintf(stderr, "%s: sample time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample); + fprintf(stderr, "%s: sample time = %8.2f ms / %5d runs (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample); fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval); - fprintf(stderr, "%s: eval time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3 * ctx->t_eval_us, n_eval, 1e-3 * ctx->t_eval_us / n_eval); + fprintf(stderr, "%s: eval time = %8.2f ms / %5d runs (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_eval_us, n_eval, 1e-3 * ctx->t_eval_us / n_eval); fprintf(stderr, "%s: total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0); } diff --git a/llama.h b/llama.h index 208b03d18..21cba8cf6 100644 --- a/llama.h +++ b/llama.h @@ -19,9 +19,11 @@ # define LLAMA_API #endif -#define LLAMA_FILE_VERSION 1 -#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex -#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files +#define LLAMA_FILE_VERSION 2 +#define LLAMA_FILE_MAGIC 'ggjt' +#define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml' +#define LLAMA_SESSION_MAGIC 'ggsn' +#define LLAMA_SESSION_VERSION 1 #ifdef __cplusplus extern "C" { @@ -39,18 +41,23 @@ extern "C" { typedef struct llama_token_data { llama_token id; // token id - + float logit; // log-odds of the token float p; // probability of the token - float plog; // log probability of the token - } llama_token_data; + typedef struct llama_token_data_array { + llama_token_data * data; + size_t size; + bool sorted; + } llama_token_data_array; + typedef void (*llama_progress_callback)(float progress, void *ctx); struct llama_context_params { - int n_ctx; // text context - int n_parts; // -1 for default - int seed; // RNG seed, 0 for random + int n_ctx; // text context + int n_parts; // -1 for default + int n_gpu_layers; // number of layers to store in VRAM + int seed; // RNG seed, -1 for random bool f16_kv; // use fp16 for KV cache bool logits_all; // the llama_eval() call computes all logits, not just the last one @@ -72,7 +79,11 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 - LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // except 1d tensors + // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed + // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed + LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors }; LLAMA_API struct llama_context_params llama_context_default_params(); @@ -92,10 +103,12 @@ extern "C" { // TODO: not great API - very likely to change // Returns 0 on success + // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given LLAMA_API int llama_model_quantize( const char * fname_inp, const char * fname_out, - enum llama_ftype ftype); + enum llama_ftype ftype, + int nthread); // Apply a LoRA adapter to a loaded model // path_base_model is the path to a higher quality model to use as a base for @@ -109,22 +122,28 @@ extern "C" { const char * path_base_model, int n_threads); - // Returns the KV cache that will contain the context for the - // ongoing prediction with the model. - LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx); - - // Returns the size of the KV cache - LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx); - // Returns the number of tokens in the KV cache - LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx); + LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx); - // Sets the KV cache containing the current context for the model - LLAMA_API void llama_set_kv_cache( - struct llama_context * ctx, - const uint8_t * kv_cache, - size_t n_size, - int n_token_count); + // Sets the current rng seed. + LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed); + + // Returns the maximum size in bytes of the state (rng, logits, embedding + // and kv_cache) - will often be smaller after compacting tokens + LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx); + + // Copies the state to the specified destination address. + // Destination needs to have allocated enough memory. + // Returns the number of bytes copied + LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst); + + // Set the state reading from the specified address + // Returns the number of bytes read + LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src); + + // Save/load session file + LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out); + LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count); // Run the llama inference to obtain the logits and probabilities for the next token. // tokens + n_tokens is the provided batch of new tokens to process @@ -149,9 +168,9 @@ extern "C" { int n_max_tokens, bool add_bos); - LLAMA_API int llama_n_vocab(struct llama_context * ctx); - LLAMA_API int llama_n_ctx (struct llama_context * ctx); - LLAMA_API int llama_n_embd (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_embd (const struct llama_context * ctx); // Token logits obtained from the last call to llama_eval() // The logits for the last token are stored in the last row @@ -165,21 +184,57 @@ extern "C" { LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); // Token Id -> String. Uses the vocabulary in the provided context - LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token); + LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token); // Special tokens LLAMA_API llama_token llama_token_bos(); LLAMA_API llama_token llama_token_eos(); + LLAMA_API llama_token llama_token_nl(); - // TODO: improve the last_n_tokens interface ? - LLAMA_API llama_token llama_sample_top_p_top_k( - struct llama_context * ctx, - const llama_token * last_n_tokens_data, - int last_n_tokens_size, - int top_k, - float top_p, - float temp, - float repeat_penalty); + // Sampling functions + + /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. + LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty); + + /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. + LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence); + + /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. + LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates); + + /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 + LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep); + + /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 + LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep); + + /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. + LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep); + + /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. + LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep); + LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp); + + /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. + /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. + /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. + /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. + /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. + /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. + LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu); + + /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. + /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. + /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. + /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. + /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. + LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu); + + /// @details Selects the token with the highest probability. + LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates); + + /// @details Randomly selects a token from the candidates based on their probabilities. + LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates); // Performance information LLAMA_API void llama_print_timings(struct llama_context * ctx); diff --git a/pocs/vdot/CMakeLists.txt b/pocs/vdot/CMakeLists.txt index cbc852236..fb89a1cd4 100644 --- a/pocs/vdot/CMakeLists.txt +++ b/pocs/vdot/CMakeLists.txt @@ -2,3 +2,8 @@ set(TARGET vdot) add_executable(${TARGET} vdot.cpp) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) + +set(TARGET q8dot) +add_executable(${TARGET} q8dot.cpp) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/pocs/vdot/q8dot.cpp b/pocs/vdot/q8dot.cpp new file mode 100644 index 000000000..5748c8ac2 --- /dev/null +++ b/pocs/vdot/q8dot.cpp @@ -0,0 +1,172 @@ +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +constexpr int kVecSize = 1 << 16; + +// Copy-pasted from ggml.c +#define QK4_0 32 +typedef struct { + float d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +typedef struct { + float d; // delta + float m; // min + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); + +// Copy-pasted from ggml.c +#define QK8_0 32 +typedef struct { + float d; // delta + float s; // d * sum(qs[i]) + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); + +static_assert(QK4_1 == QK8_0, "QK4_1 and QK8_0 must be the same"); +static_assert(QK4_0 == QK8_0, "QK4_0 and QK8_0 must be the same"); + +template +void fillQ4blocks(std::vector& blocks, std::mt19937& rndm) { + for (auto& b : blocks) { + b.d = 1; + for (int i=0; i> 28; + uint8_t v2 = rndm() >> 28; + b.qs[i] = v1 | (v2 << 4); + } + } +} + +void fillQ80blocks(std::vector& blocks, std::mt19937& rndm) { + for (auto& b : blocks) { + b.d = 1; + int sum = 0; + for (int i=0; i> 24) - 128; + sum += b.qs[i]; + } + b.s = b.d * sum; + } +} + +float simpleDot(const block_q4_0& x, const block_q8_0& y) { + int s1 = 0; //, s2 = 0; + for (int i=0; i> 4; + int v3 = x.qs[i+1] & 0xf; + int v4 = x.qs[i+1] >> 4; + int j = 2*i; + s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3]; + //s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3]; + } + return y.d * x.d * s1 - 8 * x.d * y.s; + //return y.d * x.d * (s1 - 8 * s2); +} + +float simpleDot(const block_q4_1& x, const block_q8_0& y) { + int s1 = 0; //, s2 = 0; + for (int i=0; i> 4; + int v3 = x.qs[i+1] & 0xf; + int v4 = x.qs[i+1] >> 4; + int j = 2*i; + s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3]; + //s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3]; + } + return y.d * x.d * s1 + y.s * x.m; + //return y.d * (x.d * s1 + x.m * s2); +} + +struct Stat { + double sum = 0, sumt = 0, sumt2 = 0, maxt = 0; + int nloop = 0; + void addResult(double s, double t) { + sum += s; + sumt += t; sumt2 += t*t; maxt = std::max(maxt, t); + ++nloop; + } + void reportResult(const char* title) const { + if (nloop < 1) { + printf("%s(%s): no result\n",__func__,title); + return; + } + printf("============ %s\n",title); + printf(" = %g\n",sum/nloop); + auto t = sumt/nloop, dt = sumt2/nloop - t*t; + if (dt > 0) dt = sqrt(dt); + printf("