vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'ggerganov:master' into fix-cmake-pkg-cross-compile

Browse source
This commit is contained in:
bandoti 2024-11-18 09:05:16 -04:00 • committed by GitHub
commit 03acc8d39b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
232 changed files with 23361 additions and 24937 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.devops/full-cuda.Dockerfile
View file

@ -26,7 +26,7 @@ COPY . .
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release -j$(nproc) && \
cp build/bin/* .

2
.devops/full-musa.Dockerfile
View file

@ -19,7 +19,7 @@ WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release -j$(nproc) && \
cp build/bin/* .

6
.devops/llama-cli-cann.Dockerfile
View file

@ -1,6 +1,6 @@
ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
FROM cosdt/cann:$ASCEND_VERSION AS build
FROM ascendai/cann:$ASCEND_VERSION AS build
WORKDIR /app
@ -22,11 +22,11 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
RUN echo "Building with static libs" && \
source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
cmake -B build -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF && \
cmake --build build --config Release --target llama-cli
# TODO: use image with NNRT
FROM cosdt/cann:$ASCEND_VERSION AS runtime
FROM ascendai/cann:$ASCEND_VERSION AS runtime
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENV LC_ALL=C.utf8

11
.devops/llama-cli-cuda.Dockerfile
View file

@ -22,16 +22,17 @@ COPY . .
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc)
cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/build/bin/llama-cli /llama-cli
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-cli /
ENTRYPOINT [ "/llama-cli" ]

4
.devops/llama-cli-intel.Dockerfile
View file

@ -1,4 +1,4 @@
ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
@ -15,7 +15,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
fi && \
echo "Building with static libs" && \
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
cmake --build build --config Release --target llama-cli

9
.devops/llama-cli-musa.Dockerfile
View file

@ -15,16 +15,17 @@ WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc)
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENTRYPOINT [ "/llama-cli" ]

2
.devops/llama-cli-vulkan.Dockerfile
View file

@ -14,7 +14,7 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
# Build it
WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_VULKAN=1 && \
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 && \
cmake --build build --config Release --target llama-cli
# Clean up

9
.devops/llama-server-cuda.Dockerfile
View file

@ -22,16 +22,17 @@ COPY . .
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc)
cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-server /llama-server
# Must be set to 0.0.0.0 so it can listen to requests from host machine

4
.devops/llama-server-intel.Dockerfile
View file

@ -1,4 +1,4 @@
ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
@ -15,7 +15,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
fi && \
echo "Building with dynamic libs" && \
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
cmake --build build --config Release --target llama-server
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime

9
.devops/llama-server-musa.Dockerfile
View file

@ -15,16 +15,17 @@ WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc)
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-server /llama-server
# Must be set to 0.0.0.0 so it can listen to requests from host machine

2
.devops/llama-server-vulkan.Dockerfile
View file

@ -14,7 +14,7 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
# Build it
WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
cmake --build build --config Release --target llama-server
# Clean up

6
.devops/nix/package.nix
View file

@ -126,9 +126,9 @@ effectiveStdenv.mkDerivation (finalAttrs: {
};
postPatch = ''
substituteInPlace ./ggml/src/ggml-metal.m \
substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
--replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
substituteInPlace ./ggml/src/ggml-metal.m \
substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
--replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
'';
@ -173,7 +173,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
(cmakeBool "GGML_NATIVE" false)
(cmakeBool "GGML_BLAS" useBlas)
(cmakeBool "GGML_CUDA" useCuda)
(cmakeBool "GGML_HIPBLAS" useRocm)
(cmakeBool "GGML_HIP" useRocm)
(cmakeBool "GGML_METAL" useMetalKit)
(cmakeBool "GGML_VULKAN" useVulkan)
(cmakeBool "GGML_STATIC" enableStatic)

33
.github/workflows/build.yml vendored
View file

@ -405,15 +405,36 @@ jobs:
- name: Build with native CMake HIP support
id: cmake_build
run: |
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIP=ON
cmake --build build --config Release -j $(nproc)
- name: Build with legacy HIP support
id: cmake_build_legacy_hip
run: |
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=ON
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIP=ON
cmake --build build2 --config Release -j $(nproc)
ubuntu-22-cmake-musa:
runs-on: ubuntu-22.04
container: mthreads/musa:rc3.1.0-devel-ubuntu22.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Dependencies
id: depends
run: |
apt-get update
apt-get install -y build-essential git cmake libcurl4-openssl-dev
- name: Build with native CMake MUSA support
id: cmake_build
run: |
cmake -B build -S . -DGGML_MUSA=ON
cmake --build build --config Release -j $(nproc)
ubuntu-22-cmake-sycl:
runs-on: ubuntu-22.04
@ -747,7 +768,7 @@ jobs:
id: clone_kompute
if: ${{ matrix.build == 'kompute-x64' }}
run: |
git submodule update --init ggml/src/kompute
git submodule update --init ggml/src/ggml-kompute/kompute
- name: Download OpenBLAS
id: get_openblas
@ -930,7 +951,7 @@ jobs:
shell: bash
env:
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
steps:
@ -1014,7 +1035,7 @@ jobs:
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
windows-latest-cmake-hip-release:
@ -1050,7 +1071,7 @@ jobs:
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
md "build\bin\rocblas\library\"
cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"

5
.gitignore vendored
View file

@ -3,6 +3,7 @@
*.a
*.bat
*.bin
*.d
*.dll
*.dot
*.etag
@ -133,3 +134,7 @@ poetry.toml
# Test models for lora adapters
/lora-tests
# Local scripts
/run-vim.sh
/run-chat.sh

2
.gitmodules vendored
View file

@ -1,3 +1,3 @@
[submodule "kompute"]
path = ggml/src/kompute
path = ggml/src/ggml-kompute/kompute
url = https://github.com/nomic-ai/kompute.git

1
CMakeLists.txt
View file

@ -140,7 +140,6 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o
set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files")
set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files")
# At the moment some compile definitions are placed within the ggml/src
# directory but not exported on the `ggml` target. This could be improved by
# determining _precisely_ which defines are necessary for the llama-config

34
CMakePresets.json
View file

@ -24,11 +24,12 @@
"CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
}
},
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{ "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{ "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
{ "name": "vulkan", "hidden": true, "cacheVariables": { "GGML_VULKAN": "ON" } },
{
"name": "arm64-windows-msvc", "hidden": true,
@ -57,25 +58,28 @@
}
},
{ "name": "arm64-windows-llvm-debug" , "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-apple-clang-debug" , "inherits": [ "base", "arm64-apple-clang", "debug" ] },
{ "name": "arm64-apple-clang-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
{ "name": "arm64-apple-clang+static-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] },
{ "name": "arm64-apple-clang-debug", "inherits": [ "base", "arm64-apple-clang", "debug" ] },
{ "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
{ "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] },
{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] },
{ "name": "x64-windows-msvc-debug" , "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
{ "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
{ "name": "x64-windows-sycl-debug" , "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-debug", "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
{ "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
{ "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
{ "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] },
{ "name": "x64-windows-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
{ "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
]
}

587
Makefile
View file

@ -48,7 +48,6 @@ TEST_TARGETS = \
tests/test-backend-ops \
tests/test-chat-template \
tests/test-double-float \
tests/test-grad0 \
tests/test-grammar-integration \
tests/test-grammar-parser \
tests/test-json-schema-to-grammar \
@ -359,6 +358,10 @@ ifdef LLAMA_SERVER_SSL
MK_LDFLAGS += -lssl -lcrypto
endif
ifndef GGML_NO_CPU_AARCH64
MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
endif
# warnings
WARN_FLAGS = \
-Wall \
@ -523,70 +526,59 @@ ifndef GGML_NO_ACCELERATE
# Mac OS - include Accelerate framework.
# `-framework Accelerate` works both with Apple Silicon and Mac Intel
ifeq ($(UNAME_S),Darwin)
MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
MK_LDFLAGS += -framework Accelerate
OBJ_GGML += ggml/src/ggml-blas.o
MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
MK_LDFLAGS += -framework Accelerate
OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
endif
endif # GGML_NO_ACCELERATE
ifdef GGML_MUSA
CC := clang
CXX := clang++
GGML_CUDA := 1
MK_CPPFLAGS += -DGGML_USE_MUSA
endif
ifndef GGML_NO_OPENMP
MK_CPPFLAGS += -DGGML_USE_OPENMP
MK_CFLAGS += -fopenmp
MK_CXXFLAGS += -fopenmp
ifdef GGML_MUSA
MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
MK_LDFLAGS += -L/usr/lib/llvm-10/lib
endif # GGML_MUSA
endif # GGML_NO_OPENMP
ifdef GGML_OPENBLAS
MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas)
MK_LDFLAGS += $(shell pkg-config --libs openblas)
OBJ_GGML += ggml/src/ggml-blas.o
MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas)
MK_LDFLAGS += $(shell pkg-config --libs openblas)
OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_OPENBLAS
ifdef GGML_OPENBLAS64
MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas64)
MK_LDFLAGS += $(shell pkg-config --libs openblas64)
OBJ_GGML += ggml/src/ggml-blas.o
MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas64)
MK_LDFLAGS += $(shell pkg-config --libs openblas64)
OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_OPENBLAS64
ifdef GGML_BLIS
MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
MK_LDFLAGS += -lblis -L/usr/local/lib
OBJ_GGML += ggml/src/ggml-blas.o
MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
MK_LDFLAGS += -lblis -L/usr/local/lib
OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_BLIS
ifdef GGML_NVPL
MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
MK_LDFLAGS += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
OBJ_GGML += ggml/src/ggml-blas.o
MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
MK_LDFLAGS += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_NVPL
ifndef GGML_NO_LLAMAFILE
MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
OBJ_GGML += ggml/src/llamafile/sgemm.o
MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
OBJ_GGML_EXT += ggml/src/ggml-cpu/llamafile/sgemm.o
endif
ifndef GGML_NO_AMX
MK_CPPFLAGS += -DGGML_USE_AMX
OBJ_GGML += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o
OBJ_GGML_EXT += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
endif
ifdef GGML_RPC
MK_CPPFLAGS += -DGGML_USE_RPC
OBJ_GGML += ggml/src/ggml-rpc.o
MK_CPPFLAGS += -DGGML_USE_RPC
OBJ_GGML_EXT += ggml/src/ggml-rpc.o
endif # GGML_RPC
OBJ_CUDA_TMPL = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
@ -601,41 +593,27 @@ else
endif # GGML_CUDA_FA_ALL_QUANTS
ifdef GGML_CUDA
ifdef GGML_MUSA
ifneq ('', '$(wildcard /opt/musa)')
CUDA_PATH ?= /opt/musa
else
CUDA_PATH ?= /usr/local/musa
endif
MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
MK_LDFLAGS += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
ifneq ('', '$(wildcard /opt/cuda)')
CUDA_PATH ?= /opt/cuda
else
ifneq ('', '$(wildcard /opt/cuda)')
CUDA_PATH ?= /opt/cuda
else
CUDA_PATH ?= /usr/local/cuda
endif
CUDA_PATH ?= /usr/local/cuda
endif
MK_CPPFLAGS += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
MK_NVCCFLAGS += -use_fast_math
endif # GGML_MUSA
MK_CPPFLAGS += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
MK_NVCCFLAGS += -use_fast_math
OBJ_GGML += ggml/src/ggml-cuda.o
OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML += $(OBJ_CUDA_TMPL)
OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
ifdef LLAMA_FATAL_WARNINGS
MK_NVCCFLAGS += -Werror all-warnings
endif # LLAMA_FATAL_WARNINGS
ifndef GGML_MUSA
ifndef JETSON_EOL_MODULE_DETECT
MK_NVCCFLAGS += --forward-unknown-to-host-compiler
endif # JETSON_EOL_MODULE_DETECT
endif # GGML_MUSA
ifdef LLAMA_DEBUG
MK_NVCCFLAGS += -lineinfo
@ -648,11 +626,7 @@ endif # GGML_CUDA_DEBUG
ifdef GGML_CUDA_NVCC
NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
else
ifdef GGML_MUSA
NVCC = $(CCACHE) mcc
else
NVCC = $(CCACHE) nvcc
endif # GGML_MUSA
NVCC = $(CCACHE) nvcc
endif # GGML_CUDA_NVCC
ifdef CUDA_DOCKER_ARCH
@ -661,10 +635,6 @@ else ifndef CUDA_POWER_ARCH
MK_NVCCFLAGS += -arch=native
endif # CUDA_DOCKER_ARCH
ifdef GGML_CUDA_FORCE_DMMV
MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
endif # GGML_CUDA_FORCE_DMMV
ifdef GGML_CUDA_FORCE_MMQ
MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
endif # GGML_CUDA_FORCE_MMQ
@ -673,20 +643,6 @@ ifdef GGML_CUDA_FORCE_CUBLAS
MK_NVCCFLAGS += -DGGML_CUDA_FORCE_CUBLAS
endif # GGML_CUDA_FORCE_CUBLAS
ifdef GGML_CUDA_DMMV_X
MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
else
MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
endif # GGML_CUDA_DMMV_X
ifdef GGML_CUDA_MMV_Y
MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
else ifdef GGML_CUDA_DMMV_Y
MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility
else
MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
endif # GGML_CUDA_MMV_Y
ifdef GGML_CUDA_F16
MK_NVCCFLAGS += -DGGML_CUDA_F16
endif # GGML_CUDA_F16
@ -695,12 +651,6 @@ ifdef GGML_CUDA_DMMV_F16
MK_NVCCFLAGS += -DGGML_CUDA_F16
endif # GGML_CUDA_DMMV_F16
ifdef GGML_CUDA_KQUANTS_ITER
MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
else
MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
endif
ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
else
@ -724,15 +674,9 @@ define NVCC_COMPILE
$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
endef # NVCC_COMPILE
else
ifdef GGML_MUSA
define NVCC_COMPILE
$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
endef # NVCC_COMPILE
else
define NVCC_COMPILE
$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
endef # NVCC_COMPILE
endif # GGML_MUSA
endif # JETSON_EOL_MODULE_DETECT
ggml/src/ggml-cuda/%.o: \
@ -742,8 +686,8 @@ ggml/src/ggml-cuda/%.o: \
ggml/src/ggml-cuda/common.cuh
$(NVCC_COMPILE)
ggml/src/ggml-cuda.o: \
ggml/src/ggml-cuda.cu \
ggml/src/ggml-cuda/ggml-cuda.o: \
ggml/src/ggml-cuda/ggml-cuda.cu \
ggml/include/ggml-cuda.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h \
@ -754,9 +698,9 @@ ggml/src/ggml-cuda.o: \
endif # GGML_CUDA
ifdef GGML_VULKAN
MK_CPPFLAGS += -DGGML_USE_VULKAN
MK_LDFLAGS += $(shell pkg-config --libs vulkan)
OBJ_GGML += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
MK_CPPFLAGS += -DGGML_USE_VULKAN
MK_LDFLAGS += $(shell pkg-config --libs vulkan)
OBJ_GGML_EXT += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
ifdef GGML_VULKAN_CHECK_RESULTS
MK_CPPFLAGS += -DGGML_VULKAN_CHECK_RESULTS
@ -815,11 +759,7 @@ ifdef GGML_HIPBLAS
AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
endif
GGML_CUDA_DMMV_X ?= 32
GGML_CUDA_MMV_Y ?= 1
GGML_CUDA_KQUANTS_ITER ?= 2
MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
ifdef GGML_HIP_UMA
MK_CPPFLAGS += -DGGML_HIP_UMA
@ -832,13 +772,6 @@ endif # GGML_HIP_UMA
HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
ifdef GGML_CUDA_FORCE_DMMV
HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
endif # GGML_CUDA_FORCE_DMMV
ifdef GGML_CUDA_FORCE_MMQ
HIPFLAGS += -DGGML_CUDA_FORCE_MMQ
@ -852,12 +785,12 @@ ifdef GGML_CUDA_NO_PEER_COPY
HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
endif # GGML_CUDA_NO_PEER_COPY
OBJ_GGML += ggml/src/ggml-cuda.o
OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML += $(OBJ_CUDA_TMPL)
OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
ggml/src/ggml-cuda.o: \
ggml/src/ggml-cuda.cu \
ggml/src/ggml-cuda/ggml-cuda.o: \
ggml/src/ggml-cuda/ggml-cuda.cu \
ggml/include/ggml-cuda.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h \
@ -874,10 +807,89 @@ ggml/src/ggml-cuda/%.o: \
$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
endif # GGML_HIPBLAS
ifdef GGML_MUSA
ifeq ($(wildcard /opt/musa),)
MUSA_PATH ?= /usr/local/musa
else
MUSA_PATH ?= /opt/musa
endif
MTGPU_TARGETS ?= mp_21 mp_22
MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib
MK_LDFLAGS += -lmusa -lmusart -lmublas
ifndef GGML_NO_OPENMP
# For Ubuntu Focal
MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
MK_LDFLAGS += -L/usr/lib/llvm-10/lib
# For Ubuntu Jammy
MK_CPPFLAGS += -I/usr/lib/llvm-14/lib/clang/14.0.0/include
MK_LDFLAGS += -L/usr/lib/llvm-14/lib
endif # GGML_NO_OPENMP
CC := $(MUSA_PATH)/bin/clang
CXX := $(MUSA_PATH)/bin/clang++
MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc
MUSAFLAGS += $(addprefix --cuda-gpu-arch=, $(MTGPU_TARGETS))
ifdef GGML_CUDA_FORCE_MMQ
MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
endif # GGML_CUDA_FORCE_MMQ
ifdef GGML_CUDA_FORCE_CUBLAS
MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
endif # GGML_CUDA_FORCE_CUBLAS
ifdef GGML_CUDA_F16
MUSAFLAGS += -DGGML_CUDA_F16
endif # GGML_CUDA_F16
ifdef GGML_CUDA_DMMV_F16
MUSAFLAGS += -DGGML_CUDA_F16
endif # GGML_CUDA_DMMV_F16
ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
else
MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
ifdef GGML_CUDA_NO_PEER_COPY
MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY
endif # GGML_CUDA_NO_PEER_COPY
ifdef GGML_CUDA_FA_ALL_QUANTS
MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
endif # GGML_CUDA_FA_ALL_QUANTS
OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
ggml/src/ggml-cuda/ggml-cuda.o: \
ggml/src/ggml-cuda/ggml-cuda.cu \
ggml/include/ggml-cuda.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h \
ggml/src/ggml-backend-impl.h \
ggml/src/ggml-common.h \
$(wildcard ggml/src/ggml-cuda/*.cuh)
$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
ggml/src/ggml-cuda/%.o: \
ggml/src/ggml-cuda/%.cu \
ggml/include/ggml.h \
ggml/src/ggml-common.h \
ggml/src/ggml-cuda/common.cuh
$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
endif # GGML_MUSA
ifdef GGML_METAL
MK_CPPFLAGS += -DGGML_USE_METAL
MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
OBJ_GGML += ggml/src/ggml-metal.o
MK_CPPFLAGS += -DGGML_USE_METAL
MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
OBJ_GGML_EXT += ggml/src/ggml-metal/ggml-metal.o
ifdef GGML_METAL_USE_BF16
MK_CPPFLAGS += -DGGML_METAL_USE_BF16
@ -886,62 +898,76 @@ ifdef GGML_METAL_NDEBUG
MK_CPPFLAGS += -DGGML_METAL_NDEBUG
endif
ifdef GGML_METAL_EMBED_LIBRARY
MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
OBJ_GGML += ggml/src/ggml-metal-embed.o
MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
OBJ_GGML_EXT += ggml/src/ggml-metal-embed.o
endif
endif # GGML_METAL
ifdef GGML_METAL
ggml/src/ggml-metal.o: \
ggml/src/ggml-metal.m \
ggml/src/ggml-metal/ggml-metal.o: \
ggml/src/ggml-metal/ggml-metal.m \
ggml/src/ggml-metal/ggml-metal-impl.h \
ggml/include/ggml-metal.h \
ggml/include/ggml.h
$(CC) $(CFLAGS) -c $< -o $@
ifdef GGML_METAL_EMBED_LIBRARY
ggml/src/ggml-metal-embed.o: \
ggml/src/ggml-metal.metal \
ggml/src/ggml-metal/ggml-metal.metal \
ggml/src/ggml-metal/ggml-metal-impl.h \
ggml/src/ggml-common.h
@echo "Embedding Metal library"
@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
@sed -e '/__embed_ggml-common.h__/r ggml/src/ggml-common.h' -e '/__embed_ggml-common.h__/d' < ggml/src/ggml-metal/ggml-metal.metal > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
@rmdir ${TEMP_ASSEMBLY}
endif
endif # GGML_METAL
OBJ_GGML += \
ggml/src/ggml.o \
ggml/src/ggml-cpu.o \
ggml/src/ggml-alloc.o \
ggml/src/ggml-backend.o \
ggml/src/ggml-quants.o \
ggml/src/ggml-aarch64.o
DIR_GGML = ggml
DIR_LLAMA = src
DIR_COMMON = common
OBJ_GGML = \
$(DIR_GGML)/src/ggml.o \
$(DIR_GGML)/src/ggml-aarch64.o \
$(DIR_GGML)/src/ggml-alloc.o \
$(DIR_GGML)/src/ggml-backend.o \
$(DIR_GGML)/src/ggml-backend-reg.o \
$(DIR_GGML)/src/ggml-opt.o \
$(DIR_GGML)/src/ggml-quants.o \
$(DIR_GGML)/src/ggml-threading.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
$(OBJ_GGML_EXT)
OBJ_LLAMA = \
src/llama.o \
src/llama-vocab.o \
src/llama-grammar.o \
src/llama-sampling.o \
src/unicode.o \
src/unicode-data.o
$(DIR_LLAMA)/llama.o \
$(DIR_LLAMA)/llama-vocab.o \
$(DIR_LLAMA)/llama-grammar.o \
$(DIR_LLAMA)/llama-sampling.o \
$(DIR_LLAMA)/unicode.o \
$(DIR_LLAMA)/unicode-data.o
OBJ_COMMON = \
common/common.o \
common/arg.o \
common/log.o \
common/console.o \
common/ngram-cache.o \
common/sampling.o \
common/build-info.o \
common/json-schema-to-grammar.o
$(DIR_COMMON)/common.o \
$(DIR_COMMON)/arg.o \
$(DIR_COMMON)/log.o \
$(DIR_COMMON)/console.o \
$(DIR_COMMON)/ngram-cache.o \
$(DIR_COMMON)/sampling.o \
$(DIR_COMMON)/build-info.o \
$(DIR_COMMON)/json-schema-to-grammar.o
OBJ_ALL = $(OBJ_GGML) $(OBJ_LLAMA) $(OBJ_COMMON)
@ -997,7 +1023,6 @@ $(info I CXX: $(shell $(CXX) --version | head -n 1))
ifdef GGML_CUDA
$(info I NVCC: $(shell $(NVCC) --version | tail -n 1))
CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
ifndef GGML_MUSA
ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
ifndef CUDA_DOCKER_ARCH
@ -1007,7 +1032,6 @@ endif # CUDA_POWER_ARCH
endif # CUDA_DOCKER_ARCH
endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
endif # GGML_MUSA
endif # GGML_CUDA
$(info )
@ -1044,224 +1068,78 @@ endif
# Build libraries
#
# ggml
# Libraries
LIB_GGML = libggml.so
LIB_GGML_S = libggml.a
ggml/src/ggml.o: \
ggml/src/ggml.c \
ggml/include/ggml.h
$(CC) $(CFLAGS) -c $< -o $@
LIB_LLAMA = libllama.so
LIB_LLAMA_S = libllama.a
ggml/src/ggml-cpu.o: \
ggml/src/ggml-cpu.c \
ggml/include/ggml.h \
ggml/src/ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@
LIB_COMMON = libcommon.so
LIB_COMMON_S = libcommon.a
ggml/src/ggml-alloc.o: \
ggml/src/ggml-alloc.c \
ggml/include/ggml.h \
ggml/include/ggml-alloc.h
$(CC) $(CFLAGS) -c $< -o $@
# Targets
BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
ggml/src/ggml-backend.o: \
ggml/src/ggml-backend.cpp \
ggml/src/ggml-backend-impl.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h
$(CXX) $(CXXFLAGS) -c $< -o $@
# Dependency files
DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
ggml/src/ggml-quants.o: \
ggml/src/ggml-quants.c \
ggml/include/ggml.h \
ggml/src/ggml-quants.h \
ggml/src/ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@
# Default target
all: $(BUILD_TARGETS)
ggml/src/ggml-aarch64.o: \
ggml/src/ggml-aarch64.c \
ggml/include/ggml.h \
ggml/src/ggml-aarch64.h \
ggml/src/ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@
ggml/src/ggml-blas.o: \
ggml/src/ggml-blas.cpp \
ggml/include/ggml-blas.h
$(CXX) $(CXXFLAGS) -c $< -o $@
ifndef GGML_NO_LLAMAFILE
ggml/src/llamafile/sgemm.o: \
ggml/src/llamafile/sgemm.cpp \
ggml/src/llamafile/sgemm.h \
ggml/include/ggml.h
$(CXX) $(CXXFLAGS) -c $< -o $@
endif # GGML_NO_LLAMAFILE
ifndef GGML_NO_AMX
ggml/src/ggml-amx.o: \
ggml/src/ggml-amx.cpp \
ggml/include/ggml-amx.h
$(CXX) $(CXXFLAGS) -c $< -o $@
ggml/src/ggml-amx/mmq.o: \
ggml/src/ggml-amx/mmq.cpp \
ggml/src/ggml-amx/mmq.h \
ggml/include/ggml.h
$(CXX) $(CXXFLAGS) -c $< -o $@
endif
ifdef GGML_RPC
ggml/src/ggml-rpc.o: \
ggml/src/ggml-rpc.cpp \
ggml/include/ggml-rpc.h
$(CXX) $(CXXFLAGS) -c $< -o $@
endif # GGML_RPC
$(LIB_GGML): \
$(OBJ_GGML)
$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
$(LIB_GGML_S): \
$(OBJ_GGML)
ar rcs $(LIB_GGML_S) $^
# llama
src/unicode.o: \
src/unicode.cpp \
src/unicode.h
$(CXX) $(CXXFLAGS) -c $< -o $@
src/unicode-data.o: \
src/unicode-data.cpp \
src/unicode-data.h
$(CXX) $(CXXFLAGS) -c $< -o $@
src/llama.o: \
src/llama.cpp \
src/llama-impl.h \
src/llama-vocab.h \
src/llama-grammar.h \
src/llama-sampling.h \
src/unicode.h \
include/llama.h \
ggml/include/ggml-cuda.h \
ggml/include/ggml-metal.h \
# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
# g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \
ggml/src/ggml-cpu/ggml-cpu.cpp \
ggml/include/ggml-backend.h \
ggml/include/ggml.h \
ggml/include/ggml-alloc.h \
ggml/include/ggml-backend.h
$(CXX) $(CXXFLAGS) -c $< -o $@
ggml/src/ggml-backend-impl.h \
ggml/include/ggml-cpu.h \
ggml/src/ggml-impl.h
$(CXX) $(CXXFLAGS) -c $< -o $@
src/llama-vocab.o: \
src/llama-vocab.cpp \
src/llama-vocab.h \
src/llama-impl.h \
include/llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@
# Rules for building object files
$(DIR_GGML)/%.o: $(DIR_GGML)/%.c
$(CC) $(CFLAGS) -MMD -c $< -o $@
src/llama-grammar.o: \
src/llama-grammar.cpp \
src/llama-grammar.h \
src/llama-impl.h \
src/llama-vocab.h \
src/llama-sampling.h \
include/llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@
$(DIR_GGML)/%.o: $(DIR_GGML)/%.cpp
$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
src/llama-sampling.o: \
src/llama-sampling.cpp \
src/llama-sampling.h \
src/llama-impl.h \
include/llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@
$(DIR_LLAMA)/%.o: $(DIR_LLAMA)/%.cpp
$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
$(LIB_LLAMA): \
$(OBJ_LLAMA) \
$(LIB_GGML)
$(DIR_COMMON)/%.o: $(DIR_COMMON)/%.cpp
$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
# Rules for building libraries
$(LIB_GGML): $(OBJ_GGML)
$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
$(LIB_LLAMA_S): \
$(OBJ_LLAMA)
$(LIB_GGML_S): $(OBJ_GGML)
ar rcs $(LIB_GGML_S) $^
$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
$(LIB_LLAMA_S): $(OBJ_LLAMA)
ar rcs $(LIB_LLAMA_S) $^
# common
common/common.o: \
common/common.cpp \
common/common.h \
common/console.h \
common/sampling.h \
common/json.hpp \
common/json-schema-to-grammar.h \
include/llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/arg.o: \
common/arg.cpp \
common/arg.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/log.o: \
common/log.cpp \
common/log.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/sampling.o: \
common/sampling.cpp \
common/sampling.h \
include/llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/console.o: \
common/console.cpp \
common/console.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/json-schema-to-grammar.o: \
common/json-schema-to-grammar.cpp \
common/json-schema-to-grammar.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/ngram-cache.o: \
common/ngram-cache.cpp \
common/ngram-cache.h
$(CXX) $(CXXFLAGS) -c $< -o $@
$(LIB_COMMON): \
$(OBJ_COMMON) \
$(LIB_LLAMA) \
$(LIB_GGML)
$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML)
$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
$(LIB_COMMON_S): \
$(OBJ_COMMON)
$(LIB_COMMON_S): $(OBJ_COMMON)
ar rcs $(LIB_COMMON_S) $^
# Include dependency files
-include $(DEP_FILES)
# Clean rule
clean:
rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
rm -rvf src/*.o
rm -rvf tests/*.o
rm -rvf examples/*.o
rm -rvf common/*.o
rm -rvf *.a
rm -rvf *.dll
rm -rvf *.so
rm -rvf *.dot
rm -rvf ggml/*.a
rm -rvf ggml/*.dll
rm -rvf ggml/*.so
rm -vrf ggml/src/*.o
rm -rvf ggml/src/llamafile/*.o
rm -rvf common/build-info.cpp
rm -vrf ggml/src/ggml-metal-embed.metal
rm -vrf ggml/src/ggml-cuda/*.o
rm -vrf ggml/src/ggml-cuda/template-instances/*.o
rm -vrf ggml/src/ggml-amx/*.o
rm -rvf $(BUILD_TARGETS)
rm -rvf $(TEST_TARGETS)
rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
rm -rvf $(LEGACY_TARGETS_CLEAN)
find examples pocs -type f -name "*.o" -delete
rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS)
rm -rvf *.a *.dll *.so *.dot
find ggml src common tests examples pocs -type f -name "*.o" -delete
find ggml src common tests examples pocs -type f -name "*.d" -delete
#
# Examples
@ -1567,11 +1445,6 @@ tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
tests/test-grad0: tests/test-grad0.cpp \
$(OBJ_GGML)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
tests/test-opt: tests/test-opt.cpp \
$(OBJ_GGML)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

15
Package.swift
View file

@ -10,11 +10,16 @@ var sources = [
"src/unicode.cpp",
"src/unicode-data.cpp",
"ggml/src/ggml.c",
"ggml/src/ggml-cpu.c",
"ggml/src/ggml-aarch64.c",
"ggml/src/ggml-alloc.c",
"ggml/src/ggml-backend.cpp",
"ggml/src/ggml-backend-reg.cpp",
"ggml/src/ggml-cpu/ggml-cpu.c",
"ggml/src/ggml-cpu/ggml-cpu.cpp",
"ggml/src/ggml-cpu/ggml-cpu-aarch64.c",
"ggml/src/ggml-cpu/ggml-cpu-quants.c",
"ggml/src/ggml-threading.cpp",
"ggml/src/ggml-quants.c",
"ggml/src/ggml-aarch64.c",
]
var resources: [Resource] = []
@ -22,6 +27,7 @@ var linkerSettings: [LinkerSetting] = []
var cSettings: [CSetting] = [
.unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
.unsafeFlags(["-fno-objc-arc"]),
.headerSearchPath("ggml/src"),
// NOTE: NEW_LAPACK will required iOS version 16.4+
// We should consider add this in the future when we drop support for iOS 14
// (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
@ -30,8 +36,9 @@ var cSettings: [CSetting] = [
]
#if canImport(Darwin)
sources.append("ggml/src/ggml-metal.m")
resources.append(.process("ggml/src/ggml-metal.metal"))
sources.append("ggml/src/ggml-common.h")
sources.append("ggml/src/ggml-metal/ggml-metal.m")
resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal"))
linkerSettings.append(.linkedFramework("Accelerate"))
cSettings.append(
contentsOf: [

4
README.md
View file

@ -459,14 +459,14 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
## Other documentations
## Other documentation
- [main (cli)](./examples/main/README.md)
- [server](./examples/server/README.md)
- [jeopardy](./examples/jeopardy/README.md)
- [GBNF grammars](./grammars/README.md)
**Development documentations**
**Development documentation**
- [How to build](./docs/build.md)
- [Running on Docker](./docs/docker.md)

2
cmake/llama-config.cmake.in
View file

@ -6,7 +6,7 @@ set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(GGML_BLAS @GGML_BLAS@)
set(GGML_CUDA @GGML_CUDA@)
set(GGML_METAL @GGML_METAL@)
set(GGML_HIPBLAS @GGML_HIPBLAS@)
set(GGML_HIP @GGML_HIP@)
set(GGML_ACCELERATE @GGML_ACCELERATE@)
set(GGML_VULKAN @GGML_VULKAN@)
set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)

11
common/arg.cpp
View file

@ -1939,17 +1939,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.simple_io = true;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
add_opt(common_arg(
{"-ld", "--logdir"}, "LOGDIR",
"path under which to save YAML logs (no logging if unset)",
[](common_params & params, const std::string & value) {
params.logdir = value;
if (params.logdir.back() != DIRECTORY_SEPARATOR) {
params.logdir += DIRECTORY_SEPARATOR;
}
}
));
add_opt(common_arg(
{"--positive-file"}, "FNAME",
string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()),

215
common/common.cpp
View file

@ -1890,218 +1890,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
return result;
}
//
// YAML utils
//
void yaml_dump_vector_float(FILE * stream, const char * prop_name, const std::vector<float> & data) {
if (data.empty()) {
fprintf(stream, "%s:\n", prop_name);
return;
}
fprintf(stream, "%s: [", prop_name);
for (size_t i = 0; i < data.size() - 1; ++i) {
fprintf(stream, "%e, ", data[i]);
}
fprintf(stream, "%e]\n", data.back());
}
void yaml_dump_vector_int(FILE * stream, const char * prop_name, const std::vector<int> & data) {
if (data.empty()) {
fprintf(stream, "%s:\n", prop_name);
return;
}
fprintf(stream, "%s: [", prop_name);
for (size_t i = 0; i < data.size() - 1; ++i) {
fprintf(stream, "%d, ", data[i]);
}
fprintf(stream, "%d]\n", data.back());
}
void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data) {
std::string data_str(data == NULL ? "" : data);
if (data_str.empty()) {
fprintf(stream, "%s:\n", prop_name);
return;
}
size_t pos_start = 0;
size_t pos_found = 0;
if (std::isspace(data_str[0]) || std::isspace(data_str.back())) {
data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
data_str = std::regex_replace(data_str, std::regex(R"(\\[^n"])"), R"(\$&)");
data_str = "\"" + data_str + "\"";
fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
return;
}
if (data_str.find('\n') == std::string::npos) {
fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
return;
}
fprintf(stream, "%s: |\n", prop_name);
while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
fprintf(stream, " %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
pos_start = pos_found + 1;
}
}
void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
ggml_cpu_init(); // some ARM features are detected at runtime
const auto & sparams = params.sparams;
fprintf(stream, "build_commit: %s\n", LLAMA_COMMIT);
fprintf(stream, "build_number: %d\n", LLAMA_BUILD_NUMBER);
fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false");
fprintf(stream, "cpu_has_avx_vnni: %s\n", ggml_cpu_has_avx_vnni() ? "true" : "false");
fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false");
fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
fprintf(stream, "cpu_has_cuda: %s\n", ggml_cpu_has_cuda() ? "true" : "false");
fprintf(stream, "cpu_has_vulkan: %s\n", ggml_cpu_has_vulkan() ? "true" : "false");
fprintf(stream, "cpu_has_kompute: %s\n", ggml_cpu_has_kompute() ? "true" : "false");
fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
fprintf(stream, "cpu_has_sve: %s\n", ggml_cpu_has_sve() ? "true" : "false");
fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
fprintf(stream, "cpu_has_riscv_v: %s\n", ggml_cpu_has_riscv_v() ? "true" : "false");
fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");
#ifdef NDEBUG
fprintf(stream, "debug: false\n");
#else
fprintf(stream, "debug: true\n");
#endif // NDEBUG
fprintf(stream, "model_desc: %s\n", model_desc);
fprintf(stream, "n_vocab: %d # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
#ifdef __OPTIMIZE__
fprintf(stream, "optimize: true\n");
#else
fprintf(stream, "optimize: false\n");
#endif // __OPTIMIZE__
fprintf(stream, "time: %s\n", timestamp.c_str());
fprintf(stream, "\n");
fprintf(stream, "###############\n");
fprintf(stream, "# User Inputs #\n");
fprintf(stream, "###############\n");
fprintf(stream, "\n");
fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
fprintf(stream, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
yaml_dump_string_multiline(stream, "grammar", sparams.grammar.c_str());
fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
fprintf(stream, "ignore_eos: %s # default: false\n", sparams.ignore_eos ? "true" : "false");
yaml_dump_string_multiline(stream, "in_prefix", params.input_prefix.c_str());
fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
yaml_dump_string_multiline(stream, "in_suffix", params.input_prefix.c_str());
fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
fprintf(stream, "logit_bias:\n");
for (const auto & logit_bias : sparams.logit_bias) {
fprintf(stream, " %d: %f", logit_bias.token, logit_bias.bias);
}
fprintf(stream, "lora:\n");
for (auto & la : params.lora_adapters) {
if (la.scale == 1.0f) {
fprintf(stream, " - %s\n", la.path.c_str());
}
}
fprintf(stream, "lora_scaled:\n");
for (auto & la : params.lora_adapters) {
if (la.scale != 1.0f) {
fprintf(stream, " - %s: %f\n", la.path.c_str(), la.scale);
}
}
fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
fprintf(stream, "model: %s # default: %s\n", params.model.c_str(), DEFAULT_MODEL_PATH);
fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
fprintf(stream, "penalize_nl: %s # default: false\n", sparams.penalize_nl ? "true" : "false");
fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
yaml_dump_string_multiline(stream, "prompt", params.prompt.c_str());
fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
yaml_dump_vector_int(stream, "prompt_tokens", prompt_tokens);
fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
fprintf(stream, "reverse_prompt:\n");
for (std::string ap : params.antiprompt) {
size_t pos = 0;
while ((pos = ap.find('\n', pos)) != std::string::npos) {
ap.replace(pos, 1, "\\n");
pos += 1;
}
fprintf(stream, " - %s\n", ap.c_str());
}
fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false");
fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
fprintf(stream, "xtc_probability: %f # default: 0.0\n", sparams.xtc_probability);
fprintf(stream, "xtc_threshold: %f # default: 0.1\n", sparams.xtc_threshold);
fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
}

13
common/common.h
View file

@ -209,7 +209,6 @@ struct common_params {
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT
std::string input_prefix = ""; // string to prefix user inputs with // NOLINT
std::string input_suffix = ""; // string to suffix user inputs with // NOLINT
std::string logdir = ""; // directory in which to save YAML log files // NOLINT
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT
std::string logits_file = ""; // file for saving *all* logits // NOLINT
@ -584,15 +583,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
static const char * const LLM_KV_SPLIT_NO = "split.no";
static const char * const LLM_KV_SPLIT_COUNT = "split.count";
static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
//
// YAML utils
//
void yaml_dump_vector_float (FILE * stream, const char * prop_name, const std::vector<float> & data);
void yaml_dump_vector_int (FILE * stream, const char * prop_name, const std::vector<int> & data);
void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
void yaml_dump_non_result_info(
FILE * stream, const common_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

2
docs/backend/SYCL.md
View file

@ -41,6 +41,8 @@ The following release is verified with good quality:
## News
- 2024.11
- Use syclcompat to improve the performance on some platforms. This requires to use oneAPI 2025.0 or newer.
- 2024.8
- Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.

46
docs/build.md
View file

@ -186,13 +186,9 @@ The following compilation options are also available to tweak performance:
| Option | Legal values | Default | Description |
|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models |
| GGML_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
| GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. |
@ -230,7 +226,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build --config Release -- -j 16
```
On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
@ -247,7 +243,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build -- -j 16
```
@ -259,7 +255,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
```bash
set PATH=%HIP_PATH%\bin;%PATH%
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake --build build
```
Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
@ -268,13 +264,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
| Option | Legal values | Default | Description |
|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
### Vulkan
@ -282,9 +271,9 @@ The following compilation options are also available to tweak performance (yes,
#### w64devkit
Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
```sh
@ -302,6 +291,29 @@ EOF
```
Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
#### Git Bash MINGW64
Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
Download and install [`Visual Studio Community Edition`](https://visualstudio.microsoft.com/) and make sure you select `C++`
Download and install [`CMake`](https://cmake.org/download/) with the default settings
Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
Go into your `llama.cpp` directory and right click, select `Open Git Bash Here` and then run the following commands
```
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release
```
Now you can load the model in conversation mode using `Vulkan`
```
build/bin/release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
```
#### MSYS2
Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
```sh
@ -375,7 +387,7 @@ cmake --build build --config release
You can test with:
`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
`./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
```bash

46
examples/infill/infill.cpp
View file

@ -43,50 +43,6 @@ static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false;
static void write_logfile(
const llama_context * ctx, const common_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens
) {
if (params.logdir.empty()) {
return;
}
const std::string timestamp = string_get_sortable_timestamp();
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
LOG_ERR("%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
const std::string logfile_path = params.logdir + timestamp + ".yml";
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
fprintf(logfile, "binary: infill\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "# Generation Results #\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_perf_dump_yaml(logfile, ctx);
fclose(logfile);
}
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
@ -96,7 +52,6 @@ static void sigint_handler(int signo) {
console::cleanup();
LOG("\n");
common_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
// make sure all logs are flushed
LOG("Interrupted by user\n");
@ -625,7 +580,6 @@ int main(int argc, char ** argv) {
LOG("\n");
common_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
llama_free(ctx);
llama_free_model(model);

32
examples/llama-bench/llama-bench.cpp
View file

@ -774,13 +774,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
struct test {
static const std::string build_commit;
static const int build_number;
static const bool cuda;
static const bool vulkan;
static const bool kompute;
static const bool metal;
static const bool sycl;
static const bool gpu_blas;
static const bool blas;
static const std::string cpu_info;
static const std::string gpu_info;
std::string model_filename;
@ -793,7 +786,6 @@ struct test {
std::string cpu_mask;
bool cpu_strict;
int poll;
bool has_rpc;
ggml_type type_k;
ggml_type type_v;
int n_gpu_layers;
@ -822,7 +814,6 @@ struct test {
cpu_mask = inst.cpu_mask;
cpu_strict = inst.cpu_strict;
poll = inst.poll;
has_rpc = !inst.rpc_servers.empty();
type_k = inst.type_k;
type_v = inst.type_v;
n_gpu_layers = inst.n_gpu_layers;
@ -881,8 +872,7 @@ struct test {
static const std::vector<std::string> & get_fields() {
static const std::vector<std::string> fields = {
"build_commit", "build_number",
"cuda", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
"cpu_info", "gpu_info",
"cpu_info", "gpu_info", "backends",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_ubatch",
"n_threads", "cpu_mask", "cpu_strict", "poll",
@ -908,8 +898,7 @@ struct test {
field == "avg_ns" || field == "stddev_ns") {
return INT;
}
if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
if (field == "f16_kv" || field == "no_kv_offload" ||
field == "cpu_strict" ||
field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
return BOOL;
@ -938,9 +927,7 @@ struct test {
}
std::vector<std::string> values = {
build_commit, std::to_string(build_number),
std::to_string(cuda), std::to_string(vulkan), std::to_string(vulkan),
std::to_string(metal), std::to_string(sycl), std::to_string(has_rpc), std::to_string(gpu_blas), std::to_string(blas),
cpu_info, gpu_info,
cpu_info, gpu_info, get_backend(),
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_ubatch),
std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll),
@ -967,13 +954,6 @@ struct test {
const std::string test::build_commit = LLAMA_COMMIT;
const int test::build_number = LLAMA_BUILD_NUMBER;
const bool test::cuda = !!ggml_cpu_has_cuda();
const bool test::vulkan = !!ggml_cpu_has_vulkan();
const bool test::kompute = !!ggml_cpu_has_kompute();
const bool test::metal = !!ggml_cpu_has_metal();
const bool test::gpu_blas = !!ggml_cpu_has_gpublas();
const bool test::blas = !!ggml_cpu_has_blas();
const bool test::sycl = !!ggml_cpu_has_sycl();
const std::string test::cpu_info = get_cpu_info();
const std::string test::gpu_info = get_gpu_info();
@ -1178,7 +1158,8 @@ struct markdown_printer : public printer {
fields.emplace_back("size");
fields.emplace_back("params");
fields.emplace_back("backend");
bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
bool is_cpu_backend = test::get_backend().find("CPU") != std::string::npos ||
test::get_backend().find("BLAS") != std::string::npos;
if (!is_cpu_backend) {
fields.emplace_back("n_gpu_layers");
}
@ -1268,9 +1249,6 @@ struct markdown_printer : public printer {
value = buf;
} else if (field == "backend") {
value = test::get_backend();
if (t.has_rpc) {
value += "+RPC";
}
} else if (field == "test") {
if (t.n_prompt > 0 && t.n_gen == 0) {
snprintf(buf, sizeof(buf), "pp%d", t.n_prompt);

45
examples/main/main.cpp
View file

@ -62,49 +62,6 @@ static bool file_is_empty(const std::string & path) {
return f.tellg() == 0;
}
static void write_logfile(
const llama_context * ctx, const common_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens
) {
if (params.logdir.empty()) {
return;
}
const std::string timestamp = string_get_sortable_timestamp();
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
LOG_ERR("%s: failed to create logdir %s, cannot write logfile\n", __func__, params.logdir.c_str());
return;
}
const std::string logfile_path = params.logdir + timestamp + ".yml";
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
fprintf(logfile, "binary: main\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "# Generation Results #\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_perf_dump_yaml(logfile, ctx);
fclose(logfile);
}
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
@ -115,7 +72,6 @@ static void sigint_handler(int signo) {
console::cleanup();
LOG("\n");
common_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
// make sure all logs are flushed
LOG("Interrupted by user\n");
@ -926,7 +882,6 @@ int main(int argc, char ** argv) {
LOG("\n\n");
common_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
common_sampler_free(smpl);

51
examples/perplexity/perplexity.cpp
View file

@ -34,55 +34,6 @@ struct results_log_softmax {
float prob;
};
static void write_logfile(
const llama_context * ctx, const common_params & params, const llama_model * model,
const struct results_perplexity & results
) {
if (params.logdir.empty()) {
return;
}
if (params.hellaswag) {
LOG_WRN("%s: logging results is not implemented for HellaSwag. No files will be written.\n", __func__);
return;
}
const std::string timestamp = string_get_sortable_timestamp();
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
LOG_WRN("%s: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
const std::string logfile_path = params.logdir + timestamp + ".yml";
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
fprintf(logfile, "binary: main\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "# Perplexity Results #\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
yaml_dump_vector_float(logfile, "logits", results.logits);
fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
yaml_dump_vector_float(logfile, "probs", results.probs);
llama_perf_dump_yaml(logfile, ctx);
fclose(logfile);
}
static std::vector<float> softmax(const std::vector<float>& logits) {
std::vector<float> probs(logits.size());
float max_logit = logits[0];
@ -2072,8 +2023,6 @@ int main(int argc, char ** argv) {
LOG("\n");
llama_perf_context_print(ctx);
write_logfile(ctx, params, model, results);
llama_free(ctx);
llama_free_model(model);

19
examples/quantize-stats/quantize-stats.cpp
View file

@ -142,7 +142,7 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
}
static void test_roundtrip_on_chunk(
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, bool use_reference,
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
) {
if (layer->type == GGML_TYPE_F16) {
@ -156,7 +156,7 @@ static void test_roundtrip_on_chunk(
if (use_reference) {
qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
} else {
qfns.from_float(input_scratch, quantized_scratch, chunk_size);
qfns_cpu.from_float(input_scratch, quantized_scratch, chunk_size);
}
qfns.to_float(quantized_scratch, output_scratch, chunk_size);
@ -166,7 +166,7 @@ static void test_roundtrip_on_chunk(
// Run quantization function for a single layer and update error stats
static void test_roundtrip_on_layer(
std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, bool use_reference,
std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
) {
@ -187,13 +187,13 @@ static void test_roundtrip_on_layer(
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(),
test_roundtrip_on_chunk(layer, 0, nelements, qfns, qfns_cpu, 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,
auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, nelements, layer, use_reference, input_scratch_ptr,
&quantized_scratch, &output_scratch, chunk_size] () {
error_stats local_stats {};
while (true) {
@ -205,7 +205,7 @@ static void test_roundtrip_on_layer(
}
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,
test_roundtrip_on_chunk(layer, offset, chunk, qfns, qfns_cpu, use_reference, input_scratch_ptr + offset,
quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
}
};
@ -371,8 +371,9 @@ int main(int argc, char ** argv) {
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
continue;
}
const auto * qfns = ggml_get_type_traits(type);
if (qfns->from_float && qfns->to_float) {
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
if (qfns_cpu->from_float && qfns->to_float) {
if (params.verbose) {
printf("testing %s ...\n", ggml_type_name(type));
}
@ -393,7 +394,7 @@ int main(int argc, char ** argv) {
test_roundtrip_on_layer(
layer_name,
params.per_layer_stats,
*qfns,
*qfns, *qfns_cpu,
params.reference,
kv_tensor.second,
input_scratch,

1
examples/server/README.md
View file

@ -85,7 +85,6 @@ The project is under active development, and we are [looking for feedback and co
| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
| `--log-disable` | Log disable |
| `--log-file FNAME` | Log to file |
| `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |

81
examples/server/public/index.html
View file

@ -12,7 +12,7 @@
.markdown {
h1, h2, h3, h4, h5, h6, ul, ol, li { all: revert; }
pre {
@apply whitespace-pre-wrap my-4 rounded-lg p-2;
@apply whitespace-pre-wrap rounded-lg p-2;
border: 1px solid currentColor;
}
/* TODO: fix markdown table */
@ -25,8 +25,11 @@
.bg-base-200 {background-color: var(--fallback-b2,oklch(var(--b2)/1))}
.bg-base-300 {background-color: var(--fallback-b3,oklch(var(--b3)/1))}
.text-base-content {color: var(--fallback-bc,oklch(var(--bc)/1))}
.show-on-hover {
@apply opacity-0 group-hover:opacity-100;
}
.btn-mini {
@apply cursor-pointer opacity-0 group-hover:opacity-100 hover:shadow-md;
@apply cursor-pointer hover:shadow-md;
}
.chat-screen { max-width: 900px; }
/* because the default bubble color is quite dark, we will make a custom one using bg-base-300 */
@ -152,14 +155,14 @@
<!-- actions for each message -->
<div :class="{'text-right': msg.role === 'user'}" class="mx-4 mt-2 mb-2">
<!-- user message -->
<button v-if="msg.role === 'user'" class="badge btn-mini" @click="editingMsg = msg" :disabled="isGenerating">
<button v-if="msg.role === 'user'" class="badge btn-minishow-on-hover " @click="editingMsg = msg" :disabled="isGenerating">
✍️ Edit
</button>
<!-- assistant message -->
<button v-if="msg.role === 'assistant'" class="badge btn-mini mr-2" @click="regenerateMsg(msg)" :disabled="isGenerating">
<button v-if="msg.role === 'assistant'" class="badge btn-mini show-on-hover mr-2" @click="regenerateMsg(msg)" :disabled="isGenerating">
🔄 Regenerate
</button>
<button v-if="msg.role === 'assistant'" class="badge btn-mini mr-2" @click="copyMsg(msg)" :disabled="isGenerating">
<button v-if="msg.role === 'assistant'" class="badge btn-mini show-on-hover mr-2" @click="copyMsg(msg)" :disabled="isGenerating">
📋 Copy
</button>
</div>
@ -196,20 +199,24 @@
<h3 class="text-lg font-bold mb-6">Settings</h3>
<div class="h-[calc(90vh-12rem)] overflow-y-auto">
<p class="opacity-40 mb-6">Settings below are saved in browser's localStorage</p>
<settings-modal-short-input :config-key="'apiKey'" :config-default="configDefault" :config-info="configInfo" v-model="config.apiKey"></settings-modal-short-input>
<label class="form-control mb-2">
<div class="label">System Message</div>
<textarea class="textarea textarea-bordered h-24" :placeholder="'Default: ' + configDefault.systemMessage" v-model="config.systemMessage"></textarea>
</label>
<template v-for="configKey in ['temperature', 'top_k', 'top_p', 'min_p', 'max_tokens']">
<settings-modal-numeric-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
<settings-modal-short-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
</template>
<!-- TODO: add more sampling-related configs, please regroup them into different "collapse" sections -->
<!-- Section: Other sampler settings -->
<details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
<summary class="collapse-title font-bold">Other sampler settings</summary>
<div class="collapse-content">
<!-- Samplers queue -->
<settings-modal-short-input label="Samplers queue" :config-key="'samplers'" :config-default="configDefault" :config-info="configInfo" v-model="config.samplers"></settings-modal-short-input>
<!-- Samplers -->
<template v-for="configKey in ['dynatemp_range', 'dynatemp_exponent', 'typical_p', 'xtc_probability', 'xtc_threshold']">
<settings-modal-numeric-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
<settings-modal-short-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
</template>
</div>
</details>
@ -218,7 +225,7 @@
<summary class="collapse-title font-bold">Penalties settings</summary>
<div class="collapse-content">
<template v-for="configKey in ['repeat_last_n', 'repeat_penalty', 'presence_penalty', 'frequency_penalty', 'dry_multiplier', 'dry_base', 'dry_allowed_length', 'dry_penalty_last_n']">
<settings-modal-numeric-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
<settings-modal-short-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
</template>
</div>
</details>
@ -227,6 +234,7 @@
<summary class="collapse-title font-bold">Advanced config</summary>
<div class="collapse-content">
<label class="form-control mb-2">
<!-- Custom parameters input -->
<div class="label inline">Custom JSON config (For more info, refer to <a class="underline" href="https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md" target="_blank" rel="noopener noreferrer">server documentation</a>)</div>
<textarea class="textarea textarea-bordered h-24" placeholder="Example: { &quot;mirostat&quot;: 1, &quot;min_p&quot;: 0.1 }" v-model="config.custom"></textarea>
</label>
@ -245,11 +253,11 @@
</div>
<!-- Template to be used by settings modal -->
<template id="settings-modal-numeric-input">
<template id="settings-modal-short-input">
<label class="input input-bordered join-item grow flex items-center gap-2 mb-2">
<!-- Show help message on hovering on the input label -->
<div class="dropdown dropdown-hover">
<div tabindex="0" role="button" class="font-bold">{{ configKey }}</div>
<div tabindex="0" role="button" class="font-bold">{{ label || configKey }}</div>
<div class="dropdown-content menu bg-base-100 rounded-box z-10 w-64 p-2 shadow mt-4">
{{ configInfo[configKey] || '(no help message available)' }}
</div>
@ -264,9 +272,13 @@
import { createApp, defineComponent, shallowRef, computed, h } from './deps_vue.esm-browser.js';
import { llama } from './completion.js';
// utility functions
const isString = (x) => !!x.toLowerCase;
const isNumeric = (n) => !isString(n) && !isNaN(n);
const escapeAttr = (str) => str.replace(/>/g, '&gt;').replace(/"/g, '&quot;');
const copyStr = (str) => navigator.clipboard.writeText(str);
// constants
const BASE_URL = localStorage.getItem('base') // for debugging
|| (new URL('.', document.baseURI).href).toString(); // for production
const CONFIG_DEFAULT = {
@ -274,6 +286,7 @@
apiKey: '',
systemMessage: 'You are a helpful assistant.',
// make sure these default values are in sync with `common.h`
samplers: 'dkypmxt',
temperature: 0.8,
dynatemp_range: 0.0,
dynatemp_exponent: 1.0,
@ -295,8 +308,9 @@
custom: '', // custom json-stringified object
};
const CONFIG_INFO = {
apiKey: '',
apiKey: 'Set the API Key if you are using --api-key option for the server.',
systemMessage: 'The starting message that defines how model should behave.',
samplers: 'The order at which samplers are applied, in simplified way. Default is "dkypmxt": dry->top_k->typ_p->top_p->min_p->xtc->temperature',
temperature: 'Controls the randomness of the generated text by affecting the probability distribution of the output tokens. Higher = more random, lower = more focused.',
dynatemp_range: 'Addon for the temperature sampler. The added value to the range of dynamic temperature, which adjusts probabilities by entropy of tokens.',
dynatemp_exponent: 'Addon for the temperature sampler. Smoothes out the probability redistribution based on the most probable token.',
@ -325,20 +339,35 @@
// markdown support
const VueMarkdown = defineComponent(
(props) => {
const md = shallowRef(new markdownit(props.options ?? { breaks: true }));
for (const plugin of props.plugins ?? []) {
md.value.use(plugin);
}
const md = shallowRef(new markdownit({ breaks: true }));
const origFenchRenderer = md.value.renderer.rules.fence;
md.value.renderer.rules.fence = (tokens, idx, ...args) => {
const content = tokens[idx].content;
const origRendered = origFenchRenderer(tokens, idx, ...args);
return `<div class="relative my-4">
<div class="text-right sticky top-4 mb-2 mr-2 h-0">
<button class="badge btn-mini" onclick="copyStr(${escapeAttr(JSON.stringify(content))})">📋 Copy</button>
</div>
${origRendered}
</div>`;
};
window.copyStr = copyStr;
const content = computed(() => md.value.render(props.source));
return () => h("div", { innerHTML: content.value });
},
{ props: ["source", "options", "plugins"] }
{ props: ["source"] }
);
// inout field to be used by settings modal
const SettingsModalNumericInput = defineComponent({
template: document.getElementById('settings-modal-numeric-input').innerHTML,
props: ['configKey', 'configDefault', 'configInfo', 'modelValue'],
// input field to be used by settings modal
const SettingsModalShortInput = defineComponent({
template: document.getElementById('settings-modal-short-input').innerHTML,
props: {
label: { type: String, required: false },
configKey: String,
configDefault: Object,
configInfo: Object,
modelValue: [Object, String, Number],
},
});
// coversations is stored in localStorage
@ -390,7 +419,11 @@
if (!conv) return;
const msg = conv.messages.pop();
conv.lastModified = Date.now();
localStorage.setItem(convId, JSON.stringify(conv));
if (conv.messages.length === 0) {
StorageUtils.remove(convId);
} else {
localStorage.setItem(convId, JSON.stringify(conv));
}
return msg;
},
@ -431,7 +464,7 @@
const mainApp = createApp({
components: {
VueMarkdown,
SettingsModalNumericInput,
SettingsModalShortInput,
},
data() {
return {
@ -525,6 +558,7 @@
],
stream: true,
cache_prompt: true,
samplers: this.config.samplers,
temperature: this.config.temperature,
dynatemp_range: this.config.dynatemp_range,
dynatemp_exponent: this.config.dynatemp_exponent,
@ -587,6 +621,7 @@
this.isGenerating = false;
this.stopGeneration = () => {};
this.fetchMessages();
chatScrollToBottom();
},
// message actions
@ -600,7 +635,7 @@
this.generateMessage(currConvId);
},
copyMsg(msg) {
navigator.clipboard.writeText(msg.content);
copyStr(msg.content);
},
editUserMsgAndRegenerate(msg) {
if (this.isGenerating) return;

62
examples/server/server.cpp
View file

@ -102,6 +102,12 @@ struct server_task_result {
bool error;
};
struct server_static_file {
const unsigned char * data;
unsigned int size;
const char * mime_type;
};
struct slot_params {
bool stream = true;
bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
@ -921,14 +927,22 @@ struct server_context {
{
const auto & samplers = data.find("samplers");
if (samplers != data.end() && samplers->is_array()) {
std::vector<std::string> sampler_names;
for (const auto & name : *samplers) {
if (name.is_string()) {
sampler_names.emplace_back(name);
if (samplers != data.end()) {
if (samplers->is_array()) {
std::vector<std::string> sampler_names;
for (const auto & name : *samplers) {
if (name.is_string()) {
sampler_names.emplace_back(name);
}
}
slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
} else if (samplers->is_string()){
std::string sampler_string;
for (const auto & name : *samplers) {
sampler_string += name;
}
slot.sparams.samplers = common_sampler_types_from_chars(sampler_string);
}
slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
} else {
slot.sparams.samplers = default_sparams.samplers;
}
@ -2259,6 +2273,16 @@ int main(int argc, char ** argv) {
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n");
// static files
std::map<std::string, server_static_file> static_files = {
{ "/", { index_html, index_html_len, "text/html; charset=utf-8" }},
{ "/completion.js", { completion_js, completion_js_len, "text/javascript; charset=utf-8" }},
{ "/deps_daisyui.min.css", { deps_daisyui_min_css, deps_daisyui_min_css_len, "text/css; charset=utf-8" }},
{ "/deps_markdown-it.js", { deps_markdown_it_js, deps_markdown_it_js_len, "text/javascript; charset=utf-8" }},
{ "/deps_tailwindcss.js", { deps_tailwindcss_js, deps_tailwindcss_js_len, "text/javascript; charset=utf-8" }},
{ "/deps_vue.esm-browser.js", { deps_vue_esm_browser_js, deps_vue_esm_browser_js_len, "text/javascript; charset=utf-8" }},
};
std::unique_ptr<httplib::Server> svr;
#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
if (params.ssl_file_key != "" && params.ssl_file_cert != "") {
@ -2339,7 +2363,7 @@ int main(int argc, char ** argv) {
// Middlewares
//
auto middleware_validate_api_key = [&params, &res_error](const httplib::Request & req, httplib::Response & res) {
auto middleware_validate_api_key = [&params, &res_error, &static_files](const httplib::Request & req, httplib::Response & res) {
static const std::unordered_set<std::string> public_endpoints = {
"/health",
"/models",
@ -2351,8 +2375,8 @@ int main(int argc, char ** argv) {
return true;
}
// If path is public, skip validation
if (public_endpoints.find(req.path) != public_endpoints.end()) {
// If path is public or is static file, skip validation
if (public_endpoints.find(req.path) != public_endpoints.end() || static_files.find(req.path) != static_files.end()) {
return true;
}
@ -3096,13 +3120,6 @@ int main(int argc, char ** argv) {
res.status = 200; // HTTP OK
};
auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
return false;
};
};
//
// Router
//
@ -3117,12 +3134,13 @@ int main(int argc, char ** argv) {
}
} else {
// using embedded static files
svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
svr->Get("/deps_daisyui.min.css", handle_static_file(deps_daisyui_min_css, deps_daisyui_min_css_len, "text/css; charset=utf-8"));
svr->Get("/deps_markdown-it.js", handle_static_file(deps_markdown_it_js, deps_markdown_it_js_len, "text/javascript; charset=utf-8"));
svr->Get("/deps_tailwindcss.js", handle_static_file(deps_tailwindcss_js, deps_tailwindcss_js_len, "text/javascript; charset=utf-8"));
svr->Get("/deps_vue.esm-browser.js", handle_static_file(deps_vue_esm_browser_js, deps_vue_esm_browser_js_len, "text/javascript; charset=utf-8"));
for (const auto & it : static_files) {
const server_static_file & static_file = it.second;
svr->Get(it.first.c_str(), [&static_file](const httplib::Request &, httplib::Response & res) {
res.set_content(reinterpret_cast<const char*>(static_file.data), static_file.size, static_file.mime_type);
return false;
});
}
}
// register API routes

17
ggml/CMakeLists.txt
View file

@ -92,6 +92,7 @@ else()
endif()
option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF)
option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
option(GGML_AVX "ggml: enable AVX" ${INS_ENB})
option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB})
@ -116,6 +117,7 @@ endif()
# ggml core
set(GGML_SCHED_MAX_COPIES "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
option(GGML_CPU "ggml: enable CPU backend" ON)
# 3rd party libs / backends
option(GGML_ACCELERATE "ggml: enable Accelerate framework" ON)
@ -126,14 +128,9 @@ option(GGML_LLAMAFILE "ggml: use LLAMAFILE"
option(GGML_CUDA "ggml: use CUDA" OFF)
option(GGML_MUSA "ggml: use MUSA" OFF)
option(GGML_CUDA_FORCE_DMMV "ggml: use dmmv instead of mmvq CUDA kernels" OFF)
option(GGML_CUDA_FORCE_MMQ "ggml: use mmq kernels instead of cuBLAS" OFF)
option(GGML_CUDA_FORCE_CUBLAS "ggml: always use cuBLAS instead of mmq kernels" OFF)
set (GGML_CUDA_DMMV_X "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
set (GGML_CUDA_MMV_Y "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
option(GGML_CUDA_F16 "ggml: use 16 bit floats for some calculations" OFF)
set (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
"ggml: iters./thread per block for Q2_K/Q6_K")
set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
"ggml: max. batch size for using peer access")
option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF)
@ -141,7 +138,7 @@ option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM"
option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF)
option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT})
option(GGML_HIPBLAS "ggml: use hipBLAS" OFF)
option(GGML_HIP "ggml: use HIP" OFF)
option(GGML_HIP_UMA "ggml: use HIP unified memory architecture" OFF)
option(GGML_VULKAN "ggml: use Vulkan" OFF)
option(GGML_VULKAN_CHECK_RESULTS "ggml: run Vulkan op checks" OFF)
@ -226,6 +223,7 @@ set(GGML_PUBLIC_HEADERS
include/ggml-cann.h
include/ggml-cuda.h
include/ggml-kompute.h
include/ggml-opt.h
include/ggml-metal.h
include/ggml-rpc.h
include/ggml-sycl.h
@ -238,12 +236,15 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
install(TARGETS ggml PUBLIC_HEADER)
if (BUILD_SHARED_LIBS)
install(TARGETS ggml LIBRARY)
install(TARGETS ggml LIBRARY)
install(TARGETS ggml-base LIBRARY)
endif()
# FIXME: this should be done in the backend cmake files
if (GGML_METAL)
# FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
install(
FILES src/ggml-metal.metal
FILES src/ggml-metal/ggml-metal.metal
PERMISSIONS
OWNER_READ
OWNER_WRITE

10
ggml/include/ggml-amx.h
View file

@ -9,16 +9,16 @@ extern "C" {
#endif
// buffer_type API
GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
// backend API
GGML_API ggml_backend_t ggml_backend_amx_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
#ifdef __cplusplus
}

40
ggml/include/ggml-backend.h
View file

@ -3,6 +3,20 @@
#include "ggml.h"
#include "ggml-alloc.h"
#ifdef GGML_BACKEND_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef GGML_BACKEND_BUILD
# define GGML_BACKEND_API __declspec(dllexport) extern
# else
# define GGML_BACKEND_API __declspec(dllimport) extern
# endif
# else
# define GGML_BACKEND_API __attribute__ ((visibility ("default"))) extern
# endif
#else
# define GGML_BACKEND_API extern
#endif
#ifdef __cplusplus
extern "C" {
#endif
@ -72,7 +86,7 @@ extern "C" {
GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
// "offset" refers to the offset of the tensor data for setting/getting data
// "offset" refers to the offset in tensor->data for setting/getting data
GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
GGML_API void ggml_backend_tensor_memset( struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size);
@ -228,14 +242,20 @@ extern "C" {
ggml_backend_sched_reserve(sched, reserve_graph);
// compute
graph = build_graph(sched);
ggml_backend_sched_graph_compute(sched, graph);
graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
for (int i = 0; i < 10; ++i) {
ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
}
// if there are graph inputs:
ggml_backend_sched_reset(sched);
ggml_backend_sched_alloc_graph(sched, graph);
ggml_backend_tensor_set(input_tensor, ...);
ggml_backend_sched_graph_compute(sched, graph);
graph = build_graph(sched); // get a new graph that is not allocated (the metadata for the old graph is freed once ggml_free is called)
ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
ggml_backend_sched_graph_compute(sched, graph); // execute the graph
// as an alternative to the above it is also possible to assign the inputs to a dedicated context and
// allocate them statically via ggml_backend_alloc_ctx_tensors
}
*/
@ -250,7 +270,7 @@ extern "C" {
//
typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
// Initialize a backend scheduler
// Initialize a backend scheduler, backends with low index are given priority over backends with high index
GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
@ -275,7 +295,9 @@ extern "C" {
GGML_API enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
GGML_API void ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
// Reset all assignments and allocators - must be called before changing the node backends
// Reset all assignments and allocators - must be called before changing the node backends or allocating a new graph.
// This in effect deallocates all tensors that were previously allocated and leaves them with dangling pointers.
// The correct way to use this API is to discard the deallocated tensors and create new ones.
GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched);
// Set a callback to be called for each resulting node during graph compute

8
ggml/include/ggml-blas.h
View file

@ -9,15 +9,15 @@ extern "C" {
#endif
// backend API
GGML_API ggml_backend_t ggml_backend_blas_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
// number of threads used for conversion to float
// for openblas and blis, this will also set the number of threads used for blas operations
GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
GGML_API ggml_backend_reg_t ggml_backend_blas_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
#ifdef __cplusplus

16
ggml/include/ggml-cann.h
View file

@ -34,7 +34,7 @@ extern "C" {
*/
#define GGML_CANN_MAX_DEVICES 16
GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
/**
* @brief Initializes the CANN backend for a specified device.
@ -46,7 +46,7 @@ GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
* @param device The index of the device to initialize.
* @return A pointer to the initialized backend instance, or nullptr on failure.
*/
GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
GGML_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
/**
* @brief Checks if a given backend is a CANN backend.
@ -57,7 +57,7 @@ GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
* @param backend The backend instance to check.
* @return True if the backend is a CANN backend, false otherwise.
*/
GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
/**
* @brief Retrieves the CANN buffer type for a specified device.
@ -69,7 +69,7 @@ GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
* @return A pointer to the buffer type interface for the specified device, or
* nullptr if the device index is out of range.
*/
GGML_API ggml_backend_buffer_type_t
GGML_BACKEND_API ggml_backend_buffer_type_t
ggml_backend_cann_buffer_type(int32_t device);
/**
@ -80,14 +80,14 @@ ggml_backend_cann_buffer_type(int32_t device);
*
* @return The number of CANN devices available.
*/
GGML_API int32_t ggml_backend_cann_get_device_count(void);
GGML_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
/**
* @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
*
* @return A pointer to the host buffer type interface.
*/
GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
/**
* @brief Retrieves the description of a specific CANN device.
@ -99,7 +99,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
* @param description Pointer to a buffer where the description will be written.
* @param description_size Size of the description buffer.
*/
GGML_API void ggml_backend_cann_get_device_description(
GGML_BACKEND_API void ggml_backend_cann_get_device_description(
int32_t device, char* description, size_t description_size);
/**
@ -114,7 +114,7 @@ GGML_API void ggml_backend_cann_get_device_description(
* @param total Pointer to a variable where the total memory size will be
* stored.
*/
GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
size_t* free,
size_t* total);

107
ggml/include/ggml-cpu.h
View file

@ -54,54 +54,77 @@ extern "C" {
GGML_NUMA_STRATEGY_COUNT
};
GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
GGML_BACKEND_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
GGML_BACKEND_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
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_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
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_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
GGML_BACKEND_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_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
GGML_BACKEND_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_BACKEND_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, 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_BACKEND_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
GGML_BACKEND_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
GGML_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
GGML_BACKEND_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_BACKEND_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads);
GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
GGML_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
GGML_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
GGML_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
GGML_BACKEND_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads);
GGML_BACKEND_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
GGML_BACKEND_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
GGML_BACKEND_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
GGML_BACKEND_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
GGML_BACKEND_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
GGML_BACKEND_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan(
GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
const struct ggml_cgraph * cgraph,
int n_threads, /* = GGML_DEFAULT_N_THREADS */
struct ggml_threadpool * threadpool /* = NULL */ );
GGML_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
GGML_BACKEND_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
GGML_BACKEND_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
// TODO: move to backend interface
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);
GGML_API int ggml_cpu_has_matmul_int8(void);
// get the sve vector length in bytes
GGML_API int ggml_cpu_get_sve_cnt(void);
//
// system info
//
// x86
GGML_BACKEND_API int ggml_cpu_has_sse3 (void);
GGML_BACKEND_API int ggml_cpu_has_ssse3 (void);
GGML_BACKEND_API int ggml_cpu_has_avx (void);
GGML_BACKEND_API int ggml_cpu_has_avx2 (void);
GGML_BACKEND_API int ggml_cpu_has_f16c (void);
GGML_BACKEND_API int ggml_cpu_has_fma (void);
GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void);
GGML_BACKEND_API int ggml_cpu_has_avx512 (void);
GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
GGML_BACKEND_API int ggml_cpu_has_amx_int8 (void);
// ARM
GGML_BACKEND_API int ggml_cpu_has_neon (void);
GGML_BACKEND_API int ggml_cpu_has_arm_fma (void);
GGML_BACKEND_API int ggml_cpu_has_fp16_va (void);
GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
GGML_BACKEND_API int ggml_cpu_has_sve (void);
GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes
// other
GGML_BACKEND_API int ggml_cpu_has_riscv_v (void);
GGML_BACKEND_API int ggml_cpu_has_vsx (void);
GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void);
GGML_BACKEND_API int ggml_cpu_has_llamafile (void);
// Internal types and functions exposed for tests and benchmarks
@ -115,6 +138,7 @@ extern "C" {
const void * GGML_RESTRICT y, int nr, int nc);
struct ggml_type_traits_cpu {
ggml_from_float_t from_float;
ggml_from_float_to_mat_t from_float_to_mat;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
@ -124,27 +148,30 @@ extern "C" {
ggml_gemm_t gemm;
};
GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
GGML_API void ggml_cpu_init(void);
GGML_BACKEND_API void ggml_cpu_init(void);
//
// CPU backend
//
GGML_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_API bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
GGML_BACKEND_API bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_BACKEND_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_BACKEND_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
GGML_BACKEND_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft);
#ifdef __cplusplus
}
#endif

24
ggml/include/ggml-cuda.h
View file

@ -7,7 +7,7 @@
extern "C" {
#endif
#ifdef GGML_USE_HIPBLAS
#ifdef GGML_USE_HIP
#define GGML_CUDA_NAME "ROCm"
#define GGML_CUBLAS_NAME "hipBLAS"
#elif defined(GGML_USE_MUSA)
@ -20,27 +20,27 @@ extern "C" {
#define GGML_CUDA_MAX_DEVICES 16
// backend API
GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
// device buffer
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
GGML_API int ggml_backend_cuda_get_device_count(void);
GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
GGML_BACKEND_API int ggml_backend_cuda_get_device_count(void);
GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
#ifdef __cplusplus
}

8
ggml/include/ggml-kompute.h
View file

@ -37,13 +37,13 @@ struct ggml_vk_device ggml_vk_current_device(void);
// forward declaration
typedef struct ggml_backend * ggml_backend_t;
GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
GGML_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
#ifdef __cplusplus
}

16
ggml/include/ggml-metal.h
View file

@ -39,27 +39,27 @@ extern "C" {
// user-code should use only these functions
//
GGML_API ggml_backend_t ggml_backend_metal_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_DEPRECATED(
GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
"obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
// helper to check if the device supports a specific family
// ideally, the user code should be doing these checks
// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
#ifdef __cplusplus
}

216
ggml/include/ggml-opt.h Normal file
View file

@ -0,0 +1,216 @@
// This file contains functionality for training models using GGML.
// It is not strictly needed vs. just vanilla GGML but it provides a more high-level interface for common needs such as datasets.
// At the bottom of this file especially there are relatively high-level functions that are suitable use or adaptation in user code.
//
// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_opt_dataset;
struct ggml_opt_context;
struct ggml_opt_result;
typedef struct ggml_opt_dataset * ggml_opt_dataset_t;
typedef struct ggml_opt_context * ggml_opt_context_t;
typedef struct ggml_opt_result * ggml_opt_result_t;
// ====== Loss ======
// built-in loss types, i.e. the built-in quantities minimized by the optimizer
// custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
enum ggml_opt_loss_type {
GGML_OPT_LOSS_TYPE_MEAN,
GGML_OPT_LOSS_TYPE_SUM,
GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
};
// ====== Dataset ======
GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
int64_t ne_datapoint, // number of elements per datapoint
int64_t ne_label, // number of elements per label
int64_t ndata, // total number of datapoints/labels
int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
// get underlying tensors that store the data
GGML_API struct ggml_tensor * ggml_opt_dataset_data (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label, ndata]
// shuffle idata first datapoints from dataset with RNG from opt_ctx, shuffle all datapoints if idata is negative
GGML_API void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata);
// get batch at position ibatch from dataset and copy the data to data_batch and labels_batch
GGML_API void ggml_opt_dataset_get_batch(
ggml_opt_dataset_t dataset,
struct ggml_tensor * data_batch, // shape = [ne_datapoint, ndata_batch]
struct ggml_tensor * labels_batch, // shape = [ne_label, ndata_batch]
int64_t ibatch);
// ====== Model / Context ======
enum ggml_opt_build_type {
GGML_OPT_BUILD_TYPE_FORWARD,
GGML_OPT_BUILD_TYPE_GRAD,
GGML_OPT_BUILD_TYPE_OPT,
};
// parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
struct ggml_opt_optimizer_params {
// AdamW optimizer parameters
struct {
float alpha; // learning rate
float beta1;
float beta2;
float eps; // epsilon for numerical stability
float wd; // weight decay for AdamW, use 0.0f to disable
} adamw;
};
// callback to calculate optimizer parameters prior to a backward pass
// userdata can be used to pass arbitrary data
typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
// returns the default optimizer params (constant)
// userdata is not used
GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
// parameters for initializing a new optimization context
struct ggml_opt_params {
ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
// the forward graph is defined by inputs and outputs
// those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
struct ggml_tensor * inputs;
struct ggml_tensor * outputs;
enum ggml_opt_loss_type loss_type;
enum ggml_opt_build_type build_type;
int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
void * get_opt_pars_ud; // userdata for calculating optimizer parameters
};
// get parameters for an optimization context with defaults set where possible
// parameters for which no sensible defaults exist are supplied as arguments to this function
GGML_API ggml_opt_params ggml_opt_default_params(
ggml_backend_sched_t backend_sched,
struct ggml_context * ctx_compute,
struct ggml_tensor * inputs,
struct ggml_tensor * outputs,
enum ggml_opt_loss_type loss_type);
GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
// set gradients to zero, initilize loss, and optionally reset the optimizer
GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
// get underlying tensors that store data
GGML_API struct ggml_tensor * ggml_opt_inputs( ggml_opt_context_t opt_ctx); // forward graph input tensor
GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
GGML_API struct ggml_tensor * ggml_opt_labels( ggml_opt_context_t opt_ctx); // labels to compare outputs against
GGML_API struct ggml_tensor * ggml_opt_loss( ggml_opt_context_t opt_ctx); // scalar tensor that contains the loss
GGML_API struct ggml_tensor * ggml_opt_pred( ggml_opt_context_t opt_ctx); // predictions made by outputs
GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
// ====== Optimization Result ======
GGML_API ggml_opt_result_t ggml_opt_result_init();
GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
// get data from result, uncertainties are optional and can be ignored by passing NULL
GGML_API void ggml_opt_result_ndata( ggml_opt_result_t result, int64_t * ndata); // writes 1 value, number of datapoints
GGML_API void ggml_opt_result_loss( ggml_opt_result_t result, double * loss, double * unc); // writes 1 value
GGML_API void ggml_opt_result_pred( ggml_opt_result_t result, int32_t * pred); // writes ndata values
GGML_API void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, double * unc); // writes 1 value
// ====== Computation ======
// do forward pass, increment result if not NULL
GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
// do forward pass, increment result if not NULL, do backward pass
GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
// ############################################################################
// ## The high-level functions start here. They do not depend on any private ##
// ## functions or structs and can be copied to and adapted for user code. ##
// ############################################################################
// ====== Intended Usage ======
//
// 1. Select the appropriate loss for your problem.
// 2. Create a dataset and set the data for the "data" tensor. Also set the "labels" tensor if your loss needs them.
// Setting the shard size to 1 will be fine, it's the granularity with which data is shuffled/loaded (bigger values are faster).
// 3. Create a GGML graph for your model with no_alloc == true. Use two separate contexts for the tensors.
// The first context should contain the model parameters and inputs and be allocated statically in user code.
// The second context should contain all other tensors and will be (re)allocated automatically.
// Due to this automated allocation the data of the second context is not defined when accessed in user code.
// Note that the second dimension of the inputs/outputs are interpreted as the number of datapoints in those tensors.
// 4. Call ggml_opt_fit. If you need more control you can use ggml_opt_epoch instead.
// signature for a callback while evaluating opt_ctx on dataset, called after an evaluation
typedef void (*ggml_opt_epoch_callback)(
bool train, // true after training evaluation, false after validation evaluation
ggml_opt_context_t opt_ctx,
ggml_opt_dataset_t dataset,
ggml_opt_result_t result, // result associated with the dataset subsection
int64_t ibatch, // number of batches that have been evaluated so far
int64_t ibatch_max, // total number of batches in this dataset subsection
int64_t t_start_us); // time at which the evaluation on the dataset subsection was started
// do training on front of dataset, do evaluation only on back of dataset
GGML_API void ggml_opt_epoch(
ggml_opt_context_t opt_ctx,
ggml_opt_dataset_t dataset,
ggml_opt_result_t result_train, // result to increment during training, ignored if NULL
ggml_opt_result_t result_eval, // result to increment during evaluation, ignored if NULL
int64_t idata_split, // data index at which to split training and evaluation
ggml_opt_epoch_callback callback_train,
ggml_opt_epoch_callback callback_eval);
// callback that prints a progress bar on stderr
GGML_API void ggml_opt_epoch_callback_progress_bar(
bool train,
ggml_opt_context_t opt_ctx,
ggml_opt_dataset_t dataset,
ggml_opt_result_t result,
int64_t ibatch,
int64_t ibatch_max,
int64_t t_start_us);
// fit model defined by inputs and outputs to dataset
GGML_API void ggml_opt_fit(
ggml_backend_sched_t backend_sched, // backend scheduler for constructing the compute graphs
ggml_context * ctx_compute, // context with temporarily allocated tensors to calculate the outputs
ggml_tensor * inputs, // input tensor with shape [ne_datapoint, ndata_batch]
ggml_tensor * outputs, // output tensor, must have shape [ne_label, ndata_batch] if labels are used
ggml_opt_dataset_t dataset, // dataset with data and optionally also labels
enum ggml_opt_loss_type loss_type, // loss to minimize
ggml_opt_get_optimizer_params get_opt_pars, // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
int64_t nepoch, // how many times the dataset should be iterated over
int64_t nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
float val_split, // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
bool silent); // whether or not info prints to stderr should be suppressed
#ifdef __cplusplus
}
#endif

14
ggml/include/ggml-rpc.h
View file

@ -10,18 +10,18 @@ extern "C" {
#define GGML_RPC_MAX_SERVERS 16
// backend API
GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
GGML_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
GGML_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
GGML_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
#ifdef __cplusplus
}

26
ggml/include/ggml-sycl.h
View file

@ -17,32 +17,32 @@ extern "C" {
#endif
// backend API
GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
// devide buffer
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
GGML_API void ggml_backend_sycl_get_device_description(int device,
GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device,
char *description,
size_t description_size);
GGML_API int ggml_backend_sycl_get_device_count();
GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
GGML_BACKEND_API int ggml_backend_sycl_get_device_count();
GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
// SYCL doesn't support registering host memory, keep here for reference
// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
// GGML_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
// GGML_BACKEND_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
GGML_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
#ifdef __cplusplus
}

18
ggml/include/ggml-vulkan.h
View file

@ -10,21 +10,21 @@ extern "C" {
#define GGML_VK_NAME "Vulkan"
#define GGML_VK_MAX_DEVICES 16
GGML_API void ggml_vk_instance_init(void);
GGML_BACKEND_API void ggml_vk_instance_init(void);
// backend API
GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
GGML_API int ggml_backend_vk_get_device_count(void);
GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
GGML_BACKEND_API int ggml_backend_vk_get_device_count(void);
GGML_BACKEND_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
GGML_BACKEND_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
#ifdef __cplusplus
}

261
ggml/include/ggml.h
View file

@ -176,15 +176,15 @@
#ifdef GGML_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef GGML_BUILD
# define GGML_API __declspec(dllexport)
# define GGML_API __declspec(dllexport) extern
# else
# define GGML_API __declspec(dllimport)
# define GGML_API __declspec(dllimport) extern
# endif
# else
# define GGML_API __attribute__ ((visibility ("default")))
# define GGML_API __attribute__ ((visibility ("default"))) extern
# endif
#else
# define GGML_API
# define GGML_API extern
#endif
// TODO: support for clang
@ -602,7 +602,6 @@ extern "C" {
int32_t flags;
struct ggml_tensor * grad;
struct ggml_tensor * src[GGML_MAX_SRC];
// source tensor and offset for views
@ -615,7 +614,7 @@ extern "C" {
void * extra; // extra things e.g. for ggml-cuda.cu
// char padding[4];
char padding[8];
};
static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@ -1490,7 +1489,7 @@ extern "C" {
"use ggml_rope_ext_inplace instead");
// compute correction dims for YaRN RoPE scaling
void ggml_rope_yarn_corr_dims(
GGML_API void ggml_rope_yarn_corr_dims(
int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// rotary position embedding backward, i.e compute dx from dy
@ -1985,28 +1984,20 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * grad,
float alpha,
float beta1,
float beta2,
float eps,
float wd); // weight decay
struct ggml_tensor * m,
struct ggml_tensor * v,
struct ggml_tensor * adamw_params); // parameters such a the learning rate
//
// automatic differentiation
//
GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
GGML_API void ggml_build_opt_adamw(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
float alpha,
float beta1,
float beta2,
float eps,
float wd); // weight decay
GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_backward_expand(
struct ggml_context * ctx_static, // context for static gradients (loss + gradient accumulation)
struct ggml_context * ctx_compute, // context for gradient computation
struct ggml_cgraph * cgraph,
bool accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
// graph allocation in a context
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@ -2026,7 +2017,9 @@ extern "C" {
GGML_API size_t ggml_graph_overhead(void);
GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
GGML_API struct ggml_tensor * ggml_graph_get_tensor (const struct ggml_cgraph * cgraph, const char * name);
GGML_API struct ggml_tensor * ggml_graph_get_grad (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
@ -2037,198 +2030,15 @@ extern "C" {
// 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);
// build gradient checkpointing backward graph gb for gf using provided checkpoints
// gb_tmp will contain original backward graph with rewritten backward process nodes,
// but without the second forward pass nodes.
GGML_API void ggml_build_backward_gradient_checkpointing(
struct ggml_context * ctx,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
struct ggml_cgraph * gb_tmp,
struct ggml_tensor * * checkpoints,
int n_checkpoints);
//
// optimization
//
// optimization methods
enum ggml_opt_type {
GGML_OPT_TYPE_ADAM,
GGML_OPT_TYPE_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_RESULT_OK = 0,
GGML_OPT_RESULT_DID_NOT_CONVERGE,
GGML_OPT_RESULT_NO_CONTEXT,
GGML_OPT_RESULT_INVALID_WOLFE,
GGML_OPT_RESULT_FAIL,
GGML_OPT_RESULT_CANCEL,
GGML_LINESEARCH_FAIL = -128,
GGML_LINESEARCH_MINIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_STEP,
GGML_LINESEARCH_MAXIMUM_ITERATIONS,
GGML_LINESEARCH_INVALID_PARAMETERS,
};
typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
// TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
// Set callback for all future logging events.
// If this is not called, or NULL is supplied, everything is output on stderr.
GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
// optimization parameters
//
// see ggml.c (ggml_opt_default_params) for default values
//
struct ggml_opt_params {
enum ggml_opt_type type;
size_t graph_size;
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;
int n_gradient_accumulation;
// ADAM parameters
struct {
int n_iter;
float sched; // schedule multiplier (fixed, decay or warmup)
float decay; // weight decay for AdamW, use 0.0f to disable
int decay_min_ndim; // minimum number of tensor dimension to apply weight decay
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
float gclip; // gradient clipping
} 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_context {
struct ggml_context * ctx;
struct ggml_opt_params params;
int iter;
int64_t nx; // number of parameter elements
bool just_initialized;
float loss_before;
float loss_after;
struct {
struct ggml_tensor * g; // current gradient
struct ggml_tensor * m; // first moment
struct ggml_tensor * v; // second moment
struct ggml_tensor * pf; // past function values
float fx_best;
float fx_prev;
int n_no_improvement;
} adam;
struct {
struct ggml_tensor * x; // current parameters
struct ggml_tensor * xp; // previous parameters
struct ggml_tensor * g; // current gradient
struct ggml_tensor * gp; // previous gradient
struct ggml_tensor * d; // search direction
struct ggml_tensor * pf; // past function values
struct ggml_tensor * lmal; // the L-BFGS memory alpha
struct ggml_tensor * lmys; // the L-BFGS memory ys
struct ggml_tensor * lms; // the L-BFGS memory s
struct ggml_tensor * lmy; // the L-BFGS memory y
float fx_best;
float step;
int j;
int k;
int end;
int n_no_improvement;
} lbfgs;
};
GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
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);
// initialize optimizer context
GGML_API void ggml_opt_init(
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_opt_params params,
int64_t nx);
// continue optimizing the function defined by the tensor f
GGML_API enum ggml_opt_result ggml_opt_resume(
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_tensor * f);
// continue optimizing the function defined by the tensor f
GGML_API enum ggml_opt_result ggml_opt_resume_g(
struct ggml_context * ctx,
struct ggml_opt_context * opt,
struct ggml_tensor * f,
struct ggml_cgraph * gf,
struct ggml_cgraph * gb,
ggml_opt_callback callback,
void * callback_data);
//
// quantization
//
@ -2384,38 +2194,6 @@ extern "C" {
GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data);
//
// system info
//
GGML_API int ggml_cpu_has_avx (void);
GGML_API int ggml_cpu_has_avx_vnni (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_avx512_bf16(void);
GGML_API int ggml_cpu_has_amx_int8 (void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_metal (void);
GGML_API int ggml_cpu_has_f16c (void);
GGML_API int ggml_cpu_has_fp16_va (void);
GGML_API int ggml_cpu_has_wasm_simd (void);
GGML_API int ggml_cpu_has_blas (void);
GGML_API int ggml_cpu_has_cuda (void);
GGML_API int ggml_cpu_has_vulkan (void);
GGML_API int ggml_cpu_has_kompute (void);
GGML_API int ggml_cpu_has_gpublas (void);
GGML_API int ggml_cpu_has_sse3 (void);
GGML_API int ggml_cpu_has_ssse3 (void);
GGML_API int ggml_cpu_has_riscv_v (void);
GGML_API int ggml_cpu_has_sycl (void);
GGML_API int ggml_cpu_has_rpc (void);
GGML_API int ggml_cpu_has_vsx (void);
GGML_API int ggml_cpu_has_cann (void);
GGML_API int ggml_cpu_has_llamafile (void);
#ifdef __cplusplus
// restrict not standard in C++
#define GGML_RESTRICT
@ -2432,7 +2210,6 @@ extern "C" {
size_t type_size;
bool is_quantized;
ggml_to_float_t to_float;
ggml_from_float_t from_float;
ggml_from_float_t from_float_ref;
};

1269
ggml/src/CMakeLists.txt
View file

File diff suppressed because it is too large Load diff

3443
ggml/src/ggml-aarch64.c
View file

File diff suppressed because it is too large Load diff

20
ggml/src/ggml-aarch64.h
View file

@ -1,9 +1,5 @@
// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML internal header
@ -12,27 +8,11 @@
extern "C" {
#endif
// Quantization
void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
// GEMV
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// GEMM
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#ifdef __cplusplus
}
#endif

14
ggml/src/ggml-alloc.c
View file

@ -466,18 +466,12 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
return ggml_gallocr_hash_get(galloc, t)->allocated;
}
static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
hn->buffer_id = buffer_id;
hn->offset = offset;
hn->allocated = true;
}
static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
}
static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
GGML_ASSERT(buffer_id >= 0);
struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
@ -816,7 +810,11 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
}
static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
size_t node_size = 0;
if (!node->data && !node->view_src) {
GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
}
return talloc->size_max >= node_size;
}

107
ggml/src/ggml-amx/CMakeLists.txt Normal file
View file

@ -0,0 +1,107 @@
if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
message(STATUS "Using AMX")
file(GLOB GGML_HEADERS_AMX "*.h")
list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
file(GLOB GGML_SOURCES_AMX "*.cpp")
add_library(ggml-amx
${GGML_HEADERS_AMX}
${GGML_SOURCES_AMX})
target_link_libraries(ggml-amx PRIVATE ggml-base)
target_include_directories(ggml-amx PRIVATE . ..)
# this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
# TODO: integrate AMX backend into the CPU backend
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
# TODO: improve, should not reference files from the parent folder
include(../ggml-cpu/cmake/FindSIMD.cmake)
endif ()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS /arch:AVX512)
# MSVC has no compile-time flags enabling specific
# AVX512 extensions, neither it defines the
# macros corresponding to the extensions.
# Do it manually.
if (GGML_AVX512_VBMI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
endif()
if (GGML_AVX512_VNNI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
endif()
if (GGML_AVX512_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
endif()
if (GGML_AMX_TILE)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
endif()
if (GGML_AMX_INT8)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
endif()
if (GGML_AMX_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (GGML_AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
endif()
else()
if (GGML_NATIVE)
list(APPEND ARCH_FLAGS -march=native)
endif()
if (GGML_F16C)
list(APPEND ARCH_FLAGS -mf16c)
endif()
if (GGML_FMA)
list(APPEND ARCH_FLAGS -mfma)
endif()
if (GGML_AVX)
list(APPEND ARCH_FLAGS -mavx)
endif()
if (GGML_AVX2)
list(APPEND ARCH_FLAGS -mavx2)
endif()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512bw)
endif()
if (GGML_AVX512_VBMI)
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
endif()
endif()
target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
else()
set(GGML_AMX OFF PARENT_SCOPE)
message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
endif()

3
ggml/src/ggml-amx/common.h
View file

@ -1,7 +1,8 @@
#pragma once
#include "ggml.h"
#include "ggml-cpu-impl.h" // <immintrin.h>
// hack until AMX is moved into the CPU backend
#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
#include <algorithm>
#include <memory>

18
ggml/src/ggml-amx.cpp → ggml/src/ggml-amx/ggml-amx.cpp
View file

@ -317,8 +317,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
const enum ggml_type type = src0->type;
const int64_t ne0 = op->ne[0];
bool is_training = src0->grad || src1->grad;
// amx kernels enables for Q4_0, Q4_1, Q8_0, F16
// Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
@ -326,7 +324,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
bool can_use_amx =
is_contiguous_2d(src0) && // src0 must be contiguous
is_contiguous_2d(src1) && // src1 must be contiguous
!is_training && // inference only
src1->type == GGML_TYPE_F32 && // src1 must be float32
has_amx_kernels && // with amx kernel impls
ne0 % (TILE_N * 2) == 0; // out_features is 32x
@ -421,9 +418,18 @@ ggml_backend_reg_t ggml_backend_amx_reg(void) {
#else // if defined(__AMX_INT8__)
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
return nullptr;
}
bool ggml_backend_is_amx(ggml_backend_t backend) {
GGML_UNUSED(backend);
return false;
}
ggml_backend_t ggml_backend_amx_init(void) {
fprintf(stderr, "GGML is not compiled with AMX support!\n");
return ggml_backend_t{};
return nullptr;
}
void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
@ -433,4 +439,8 @@ void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
GGML_UNUSED(n_threads);
}
ggml_backend_reg_t ggml_backend_amx_reg(void) {
return nullptr;
}
#endif

7
ggml/src/ggml-amx/mmq.cpp
View file

@ -496,19 +496,20 @@ inline void from_float(const float * x, char * vy, int64_t k);
template <>
inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
quantize_row_q8_0(x, vy, k);
// FIXME: using unoptimized reference impl until moved to CPU backend
quantize_row_q8_0_ref(x, (block_q8_0 *)vy, k);
}
template <>
inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
quantize_row_q8_1(x, vy, k);
quantize_row_q8_1_ref(x, (block_q8_1 *)vy, k);
}
template <>
inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
#if 1
// TODO: this is reference impl!
quantize_row_q8_K(x, vy, k);
quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
#else
quantize_row_q8_K_vnni(x, vy, k);
#endif

195
ggml/src/ggml-backend-reg.cpp Normal file
View file

@ -0,0 +1,195 @@
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include <cstring>
#include <vector>
// Backend registry
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
#ifdef GGML_USE_RPC
#include "ggml-rpc.h"
#endif
#ifdef GGML_USE_AMX
# include "ggml-amx.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_KOMPUTE
#include "ggml-kompute.h"
#endif
struct ggml_backend_registry {
std::vector<ggml_backend_reg_t> backends;
std::vector<ggml_backend_dev_t> devices;
ggml_backend_registry() {
#ifdef GGML_USE_CUDA
register_backend(ggml_backend_cuda_reg());
#endif
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_SYCL
register_backend(ggml_backend_sycl_reg());
#endif
#ifdef GGML_USE_VULKAN
register_backend(ggml_backend_vk_reg());
#endif
#ifdef GGML_USE_CANN
register_backend(ggml_backend_cann_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_AMX
register_backend(ggml_backend_amx_reg());
#endif
#ifdef GGML_USE_KOMPUTE
register_backend(ggml_backend_kompute_reg());
#endif
register_backend(ggml_backend_cpu_reg());
}
void register_backend(ggml_backend_reg_t reg) {
if (!reg) {
return;
}
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
#endif
backends.push_back(reg);
for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
register_device(ggml_backend_reg_dev_get(reg, i));
}
}
void register_device(ggml_backend_dev_t device) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
devices.push_back(device);
}
};
static ggml_backend_registry & get_reg() {
static ggml_backend_registry reg;
return reg;
}
// Internal API
void ggml_backend_register(ggml_backend_reg_t reg) {
get_reg().register_backend(reg);
}
void ggml_backend_device_register(ggml_backend_dev_t device) {
get_reg().register_device(device);
}
// Backend (reg) enumeration
size_t ggml_backend_reg_count() {
return get_reg().backends.size();
}
ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
GGML_ASSERT(index < ggml_backend_reg_count());
return get_reg().backends[index];
}
ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
ggml_backend_reg_t reg = ggml_backend_reg_get(i);
if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
return reg;
}
}
return NULL;
}
// Device enumeration
size_t ggml_backend_dev_count() {
return get_reg().devices.size();
}
ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
GGML_ASSERT(index < ggml_backend_dev_count());
return get_reg().devices[index];
}
ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
return dev;
}
}
return NULL;
}
ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == type) {
return dev;
}
}
return NULL;
}
// Convenience functions
ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_best(void) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
if (!dev) {
dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
}
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, NULL);
}

690
ggml/src/ggml-backend.cpp
View file

@ -279,7 +279,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
buf->iface.get_tensor(buf, tensor, data, offset, size);
}
GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
if (size == 0) {
@ -525,197 +525,6 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
return reg->iface.get_proc_address(reg, name);
}
// Backend registry
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
#ifdef GGML_USE_RPC
#include "ggml-rpc.h"
#endif
#ifndef __AMX_INT8__
#undef GGML_USE_AMX
#endif
#ifdef GGML_USE_AMX
# include "ggml-amx.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_KOMPUTE
#include "ggml-kompute.h"
#endif
#include "ggml-cpu.h"
struct ggml_backend_registry {
std::vector<ggml_backend_reg_t> backends;
std::vector<ggml_backend_dev_t> devices;
ggml_backend_registry() {
#ifdef GGML_USE_CUDA
register_backend(ggml_backend_cuda_reg());
#endif
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_SYCL
register_backend(ggml_backend_sycl_reg());
#endif
#ifdef GGML_USE_VULKAN
register_backend(ggml_backend_vk_reg());
#endif
#ifdef GGML_USE_CANN
register_backend(ggml_backend_cann_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_AMX
register_backend(ggml_backend_amx_reg());
#endif
#ifdef GGML_USE_KOMPUTE
register_backend(ggml_backend_kompute_reg());
#endif
register_backend(ggml_backend_cpu_reg());
}
void register_backend(ggml_backend_reg_t reg) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
#endif
backends.push_back(reg);
for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
register_device(ggml_backend_reg_dev_get(reg, i));
}
}
void register_device(ggml_backend_dev_t device) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
devices.push_back(device);
}
};
static ggml_backend_registry & get_reg() {
static ggml_backend_registry reg;
return reg;
}
// Internal API
void ggml_backend_register(ggml_backend_reg_t reg) {
get_reg().register_backend(reg);
}
void ggml_backend_device_register(ggml_backend_dev_t device) {
get_reg().register_device(device);
}
// Backend (reg) enumeration
size_t ggml_backend_reg_count() {
return get_reg().backends.size();
}
ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
GGML_ASSERT(index < ggml_backend_reg_count());
return get_reg().backends[index];
}
ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
ggml_backend_reg_t reg = ggml_backend_reg_get(i);
if (strcmp(ggml_backend_reg_name(reg), name) == 0) {
return reg;
}
}
return NULL;
}
// Device enumeration
size_t ggml_backend_dev_count() {
return get_reg().devices.size();
}
ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
GGML_ASSERT(index < ggml_backend_dev_count());
return get_reg().devices[index];
}
ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
return dev;
}
}
return NULL;
}
ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == type) {
return dev;
}
}
return NULL;
}
// Convenience functions
ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_best(void) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
if (!dev) {
dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
}
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, NULL);
}
// multi-buffer buffer
struct ggml_backend_multi_buffer_context {
@ -880,7 +689,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
}
static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
ggml_backend_buffer_t buffer = tensor->buffer;
ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
if (buffer == NULL) {
return -1;
}
@ -913,8 +722,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
// returns the backend that should be used for the node based on the current locations
static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
// TODO: use supports_op to check if the backend supports the op
// assign pre-allocated nodes to their backend
int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
if (cur_backend_id != -1) {
@ -933,7 +740,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
// since the tensor is pre-allocated, it cannot be moved to another backend
GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
}
// graph input
@ -1077,6 +884,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
int * node_backend_id = &tensor_backend_id(node);
if (ggml_is_view_op(node->op)) {
continue;
}
// do not overwrite user assignments
if (*node_backend_id == -1) {
*node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
@ -1640,7 +1450,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
bool parallel) {
GGML_ASSERT(n_backends > 0);
GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
@ -1729,12 +1539,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
ggml_backend_sched_split_graph(sched, measure_graph);
ggml_backend_sched_synchronize(sched);
if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
return false;
}
ggml_backend_sched_reset(sched);
ggml_backend_sched_synchronize(sched);
return true;
}
@ -2036,17 +1847,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
return true;
}
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include <cctype>
#include <string>
// ggml-backend interface
// CPU backend - buffer
static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
@ -2120,7 +1920,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
/* .reset = */ NULL,
};
// CPU backend - buffer type
// CPU backend buffer type
// this buffer type is defined here to make it available to all backends
static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU";
@ -2161,7 +1963,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .device = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
@ -2184,478 +1986,14 @@ static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .device = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type;
}
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
static ggml_backend_buffer_type_t bufts[] = {
#ifdef GGML_USE_CPU_HBM
ggml_backend_cpu_hbm_buffer_type(),
#endif
NULL
};
return bufts;
GGML_UNUSED(device);
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
int n_threads;
ggml_threadpool_t threadpool;
uint8_t * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
delete[] cpu_ctx->work_data;
delete cpu_ctx;
delete backend;
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
if (cpu_plan->cplan.work_data == NULL) {
delete cpu_plan;
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
delete[] cpu_plan->cplan.work_data;
delete cpu_plan;
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
cpu_ctx->work_data = new uint8_t[cplan.work_size];
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = (uint8_t *)cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .get_name = */ ggml_backend_cpu_get_name,
/* .free = */ ggml_backend_cpu_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->threadpool = NULL;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
if (ctx->threadpool && ctx->threadpool != threadpool) {
// already had a different threadpool, pause/suspend it before switching
ggml_threadpool_pause(ctx->threadpool);
}
ctx->threadpool = threadpool;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
}
// CPU backend - device
struct ggml_backend_cpu_device_context {
std::string description = "CPU";
ggml_backend_cpu_device_context() {
#ifdef __APPLE__
size_t len = 0;
if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
description.resize(len);
sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
}
#elif defined(__linux__)
FILE * f = fopen("/proc/cpuinfo", "r");
if (f) {
char buf[1024];
while (fgets(buf, sizeof(buf), f)) {
if (strncmp(buf, "model name", 10) == 0) {
char * p = strchr(buf, ':');
if (p) {
p++;
while (std::isspace(*p)) {
p++;
}
while (std::isspace(p[strlen(p) - 1])) {
p[strlen(p) - 1] = '\0';
}
description = p;
break;
}
}
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) == ERROR_SUCCESS) {
DWORD cpu_brand_size = 0;
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
NULL,
&cpu_brand_size) == ERROR_SUCCESS) {
description.resize(cpu_brand_size);
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
(LPBYTE)&description[0], // NOLINT
&cpu_brand_size) == ERROR_SUCCESS) {
if (description.find('\0') != std::string::npos) {
description.resize(description.find('\0'));
}
}
}
RegCloseKey(hKey);
}
#endif
}
};
static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
return "CPU";
GGML_UNUSED(dev);
}
static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// TODO
*free = 0;
*total = 0;
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_cpu_device_get_name(dev);
props->description = ggml_backend_cpu_device_get_description(dev);
props->type = ggml_backend_cpu_device_get_type(dev);
ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_cpu_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_CPY:
return
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32;// FIXME || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
case GGML_OP_ROPE_BACK:
return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
case GGML_OP_IM2COL_BACK:
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
case GGML_OP_OUT_PROD:
return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
default:
return true;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
/* .get_name = */ ggml_backend_cpu_device_get_name,
/* .get_description = */ ggml_backend_cpu_device_get_description,
/* .get_memory = */ ggml_backend_cpu_device_get_memory,
/* .get_type = */ ggml_backend_cpu_device_get_type,
/* .get_props = */ ggml_backend_cpu_device_get_props,
/* .init_backend = */ ggml_backend_cpu_device_init_backend,
/* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_cpu_device_supports_op,
/* .supports_buft = */ ggml_backend_cpu_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// CPU backend - backend (reg)
static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
return "CPU";
GGML_UNUSED(reg);
}
static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_cpu_device_context ctx;
static ggml_backend_device ggml_backend_cpu_device = {
/* .iface = */ ggml_backend_cpu_device_i,
/* .reg = */ reg,
/* .context = */ &ctx,
};
return &ggml_backend_cpu_device;
}
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
return (void *)ggml_backend_cpu_set_n_threads;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
return (void *)ggml_backend_cpu_get_extra_bufts;
}
return NULL;
GGML_UNUSED(reg);
}
static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
/* .get_name = */ ggml_backend_cpu_reg_get_name,
/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
/* .get_device = */ ggml_backend_cpu_reg_get_device,
/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}

91
ggml/src/ggml-blas/CMakeLists.txt Normal file
View file

@ -0,0 +1,91 @@
if (GGML_STATIC)
set(BLA_STATIC ON)
endif()
#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
# set(BLA_SIZEOF_INTEGER 8)
#endif()
set(BLA_VENDOR ${GGML_BLAS_VENDOR})
find_package(BLAS)
if (BLAS_FOUND)
message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
add_library(ggml-blas
ggml-blas.cpp
)
target_link_libraries(ggml-blas PRIVATE ggml-base)
target_include_directories(ggml-blas PRIVATE . ..)
if (${GGML_BLAS_VENDOR} MATCHES "Apple")
add_compile_definitions(ACCELERATE_NEW_LAPACK)
add_compile_definitions(ACCELERATE_LAPACK_ILP64)
add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
# BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
# see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
find_package(PkgConfig REQUIRED)
if (${GGML_BLAS_VENDOR} MATCHES "Generic")
pkg_check_modules(DepBLAS blas)
elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
# As of openblas v0.3.22, the 64-bit is named openblas64.pc
pkg_check_modules(DepBLAS openblas64)
if (NOT DepBLAS_FOUND)
pkg_check_modules(DepBLAS openblas)
endif()
elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
add_compile_definitions(GGML_BLAS_USE_BLIS)
pkg_check_modules(DepBLAS blis)
elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
pkg_check_modules(DepBLAS blas-atlas)
elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
pkg_check_modules(DepBLAS flexiblas_api)
elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
add_compile_definitions(GGML_BLAS_USE_MKL)
# all Intel* libraries share the same include path
pkg_check_modules(DepBLAS mkl-sdl)
elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
# this doesn't provide pkg-config
# suggest to assign BLAS_INCLUDE_DIRS on your own
if ("${NVHPC_VERSION}" STREQUAL "")
message(WARNING "Better to set NVHPC_VERSION")
else()
set(DepBLAS_FOUND ON)
set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
endif()
endif()
if (DepBLAS_FOUND)
set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
else()
message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
" detected by pkgconfig, trying to find cblas.h from possible paths...")
find_path(BLAS_INCLUDE_DIRS
NAMES cblas.h
HINTS
/usr/include
/usr/local/include
/usr/include/openblas
/opt/homebrew/opt/openblas/include
/usr/local/opt/openblas/include
/usr/include/x86_64-linux-gnu/openblas/include
)
endif()
endif()
message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
#add_compile_options(${BLAS_LINKER_FLAGS})
target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
add_compile_definitions(GGML_BLAS_USE_MKL)
endif()
target_link_libraries (ggml-blas PRIVATE ${BLAS_LIBRARIES})
target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
else()
message(ERROR "BLAS not found, please refer to "
"https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
" to set correct GGML_BLAS_VENDOR")
endif()

4
ggml/src/ggml-blas.cpp → ggml/src/ggml-blas/ggml-blas.cpp
View file

@ -6,7 +6,7 @@
#include <vector>
#include <cstring>
#if defined(GGML_USE_ACCELERATE)
#if defined(GGML_BLAS_USE_ACCELERATE)
# include <Accelerate/Accelerate.h>
#elif defined(GGML_BLAS_USE_MKL)
# include <mkl.h>
@ -320,7 +320,7 @@ static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
}
static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
#if defined(GGML_USE_ACCELERATE)
#if defined(GGML_BLAS_USE_ACCELERATE)
return "Accelerate";
#elif defined(GGML_BLAS_USE_MKL)
return "MKL";

46
ggml/src/ggml-cann/CMakeLists.txt Normal file
View file

@ -0,0 +1,46 @@
if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
endif()
if (CANN_INSTALL_DIR)
# Only Support Linux.
if (NOT UNIX)
message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
endif()
# Supported platforms: x86-64, arm64
if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
else()
message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
endif()
# Set header and libs
set(CANN_INCLUDE_DIRS
${CANN_INSTALL_DIR}/include
${CANN_INSTALL_DIR}/include/aclnn
${CANN_INSTALL_DIR}/acllib/include
)
add_subdirectory(kernels)
list(APPEND CANN_LIBRARIES
ascendcl
nnopbase
opapi
acl_op_compiler
ascendc_kernels
)
file(GLOB GGML_SOURCES_CANN "*.cpp")
add_library(ggml-cann ${GGML_SOURCES_CANN})
target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES})
target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
message(STATUS "CANN: CANN_INCLUDE_DIRS = ${CANN_INCLUDE_DIRS}")
message(STATUS "CANN: CANN_LIBRARIES = ${CANN_LIBRARIES}")
else()
message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
endif()

0
ggml/src/ggml-cann.cpp → ggml/src/ggml-cann/ggml-cann.cpp
View file

261
ggml/src/ggml-cpu/CMakeLists.txt Normal file
View file

@ -0,0 +1,261 @@
add_library(ggml-cpu
ggml-cpu.c
ggml-cpu.cpp
ggml-cpu-aarch64.c
ggml-cpu-aarch64.h
ggml-cpu-quants.c
ggml-cpu-quants.h
)
target_link_libraries(ggml-cpu PRIVATE ggml-base)
target_include_directories(ggml-cpu PRIVATE . ..)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if (ACCELERATE_FRAMEWORK)
message(STATUS "Accelerate framework found")
add_compile_definitions(GGML_USE_ACCELERATE)
add_compile_definitions(ACCELERATE_NEW_LAPACK)
add_compile_definitions(ACCELERATE_LAPACK_ILP64)
target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK})
else()
message(WARNING "Accelerate framework not found")
endif()
endif()
if (GGML_OPENMP)
find_package(OpenMP)
if (OpenMP_FOUND)
message(STATUS "OpenMP found")
add_compile_definitions(GGML_USE_OPENMP)
target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
# FIXME: should be replaced with a compiler id check
#if (GGML_MUSA)
# list(APPEND GGML_CPU_EXTRA_INCLUDES "/usr/lib/llvm-14/lib/clang/14.0.0/include")
# list(APPEND GGML_CPU_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so")
#endif()
else()
message(WARNING "OpenMP not found")
endif()
endif()
if (GGML_LLAMAFILE)
message(STATUS "Using llamafile")
add_compile_definitions(GGML_USE_LLAMAFILE)
target_sources(ggml-cpu PRIVATE
llamafile/sgemm.cpp
llamafile/sgemm.h)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind REQUIRED)
message(STATUS "Using memkind for CPU HBM")
add_compile_definitions(GGML_USE_CPU_HBM)
target_link_libraries(ggml-cpu PUBLIC memkind)
endif()
if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR
CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
(NOT CMAKE_OSX_ARCHITECTURES AND
NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
message(STATUS "ARM detected")
if (MSVC)
add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
add_compile_definitions(__ARM_NEON)
add_compile_definitions(__ARM_FEATURE_FMA)
set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
if (GGML_COMPILER_SUPPORT_DOTPROD)
add_compile_definitions(__ARM_FEATURE_DOTPROD)
endif ()
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
endif ()
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
endif ()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
else()
check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
list(APPEND ARCH_FLAGS -mfp16-format=ieee)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
# Raspberry Pi 1, Zero
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
# Android armeabi-v7a
list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
else()
# Raspberry Pi 2
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
endif()
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
# Android arm64-v8a
# Raspberry Pi 3, 4, Zero 2 (32-bit)
list(APPEND ARCH_FLAGS -mno-unaligned-access)
endif()
if (GGML_SVE)
list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
endif()
endif()
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
message(STATUS "x86 detected")
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
# TODO: improve, should not reference files from the parent folder
include(cmake/FindSIMD.cmake)
endif ()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS /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 (GGML_AVX512_VBMI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
endif()
if (GGML_AVX512_VNNI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
endif()
if (GGML_AVX512_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
endif()
if (GGML_AMX_TILE)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
endif()
if (GGML_AMX_INT8)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
endif()
if (GGML_AMX_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (GGML_AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
endif()
else()
if (GGML_NATIVE)
list(APPEND ARCH_FLAGS -march=native)
endif()
if (GGML_F16C)
list(APPEND ARCH_FLAGS -mf16c)
endif()
if (GGML_FMA)
list(APPEND ARCH_FLAGS -mfma)
endif()
if (GGML_AVX)
list(APPEND ARCH_FLAGS -mavx)
endif()
if (GGML_AVX2)
list(APPEND ARCH_FLAGS -mavx2)
endif()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512bw)
endif()
if (GGML_AVX512_VBMI)
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
endif()
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
message(STATUS "PowerPC detected")
execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
string(FIND "${POWER10_M}" "POWER10" substring_index)
if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
set(substring_index -1)
endif()
if (${substring_index} GREATER_EQUAL 0)
list(APPEND ARCH_FLAGS -mcpu=power10)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
else()
list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
#TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
message(STATUS "loongarch64 detected")
list(APPEND ARCH_FLAGS -march=loongarch64)
if (GGML_LASX)
list(APPEND ARCH_FLAGS -mlasx)
endif()
if (GGML_LSX)
list(APPEND ARCH_FLAGS -mlsx)
endif()
else()
message(STATUS "Unknown architecture")
endif()
if (GGML_CPU_AARCH64)
message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels")
add_compile_definitions(GGML_USE_CPU_AARCH64)
endif()
target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
if (EMSCRIPTEN)
set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")
endif()

0
ggml/cmake/FindSIMD.cmake → ggml/src/ggml-cpu/cmake/FindSIMD.cmake
View file

3560
ggml/src/ggml-cpu/ggml-cpu-aarch64.c Normal file
View file

File diff suppressed because it is too large Load diff

30
ggml/src/ggml-cpu/ggml-cpu-aarch64.h Normal file
View file

@ -0,0 +1,30 @@
#pragma once
#include "ggml.h"
// GGML internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
// GEMV
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// GEMM
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
#ifdef __cplusplus
}
#endif

243
ggml/src/ggml-cpu-impl.h → ggml/src/ggml-cpu/ggml-cpu-impl.h
View file

@ -27,80 +27,6 @@ extern "C" {
#endif
/**
* Converts brain16 to float32.
*
* The bfloat16 floating point format has the following structure:
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───┐
* 0b0000000000000000 brain16
*
* Since bf16 has the same number of exponent bits as a 32bit float,
* encoding and decoding numbers becomes relatively straightforward.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───────────────────┐
* 0b00000000000000000000000000000000 IEEE binary32
*
* For comparison, the standard fp16 format has fewer exponent bits.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌─┴─┐┌─┴──────┐
* 0b0000000000000000 IEEE binary16
*
* @see IEEE 754-2008
*/
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
/**
* Converts float32 to brain16.
*
* This is binary identical with Google Brain float conversion.
* Floats shall round to nearest even, and NANs shall be quiet.
* Subnormals aren't flushed to zero, except perhaps when used.
* This code should vectorize nicely if using modern compilers.
*/
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
ggml_bf16_t h;
union {
float f;
uint32_t i;
} u;
u.f = s;
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
h.bits = (u.i >> 16) | 64; /* force to quiet */
return h;
}
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
return h;
}
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
@ -388,28 +314,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
#endif // defined(__ARM_NEON)
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
@ -462,153 +366,6 @@ static __m256 __lasx_xvreplfr2vr_s(float val) {
}
#endif
#ifdef __F16C__
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // __F16C__
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
#ifdef __cplusplus
}
#endif

10822
ggml/src/ggml-cpu/ggml-cpu-quants.c Normal file
View file

File diff suppressed because it is too large Load diff

63
ggml/src/ggml-cpu/ggml-cpu-quants.h Normal file
View file

@ -0,0 +1,63 @@
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML CPU internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
#ifdef __cplusplus
}
#endif

320
ggml/src/ggml-cpu.c → ggml/src/ggml-cpu/ggml-cpu.c
View file

@ -1,13 +1,15 @@
#define _CRT_SECURE_NO_DEPRECATE // Disables "unsafe" warnings on Windows
#define _USE_MATH_DEFINES // For M_PI on MSVC
#include "ggml-aarch64.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-cpu-aarch64.h"
#include "ggml-cpu-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "ggml-quants.h"
#include "ggml-cpu-quants.h"
#include "ggml-threading.h"
#include "ggml.h"
#if defined(_MSC_VER) || defined(__MINGW32__)
@ -42,7 +44,7 @@
#endif
#ifdef GGML_USE_LLAMAFILE
#include <llamafile/sgemm.h>
#include "llamafile/sgemm.h"
#endif
#if defined(_MSC_VER)
@ -104,9 +106,6 @@ static ggml_fp16_t ggml_table_gelu_f16[1 << 16];
// precomputed quick gelu table for f16 (128 KB)
static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
float ggml_table_f32_f16[1 << 16];
#if defined(__ARM_ARCH)
struct ggml_arm_arch_features_type {
int has_neon;
@ -261,11 +260,13 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.nrows = 1,
},
[GGML_TYPE_F16] = {
.from_float = (ggml_from_float_t) ggml_fp32_to_fp16_row,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f16,
.vec_dot_type = GGML_TYPE_F16,
.nrows = 1,
},
[GGML_TYPE_Q4_0] = {
.from_float = quantize_row_q4_0,
.vec_dot = ggml_vec_dot_q4_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
#if defined (__ARM_FEATURE_MATMUL_INT8)
@ -275,6 +276,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
#endif
},
[GGML_TYPE_Q4_1] = {
.from_float = quantize_row_q4_1,
.vec_dot = ggml_vec_dot_q4_1_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1,
#if defined (__ARM_FEATURE_MATMUL_INT8)
@ -283,27 +285,20 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.nrows = 1,
#endif
},
[4] = { // GGML_TYPE_Q4_2
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_COUNT,
.nrows = 1,
},
[5] = { // GGML_TYPE_Q4_3
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_COUNT,
.nrows = 1,
},
[GGML_TYPE_Q5_0] = {
.from_float = quantize_row_q5_0,
.vec_dot = ggml_vec_dot_q5_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
},
[GGML_TYPE_Q5_1] = {
.from_float = quantize_row_q5_1,
.vec_dot = ggml_vec_dot_q5_1_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1,
},
[GGML_TYPE_Q8_0] = {
.from_float = quantize_row_q8_0,
.from_float_to_mat = quantize_mat_q8_0,
.vec_dot = ggml_vec_dot_q8_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
@ -314,85 +309,106 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
#endif
},
[GGML_TYPE_Q8_1] = {
.from_float = quantize_row_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1,
},
[GGML_TYPE_Q2_K] = {
.from_float = quantize_row_q2_K,
.vec_dot = ggml_vec_dot_q2_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q3_K] = {
.from_float = quantize_row_q3_K,
.vec_dot = ggml_vec_dot_q3_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q4_K] = {
.from_float = quantize_row_q4_K,
.vec_dot = ggml_vec_dot_q4_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q5_K] = {
.from_float = quantize_row_q5_K,
.vec_dot = ggml_vec_dot_q5_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q6_K] = {
.from_float = quantize_row_q6_K,
.vec_dot = ggml_vec_dot_q6_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ2_XXS] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq2_xxs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ2_XS] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq2_xs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ3_XXS] = {
// NOTE: from_float for iq3 and iq2_s was removed because these quants require initialization in ggml_quantize_init
//.from_float = quantize_row_iq3_xxs,
.vec_dot = ggml_vec_dot_iq3_xxs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ3_S] = {
//.from_float = quantize_row_iq3_s,
.vec_dot = ggml_vec_dot_iq3_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ2_S] = {
//.from_float = quantize_row_iq2_s,
.vec_dot = ggml_vec_dot_iq2_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ1_S] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq1_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ1_M] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq1_m_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ4_NL] = {
.from_float = quantize_row_iq4_nl,
.vec_dot = ggml_vec_dot_iq4_nl_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
},
[GGML_TYPE_IQ4_XS] = {
.from_float = quantize_row_iq4_xs,
.vec_dot = ggml_vec_dot_iq4_xs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q8_K] = {
.from_float = quantize_row_q8_K,
},
[GGML_TYPE_BF16] = {
.from_float = (ggml_from_float_t) ggml_fp32_to_bf16_row,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16,
.vec_dot_type = GGML_TYPE_BF16,
.nrows = 1,
},
[GGML_TYPE_Q4_0_4_4] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
@ -401,6 +417,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.gemm = ggml_gemm_q4_0_4x4_q8_0,
},
[GGML_TYPE_Q4_0_4_8] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
@ -409,6 +426,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.gemm = ggml_gemm_q4_0_4x8_q8_0,
},
[GGML_TYPE_Q4_0_8_8] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
@ -417,11 +435,13 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.gemm = ggml_gemm_q4_0_8x8_q8_0,
},
[GGML_TYPE_TQ1_0] = {
.from_float = quantize_row_tq1_0,
.vec_dot = ggml_vec_dot_tq1_0_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_TQ2_0] = {
.from_float = quantize_row_tq2_0,
.vec_dot = ggml_vec_dot_tq2_0_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
@ -1449,8 +1469,12 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#undef LOAD
#elif defined(__AVX2__)
#elif defined(__AVX2__) || defined(__AVX__)
#if defined(__AVX2__)
#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
#else
#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
#endif
__m256 c1 = _mm256_setzero_ps();
__m256 c2 = _mm256_setzero_ps();
__m256 c3 = _mm256_setzero_ps();
@ -2250,22 +2274,7 @@ struct ggml_state {
struct ggml_numa_nodes numa;
};
// global state
static struct ggml_state g_state = {0};
static atomic_flag g_state_critical = ATOMIC_FLAG_INIT;
// TODO: move to threading file
// critical section via spin lock
void ggml_critical_section_start(void) {
while (atomic_flag_test_and_set(&g_state_critical)) {
// spin
sched_yield();
}
}
void ggml_critical_section_end(void) {
atomic_flag_clear(&g_state_critical);
}
static void ggml_barrier(struct ggml_threadpool * tp) {
int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
@ -2360,7 +2369,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
// figure out which node we're on
uint current_cpu;
int getcpu_ret = 0;
#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 33) || defined(__COSMOPOLITAN__)
getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
#else
// old glibc doesn't have a wrapper for this call. Fall back on direct syscall
@ -2997,8 +3006,8 @@ static void ggml_compute_forward_dup_f16(
id += ne00 * (ne01 - ir1);
}
}
} else if (ggml_get_type_traits(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
} else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
size_t id = 0;
@ -3278,8 +3287,8 @@ static void ggml_compute_forward_dup_bf16(
id += ne00 * (ne01 - ir1);
}
}
} else if (ggml_get_type_traits(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
} else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
size_t id = 0;
@ -3594,8 +3603,8 @@ static void ggml_compute_forward_dup_f32(
id += rs * (ne01 - ir1);
}
}
} else if (ggml_get_type_traits(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
} else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
size_t id = 0;
size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@ -4377,7 +4386,7 @@ static void ggml_compute_forward_add_q_f32(
const enum ggml_type type = src0->type;
const enum ggml_type dtype = dst->type;
ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dtype)->from_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dtype)->from_float;
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type));
@ -4679,7 +4688,7 @@ static void ggml_compute_forward_add1_q_f32(
const enum ggml_type type = src0->type;
ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(type)->from_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(type)->from_float;
// we don't support permuted src0
GGML_ASSERT(nb00 == ggml_type_size(type));
@ -7325,6 +7334,7 @@ static void ggml_compute_forward_group_norm(
static void ggml_compute_forward_mul_mat_one_chunk(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const enum ggml_type type,
const int64_t num_rows_per_vec_dot,
const int64_t ir0_start,
const int64_t ir0_end,
@ -7336,8 +7346,6 @@ static void ggml_compute_forward_mul_mat_one_chunk(
GGML_TENSOR_BINARY_OP_LOCALS
const enum ggml_type type = src0->type;
const bool src1_cont = ggml_is_contiguous(src1);
ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot;
@ -7425,10 +7433,14 @@ static void ggml_compute_forward_mul_mat(
const int ith = params->ith;
const int nth = params->nth;
const enum ggml_type type = src0->type;
enum ggml_type type = src0->type;
if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
type = (enum ggml_type)(intptr_t)src0->extra;
}
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
ggml_from_float_to_mat_t const from_float_to_mat = type_traits_cpu[vec_dot_type].from_float_to_mat;
int64_t const vec_dot_num_rows = type_traits_cpu[type].nrows;
int64_t const matmul_num_cols = type_traits_cpu[type].ncols;
@ -7464,15 +7476,15 @@ static void ggml_compute_forward_mul_mat(
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
nb01/ggml_type_size(type),
(const char *)src1->data + i12*nb12 + i13*nb13,
nb11/ggml_type_size(src1->type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
type,
src1->type,
dst->type))
goto UseGgmlGemm1;
@ -7525,15 +7537,15 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
nb01/ggml_type_size(type),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
type,
vec_dot_type,
dst->type))
goto UseGgmlGemm2;
@ -7618,7 +7630,7 @@ UseGgmlGemm2:;
const int64_t ir1_start = dr1 * ith1;
const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
if (nth >= nchunk0 * nchunk1) {
break;
@ -7649,7 +7661,7 @@ static void ggml_compute_forward_mul_mat_id(
ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot;
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
int64_t const matmul_num_cols = type_traits_cpu[type].ncols;
ggml_gemv_t const gemv = type_traits_cpu[type].gemv;
@ -9159,12 +9171,6 @@ static void rope_yarn(
*sin_theta = sinf(theta) * mscale;
}
// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
}
static void ggml_rope_cache_init(
float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
float * cache, float sin_sign, float theta_scale) {
@ -9181,16 +9187,6 @@ static void ggml_rope_cache_init(
}
}
void ggml_rope_yarn_corr_dims(
int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
) {
// start and end correction dims
float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
float end = ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
dims[0] = MAX(0, start);
dims[1] = MIN(n_dims - 1, end);
}
static void ggml_compute_forward_rope_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
@ -10668,7 +10664,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
enum ggml_type const k_vec_dot_type = type_traits_cpu[k->type].vec_dot_type;
ggml_from_float_t const q_to_vec_dot = ggml_get_type_traits(k_vec_dot_type)->from_float;
ggml_from_float_t const q_to_vec_dot = type_traits_cpu[k_vec_dot_type].from_float;
ggml_vec_dot_t const kq_vec_dot = type_traits_cpu[k->type].vec_dot;
ggml_to_float_t const v_to_float = ggml_get_type_traits(v->type)->to_float;
@ -12220,11 +12216,16 @@ static void ggml_compute_forward_opt_step_adamw_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src0_grad = dst->src[1];
const struct ggml_tensor * src0_grad_m = dst->src[2];
const struct ggml_tensor * src0_grad_v = dst->src[3];
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src0_grad = dst->src[1];
const struct ggml_tensor * src0_grad_m = dst->src[2];
const struct ggml_tensor * src0_grad_v = dst->src[3];
const struct ggml_tensor * adamw_params = dst->src[4];
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
GGML_ASSERT(ggml_nelements(adamw_params) == 7);
const int ith = params->ith;
const int nth = params->nth;
@ -12241,16 +12242,14 @@ static void ggml_compute_forward_opt_step_adamw_f32(
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
/* const float gnorm = 1.0f; */
int64_t iter; memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
const float alpha = ggml_get_op_params_f32(dst, 2);
const float beta1 = ggml_get_op_params_f32(dst, 3);
const float beta2 = ggml_get_op_params_f32(dst, 4);
const float eps = ggml_get_op_params_f32(dst, 5);
const float wd = ggml_get_op_params_f32(dst, 6);
const float beta1h = alpha/(1.0f - powf(beta1, iter));
const float beta2h = 1.0f/(1.0f - powf(beta2, iter));
const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
const float alpha = adamw_params_ptr[0];
const float beta1 = adamw_params_ptr[1];
const float beta2 = adamw_params_ptr[2];
const float eps = adamw_params_ptr[3];
const float wd = adamw_params_ptr[4];
const float beta1h = adamw_params_ptr[5];
const float beta2h = adamw_params_ptr[6];
for (int ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
@ -12274,17 +12273,9 @@ static void ggml_compute_forward_opt_step_adamw_f32(
// The weight decay is applied independently of the Adam momenta m and v.
// This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
// See: https://arxiv.org/pdf/1711.05101v3.pdf
w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh;
w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
}
}
ggml_barrier(params->threadpool);
if (ith != 0) {
return;
}
iter++;
memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
}
static void ggml_compute_forward_opt_step_adamw(
@ -13759,6 +13750,151 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
return ggml_graph_compute(cgraph, &cplan);
}
int ggml_cpu_has_avx(void) {
#if defined(__AVX__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx_vnni(void) {
#if defined(__AVXVNNI__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx2(void) {
#if defined(__AVX2__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512(void) {
#if defined(__AVX512F__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512_vbmi(void) {
#if defined(__AVX512VBMI__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512_vnni(void) {
#if defined(__AVX512VNNI__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512_bf16(void) {
#if defined(__AVX512BF16__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_amx_int8(void) {
#if defined(__AMX_INT8__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_fma(void) {
#if defined(__FMA__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_arm_fma(void) {
#if defined(__ARM_FEATURE_FMA)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_riscv_v(void) {
#if defined(__riscv_v_intrinsic)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_f16c(void) {
#if defined(__F16C__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_fp16_va(void) {
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_wasm_simd(void) {
#if defined(__wasm_simd128__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_llamafile(void) {
#if defined(GGML_USE_LLAMAFILE)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_sse3(void) {
#if defined(__SSE3__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_ssse3(void) {
#if defined(__SSSE3__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_vsx(void) {
#if defined(__POWER9_VECTOR__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_neon(void) {
#if defined(__ARM_ARCH)
return ggml_arm_arch_features.has_neon;

663
ggml/src/ggml-cpu/ggml-cpu.cpp Normal file
View file

@ -0,0 +1,663 @@
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu-aarch64.h"
#include "ggml-impl.h"
#include <cctype>
#include <string>
#include <vector>
#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
#endif
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#endif
// ggml-backend interface
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
// buffer type AARCH64
static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *)ggml_aarch64_get_optimal_repack_type(tensor); // NOLINT
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
enum ggml_type repack_type = (enum ggml_type)(intptr_t)tensor->extra;
ggml_aarch64_repack_tensor(tensor, repack_type, data, size);
GGML_UNUSED(buffer);
}
static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_AARCH64";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
auto * buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
if (buffer == NULL) {
return NULL;
}
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ NULL,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_aarch64;
}
bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft) {
return buft == ggml_backend_cpu_aarch64_buffer_type();
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
static std::vector<ggml_backend_buffer_type_t> bufts = []() {
std::vector<ggml_backend_buffer_type_t> bufts;
#ifdef GGML_USE_CPU_HBM
bufts.push_back(ggml_backend_cpu_hbm_buffer_type());
#endif
#ifdef GGML_USE_CPU_AARCH64
bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
#endif
bufts.push_back(NULL);
return bufts;
}();
return bufts.data();
GGML_UNUSED(device);
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
int n_threads;
ggml_threadpool_t threadpool;
uint8_t * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
delete[] cpu_ctx->work_data;
delete cpu_ctx;
delete backend;
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
if (cpu_plan->cplan.work_data == NULL) {
delete cpu_plan;
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
delete[] cpu_plan->cplan.work_data;
delete cpu_plan;
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
cpu_ctx->work_data = new uint8_t[cplan.work_size];
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = (uint8_t *)cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .get_name = */ ggml_backend_cpu_get_name,
/* .free = */ ggml_backend_cpu_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->threadpool = NULL;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
if (ctx->threadpool && ctx->threadpool != threadpool) {
// already had a different threadpool, pause/suspend it before switching
ggml_threadpool_pause(ctx->threadpool);
}
ctx->threadpool = threadpool;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
// CPU backend - device
struct ggml_backend_cpu_device_context {
std::string description = "CPU";
ggml_backend_cpu_device_context() {
#ifdef __APPLE__
size_t len = 0;
if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
description.resize(len);
sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
}
#elif defined(__linux__)
FILE * f = fopen("/proc/cpuinfo", "r");
if (f) {
char buf[1024];
while (fgets(buf, sizeof(buf), f)) {
if (strncmp(buf, "model name", 10) == 0) {
char * p = strchr(buf, ':');
if (p) {
p++;
while (std::isspace(*p)) {
p++;
}
while (std::isspace(p[strlen(p) - 1])) {
p[strlen(p) - 1] = '\0';
}
description = p;
break;
}
}
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) == ERROR_SUCCESS) {
DWORD cpu_brand_size = 0;
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
NULL,
&cpu_brand_size) == ERROR_SUCCESS) {
description.resize(cpu_brand_size);
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
(LPBYTE)&description[0], // NOLINT
&cpu_brand_size) == ERROR_SUCCESS) {
if (description.find('\0') != std::string::npos) {
description.resize(description.find('\0'));
}
}
}
RegCloseKey(hKey);
}
#endif
}
};
static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
return "CPU";
GGML_UNUSED(dev);
}
static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// TODO
*free = 0;
*total = 0;
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_cpu_device_get_name(dev);
props->description = ggml_backend_cpu_device_get_description(dev);
props->type = ggml_backend_cpu_device_get_type(dev);
ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_cpu_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
if (op->op != GGML_OP_MUL_MAT || src0->type != GGML_TYPE_Q4_0 || ggml_aarch64_get_optimal_repack_type(src0) == GGML_TYPE_Q4_0) {
return false;
}
}
for (int i = 1; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) {
return false;
}
}
switch (op->op) {
case GGML_OP_CPY:
return
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
case GGML_OP_ROPE_BACK:
return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
case GGML_OP_IM2COL_BACK:
return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
case GGML_OP_OUT_PROD:
return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
default:
return true;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
/* .get_name = */ ggml_backend_cpu_device_get_name,
/* .get_description = */ ggml_backend_cpu_device_get_description,
/* .get_memory = */ ggml_backend_cpu_device_get_memory,
/* .get_type = */ ggml_backend_cpu_device_get_type,
/* .get_props = */ ggml_backend_cpu_device_get_props,
/* .init_backend = */ ggml_backend_cpu_device_init_backend,
/* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_cpu_device_supports_op,
/* .supports_buft = */ ggml_backend_cpu_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// CPU backend - backend (reg)
static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
return "CPU";
GGML_UNUSED(reg);
}
static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_cpu_device_context ctx;
static ggml_backend_device ggml_backend_cpu_device = {
/* .iface = */ ggml_backend_cpu_device_i,
/* .reg = */ reg,
/* .context = */ &ctx,
};
return &ggml_backend_cpu_device;
}
struct ggml_backend_feature {
const char * name;
const char * value;
};
// Not used yet
// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
static std::vector<ggml_backend_feature> features = []() {
std::vector<ggml_backend_feature> features;
if (ggml_cpu_has_sse3()) {
features.push_back({ "SSE3", "1" });
}
if (ggml_cpu_has_ssse3()) {
features.push_back({ "SSSE3", "1" });
}
if (ggml_cpu_has_avx()) {
features.push_back({ "AVX", "1" });
}
if (ggml_cpu_has_avx2()) {
features.push_back({ "AVX2", "1" });
}
if (ggml_cpu_has_f16c()) {
features.push_back({ "F16C", "1" });
}
if (ggml_cpu_has_fma()) {
features.push_back({ "FMA", "1" });
}
if (ggml_cpu_has_avx_vnni()) {
features.push_back({ "AVX_VNNI", "1" });
}
if (ggml_cpu_has_avx512()) {
features.push_back({ "AVX512", "1" });
}
if (ggml_cpu_has_avx512_vbmi()) {
features.push_back({ "AVX512_VBMI", "1" });
}
if (ggml_cpu_has_avx512_vnni()) {
features.push_back({ "AVX512_VNNI", "1" });
}
if (ggml_cpu_has_avx512_bf16()) {
features.push_back({ "AVX512_BF16", "1" });
}
if (ggml_cpu_has_amx_int8()) {
features.push_back({ "AMX_INT8", "1" });
}
if (ggml_cpu_has_neon()) {
features.push_back({ "NEON", "1" });
}
if (ggml_cpu_has_arm_fma()) {
features.push_back({ "ARM_FMA", "1" });
}
if (ggml_cpu_has_fp16_va()) {
features.push_back({ "FP16_VA", "1" });
}
if (ggml_cpu_has_matmul_int8()) {
features.push_back({ "MATMUL_INT8", "1" });
}
if (ggml_cpu_has_sve()) {
features.push_back({ "SVE", "1" });
}
if (ggml_cpu_get_sve_cnt() > 0) {
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
if (ggml_cpu_has_vsx()) {
features.push_back({ "VSX", "1" });
}
if (ggml_cpu_has_wasm_simd()) {
features.push_back({ "WASM_SIMD", "1" });
}
if (ggml_cpu_has_llamafile()) {
features.push_back({ "LLAMAFILE", "1" });
}
features.push_back({ nullptr, nullptr });
return features;
}();
return features.data();
GGML_UNUSED(reg);
}
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
return (void *)ggml_backend_cpu_set_n_threads;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
return (void *)ggml_backend_cpu_get_extra_bufts;
}
return NULL;
GGML_UNUSED(reg);
}
static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
/* .get_name = */ ggml_backend_cpu_reg_get_name,
/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
/* .get_device = */ ggml_backend_cpu_reg_get_device,
/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
// init CPU feature detection
ggml_cpu_init();
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}

3
ggml/src/llamafile/sgemm.cpp → ggml/src/ggml-cpu/llamafile/sgemm.cpp
View file

@ -50,7 +50,8 @@
#include "sgemm.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
// hack until moved into the CPU backend
#include "../ggml-cpu-impl.h"
#include "ggml-quants.h"
#ifdef _MSC_VER

0
ggml/src/llamafile/sgemm.h → ggml/src/ggml-cpu/llamafile/sgemm.h
View file

155
ggml/src/ggml-cuda/CMakeLists.txt Normal file
View file

@ -0,0 +1,155 @@
cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES
find_package(CUDAToolkit)
if (CUDAToolkit_FOUND)
message(STATUS "CUDA Toolkit found")
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# native == GPUs available at build time
# 52 == Maxwell, lowest CUDA 12 standard
# 60 == P100, FP16 CUDA intrinsics
# 61 == Pascal, __dp4a instruction (per-byte integer dot product)
# 70 == V100, FP16 tensor cores
# 75 == Turing, int8 tensor cores
if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6")
set(CMAKE_CUDA_ARCHITECTURES "native")
elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
enable_language(CUDA)
file(GLOB GGML_HEADERS_CUDA "*.cuh")
list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
file(GLOB GGML_SOURCES_CUDA "*.cu")
file(GLOB SRCS "template-instances/fattn-wmma*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/mmq*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
if (GGML_CUDA_FA_ALL_QUANTS)
file(GLOB SRCS "template-instances/fattn-vec*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
else()
file(GLOB SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/fattn-vec*f16-f16.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
endif()
add_library(ggml-cuda
${GGML_HEADERS_CUDA}
${GGML_SOURCES_CUDA}
)
target_link_libraries(ggml-cuda PRIVATE ggml-base)
target_include_directories(ggml-cuda PRIVATE . ..)
add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
if (GGML_CUDA_GRAPHS)
add_compile_definitions(GGML_CUDA_USE_GRAPHS)
endif()
if (GGML_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_CUDA_FORCE_MMQ)
endif()
if (GGML_CUDA_FORCE_CUBLAS)
add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
endif()
if (GGML_CUDA_NO_VMM)
add_compile_definitions(GGML_CUDA_NO_VMM)
endif()
if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
add_compile_definitions(GGML_CUDA_F16)
endif()
if (GGML_CUDA_NO_PEER_COPY)
add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
endif()
if (GGML_STATIC)
if (WIN32)
# As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
else ()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
endif()
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
endif()
if (GGML_CUDA_NO_VMM)
# No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
endif()
set(CUDA_CXX_FLAGS "")
set(CUDA_FLAGS -use_fast_math)
if (GGML_FATAL_WARNINGS)
list(APPEND CUDA_FLAGS -Werror all-warnings)
endif()
if (GGML_ALL_WARNINGS AND NOT MSVC)
set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
endif()
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler --version
OUTPUT_VARIABLE CUDA_CCFULLVER
ERROR_QUIET
)
if (NOT CUDA_CCFULLVER MATCHES clang)
set(CUDA_CCID "GNU")
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
OUTPUT_VARIABLE CUDA_CCVER
ERROR_QUIET
)
else()
if (CUDA_CCFULLVER MATCHES Apple)
set(CUDA_CCID "AppleClang")
else()
set(CUDA_CCID "Clang")
endif()
string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
endif()
message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
get_flags(${CUDA_CCID} ${CUDA_CCVER})
list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later
endif()
if (NOT MSVC)
list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
endif()
list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED) # pass host compiler flags as a single argument
if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
endif()
add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
else()
message(FATAL_ERROR "CUDA Toolkit not found")
endif()

50
ggml/src/ggml-cuda/common.cuh
View file

@ -6,7 +6,7 @@
#include <cstdint>
#include <memory>
#if defined(GGML_USE_HIPBLAS)
#if defined(GGML_USE_HIP)
#define GGML_COMMON_DECL_HIP
#define GGML_COMMON_IMPL_HIP
#else
@ -26,13 +26,13 @@
#include <string>
#include <vector>
#if defined(GGML_USE_HIPBLAS)
#if defined(GGML_USE_HIP)
#include "vendors/hip.h"
#elif defined(GGML_USE_MUSA)
#include "vendors/musa.h"
#else
#include "vendors/cuda.h"
#endif // defined(GGML_USE_HIPBLAS)
#endif // defined(GGML_USE_HIP)
#define STRINGIZE_IMPL(...) #__VA_ARGS__
#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
@ -97,7 +97,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
#if !defined(GGML_USE_HIPBLAS)
#if !defined(GGML_USE_HIP)
static const char * cu_get_error_str(CUresult err) {
const char * err_str;
cuGetErrorString(err, &err_str);
@ -120,21 +120,21 @@ typedef float dfloat; // dequantize float
typedef float2 dfloat2;
#endif // GGML_CUDA_F16
#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#define FP16_AVAILABLE
#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
#define FAST_FP16_AVAILABLE
#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#define FP16_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#define INT8_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
#define FLASH_ATTN_AVAILABLE
@ -156,14 +156,14 @@ static constexpr bool int8_mma_available(const int cc) {
static __device__ void no_device_code(
const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
file_name, line, function_name, arch);
GGML_UNUSED(arch_list);
#else
printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
file_name, line, function_name, arch, arch_list);
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
__trap();
GGML_UNUSED(no_device_code); // suppress unused function warning
@ -176,7 +176,7 @@ static __device__ void no_device_code(
#endif // __CUDA_ARCH__
static __device__ __forceinline__ int warp_reduce_sum(int x) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
return __reduce_add_sync(0xffffffff, x);
#else
#pragma unroll
@ -184,7 +184,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
x += __shfl_xor_sync(0xffffffff, x, mask, 32);
}
return x;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
}
static __device__ __forceinline__ float warp_reduce_sum(float x) {
@ -207,7 +207,7 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
#ifdef FP16_AVAILABLE
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
@ -221,7 +221,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
}
return a;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#else
NO_DEVICE_CODE;
@ -240,11 +240,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
#ifdef FP16_AVAILABLE
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
return __float2half(fmaxf(__half2float(a), __half2float(b)));
#else
return __hmax(a, b);
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#else
NO_DEVICE_CODE;
@ -254,7 +254,7 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
}
static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
#if CUDART_VERSION >= CUDART_HMAX
return __hmax2(a, b);
@ -269,11 +269,11 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
GGML_UNUSED(a);
GGML_UNUSED(b);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
}
static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
@ -282,7 +282,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
#else
GGML_UNUSED(x);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
}
#if CUDART_VERSION < CUDART_HMASK
@ -294,7 +294,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
#endif // CUDART_VERSION < CUDART_HMASK
static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
c = __builtin_amdgcn_sdot4(a, b, c, false);
#elif defined(RDNA3)
@ -320,7 +320,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
#endif
return c;
#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if __CUDA_ARCH__ >= MIN_CC_DP4A
return __dp4a(a, b, c);
@ -330,7 +330,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
}
// TODO: move to ggml-common.h

699
ggml/src/ggml-cuda/dmmv.cu
View file

@ -1,699 +0,0 @@
#include "dmmv.cuh"
#include "dequantize.cuh"
#include "convert.cuh"
#ifndef K_QUANTS_PER_ITERATION
#define K_QUANTS_PER_ITERATION 2
#else
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
#endif
static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
const int row = blockIdx.x*blockDim.y + threadIdx.y;
if (row > nrows) return;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q2_K * x = (const block_q2_K *)vx + ib0;
float tmp = 0; // partial sum for thread in warp
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...15
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2
const int q_offset = 32*im + l0;
const int s_offset = 8*im;
const int y_offset = 128*im + l0;
uint32_t aux[4];
const uint8_t * d = (const uint8_t *)aux;
const uint8_t * m = (const uint8_t *)(aux + 2);
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + y_offset;
const uint8_t * q = x[i].qs + q_offset;
const float dall = __low2half(x[i].dm);
const float dmin = __high2half(x[i].dm);
const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
aux[0] = a[0] & 0x0f0f0f0f;
aux[1] = a[1] & 0x0f0f0f0f;
aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
float sum1 = 0, sum2 = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+ y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+ y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+ y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+ y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+ y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+ y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+ y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
}
tmp += dall * sum1 - dmin * sum2;
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
const int row = blockIdx.x*blockDim.y + threadIdx.y;
if (row > nrows) return;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q3_K * x = (const block_q3_K *)vx + ib0;
float tmp = 0; // partial sum for thread in warp
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0....15 or 0...7
const uint8_t m = 1 << (4*im);
const int l0 = n*in; // 0...15 or 0...14 in steps of 2
const int q_offset = 32*im + l0;
const int y_offset = 128*im + l0;
uint16_t utmp[4];
const int8_t * s = (const int8_t *)utmp;
const uint16_t s_shift = 4*im;
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + y_offset;
const uint8_t * q = x[i].qs + q_offset;
const uint8_t * h = x[i].hmask + l0;
const uint16_t * a = (const uint16_t *)x[i].scales;
utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
const float d = x[i].d;
float sum = 0;
for (int l = 0; l < n; ++l) {
sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+ y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+ y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+ y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+ y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+ y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+ y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
}
tmp += d * sum;
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
const int row = blockIdx.x*blockDim.y + threadIdx.y;
if (row > nrows) return;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q4_K * x = (const block_q4_K *)vx + ib0;
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4
const int il = tid/step; // 0...3
const int ir = tid - step*il; // 0...7 or 0...3
const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
#if K_QUANTS_PER_ITERATION == 2
uint32_t q32[4];
const uint8_t * q4 = (const uint8_t *)q32;
#else
uint16_t q16[4];
const uint8_t * q4 = (const uint8_t *)q16;
#endif
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const float * y1 = yy + i*QK_K + y_offset;
const float * y2 = y1 + 128;
const float dall = __low2half(x[i].dm);
const float dmin = __high2half(x[i].dm);
const uint16_t * a = (const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
#if K_QUANTS_PER_ITERATION == 2
const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
const uint32_t * q2 = q1 + 16;
q32[0] = q1[0] & 0x0f0f0f0f;
q32[1] = q1[0] & 0xf0f0f0f0;
q32[2] = q2[0] & 0x0f0f0f0f;
q32[3] = q2[0] & 0xf0f0f0f0;
float4 s = {0.f, 0.f, 0.f, 0.f};
float smin = 0;
for (int l = 0; l < 4; ++l) {
s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
}
tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
#else
const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
const uint16_t * q2 = q1 + 32;
q16[0] = q1[0] & 0x0f0f;
q16[1] = q1[0] & 0xf0f0;
q16[2] = q2[0] & 0x0f0f;
q16[3] = q2[0] & 0xf0f0;
float4 s = {0.f, 0.f, 0.f, 0.f};
float smin = 0;
for (int l = 0; l < 2; ++l) {
s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
}
tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
#endif
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (tid == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
const int row = blockIdx.x;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q5_K * x = (const block_q5_K *)vx + ib0;
float tmp = 0; // partial sum for thread in warp
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int tid = threadIdx.x/2; // 0...15
const int ix = threadIdx.x%2;
const int il = tid/4; // 0...3
const int ir = tid - 4*il;// 0...3
const int n = 2;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
const uint8_t hm1 = 1 << (2*im);
const uint8_t hm2 = hm1 << 4;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
uint16_t q16[8];
const uint8_t * q4 = (const uint8_t *)q16;
for (int i = ix; i < num_blocks_per_row; i += 2) {
const uint8_t * ql1 = x[i].qs + q_offset;
const uint8_t * qh = x[i].qh + l0;
const float * y1 = yy + i*QK_K + y_offset;
const float * y2 = y1 + 128;
const float dall = __low2half(x[i].dm);
const float dmin = __high2half(x[i].dm);
const uint16_t * a = (const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 sum = {0.f, 0.f, 0.f, 0.f};
float smin = 0;
const uint16_t * q1 = (const uint16_t *)ql1;
const uint16_t * q2 = q1 + 32;
q16[0] = q1[0] & 0x0f0f;
q16[1] = q1[8] & 0x0f0f;
q16[2] = (q1[0] >> 4) & 0x0f0f;
q16[3] = (q1[8] >> 4) & 0x0f0f;
q16[4] = q2[0] & 0x0f0f;
q16[5] = q2[8] & 0x0f0f;
q16[6] = (q2[0] >> 4) & 0x0f0f;
q16[7] = (q2[8] >> 4) & 0x0f0f;
for (int l = 0; l < n; ++l) {
sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+ y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+ y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+ y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+ y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+ (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
}
tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
const int row = blockIdx.x*blockDim.y + threadIdx.y;
if (row > nrows) return;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q6_K * x = (const block_q6_K *)vx + ib0;
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
#if K_QUANTS_PER_ITERATION == 1
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
const int is = 0;
#else
const int l0 = 4 * in; // 0, 4, 8, ..., 28
const int is = in / 4;
#endif
const int ql_offset = 64*im + l0;
const int qh_offset = 32*im + l0;
const int s_offset = 8*im + is;
const int y_offset = 128*im + l0;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + y_offset;
const uint8_t * ql = x[i].ql + ql_offset;
const uint8_t * qh = x[i].qh + qh_offset;
const int8_t * s = x[i].scales + s_offset;
const float d = x[i].d;
#if K_QUANTS_PER_ITERATION == 1
float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+ y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+ y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+ y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+ y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+ y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+ y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
tmp += sum;
#else
float sum = 0;
for (int l = 0; l < 4; ++l) {
sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+ y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+ y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+ y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
}
tmp += sum;
#endif
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (tid == 0) {
dst[row] = tmp;
}
}
static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const half * x = (const half *) vx;
// load 2 halfs into register in a single instruction
const half2 x_reg = *((half2 *) &(x[ib + iqs]));
// automatic half -> float type cast if dfloat == float
v.x = __low2float(x_reg);
v.y = __high2float(x_reg);
}
static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
type == GGML_TYPE_F16 ? convert_f16 :
nullptr;
}
template <ggml_type type>
static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
if (row >= nrows) {
return;
}
const int tid = threadIdx.x;
const int iter_stride = 2*GGML_CUDA_DMMV_X;
const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
const int y_offset = qr == 1 ? 1 : qk/2;
// partial sum for each thread
#ifdef GGML_CUDA_F16
half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
#else
float tmp = 0.0f;
#endif // GGML_CUDA_F16
for (int i = 0; i < ncols; i += iter_stride) {
const int col = i + vals_per_iter*tid;
const int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
const int iqs = (col%qk)/qr; // x quant index
const int iybs = col - col%qk; // y block start index
// processing >2 values per i iter is faster for fast GPUs
#pragma unroll
for (int j = 0; j < vals_per_iter; j += 2) {
// process 2 vals per j iter
// dequantize
// for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
dfloat2 v;
dequantize_kernel(vx, ib, iqs + j/qr, v);
// matrix multiplication
// for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
#ifdef GGML_CUDA_F16
if ( y_offset == 1 ) {
// load 2 dfloats into register in a single instruction
const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
tmp += __hmul2(v, y_reg);
}
else {
tmp += __hmul2(v, {
y[iybs + iqs + j/qr + 0],
y[iybs + iqs + j/qr + y_offset]
});
}
#else
if ( y_offset == 1 ) {
// load 2 dfloats into register in a single instruction
const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
tmp += v.x * y_reg.x;
tmp += v.y * y_reg.y;
}
else {
tmp += v.x * y[iybs + iqs + j/qr + 0];
tmp += v.y * y[iybs + iqs + j/qr + y_offset];
}
#endif // GGML_CUDA_F16
}
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (tid == 0) {
#ifdef GGML_CUDA_F16
dst[row] = tmp.x + tmp.y;
#else
dst[row] = tmp;
#endif // GGML_CUDA_F16
}
}
static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
// the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2 / K_QUANTS_PER_ITERATION;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2 / K_QUANTS_PER_ITERATION;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const dim3 block_dims(32, 1, 1);
dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2 / K_QUANTS_PER_ITERATION;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
const dim3 block_nums(block_num_y, 1, 1);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
dequantize_mul_mat_vec<GGML_TYPE_F16>
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
void ggml_cuda_op_dequantize_mul_mat_vec(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream) {
GGML_UNUSED(ctx);
const int64_t ne00 = src0->ne[0];
const int64_t row_diff = row_high - row_low;
GGML_ASSERT(src1->type == GGML_TYPE_F32);
// on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
#ifdef GGML_CUDA_F16
ggml_cuda_pool_alloc<half> src1_dfloat_a(ctx.pool());
half * src1_dfloat = nullptr; // dfloat == half
bool src1_convert_f16 =
src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
if (src1_convert_f16) {
src1_dfloat = src1_dfloat_a.alloc(ne00);
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
GGML_ASSERT(to_fp16_cuda != nullptr);
to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
}
#else
const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
#endif // GGML_CUDA_F16
switch (src0->type) {
case GGML_TYPE_Q4_0:
dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q4_1:
dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q5_0:
dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q5_1:
dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q8_0:
dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q2_K:
dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q3_K:
dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q4_K:
dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q5_K:
dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_Q6_K:
dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
break;
case GGML_TYPE_F16:
convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
break;
default:
GGML_ABORT("fatal error");
break;
}
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_ddq_i);
GGML_UNUSED(src1_ncols);
GGML_UNUSED(src1_padded_row_size);
}
bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
src0_type == GGML_TYPE_F16;
}

4
ggml/src/ggml-cuda/fattn-common.cuh
View file

@ -517,9 +517,9 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
}
template<int D, int parallel_blocks> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_combine_results(
const float * __restrict__ VKQ_parts,
const float2 * __restrict__ VKQ_meta,

4
ggml/src/ggml-cuda/fattn-tile-f16.cu
View file

@ -5,9 +5,9 @@
#define FATTN_KQ_STRIDE_TILE_F16 64
template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-tile-f32.cu
View file

@ -5,9 +5,9 @@
#define FATTN_KQ_STRIDE_TILE_F32 32
template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f32(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-vec-f16.cuh
View file

@ -2,9 +2,9 @@
#include "fattn-common.cuh"
template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_vec_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-vec-f32.cuh
View file

@ -2,9 +2,9 @@
#include "fattn-common.cuh"
template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_vec_ext_f32(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-wmma-f16.cuh
View file

@ -7,9 +7,9 @@
// D == head size, VKQ_stride == num VKQ rows calculated in parallel:
template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,

234
ggml/src/ggml-cuda.cu → ggml/src/ggml-cuda/ggml-cuda.cu
View file

@ -16,11 +16,11 @@
#include "ggml-cuda/cpy.cuh"
#include "ggml-cuda/cross-entropy-loss.cuh"
#include "ggml-cuda/diagmask.cuh"
#include "ggml-cuda/dmmv.cuh"
#include "ggml-cuda/fattn.cuh"
#include "ggml-cuda/getrows.cuh"
#include "ggml-cuda/im2col.cuh"
#include "ggml-cuda/mmq.cuh"
#include "ggml-cuda/mmv.cuh"
#include "ggml-cuda/mmvq.cuh"
#include "ggml-cuda/norm.cuh"
#include "ggml-cuda/opt-step-adamw.cuh"
@ -91,7 +91,7 @@ int ggml_cuda_get_device() {
static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
ggml_cuda_set_device(device);
#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
auto res = hipMallocManaged(ptr, size);
if (res == hipSuccess) {
// if error we "need" to know why...
@ -100,7 +100,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
return res;
#else
#if !defined(GGML_USE_HIPBLAS)
#if !defined(GGML_USE_HIP)
cudaError_t err;
if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
{
@ -113,7 +113,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
return err;
#else
return cudaMalloc(ptr, size);
#endif // !defined(GGML_USE_HIPBLAS)
#endif // !defined(GGML_USE_HIP)
#endif
}
@ -151,7 +151,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
for (int id = 0; id < info.device_count; ++id) {
int device_vmm = 0;
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
CUdevice device;
CU_CHECK(cuDeviceGet(&device, id));
CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@ -163,7 +163,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
alloc_prop.location.id = id;
CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
}
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
info.devices[id].vmm = !!device_vmm;
cudaDeviceProp prop;
@ -175,13 +175,13 @@ static ggml_cuda_device_info ggml_cuda_init() {
info.devices[id].nsm = prop.multiProcessorCount;
info.devices[id].smpb = prop.sharedMemPerBlock;
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
info.devices[id].smpbo = prop.sharedMemPerBlock;
info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
#else
info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
info.devices[id].cc = 100*prop.major + 10*prop.minor;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
}
for (int id = 0; id < info.device_count; ++id) {
@ -299,7 +299,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
};
// pool with virtual memory
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
@ -393,14 +393,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
}
};
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
if (ggml_cuda_info().devices[device].vmm) {
return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
}
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
}
@ -1020,114 +1020,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
#define MUL_MAT_SRC1_COL_STRIDE 128
static __global__ void mul_mat_p021_f16_f32(
const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
const half * x = (const half *) vx;
const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
const int channel = blockDim.z*blockIdx.z + threadIdx.z;
const int channel_x = channel / (nchannels_y / nchannels_x);
const int nrows_y = ncols_x;
const int nrows_dst = nrows_x;
const int row_dst = row_x;
float tmp = 0.0f;
for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
const int col_x = col_x0 + threadIdx.x;
if (col_x >= ncols_x) {
break;
}
// x is transposed and permuted
const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
const float xi = __half2float(x[ix]);
const int row_y = col_x;
// y is not transposed but permuted
const int iy = channel*nrows_y + row_y;
tmp += xi * y[iy];
}
// dst is not transposed and not permuted
const int idst = channel*nrows_dst + row_dst;
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[idst] = tmp;
}
}
static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
const half * x = (const half *) vx;
const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
const int channel = blockDim.z*blockIdx.z + threadIdx.z;
const int channel_x = channel / channel_x_divisor;
const int nrows_y = ncols_x;
const int nrows_dst = nrows_x;
const int row_dst = row_x;
const int idst = channel*nrows_dst + row_dst;
float tmp = 0.0f;
for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
const int col_x = col_x0 + threadIdx.x;
if (col_x >= ncols_x) {
break;
}
const int row_y = col_x;
const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
const int iy = channel*nrows_y + row_y;
const float xi = __half2float(x[ix]);
tmp += xi * y[iy];
}
// sum up partial sums and write back result
tmp = warp_reduce_sum(tmp);
if (threadIdx.x == 0) {
dst[idst] = tmp;
}
}
static void ggml_mul_mat_p021_f16_f32_cuda(
const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
const dim3 block_nums(1, nrows_x, nchannels_y);
const dim3 block_dims(WARP_SIZE, 1, 1);
mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
}
static void ggml_mul_mat_vec_nc_f16_f32_cuda(
const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
const dim3 block_nums(1, nrows_x, nchannels_y);
const dim3 block_dims(WARP_SIZE, 1, 1);
mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
(vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
}
static cudaError_t ggml_cuda_cpy_tensor_2d(
void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
@ -1325,7 +1217,7 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
// cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
cudaMemcpy3DPeerParms p = {};
p.dstDevice = dstDevice;
@ -1339,7 +1231,7 @@ static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
GGML_UNUSED(dstDevice);
GGML_UNUSED(srcDevice);
return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
}
static void ggml_cuda_op_mul_mat(
@ -1654,58 +1546,6 @@ static void ggml_cuda_op_mul_mat(
}
}
static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne12 = src1->ne[2];
cudaStream_t main_stream = ctx.stream();
void * src0_ddq = src0->data;
float * src1_ddf = (float *) src1->data;
float * dst_ddf = (float *) dst->data;
ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
}
static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(!ggml_is_permuted(src0));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t nb01 = src0->nb[1];
const int64_t nb02 = src0->nb[2];
const int64_t ne12 = src1->ne[2];
cudaStream_t main_stream = ctx.stream();
void * src0_ddq = src0->data;
float * src1_ddf = (float *) src1->data;
float * dst_ddf = (float *) dst->data;
const int64_t row_stride_x = nb01 / sizeof(half);
const int64_t channel_stride_x = nb02 / sizeof(half);
ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
}
static __global__ void k_compute_batched_ptrs(
const half * src0_as_f16, const half * src1_as_f16, char * dst,
const void ** ptrs_src, void ** ptrs_dst,
@ -1879,21 +1719,17 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
bool use_mul_mat_vec = src0->type == GGML_TYPE_F16
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
&& src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
bool use_mul_mat_q = ggml_is_quantized(src0->type)
bool use_mul_mat_q = ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
// if mmvq is available it's a better choice than dmmv:
#ifndef GGML_CUDA_FORCE_DMMV
use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
#endif // GGML_CUDA_FORCE_DMMV
bool any_gpus_with_slow_fp16 = false;
bool any_gpus_with_slow_fp16 = false;
bool any_gpus_without_fp16_mma = false;
if (split) {
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@ -1904,14 +1740,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
continue;
}
const int cc = ggml_cuda_info().devices[id].cc;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
const int cc = ggml_cuda_info().devices[id].cc;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
}
} else {
const int cc = ggml_cuda_info().devices[ctx.device].cc;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
const int cc = ggml_cuda_info().devices[ctx.device].cc;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
}
// debug helpers
@ -1922,18 +1760,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
//printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
//printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
// FP32 precision KQ single-batch for batch size 1 without FlashAttention
ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
} else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
// FP32 precision KQV single-batch for batch size 1 without FlashAttention
ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
if (!split && use_mul_mat_vec && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
// the custom F16 vector kernel can be used over batched cuBLAS GEMM
// but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
&& !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
// KQ + KQV multi-batch without FlashAttention
// general KQ + KQV multi-batch without FlashAttention
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
} else if (use_dequantize_mul_mat_vec) {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
} else if (use_mul_mat_vec) {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
} else if (use_mul_mat_vec_q) {
ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
} else if (use_mul_mat_q) {

22
ggml/src/ggml-cuda/mmq.cuh
View file

@ -100,9 +100,9 @@ static constexpr __device__ int get_mmq_x_max_device() {
return 128;
#else // INT8_MMA_AVAILABLE
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
return 128;
#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if __CUDA_ARCH__ >= CC_VOLTA
#ifdef GGML_CUDA_FORCE_MMQ
@ -115,7 +115,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
return 64;
#endif // __CUDA_ARCH__ >= CC_VOLTA
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#endif // INT8_MMA_AVAILABLE
}
@ -124,7 +124,7 @@ static constexpr int get_mmq_y_host(const int cc) {
}
static constexpr __device__ int get_mmq_y_device() {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if defined(RDNA1)
return 64;
#else
@ -136,7 +136,7 @@ static constexpr __device__ int get_mmq_y_device() {
#else
return 64;
#endif // __CUDA_ARCH__ >= CC_VOLTA
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
}
#define MMQ_DP4A_TXS_Q4_0 tile_x_sizes{mmq_y*WARP_SIZE + mmq_y, mmq_y*WARP_SIZE/QI4_0 + mmq_y/QI4_0, 0}
@ -2569,7 +2569,7 @@ static __device__ void mul_mat_q_process_tile(
// The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
template <ggml_type type, int mmq_x, int nwarps, bool need_check>
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if defined(RDNA3) || defined(RDNA2)
__launch_bounds__(WARP_SIZE*nwarps, 2)
#endif // defined(RDNA3) || defined(RDNA2)
@ -2579,7 +2579,7 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
#else
__launch_bounds__(WARP_SIZE*nwarps, 2)
#endif // __CUDA_ARCH__ >= CC_VOLTA
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
static __global__ void mul_mat_q(
const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
@ -2594,7 +2594,7 @@ static __global__ void mul_mat_q(
constexpr int mmq_y = get_mmq_y_device();
// On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
{
constexpr bool fixup = false;
mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
@ -2602,7 +2602,7 @@ static __global__ void mul_mat_q(
blockIdx.x, blockIdx.y, 0, ne00/qk);
return;
}
#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
const int64_t blocks_per_ne00 = ne00 / qk;
constexpr int blocks_per_iter = MMQ_ITER_K / qk;
@ -2765,14 +2765,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
if (!shmem_limit_raised[id]) {
CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
shmem_limit_raised[id] = true;
}
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;

223
ggml/src/ggml-cuda/mmv.cu Normal file
View file

@ -0,0 +1,223 @@
#include "common.cuh"
#include "mmv.cuh"
template <typename type_acc, int block_size>
static __global__ void mul_mat_vec(
const half * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst) {
const int64_t row = blockIdx.x;
const int64_t channel = blockIdx.z;
const int tid = threadIdx.x;
x += (channel/channel_ratio)*stride_channel_x + row*stride_row;
y += channel *stride_channel_y;
dst += channel *stride_channel_dst;
const half2 * x2 = (const half2 *) x;
const float2 * y2 = (const float2 *) y;
extern __shared__ char data_mmv[];
float * buf_iw = (float *) data_mmv;
if (block_size > WARP_SIZE) {
if (tid < WARP_SIZE) {
buf_iw[tid] = 0.0f;
}
__syncthreads();
}
float sumf;
if (std::is_same<type_acc, float>::value) {
sumf = 0.0f;
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmpx = __half22float2(x2[col2]);
const float2 tmpy = y2[col2];
sumf += tmpx.x * tmpy.x;
sumf += tmpx.y * tmpy.y;
}
} else {
#ifdef FP16_AVAILABLE
half2 sumh2 = make_half2(0.0f, 0.0f);
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmp = y2[col2];
sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
}
sumf = __low2float(sumh2) + __high2float(sumh2);
#else
NO_DEVICE_CODE;
#endif // FP16_AVAILABLE
}
sumf = warp_reduce_sum(sumf);
if (block_size > WARP_SIZE) {
buf_iw[tid/WARP_SIZE] = sumf;
__syncthreads();
if (tid > WARP_SIZE) {
return;
}
sumf = buf_iw[tid];
sumf = warp_reduce_sum(sumf);
}
if (tid != 0) {
return;
}
dst[row] = sumf;
}
template <typename type_acc>
static void launch_mul_mat_vec_cuda(
const half * x, const float * y, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
cudaStream_t stream) {
GGML_ASSERT(ncols % 2 == 0);
GGML_ASSERT(stride_row % 2 == 0);
GGML_ASSERT(nchannels_y % nchannels_x == 0);
const int64_t channel_ratio = nchannels_y / nchannels_x;
int64_t block_size_best = WARP_SIZE;
int64_t niter_best = (ncols + 2*WARP_SIZE - 1) / (2*WARP_SIZE);
for (int64_t block_size = 2*WARP_SIZE; block_size <= 256; block_size += WARP_SIZE) {
const int64_t niter = (ncols + 2*block_size - 1) / (2*block_size);
if (niter < niter_best) {
niter_best = niter;
block_size_best = block_size;
}
}
const int smem = WARP_SIZE*sizeof(float);
const dim3 block_nums(nrows, 1, nchannels_y);
const dim3 block_dims(block_size_best, 1, 1);
switch (block_size_best) {
case 32: {
mul_mat_vec<type_acc, 32><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 64: {
mul_mat_vec<type_acc, 64><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 96: {
mul_mat_vec<type_acc, 96><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 128: {
mul_mat_vec<type_acc, 128><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 160: {
mul_mat_vec<type_acc, 160><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 192: {
mul_mat_vec<type_acc, 192><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 224: {
mul_mat_vec<type_acc, 224><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
case 256: {
mul_mat_vec<type_acc, 256><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
} break;
default: {
GGML_ABORT("fatal error");
} break;
}
}
static void mul_mat_vec_cuda(
const half * x, const float * y, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
enum ggml_prec prec, cudaStream_t stream) {
switch (prec) {
case GGML_PREC_DEFAULT: {
launch_mul_mat_vec_cuda<half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
stride_channel_x, stride_channel_y, stride_channel_dst, stream);
} break;
case GGML_PREC_F32: {
launch_mul_mat_vec_cuda<float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
stride_channel_x, stride_channel_y, stride_channel_dst, stream);
} break;
}
}
void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
GGML_ASSERT(src1->ne[1] == 1);
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
const half * src0_d = (const half *) src0->data;
const float * src1_d = (const float *) src1->data;
float * dst_d = (float *) dst->data;
const int64_t ne02 = src0->ne[2];
const int64_t ne12 = src1->ne[2];
GGML_ASSERT(dst->ne[2] == ne12);
GGML_ASSERT(src0->ne[3] == 1);
GGML_ASSERT(src1->ne[3] == 1);
GGML_ASSERT( dst->ne[3] == 1);
const int64_t stride_row = src0->nb[1] / ggml_type_size(src0->type);
const int64_t channel_stride_x = src0->nb[2] / ggml_type_size(src0->type);
const int64_t channel_stride_y = src1->nb[2] / ggml_type_size(src1->type);
const int64_t channel_stride_dst = dst->nb[2] / ggml_type_size( dst->type);
mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12, channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
}
void ggml_cuda_op_mul_mat_vec(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
const int64_t ne00 = src0->ne[0];
const int64_t row_diff = row_high - row_low;
GGML_ASSERT(src1_ncols == 1);
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
// ggml_cuda_op provides single, contiguous matrices
const int64_t stride_row = ne00;
const int64_t nchannels_x = 1;
const int64_t nchannels_y = 1;
const int64_t channel_stride_x = 0;
const int64_t channel_stride_y = 0;
const int64_t channel_stride_dst = 0;
mul_mat_vec_cuda((const half *) src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
GGML_UNUSED(ctx);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_ddq_i);
GGML_UNUSED(src1_ncols);
GGML_UNUSED(src1_padded_row_size);
}

16
ggml/src/ggml-cuda/dmmv.cuh → ggml/src/ggml-cuda/mmv.cuh
View file

@ -1,20 +1,12 @@
#include "common.cuh"
// dmmv = dequantize_mul_mat_vec
// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
#define MMV_MAX_ROWS 512
// TODO: remove this?
#ifndef GGML_CUDA_DMMV_X
#define GGML_CUDA_DMMV_X 32
#endif
void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
#ifndef GGML_CUDA_MMV_Y
#define GGML_CUDA_MMV_Y 1
#endif
void ggml_cuda_op_dequantize_mul_mat_vec(
void ggml_cuda_op_mul_mat_vec(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
const int64_t src1_padded_row_size, cudaStream_t stream);
bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);

8
ggml/src/ggml-cuda/mmvq.cu
View file

@ -48,10 +48,10 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
}
template <ggml_type type, int ncols_y>
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
// tell the compiler to use as many registers as it wants, see nwarps definition below
__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void mul_mat_vec_q(
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
constexpr int nwarps = 1;
constexpr int rows_per_cuda_block = 1;
#else
constexpr int nwarps = ncols_y <= 4 ? 4 : 2;
constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
const int row0 = rows_per_cuda_block*blockIdx.x;

66
ggml/src/ggml-cuda/opt-step-adamw.cu
View file

@ -1,11 +1,11 @@
#include "ggml-impl.h"
#include "opt-step-adamw.cuh"
#include <cstdint>
static __global__ void opt_step_adamw_f32(
float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
const float alpha, const float beta1, const float beta2, const float eps, const float wd,
const float beta1h, const float beta2h) {
float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
const float * __restrict__ pars, const int64_t k) {
const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
@ -13,6 +13,14 @@ static __global__ void opt_step_adamw_f32(
return;
}
const float alpha = pars[0];
const float beta1 = pars[1];
const float beta2 = pars[2];
const float eps = pars[3];
const float wd = pars[4];
const float beta1h = pars[5];
const float beta2h = pars[6];
const float gi = g[i];
const float gmi = g_m[i]*beta1 + gi*(1.0f - beta1);
const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
@ -23,58 +31,48 @@ static __global__ void opt_step_adamw_f32(
const float mh = gmi*beta1h;
const float vh = sqrtf(gvi*beta2h) + eps;
x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
}
static void opt_step_adamw_f32_cuda(
float * x, const float * g, float * g_m, float * g_v, const int64_t k,
const float alpha, const float beta1, const float beta2, const float eps, const float wd,
const float beta1h, const float beta2h, cudaStream_t stream) {
float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
}
void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src0_grad = dst->src[1];
const ggml_tensor * src0_grad_m = dst->src[2];
const ggml_tensor * src0_grad_v = dst->src[3];
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src0_grad = dst->src[1];
const ggml_tensor * src0_grad_m = dst->src[2];
const ggml_tensor * src0_grad_v = dst->src[3];
const ggml_tensor * adamw_params = dst->src[4];
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(src0_grad));
GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
GGML_ASSERT(ggml_is_contiguous(adamw_params));
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
GGML_ASSERT(ggml_nelements(adamw_params) == 7);
float * src0_d = (float *) src0->data;
const float * src0_grad_d = (const float *) src0_grad->data;
float * src0_grad_m_d = (float *) src0_grad_m->data;
float * src0_grad_v_d = (float *) src0_grad_v->data;
float * src0_d = (float *) src0->data;
const float * src0_grad_d = (const float *) src0_grad->data;
float * src0_grad_m_d = (float *) src0_grad_m->data;
float * src0_grad_v_d = (float *) src0_grad_v->data;
const float * adamw_params_d = (const float *) adamw_params->data;
cudaStream_t stream = ctx.stream();
const int64_t ne = ggml_nelements(src0);
int64_t iter; memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
float alpha; memcpy(&alpha, &dst->op_params[2], sizeof(float));
float beta1; memcpy(&beta1, &dst->op_params[3], sizeof(float));
float beta2; memcpy(&beta2, &dst->op_params[4], sizeof(float));
float eps; memcpy(&eps, &dst->op_params[5], sizeof(float));
float wd; memcpy(&wd, &dst->op_params[6], sizeof(float));
const float beta1h = alpha/(1.0f - powf(beta1, iter));
const float beta2h = 1.0f/(1.0f - powf(beta2, iter));
opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, ne, alpha, beta1, beta2, eps, wd, beta1h, beta2h, stream);
iter++;
memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
}

4
ggml/src/ggml-cuda/sum.cu
View file

@ -1,6 +1,6 @@
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
#define USE_CUB
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
#ifdef USE_CUB
// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.

106
ggml/src/ggml-hip/CMakeLists.txt Normal file
View file

@ -0,0 +1,106 @@
if (NOT EXISTS $ENV{ROCM_PATH})
if (NOT EXISTS /opt/rocm)
set(ROCM_PATH /usr)
else()
set(ROCM_PATH /opt/rocm)
endif()
else()
set(ROCM_PATH $ENV{ROCM_PATH})
endif()
list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})
list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
# CMake on Windows doesn't support the HIP language yet
if (WIN32)
set(CXX_IS_HIPCC TRUE)
else()
string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
endif()
if (CXX_IS_HIPCC)
if (LINUX)
if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
endif()
message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
" Prefer setting the HIP compiler directly. See README for details.")
endif()
else()
# Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
endif()
cmake_minimum_required(VERSION 3.21)
enable_language(HIP)
endif()
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
message(STATUS "HIP and hipBLAS found")
file(GLOB GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
file(GLOB GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
file(GLOB SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
list(APPEND GGML_SOURCES_ROCM ${SRCS})
file(GLOB SRCS "../ggml-cuda/template-instances/mmq*.cu")
list(APPEND GGML_SOURCES_ROCM ${SRCS})
if (GGML_CUDA_FA_ALL_QUANTS)
file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
list(APPEND GGML_SOURCES_ROCM ${SRCS})
add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
else()
file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
list(APPEND GGML_SOURCES_ROCM ${SRCS})
file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
list(APPEND GGML_SOURCES_ROCM ${SRCS})
file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
list(APPEND GGML_SOURCES_ROCM ${SRCS})
endif()
add_library(ggml-hip
${GGML_HEADERS_ROCM}
${GGML_SOURCES_ROCM})
target_link_libraries(ggml-hip PRIVATE ggml-base)
target_include_directories(ggml-hip PRIVATE . ..)
# TODO: do not use CUDA definitions for HIP
target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
add_compile_definitions(GGML_USE_HIP)
if (GGML_HIP_UMA)
add_compile_definitions(GGML_HIP_UMA)
endif()
if (GGML_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_CUDA_FORCE_MMQ)
endif()
if (GGML_CUDA_FORCE_CUBLAS)
add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
endif()
if (GGML_CUDA_NO_PEER_COPY)
add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
endif()
if (CXX_IS_HIPCC)
set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
target_link_libraries(ggml-hip PRIVATE hip::device)
else()
set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
endif()
if (GGML_STATIC)
message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
endif()
target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)

297
ggml/src/ggml-impl.h
View file

@ -3,13 +3,29 @@
// GGML internal header
#include "ggml.h"
#include <assert.h>
#include <math.h>
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
#include <stdbool.h>
#include <stdint.h>
#include <string.h>
#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE
#if defined(__ARM_NEON)
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#endif
#if defined(__F16C__)
#include <immintrin.h>
#endif
#ifdef __cplusplus
extern "C" {
#endif
@ -28,13 +44,13 @@ extern "C" {
// if C99 - static_assert is noop
// ref: https://stackoverflow.com/a/53923785/4039976
#ifndef __cplusplus
#ifndef static_assert
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
#define static_assert(cond, msg) _Static_assert(cond, msg)
#else
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#endif
#ifndef static_assert
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
#define static_assert(cond, msg) _Static_assert(cond, msg)
#else
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#endif
#endif
static inline int ggml_up32(int n) {
@ -120,14 +136,12 @@ struct ggml_map_custom1_op_params {
void * userdata;
};
struct ggml_map_custom2_op_params {
ggml_custom2_op_t fun;
int n_tasks;
void * userdata;
};
struct ggml_map_custom3_op_params {
ggml_custom3_op_t fun;
int n_tasks;
@ -182,7 +196,7 @@ void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
// returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key);
// returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
@ -196,7 +210,7 @@ static inline size_t ggml_hash(const struct ggml_tensor * p) {
return (size_t)(uintptr_t)p >> 4;
}
static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key) {
size_t h = ggml_hash(key) % hash_set->size;
// linear probing
@ -267,13 +281,14 @@ enum ggml_cgraph_eval_order {
};
struct ggml_cgraph {
int size;
int n_nodes;
int n_leafs;
int size; // maximum number of nodes/leafs/grads/grad_accs
int n_nodes; // number of nodes currently in use
int n_leafs; // number of leafs currently in use
struct ggml_tensor ** nodes;
struct ggml_tensor ** grads;
struct ggml_tensor ** leafs;
struct ggml_tensor ** nodes; // tensors with data that can change if the graph is evaluated
struct ggml_tensor ** grads; // the outputs of these tensors are the gradients of the nodes
struct ggml_tensor ** grad_accs; // accumulators for node gradients
struct ggml_tensor ** leafs; // tensors with constant data
struct ggml_hash_set visited_hash_set;
@ -287,9 +302,249 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
void * ggml_aligned_malloc(size_t size);
void ggml_aligned_free(void * ptr, size_t size);
// TODO: move to threading file
void ggml_critical_section_start(void);
void ggml_critical_section_end(void);
// FP16 to FP32 conversion
#if defined(__ARM_NEON)
#ifdef _MSC_VER
typedef uint16_t ggml_fp16_internal_t;
#else
typedef __fp16 ggml_fp16_internal_t;
#endif
#endif
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#elif defined(__F16C__)
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
GGML_API float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
/**
* Converts brain16 to float32.
*
* The bfloat16 floating point format has the following structure:
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───┐
* 0b0000000000000000 brain16
*
* Since bf16 has the same number of exponent bits as a 32bit float,
* encoding and decoding numbers becomes relatively straightforward.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───────────────────┐
* 0b00000000000000000000000000000000 IEEE binary32
*
* For comparison, the standard fp16 format has fewer exponent bits.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌─┴─┐┌─┴──────┐
* 0b0000000000000000 IEEE binary16
*
* @see IEEE 754-2008
*/
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
/**
* Converts float32 to brain16.
*
* This is binary identical with Google Brain float conversion.
* Floats shall round to nearest even, and NANs shall be quiet.
* Subnormals aren't flushed to zero, except perhaps when used.
* This code should vectorize nicely if using modern compilers.
*/
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
ggml_bf16_t h;
union {
float f;
uint32_t i;
} u;
u.f = s;
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
h.bits = (u.i >> 16) | 64; /* force to quiet */
return h;
}
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
return h;
}
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
#ifdef __cplusplus
}

162
ggml/src/ggml-kompute/CMakeLists.txt Normal file
View file

@ -0,0 +1,162 @@
find_package(Vulkan COMPONENTS glslc REQUIRED)
find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
if (NOT glslc_executable)
message(FATAL_ERROR "glslc not found")
endif()
add_library(ggml-kompute
ggml-kompute.cpp
../../include/ggml-kompute.h
)
target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
target_include_directories(ggml-kompute PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
function(compile_shader)
set(options)
set(oneValueArgs)
set(multiValueArgs SOURCES)
cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
foreach(source ${compile_shader_SOURCES})
get_filename_component(filename ${source} NAME)
set(spv_file ${filename}.spv)
add_custom_command(
OUTPUT ${spv_file}
DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
COMMENT "Compiling ${source} to ${spv_file}"
)
get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
set(FILE_NAME "shader${RAW_FILE_NAME}")
string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
if(CMAKE_GENERATOR MATCHES "Visual Studio")
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
)
else()
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
)
endif()
endforeach()
endfunction()
if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
message(STATUS "Kompute found")
set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
add_subdirectory(kompute)
# Compile our shaders
compile_shader(SOURCES
kompute-shaders/op_scale.comp
kompute-shaders/op_scale_8.comp
kompute-shaders/op_add.comp
kompute-shaders/op_addrow.comp
kompute-shaders/op_mul.comp
kompute-shaders/op_silu.comp
kompute-shaders/op_relu.comp
kompute-shaders/op_gelu.comp
kompute-shaders/op_softmax.comp
kompute-shaders/op_norm.comp
kompute-shaders/op_rmsnorm.comp
kompute-shaders/op_diagmask.comp
kompute-shaders/op_mul_mat_mat_f32.comp
kompute-shaders/op_mul_mat_f16.comp
kompute-shaders/op_mul_mat_q8_0.comp
kompute-shaders/op_mul_mat_q4_0.comp
kompute-shaders/op_mul_mat_q4_1.comp
kompute-shaders/op_mul_mat_q4_k.comp
kompute-shaders/op_mul_mat_q6_k.comp
kompute-shaders/op_getrows_f32.comp
kompute-shaders/op_getrows_f16.comp
kompute-shaders/op_getrows_q4_0.comp
kompute-shaders/op_getrows_q4_1.comp
kompute-shaders/op_getrows_q6_k.comp
kompute-shaders/op_rope_f16.comp
kompute-shaders/op_rope_f32.comp
kompute-shaders/op_cpy_f16_f16.comp
kompute-shaders/op_cpy_f16_f32.comp
kompute-shaders/op_cpy_f32_f16.comp
kompute-shaders/op_cpy_f32_f32.comp
)
# Create a custom target for our generated shaders
add_custom_target(generated_shaders DEPENDS
shaderop_scale.h
shaderop_scale_8.h
shaderop_add.h
shaderop_addrow.h
shaderop_mul.h
shaderop_silu.h
shaderop_relu.h
shaderop_gelu.h
shaderop_softmax.h
shaderop_norm.h
shaderop_rmsnorm.h
shaderop_diagmask.h
shaderop_mul_mat_mat_f32.h
shaderop_mul_mat_f16.h
shaderop_mul_mat_q8_0.h
shaderop_mul_mat_q4_0.h
shaderop_mul_mat_q4_1.h
shaderop_mul_mat_q4_k.h
shaderop_mul_mat_q6_k.h
shaderop_getrows_f32.h
shaderop_getrows_f16.h
shaderop_getrows_q4_0.h
shaderop_getrows_q4_1.h
shaderop_getrows_q6_k.h
shaderop_rope_f16.h
shaderop_rope_f32.h
shaderop_cpy_f16_f16.h
shaderop_cpy_f16_f32.h
shaderop_cpy_f32_f16.h
shaderop_cpy_f32_f32.h
)
# Create a custom command that depends on the generated_shaders
add_custom_command(
OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
DEPENDS generated_shaders
COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
)
# Add the stamp to the main sources to ensure dependency tracking
target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
else()
message(WARNING "Kompute not found")
endif()

0
ggml/src/ggml-kompute.cpp → ggml/src/ggml-kompute/ggml-kompute.cpp
View file

0
ggml/src/kompute → ggml/src/ggml-kompute/kompute
View file

0
ggml/src/kompute-shaders/common.comp → ggml/src/ggml-kompute/kompute-shaders/common.comp
View file

0
ggml/src/kompute-shaders/op_add.comp → ggml/src/ggml-kompute/kompute-shaders/op_add.comp
View file

0
ggml/src/kompute-shaders/op_addrow.comp → ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
View file

0
ggml/src/kompute-shaders/op_cpy_f16_f16.comp → ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
View file

0
ggml/src/kompute-shaders/op_cpy_f16_f32.comp → ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
View file

0
ggml/src/kompute-shaders/op_cpy_f32_f16.comp → ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
View file

Some files were not shown because too many files have changed in this diff Show more