happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into dev-refactoring

This commit is contained in:
hongruichen 2025-04-16 00:39:25 +08:00
parent e4fcdd418f f8f820cc4d
commit a004951bb9
148 changed files with 19879 additions and 15862 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
.devops/cpu.Dockerfile
View File

@ -14,9 +14,9 @@ WORKDIR /app
COPY . .
RUN if [ "$TARGETARCH" = "amd64" ]; then \
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
elif [ "$TARGETARCH" = "arm64" ]; then \
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
else \
echo "Unsupported architecture"; \
exit 1; \

2
.devops/cuda.Dockerfile
View File

@ -21,7 +21,7 @@ COPY . .
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -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 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release -j$(nproc)
RUN mkdir -p /app/lib && \

2
.devops/intel.Dockerfile
View File

@ -17,7 +17,7 @@ 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_NATIVE=OFF -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 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \
cmake --build build --config Release -j$(nproc)
RUN mkdir -p /app/lib && \

4
.devops/llama-cli-cann.Dockerfile
View File

@ -1,4 +1,4 @@
ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
ARG ASCEND_VERSION=8.1.RC1.alpha001-910b-openeuler22.03-py3.10
FROM ascendai/cann:$ASCEND_VERSION AS build
@ -6,7 +6,7 @@ WORKDIR /app
COPY . .
RUN yum install -y gcc g++ cmake make
RUN yum install -y gcc g++ cmake make libcurl-devel
ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}

2
.devops/musa.Dockerfile
View File

@ -35,7 +35,7 @@ COPY . .
RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release -j$(nproc)
RUN mkdir -p /app/lib && \

6
.devops/rocm.Dockerfile
View File

@ -17,8 +17,8 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# gfx906 is deprecated
#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html
#ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
ARG ROCM_DOCKER_ARCH=gfx1100
ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
#ARG ROCM_DOCKER_ARCH=gfx1100
# Set nvcc architectured
ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
@ -40,7 +40,7 @@ WORKDIR /app
COPY . .
RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
&& cmake --build build --config Release -j$(nproc)
RUN mkdir -p /app/lib \

2
.devops/vulkan.Dockerfile
View File

@ -16,7 +16,7 @@ WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
cmake --build build --config Release -j$(nproc)
RUN mkdir -p /app/lib && \

25
.github/actions/windows-setup-curl/action.yml vendored Normal file
View File

@ -0,0 +1,25 @@
name: 'Windows - Setup CURL'
description: 'Composite action, to be reused in other workflow'
inputs:
curl_version:
description: 'CURL version'
required: false
default: '8.6.0_6'
outputs:
curl_path:
description: "Path to the downloaded libcurl"
value: ${{ steps.get_libcurl.outputs.curl_path }}
runs:
using: "composite"
steps:
- name: libCURL
id: get_libcurl
shell: powershell
env:
CURL_VERSION: ${{ inputs.curl_version }}
run: |
curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
mkdir $env:RUNNER_TEMP/libcurl
tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT

1
.github/workflows/bench.yml.disabled vendored
View File

@ -104,7 +104,6 @@ jobs:
cmake -B build \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DLLAMA_CUBLAS=ON \
-DCUDAToolkit_ROOT=/usr/local/cuda \
-DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \

124
.github/workflows/build-linux-cross.yml vendored Normal file
View File

@ -0,0 +1,124 @@
name: Build on Linux using cross-compiler
on:
workflow_dispatch:
workflow_call:
jobs:
ubuntu-latest-riscv64-cpu-cross:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Setup Riscv
run: |
sudo dpkg --add-architecture riscv64
sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
/etc/apt/sources.list /etc/apt/apt-mirrors.txt
sudo apt-get clean
sudo apt-get update
sudo apt-get install -y --no-install-recommends \
build-essential \
gcc-14-riscv64-linux-gnu \
g++-14-riscv64-linux-gnu \
libcurl4-openssl-dev:riscv64
- name: Build
run: |
cmake -B build -DCMAKE_BUILD_TYPE=Release \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_SYSTEM_NAME=Linux \
-DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)
ubuntu-latest-riscv64-vulkan-cross:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Setup Riscv
run: |
sudo dpkg --add-architecture riscv64
sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
/etc/apt/sources.list /etc/apt/apt-mirrors.txt
sudo apt-get clean
sudo apt-get update
sudo apt-get install -y --no-install-recommends \
build-essential \
glslc \
gcc-14-riscv64-linux-gnu \
g++-14-riscv64-linux-gnu \
libvulkan-dev:riscv64 \
libcurl4-openssl-dev:riscv64
- name: Build
run: |
cmake -B build -DCMAKE_BUILD_TYPE=Release \
-DGGML_VULKAN=ON \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_SYSTEM_NAME=Linux \
-DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)
ubuntu-latest-arm64-vulkan-cross:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Setup Arm64
run: |
sudo dpkg --add-architecture arm64
sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
/etc/apt/sources.list /etc/apt/apt-mirrors.txt
sudo apt-get clean
sudo apt-get update
sudo apt-get install -y --no-install-recommends \
build-essential \
glslc \
crossbuild-essential-arm64 \
libvulkan-dev:arm64 \
libcurl4-openssl-dev:arm64
- name: Build
run: |
cmake -B build -DCMAKE_BUILD_TYPE=Release \
-DGGML_VULKAN=ON \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_SYSTEM_NAME=Linux \
-DCMAKE_SYSTEM_PROCESSOR=aarch64 \
-DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
-DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-linux-gnu \
-DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)

94
.github/workflows/build.yml vendored
View File

@ -10,7 +10,7 @@ on:
push:
branches:
- master
paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
@ -54,6 +54,7 @@ jobs:
continue-on-error: true
run: |
brew update
brew install curl
- name: Build
id: cmake_build
@ -62,7 +63,6 @@ jobs:
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_RPC=ON
@ -92,7 +92,6 @@ jobs:
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
run: |
cp LICENSE ./build/bin/
cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
- name: Upload artifacts
@ -123,6 +122,7 @@ jobs:
continue-on-error: true
run: |
brew update
brew install curl
- name: Build
id: cmake_build
@ -133,7 +133,6 @@ jobs:
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
@ -162,7 +161,6 @@ jobs:
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
run: |
cp LICENSE ./build/bin/
cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
- name: Upload artifacts
@ -207,7 +205,6 @@ jobs:
run: |
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_RPC=ON
cmake --build build --config Release -j $(nproc)
@ -246,7 +243,6 @@ jobs:
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
run: |
cp LICENSE ./build/bin/
cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
- name: Upload artifacts
@ -281,7 +277,7 @@ jobs:
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Build
id: cmake_build
@ -322,7 +318,7 @@ jobs:
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Build
id: cmake_build
@ -360,7 +356,7 @@ jobs:
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Build
id: cmake_build
@ -397,7 +393,7 @@ jobs:
wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
sudo apt-get update -y
sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk
sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
- name: Build
id: cmake_build
@ -431,7 +427,6 @@ jobs:
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
run: |
cp LICENSE ./build/bin/
cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
- name: Upload artifacts
@ -454,7 +449,7 @@ jobs:
id: depends
run: |
sudo apt-get update
sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2.16
@ -530,7 +525,7 @@ jobs:
shell: bash
run: |
sudo apt update
sudo apt install intel-oneapi-compiler-dpcpp-cpp
sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
- name: install oneAPI MKL library
shell: bash
@ -578,7 +573,7 @@ jobs:
shell: bash
run: |
sudo apt update
sudo apt install intel-oneapi-compiler-dpcpp-cpp
sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev
- name: install oneAPI MKL library
shell: bash
@ -606,6 +601,9 @@ jobs:
-DGGML_SYCL_F16=ON
cmake --build build --config Release -j $(nproc)
build-linux-cross:
uses: ./.github/workflows/build-linux-cross.yml
macOS-latest-cmake-ios:
runs-on: macos-latest
@ -633,6 +631,7 @@ jobs:
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_COMMON=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
@ -668,6 +667,7 @@ jobs:
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_COMMON=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
@ -697,6 +697,7 @@ jobs:
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_COMMON=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
@ -736,6 +737,7 @@ jobs:
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_CURL=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
@ -896,10 +898,17 @@ jobs:
-DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release"
cmake --build build-arm64-release --target install --config release
- name: libCURL
id: get_libcurl
uses: ./.github/actions/windows-setup-curl
- name: Build
id: cmake_build
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
cmake -S . -B build ${{ matrix.defines }}
cmake -S . -B build ${{ matrix.defines }} `
-DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
- name: Add libopenblas.dll
@ -959,9 +968,10 @@ jobs:
- name: Pack artifacts
id: pack_artifacts
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
Copy-Item .\examples\run\linenoise.cpp\LICENSE .\build\bin\Release\linenoise.cpp.txt
Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
- name: Upload artifacts
@ -987,7 +997,7 @@ jobs:
DEBIAN_FRONTEND: noninteractive
run: |
apt update
apt install -y cmake build-essential ninja-build libgomp1 git
apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2.16
@ -1089,16 +1099,23 @@ jobs:
run: |
choco install ninja
- name: libCURL
id: get_libcurl
uses: ./.github/actions/windows-setup-curl
- name: Build
id: cmake_build
shell: cmd
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
cmake -S . -B build -G "Ninja Multi-Config" ^
-DLLAMA_BUILD_SERVER=ON ^
-DGGML_NATIVE=OFF ^
-DGGML_CUDA=ON ^
-DGGML_RPC=ON
-DGGML_RPC=ON ^
-DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include"
set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
cmake --build build --config Release -j %NINJA_JOBS% -t ggml
cmake --build build --config Release
@ -1119,7 +1136,10 @@ jobs:
- name: Pack artifacts
id: pack_artifacts
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
- name: Upload artifacts
@ -1174,6 +1194,8 @@ jobs:
run: |
scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
# TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
- name: Build
id: cmake_build
run: examples/sycl/win-build-sycl.bat
@ -1259,8 +1281,14 @@ jobs:
key: ${{ github.job }}
evict-old-files: 1d
- name: libCURL
id: get_libcurl
uses: ./.github/actions/windows-setup-curl
- name: Build
id: cmake_build
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
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}"
@ -1271,9 +1299,11 @@ jobs:
-DCMAKE_BUILD_TYPE=Release `
-DGGML_HIP=ON `
-DGGML_HIP_ROCWMMA_FATTN=ON `
-DGGML_RPC=ON
-DGGML_RPC=ON `
-DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
# TODO: reuse windows-latest-cmake-hip instead of duplicating this job
windows-latest-cmake-hip-release:
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
runs-on: windows-latest
@ -1315,8 +1345,14 @@ jobs:
run: |
& 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
- name: libCURL
id: get_libcurl
uses: ./.github/actions/windows-setup-curl
- name: Build
id: cmake_build
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
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}"
@ -1328,7 +1364,8 @@ jobs:
-DAMDGPU_TARGETS=${{ matrix.gpu_target }} `
-DGGML_HIP_ROCWMMA_FATTN=ON `
-DGGML_HIP=ON `
-DGGML_RPC=ON
-DGGML_RPC=ON `
-DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
md "build\bin\rocblas\library\"
cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
@ -1350,7 +1387,10 @@ jobs:
- name: Pack artifacts
id: pack_artifacts
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll
7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
- name: Upload artifacts
@ -1375,6 +1415,7 @@ jobs:
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_CURL=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
@ -1725,16 +1766,17 @@ jobs:
if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
defaults:
run:
shell: bash -el {0}
runs-on: ubuntu-24.04-arm
shell: bash -el {0}
strategy:
matrix:
arch: [x86, aarch64]
cann:
- '8.0.rc3.beta1-910b-openeuler22.03-py3.10'
- '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
device:
- 'ascend910b3'
build:
- 'Release'
runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
container: ascendai/cann:${{ matrix.cann }}
steps:
- name: Checkout
@ -1743,7 +1785,7 @@ jobs:
- name: Dependencies
run: |
yum update -y
yum install -y git gcc gcc-c++ make cmake
yum install -y git gcc gcc-c++ make cmake libcurl-devel
- name: Build
run: |

12
.github/workflows/docker.yml vendored
View File

@ -36,13 +36,13 @@ jobs:
matrix:
config:
# Multi-stage build
- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: false}
- { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
- { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
- { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
- { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false}
- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
- { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
# Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
#- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: true }
#- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
steps:
- name: Check out the repo
uses: actions/checkout@v4

18
.github/workflows/server.yml vendored
View File

@ -129,7 +129,6 @@ jobs:
cmake -B build \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-DGGML_OPENMP=OFF ;
@ -142,7 +141,6 @@ jobs:
cmake -B build \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
@ -154,7 +152,6 @@ jobs:
cmake -B build \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
@ -195,17 +192,14 @@ jobs:
- name: libCURL
id: get_libcurl
env:
CURL_VERSION: 8.6.0_6
run: |
curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
mkdir $env:RUNNER_TEMP/libcurl
tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
uses: ./.github/actions/windows-setup-curl
- name: Build
id: cmake_build
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
- name: Python setup
@ -221,8 +215,10 @@ jobs:
- name: Copy Libcurl
id: prepare_libcurl
env:
CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
run: |
cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
- name: Tests
id: server_integration_tests

24
CMakeLists.txt
View File

@ -81,7 +81,7 @@ option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE})
# 3rd party libs
option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
option(LLAMA_CURL "llama: use libcurl to download model from an URL" ON)
option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
# Required for relocatable CMake package
@ -169,6 +169,11 @@ add_subdirectory(src)
# utils, programs, examples and tests
#
if (NOT LLAMA_BUILD_COMMON)
message(STATUS "LLAMA_BUILD_COMMON is OFF, disabling LLAMA_CURL")
set(LLAMA_CURL OFF)
endif()
if (LLAMA_BUILD_COMMON)
add_subdirectory(common)
endif()
@ -243,3 +248,20 @@ configure_file(cmake/llama.pc.in
install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
#
# copy the license files
#
# Check if running in GitHub Actions
if(DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
message(STATUS "Running inside GitHub Actions - copying license files")
# Copy all files from licenses/ to build/bin/
file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
foreach(LICENSE_FILE ${LICENSE_FILES})
get_filename_component(FILENAME ${LICENSE_FILE} NAME)
configure_file(${LICENSE_FILE} "${CMAKE_BINARY_DIR}/bin/${FILENAME}" COPYONLY)
endforeach()
endif()

4
Makefile
View File

@ -780,10 +780,6 @@ ifdef GGML_HIP
MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
ifdef GGML_HIP_UMA
MK_CPPFLAGS += -DGGML_HIP_UMA
endif # GGML_HIP_UMA
MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas

42
README.md
View File

@ -9,13 +9,6 @@
Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
> [!IMPORTANT]
> New `llama.cpp` package location: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/pkgs/container/llama.cpp)
>
> Update your container URLs to: `ghcr.io/ggml-org/llama.cpp`
>
> More info: https://github.com/ggml-org/llama.cpp/discussions/11801
## Recent API changes
- [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)
@ -104,6 +97,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
- [x] [GLM-4-0414](https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e)
- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
@ -247,6 +241,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
| [Vulkan](docs/build.md#vulkan) | GPU |
| [CANN](docs/build.md#cann) | Ascend NPU |
| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All |
## Building the project
@ -265,7 +260,9 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
- [Trending](https://huggingface.co/models?library=gguf&sort=trending)
- [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from Hugging Face by using this CLI argument: `-hf <user>/<model>[:quant]`
You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.
After downloading a model, use the CLI tools to run it locally - see below.
@ -530,6 +527,35 @@ If your issue is with model generation quality, then please at least scan the fo
- [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
## XCFramework
The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS,
and macOS. It can be used in Swift projects without the need to compile the
library from source. For example:
```swift
// swift-tools-version: 5.10
// The swift-tools-version declares the minimum version of Swift required to build this package.
import PackageDescription
let package = Package(
name: "MyLlamaPackage",
targets: [
.executableTarget(
name: "MyLlamaPackage",
dependencies: [
"LlamaFramework"
]),
.binaryTarget(
name: "LlamaFramework",
url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip",
checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab"
)
]
)
```
The above example is using an intermediate build `b5046` of the library. This can be modified
to use a different version by changing the URL and checksum.
## Completions
Command-line completion is available for some environments.

27
build-xcframework.sh
View File

@ -41,6 +41,11 @@ COMMON_CMAKE_ARGS=(
-DGGML_OPENMP=${GGML_OPENMP}
)
XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
echo "Detected Xcode version: $XCODE_VERSION"
check_required_tool() {
local tool=$1
local install_message=$2
@ -325,21 +330,28 @@ combine_static_libraries() {
# Platform-specific post-processing for device builds
if [[ "$is_simulator" == "false" ]]; then
if command -v vtool &>/dev/null; then
if command -v xcrun vtool &>/dev/null; then
case "$platform" in
"ios")
echo "Marking binary as a framework binary for iOS..."
vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
xcrun vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
-output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
;;
"visionos")
echo "Marking binary as a framework binary for visionOS..."
vtool -set-build-version xros ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
if [[ "$MAJOR_VERSION" -gt 16 ]] || [[ "$MAJOR_VERSION" -eq 16 && "$MINOR_VERSION" -gt 2 ]]; then
echo "Xcode version greater than 16.2, using visionOS."
VISION_OS_BUILD_VERSION="visionos"
else
echo "Xcode version less than or equal to 16.2, using xros."
VISION_OS_BUILD_VERSION="xros"
fi
xcrun vtool -set-build-version ${VISION_OS_BUILD_VERSION} ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
-output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
;;
"tvos")
echo "Marking binary as a framework binary for tvOS..."
vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
xcrun vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
-output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
;;
esac
@ -399,6 +411,7 @@ cmake -B build-ios-sim -G Xcode \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
-DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-ios-sim --config Release -- -quiet
@ -411,6 +424,7 @@ cmake -B build-ios-device -G Xcode \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
-DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-ios-device --config Release -- -quiet
@ -421,6 +435,7 @@ cmake -B build-macos -G Xcode \
-DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
-DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-macos --config Release -- -quiet
@ -434,6 +449,7 @@ cmake -B build-visionos -G Xcode \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-visionos --config Release -- -quiet
@ -447,6 +463,7 @@ cmake -B build-visionos-sim -G Xcode \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-visionos-sim --config Release -- -quiet
@ -462,6 +479,7 @@ cmake -B build-tvos-sim -G Xcode \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
-DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-tvos-sim --config Release -- -quiet
@ -476,6 +494,7 @@ cmake -B build-tvos-device -G Xcode \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
-DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
-DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
-DLLAMA_CURL=OFF \
-S .
cmake --build build-tvos-device --config Release -- -quiet

2
ci/run.sh
View File

@ -39,7 +39,7 @@ sd=`dirname $0`
cd $sd/../
SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"

5
common/CMakeLists.txt
View File

@ -85,7 +85,10 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info)
# Use curl to download model url
if (LLAMA_CURL)
find_package(CURL REQUIRED)
find_package(CURL)
if (NOT CURL_FOUND)
message(FATAL_ERROR "Could NOT find CURL. Hint: to disable this feature, set -DLLAMA_CURL=OFF")
endif()
target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
include_directories(${CURL_INCLUDE_DIRS})
find_library(CURL_LIBRARY curl REQUIRED)

16
common/arg.cpp
View File

@ -18,6 +18,7 @@
#include <algorithm>
#include <climits>
#include <cstdarg>
#include <filesystem>
#include <fstream>
#include <regex>
#include <set>
@ -162,6 +163,8 @@ struct common_hf_file_res {
# if !defined(PATH_MAX)
# define PATH_MAX MAX_PATH
# endif
#elif defined(_AIX)
#include <sys/limits.h>
#else
#include <sys/syslimits.h>
#endif
@ -225,12 +228,13 @@ static bool common_download_file_single(const std::string & url, const std::stri
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
// Check if hf-token or bearer-token was specified
if (!bearer_token.empty()) {
std::string auth_header = "Authorization: Bearer " + bearer_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
}
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
#if defined(_WIN32)
// CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
@ -541,7 +545,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
std::string res_str;
std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
std::string model_endpoint = get_model_endpoint();
std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag;
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
@ -656,9 +663,8 @@ static void common_params_handle_model(
}
}
// TODO: allow custom host
model.url = "https://huggingface.co/" + model.hf_repo + "/resolve/main/" + model.hf_file;
std::string model_endpoint = get_model_endpoint();
model.url = model_endpoint + model.hf_repo + "/resolve/main/" + model.hf_file;
// make sure model path is present (for caching purposes)
if (model.path.empty()) {
// this is to avoid different repo having same file name, or same file name in different subdirs

2
common/chat.cpp
View File

@ -1622,7 +1622,7 @@ static common_chat_params common_chat_templates_apply_jinja(
}
// Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null() && params.tools.is_array() && params.json_schema.is_null()) {
return common_chat_params_init_hermes_2_pro(tmpl, params);
}

19
common/common.cpp
View File

@ -830,7 +830,7 @@ std::string fs_get_cache_directory() {
if (getenv("LLAMA_CACHE")) {
cache_directory = std::getenv("LLAMA_CACHE");
} else {
#ifdef __linux__
#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
if (std::getenv("XDG_CACHE_HOME")) {
cache_directory = std::getenv("XDG_CACHE_HOME");
} else {
@ -840,7 +840,9 @@ std::string fs_get_cache_directory() {
cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
#elif defined(_WIN32)
cache_directory = std::getenv("LOCALAPPDATA");
#endif // __linux__
#else
# error Unknown architecture
#endif
cache_directory = ensure_trailing_slash(cache_directory);
cache_directory += "llama.cpp";
}
@ -1027,6 +1029,19 @@ struct common_init_result common_init_from_params(common_params & params) {
return iparams;
}
std::string get_model_endpoint() {
const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
// We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
const char * hf_endpoint_env = getenv("HF_ENDPOINT");
const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
std::string model_endpoint = "https://huggingface.co/";
if (endpoint_env) {
model_endpoint = endpoint_env;
if (model_endpoint.back() != '/') model_endpoint += '/';
}
return model_endpoint;
}
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
llama_clear_adapter_lora(ctx);
for (auto & la : lora) {

2
common/common.h
View File

@ -543,6 +543,8 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
// clear LoRA adapters from context, then apply new list of adapters
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
std::string get_model_endpoint();
//
// Batch utils
//

22
common/minja/chat-template.hpp
View File

@ -9,10 +9,19 @@
#pragma once
#include "minja.hpp"
#include <json.hpp>
#include <chrono>
#include <cstddef>
#include <cstdio>
#include <exception>
#include <iomanip>
#include <memory>
#include <sstream>
#include <string>
#include <vector>
#include <json.hpp>
using json = nlohmann::ordered_json;
namespace minja {
@ -425,7 +434,7 @@ class chat_template {
auto obj = json {
{"tool_calls", tool_calls},
};
if (!content.is_null() && content != "") {
if (!content.is_null() && !content.empty()) {
obj["content"] = content;
}
message["content"] = obj.dump(2);
@ -435,13 +444,12 @@ class chat_template {
if (polyfill_tool_responses && role == "tool") {
message["role"] = "user";
auto obj = json {
{"tool_response", {
{"content", message.at("content")},
}},
{"tool_response", json::object()},
};
if (message.contains("name")) {
obj["tool_response"]["name"] = message.at("name");
obj["tool_response"]["tool"] = message.at("name");
}
obj["tool_response"]["content"] = message.at("content");
if (message.contains("tool_call_id")) {
obj["tool_response"]["tool_call_id"] = message.at("tool_call_id");
}
@ -510,7 +518,7 @@ class chat_template {
static nlohmann::ordered_json add_system(const nlohmann::ordered_json & messages, const std::string & system_prompt) {
json messages_with_system = messages;
if (messages_with_system.size() > 0 && messages_with_system[0].at("role") == "system") {
if (!messages_with_system.empty() && messages_with_system[0].at("role") == "system") {
std::string existing_system = messages_with_system.at(0).at("content");
messages_with_system[0] = json {
{"role", "system"},

194
common/minja/minja.hpp
View File

@ -8,14 +8,26 @@
// SPDX-License-Identifier: MIT
#pragma once
#include <algorithm>
#include <cctype>
#include <cstddef>
#include <cmath>
#include <exception>
#include <functional>
#include <iostream>
#include <string>
#include <vector>
#include <regex>
#include <iterator>
#include <limits>
#include <map>
#include <memory>
#include <stdexcept>
#include <regex>
#include <sstream>
#include <string>
#include <stdexcept>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include <json.hpp>
using json = nlohmann::ordered_json;
@ -240,7 +252,7 @@ public:
auto index = key.get<int>();
return array_->at(index < 0 ? array_->size() + index : index);
} else if (object_) {
if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
auto it = object_->find(key.primitive_);
if (it == object_->end()) return Value();
return it->second;
@ -249,7 +261,7 @@ public:
}
void set(const Value& key, const Value& value) {
if (!object_) throw std::runtime_error("Value is not an object: " + dump());
if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
(*object_)[key.primitive_] = value;
}
Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {
@ -731,51 +743,51 @@ public:
struct TextTemplateToken : public TemplateToken {
std::string text;
TextTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Text, location, pre, post), text(t) {}
TextTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Text, loc, pre, post), text(t) {}
};
struct ExpressionTemplateToken : public TemplateToken {
std::shared_ptr<Expression> expr;
ExpressionTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && e) : TemplateToken(Type::Expression, location, pre, post), expr(std::move(e)) {}
ExpressionTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && e) : TemplateToken(Type::Expression, loc, pre, post), expr(std::move(e)) {}
};
struct IfTemplateToken : public TemplateToken {
std::shared_ptr<Expression> condition;
IfTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::If, location, pre, post), condition(std::move(c)) {}
IfTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::If, loc, pre, post), condition(std::move(c)) {}
};
struct ElifTemplateToken : public TemplateToken {
std::shared_ptr<Expression> condition;
ElifTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::Elif, location, pre, post), condition(std::move(c)) {}
ElifTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && c) : TemplateToken(Type::Elif, loc, pre, post), condition(std::move(c)) {}
};
struct ElseTemplateToken : public TemplateToken {
ElseTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Else, location, pre, post) {}
ElseTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Else, loc, pre, post) {}
};
struct EndIfTemplateToken : public TemplateToken {
EndIfTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndIf, location, pre, post) {}
EndIfTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndIf, loc, pre, post) {}
};
struct MacroTemplateToken : public TemplateToken {
std::shared_ptr<VariableExpr> name;
Expression::Parameters params;
MacroTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p)
: TemplateToken(Type::Macro, location, pre, post), name(std::move(n)), params(std::move(p)) {}
MacroTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p)
: TemplateToken(Type::Macro, loc, pre, post), name(std::move(n)), params(std::move(p)) {}
};
struct EndMacroTemplateToken : public TemplateToken {
EndMacroTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndMacro, location, pre, post) {}
EndMacroTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndMacro, loc, pre, post) {}
};
struct FilterTemplateToken : public TemplateToken {
std::shared_ptr<Expression> filter;
FilterTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && filter)
: TemplateToken(Type::Filter, location, pre, post), filter(std::move(filter)) {}
FilterTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr<Expression> && filter)
: TemplateToken(Type::Filter, loc, pre, post), filter(std::move(filter)) {}
};
struct EndFilterTemplateToken : public TemplateToken {
EndFilterTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFilter, location, pre, post) {}
EndFilterTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFilter, loc, pre, post) {}
};
struct ForTemplateToken : public TemplateToken {
@ -783,38 +795,38 @@ struct ForTemplateToken : public TemplateToken {
std::shared_ptr<Expression> iterable;
std::shared_ptr<Expression> condition;
bool recursive;
ForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::vector<std::string> & vns, std::shared_ptr<Expression> && iter,
ForTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::vector<std::string> & vns, std::shared_ptr<Expression> && iter,
std::shared_ptr<Expression> && c, bool r)
: TemplateToken(Type::For, location, pre, post), var_names(vns), iterable(std::move(iter)), condition(std::move(c)), recursive(r) {}
: TemplateToken(Type::For, loc, pre, post), var_names(vns), iterable(std::move(iter)), condition(std::move(c)), recursive(r) {}
};
struct EndForTemplateToken : public TemplateToken {
EndForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, location, pre, post) {}
EndForTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, loc, pre, post) {}
};
struct GenerationTemplateToken : public TemplateToken {
GenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, location, pre, post) {}
GenerationTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, loc, pre, post) {}
};
struct EndGenerationTemplateToken : public TemplateToken {
EndGenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, location, pre, post) {}
EndGenerationTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, loc, pre, post) {}
};
struct SetTemplateToken : public TemplateToken {
std::string ns;
std::vector<std::string> var_names;
std::shared_ptr<Expression> value;
SetTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
: TemplateToken(Type::Set, location, pre, post), ns(ns), var_names(vns), value(std::move(v)) {}
SetTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
: TemplateToken(Type::Set, loc, pre, post), ns(ns), var_names(vns), value(std::move(v)) {}
};
struct EndSetTemplateToken : public TemplateToken {
EndSetTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndSet, location, pre, post) {}
EndSetTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndSet, loc, pre, post) {}
};
struct CommentTemplateToken : public TemplateToken {
std::string text;
CommentTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Comment, location, pre, post), text(t) {}
CommentTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Comment, loc, pre, post), text(t) {}
};
enum class LoopControlType { Break, Continue };
@ -830,7 +842,7 @@ public:
struct LoopControlTemplateToken : public TemplateToken {
LoopControlType control_type;
LoopControlTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, LoopControlType control_type) : TemplateToken(Type::Break, location, pre, post), control_type(control_type) {}
LoopControlTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, LoopControlType control_type) : TemplateToken(Type::Break, loc, pre, post), control_type(control_type) {}
};
class TemplateNode {
@ -868,8 +880,8 @@ public:
class SequenceNode : public TemplateNode {
std::vector<std::shared_ptr<TemplateNode>> children;
public:
SequenceNode(const Location & location, std::vector<std::shared_ptr<TemplateNode>> && c)
: TemplateNode(location), children(std::move(c)) {}
SequenceNode(const Location & loc, std::vector<std::shared_ptr<TemplateNode>> && c)
: TemplateNode(loc), children(std::move(c)) {}
void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
for (const auto& child : children) child->render(out, context);
}
@ -878,7 +890,7 @@ public:
class TextNode : public TemplateNode {
std::string text;
public:
TextNode(const Location & location, const std::string& t) : TemplateNode(location), text(t) {}
TextNode(const Location & loc, const std::string& t) : TemplateNode(loc), text(t) {}
void do_render(std::ostringstream & out, const std::shared_ptr<Context> &) const override {
out << text;
}
@ -887,7 +899,7 @@ public:
class ExpressionNode : public TemplateNode {
std::shared_ptr<Expression> expr;
public:
ExpressionNode(const Location & location, std::shared_ptr<Expression> && e) : TemplateNode(location), expr(std::move(e)) {}
ExpressionNode(const Location & loc, std::shared_ptr<Expression> && e) : TemplateNode(loc), expr(std::move(e)) {}
void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
if (!expr) throw std::runtime_error("ExpressionNode.expr is null");
auto result = expr->evaluate(context);
@ -904,8 +916,8 @@ public:
class IfNode : public TemplateNode {
std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<TemplateNode>>> cascade;
public:
IfNode(const Location & location, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<TemplateNode>>> && c)
: TemplateNode(location), cascade(std::move(c)) {}
IfNode(const Location & loc, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<TemplateNode>>> && c)
: TemplateNode(loc), cascade(std::move(c)) {}
void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
for (const auto& branch : cascade) {
auto enter_branch = true;
@ -924,7 +936,7 @@ public:
class LoopControlNode : public TemplateNode {
LoopControlType control_type_;
public:
LoopControlNode(const Location & location, LoopControlType control_type) : TemplateNode(location), control_type_(control_type) {}
LoopControlNode(const Location & loc, LoopControlType control_type) : TemplateNode(loc), control_type_(control_type) {}
void do_render(std::ostringstream &, const std::shared_ptr<Context> &) const override {
throw LoopControlException(control_type_);
}
@ -938,9 +950,9 @@ class ForNode : public TemplateNode {
bool recursive;
std::shared_ptr<TemplateNode> else_body;
public:
ForNode(const Location & location, std::vector<std::string> && var_names, std::shared_ptr<Expression> && iterable,
ForNode(const Location & loc, std::vector<std::string> && var_names, std::shared_ptr<Expression> && iterable,
std::shared_ptr<Expression> && condition, std::shared_ptr<TemplateNode> && body, bool recursive, std::shared_ptr<TemplateNode> && else_body)
: TemplateNode(location), var_names(var_names), iterable(std::move(iterable)), condition(std::move(condition)), body(std::move(body)), recursive(recursive), else_body(std::move(else_body)) {}
: TemplateNode(loc), var_names(var_names), iterable(std::move(iterable)), condition(std::move(condition)), body(std::move(body)), recursive(recursive), else_body(std::move(else_body)) {}
void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
// https://jinja.palletsprojects.com/en/3.0.x/templates/#for
@ -1025,8 +1037,8 @@ class MacroNode : public TemplateNode {
std::shared_ptr<TemplateNode> body;
std::unordered_map<std::string, size_t> named_param_positions;
public:
MacroNode(const Location & location, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p, std::shared_ptr<TemplateNode> && b)
: TemplateNode(location), name(std::move(n)), params(std::move(p)), body(std::move(b)) {
MacroNode(const Location & loc, std::shared_ptr<VariableExpr> && n, Expression::Parameters && p, std::shared_ptr<TemplateNode> && b)
: TemplateNode(loc), name(std::move(n)), params(std::move(p)), body(std::move(b)) {
for (size_t i = 0; i < params.size(); ++i) {
const auto & name = params[i].first;
if (!name.empty()) {
@ -1072,8 +1084,8 @@ class FilterNode : public TemplateNode {
std::shared_ptr<TemplateNode> body;
public:
FilterNode(const Location & location, std::shared_ptr<Expression> && f, std::shared_ptr<TemplateNode> && b)
: TemplateNode(location), filter(std::move(f)), body(std::move(b)) {}
FilterNode(const Location & loc, std::shared_ptr<Expression> && f, std::shared_ptr<TemplateNode> && b)
: TemplateNode(loc), filter(std::move(f)), body(std::move(b)) {}
void do_render(std::ostringstream & out, const std::shared_ptr<Context> & context) const override {
if (!filter) throw std::runtime_error("FilterNode.filter is null");
@ -1095,8 +1107,8 @@ class SetNode : public TemplateNode {
std::vector<std::string> var_names;
std::shared_ptr<Expression> value;
public:
SetNode(const Location & location, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
: TemplateNode(location), ns(ns), var_names(vns), value(std::move(v)) {}
SetNode(const Location & loc, const std::string & ns, const std::vector<std::string> & vns, std::shared_ptr<Expression> && v)
: TemplateNode(loc), ns(ns), var_names(vns), value(std::move(v)) {}
void do_render(std::ostringstream &, const std::shared_ptr<Context> & context) const override {
if (!value) throw std::runtime_error("SetNode.value is null");
if (!ns.empty()) {
@ -1118,8 +1130,8 @@ class SetTemplateNode : public TemplateNode {
std::string name;
std::shared_ptr<TemplateNode> template_value;
public:
SetTemplateNode(const Location & location, const std::string & name, std::shared_ptr<TemplateNode> && tv)
: TemplateNode(location), name(name), template_value(std::move(tv)) {}
SetTemplateNode(const Location & loc, const std::string & name, std::shared_ptr<TemplateNode> && tv)
: TemplateNode(loc), name(name), template_value(std::move(tv)) {}
void do_render(std::ostringstream &, const std::shared_ptr<Context> & context) const override {
if (!template_value) throw std::runtime_error("SetTemplateNode.template_value is null");
Value value { template_value->render(context) };
@ -1132,8 +1144,8 @@ class IfExpr : public Expression {
std::shared_ptr<Expression> then_expr;
std::shared_ptr<Expression> else_expr;
public:
IfExpr(const Location & location, std::shared_ptr<Expression> && c, std::shared_ptr<Expression> && t, std::shared_ptr<Expression> && e)
: Expression(location), condition(std::move(c)), then_expr(std::move(t)), else_expr(std::move(e)) {}
IfExpr(const Location & loc, std::shared_ptr<Expression> && c, std::shared_ptr<Expression> && t, std::shared_ptr<Expression> && e)
: Expression(loc), condition(std::move(c)), then_expr(std::move(t)), else_expr(std::move(e)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
if (!condition) throw std::runtime_error("IfExpr.condition is null");
if (!then_expr) throw std::runtime_error("IfExpr.then_expr is null");
@ -1150,16 +1162,16 @@ public:
class LiteralExpr : public Expression {
Value value;
public:
LiteralExpr(const Location & location, const Value& v)
: Expression(location), value(v) {}
LiteralExpr(const Location & loc, const Value& v)
: Expression(loc), value(v) {}
Value do_evaluate(const std::shared_ptr<Context> &) const override { return value; }
};
class ArrayExpr : public Expression {
std::vector<std::shared_ptr<Expression>> elements;
public:
ArrayExpr(const Location & location, std::vector<std::shared_ptr<Expression>> && e)
: Expression(location), elements(std::move(e)) {}
ArrayExpr(const Location & loc, std::vector<std::shared_ptr<Expression>> && e)
: Expression(loc), elements(std::move(e)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
auto result = Value::array();
for (const auto& e : elements) {
@ -1173,8 +1185,8 @@ public:
class DictExpr : public Expression {
std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<Expression>>> elements;
public:
DictExpr(const Location & location, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<Expression>>> && e)
: Expression(location), elements(std::move(e)) {}
DictExpr(const Location & loc, std::vector<std::pair<std::shared_ptr<Expression>, std::shared_ptr<Expression>>> && e)
: Expression(loc), elements(std::move(e)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
auto result = Value::object();
for (const auto& [key, value] : elements) {
@ -1189,8 +1201,8 @@ public:
class SliceExpr : public Expression {
public:
std::shared_ptr<Expression> start, end;
SliceExpr(const Location & location, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
: Expression(location), start(std::move(s)), end(std::move(e)) {}
SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
: Expression(loc), start(std::move(s)), end(std::move(e)) {}
Value do_evaluate(const std::shared_ptr<Context> &) const override {
throw std::runtime_error("SliceExpr not implemented");
}
@ -1200,8 +1212,8 @@ class SubscriptExpr : public Expression {
std::shared_ptr<Expression> base;
std::shared_ptr<Expression> index;
public:
SubscriptExpr(const Location & location, std::shared_ptr<Expression> && b, std::shared_ptr<Expression> && i)
: Expression(location), base(std::move(b)), index(std::move(i)) {}
SubscriptExpr(const Location & loc, std::shared_ptr<Expression> && b, std::shared_ptr<Expression> && i)
: Expression(loc), base(std::move(b)), index(std::move(i)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
if (!base) throw std::runtime_error("SubscriptExpr.base is null");
if (!index) throw std::runtime_error("SubscriptExpr.index is null");
@ -1243,8 +1255,8 @@ public:
enum class Op { Plus, Minus, LogicalNot, Expansion, ExpansionDict };
std::shared_ptr<Expression> expr;
Op op;
UnaryOpExpr(const Location & location, std::shared_ptr<Expression> && e, Op o)
: Expression(location), expr(std::move(e)), op(o) {}
UnaryOpExpr(const Location & loc, std::shared_ptr<Expression> && e, Op o)
: Expression(loc), expr(std::move(e)), op(o) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
if (!expr) throw std::runtime_error("UnaryOpExpr.expr is null");
auto e = expr->evaluate(context);
@ -1269,8 +1281,8 @@ private:
std::shared_ptr<Expression> right;
Op op;
public:
BinaryOpExpr(const Location & location, std::shared_ptr<Expression> && l, std::shared_ptr<Expression> && r, Op o)
: Expression(location), left(std::move(l)), right(std::move(r)), op(o) {}
BinaryOpExpr(const Location & loc, std::shared_ptr<Expression> && l, std::shared_ptr<Expression> && r, Op o)
: Expression(loc), left(std::move(l)), right(std::move(r)), op(o) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
if (!left) throw std::runtime_error("BinaryOpExpr.left is null");
if (!right) throw std::runtime_error("BinaryOpExpr.right is null");
@ -1427,8 +1439,8 @@ class MethodCallExpr : public Expression {
std::shared_ptr<VariableExpr> method;
ArgumentsExpression args;
public:
MethodCallExpr(const Location & location, std::shared_ptr<Expression> && obj, std::shared_ptr<VariableExpr> && m, ArgumentsExpression && a)
: Expression(location), object(std::move(obj)), method(std::move(m)), args(std::move(a)) {}
MethodCallExpr(const Location & loc, std::shared_ptr<Expression> && obj, std::shared_ptr<VariableExpr> && m, ArgumentsExpression && a)
: Expression(loc), object(std::move(obj)), method(std::move(m)), args(std::move(a)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
if (!object) throw std::runtime_error("MethodCallExpr.object is null");
if (!method) throw std::runtime_error("MethodCallExpr.method is null");
@ -1526,8 +1538,8 @@ class CallExpr : public Expression {
public:
std::shared_ptr<Expression> object;
ArgumentsExpression args;
CallExpr(const Location & location, std::shared_ptr<Expression> && obj, ArgumentsExpression && a)
: Expression(location), object(std::move(obj)), args(std::move(a)) {}
CallExpr(const Location & loc, std::shared_ptr<Expression> && obj, ArgumentsExpression && a)
: Expression(loc), object(std::move(obj)), args(std::move(a)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
if (!object) throw std::runtime_error("CallExpr.object is null");
auto obj = object->evaluate(context);
@ -1542,8 +1554,8 @@ public:
class FilterExpr : public Expression {
std::vector<std::shared_ptr<Expression>> parts;
public:
FilterExpr(const Location & location, std::vector<std::shared_ptr<Expression>> && p)
: Expression(location), parts(std::move(p)) {}
FilterExpr(const Location & loc, std::vector<std::shared_ptr<Expression>> && p)
: Expression(loc), parts(std::move(p)) {}
Value do_evaluate(const std::shared_ptr<Context> & context) const override {
Value result;
bool first = true;
@ -2460,7 +2472,7 @@ private:
static std::regex leading_space_regex(R"(^\s+)");
text = std::regex_replace(text, leading_space_regex, "");
} else if (options.trim_blocks && (it - 1) != begin && !dynamic_cast<ExpressionTemplateToken*>((*(it - 2)).get())) {
if (text.length() > 0 && text[0] == '\n') {
if (!text.empty() && text[0] == '\n') {
text.erase(0, 1);
}
}
@ -2538,7 +2550,7 @@ public:
TemplateTokenIterator begin = tokens.begin();
auto it = begin;
TemplateTokenIterator end = tokens.end();
return parser.parseTemplate(begin, it, end, /* full= */ true);
return parser.parseTemplate(begin, it, end, /* fully= */ true);
}
};
@ -2577,7 +2589,7 @@ inline std::shared_ptr<Context> Context::builtins() {
throw std::runtime_error(args.at("message").get<std::string>());
}));
globals.set("tojson", simple_function("tojson", { "value", "indent" }, [](const std::shared_ptr<Context> &, Value & args) {
return Value(args.at("value").dump(args.get<int64_t>("indent", -1), /* tojson= */ true));
return Value(args.at("value").dump(args.get<int64_t>("indent", -1), /* to_json= */ true));
}));
globals.set("items", simple_function("items", { "object" }, [](const std::shared_ptr<Context> &, Value & args) {
auto items = Value::array();
@ -2599,7 +2611,7 @@ inline std::shared_ptr<Context> Context::builtins() {
globals.set("last", simple_function("last", { "items" }, [](const std::shared_ptr<Context> &, Value & args) {
auto items = args.at("items");
if (!items.is_array()) throw std::runtime_error("object is not a list");
if (items.size() == 0) return Value();
if (items.empty()) return Value();
return items.at(items.size() - 1);
}));
globals.set("trim", simple_function("trim", { "text" }, [](const std::shared_ptr<Context> &, Value & args) {
@ -2747,12 +2759,17 @@ inline std::shared_ptr<Context> Context::builtins() {
return Value::callable([=](const std::shared_ptr<Context> & context, ArgumentsValue & args) {
args.expectArgs(is_select ? "select" : "reject", {2, (std::numeric_limits<size_t>::max)()}, {0, 0});
auto & items = args.args[0];
if (items.is_null())
if (items.is_null()) {
return Value::array();
if (!items.is_array()) throw std::runtime_error("object is not iterable: " + items.dump());
}
if (!items.is_array()) {
throw std::runtime_error("object is not iterable: " + items.dump());
}
auto filter_fn = context->get(args.args[1]);
if (filter_fn.is_null()) throw std::runtime_error("Undefined filter: " + args.args[1].dump());
if (filter_fn.is_null()) {
throw std::runtime_error("Undefined filter: " + args.args[1].dump());
}
auto filter_args = Value::array();
for (size_t i = 2, n = args.args.size(); i < n; i++) {
@ -2874,20 +2891,25 @@ inline std::shared_ptr<Context> Context::builtins() {
auto v = arg.get<int64_t>();
startEndStep[i] = v;
param_set[i] = true;
}
}
for (auto & [name, value] : args.kwargs) {
size_t i;
if (name == "start") i = 0;
else if (name == "end") i = 1;
else if (name == "step") i = 2;
else throw std::runtime_error("Unknown argument " + name + " for function range");
}
for (auto & [name, value] : args.kwargs) {
size_t i;
if (name == "start") {
i = 0;
} else if (name == "end") {
i = 1;
} else if (name == "step") {
i = 2;
} else {
throw std::runtime_error("Unknown argument " + name + " for function range");
}
if (param_set[i]) {
throw std::runtime_error("Duplicate argument " + name + " for function range");
}
startEndStep[i] = value.get<int64_t>();
param_set[i] = true;
if (param_set[i]) {
throw std::runtime_error("Duplicate argument " + name + " for function range");
}
startEndStep[i] = value.get<int64_t>();
param_set[i] = true;
}
if (!param_set[1]) {
throw std::runtime_error("Missing required argument 'end' for function range");

185
convert_hf_to_gguf.py
View File

@ -65,6 +65,7 @@ class Model:
model_name: str | None
metadata_override: Path | None
dir_model_card: Path
remote_hf_model_id: str | None
# subclasses should define this!
model_arch: gguf.MODEL_ARCH
@ -73,7 +74,7 @@ class Model:
use_temp_file: bool = False, eager: bool = False,
metadata_override: Path | None = None, model_name: str | None = None,
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
if type(self) is Model:
raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
@ -83,11 +84,24 @@ class Model:
self.is_big_endian = is_big_endian
self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
self.use_temp_file = use_temp_file
self.lazy = not eager
self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
self.is_safetensors = len(self.part_names) > 0
if not self.is_safetensors:
self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
self.lazy = not eager or (remote_hf_model_id is not None)
self.remote_hf_model_id = remote_hf_model_id
if remote_hf_model_id is not None:
self.is_safetensors = True
def get_remote_tensors() -> Iterator[tuple[str, Tensor]]:
logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}")
remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id)
self.tensor_names = set(name for name in remote_tensors.keys())
for name, remote_tensor in gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id).items():
yield (name, LazyTorchTensor.from_remote_tensor(remote_tensor))
self.get_tensors = get_remote_tensors
else:
self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
self.is_safetensors = len(self.part_names) > 0
if not self.is_safetensors:
self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
@ -393,6 +407,10 @@ class Model:
self.metadata = gguf.Metadata.load(self.metadata_override, self.dir_model_card, self.model_name, total_params)
# If we are using HF model id, set the metadata name to the model id
if self.remote_hf_model_id:
self.metadata.name = self.remote_hf_model_id
# Fallback to model directory name if metadata name is still missing
if self.metadata.name is None:
self.metadata.name = self.dir_model.name
@ -714,6 +732,12 @@ class Model:
if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224":
# ref: https://huggingface.co/inclusionAI/Ling-lite
res = "bailingmoe"
if chkhsh == "d353350c764d8c3b39c763113960e4fb4919bea5fbf208a0e3b22e8469dc7406":
# ref: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
res = "llama4"
if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
# ref: https://huggingface.co/THUDM/glm-4-9b-hf
res = "glm4"
if res is None:
logger.warning("\n")
@ -1608,6 +1632,7 @@ class StableLMModel(Model):
@Model.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
class LlamaModel(Model):
model_arch = gguf.MODEL_ARCH.LLAMA
undo_permute = True
def set_vocab(self):
try:
@ -1672,10 +1697,11 @@ class LlamaModel(Model):
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads")
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
if self.undo_permute:
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
# process the experts separately
if name.find("block_sparse_moe.experts") != -1:
@ -1727,7 +1753,7 @@ class LlamaModel(Model):
low_freq_wavelen = old_context_len / low_freq_factor
high_freq_wavelen = old_context_len / high_freq_factor
assert low_freq_wavelen != high_freq_wavelen
# assert low_freq_wavelen != high_freq_wavelen # Errors for Llama4
rope_factors = []
for freq in freqs:
@ -1752,6 +1778,57 @@ class LlamaModel(Model):
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("Llama4ForConditionalGeneration")
class Llama4Model(LlamaModel):
model_arch = gguf.MODEL_ARCH.LLAMA4
has_vision: bool = False
undo_permute = False
# TODO @ngxson : avoid duplicate this code everywhere by at least support "text_config"
# same with llama, but we need to merge the text_config into the root level of hparams
def __init__(self, *args, **kwargs):
hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
if "text_config" in hparams:
hparams = {**hparams, **hparams["text_config"]}
kwargs["hparams"] = hparams
super().__init__(*args, **kwargs)
if "vision_config" in hparams:
logger.info("Has vision encoder, but it will be ignored")
self.has_vision = True
# IMPORTANT: the normal "intermediate_size" is renamed to "intermediate_size_mlp", we need to undo this
self.hparams["intermediate_size_moe"] = self.hparams["intermediate_size"]
self.hparams["intermediate_size"] = self.hparams["intermediate_size_mlp"]
def set_vocab(self):
self._set_vocab_gpt2()
self.gguf_writer.add_add_bos_token(True)
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_interleave_moe_layer_step(self.hparams["interleave_moe_layer_step"])
self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
# split the gate_up into gate and up
if "gate_up_proj" in name:
name_up = name.replace("gate_up_proj", "up_proj.weight")
name_gate = name.replace("gate_up_proj", "gate_proj.weight")
dim_half = data_torch.shape[-1] // 2
gate_proj_weight, up_proj_weight = data_torch.transpose(-1, -2).split(dim_half, dim=-2)
return [
(self.map_tensor_name(name_gate), gate_proj_weight),
(self.map_tensor_name(name_up), up_proj_weight)
]
if name.endswith("down_proj"):
name += ".weight"
data_torch = data_torch.transpose(-1, -2)
if "multi_modal_projector" in name or "vision_model" in name:
return []
return super().modify_tensors(data_torch, name, bid)
@Model.register("Mistral3ForConditionalGeneration")
class Mistral3Model(LlamaModel):
model_arch = gguf.MODEL_ARCH.LLAMA
@ -2399,6 +2476,16 @@ class Qwen2MoeModel(Model):
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("Qwen3ForCausalLM")
class Qwen3Model(Qwen2Model):
model_arch = gguf.MODEL_ARCH.QWEN3
@Model.register("Qwen3MoeForCausalLM")
class Qwen3MoeModel(Qwen2MoeModel):
model_arch = gguf.MODEL_ARCH.QWEN3MOE
@Model.register("GPT2LMHeadModel")
class GPT2Model(Model):
model_arch = gguf.MODEL_ARCH.GPT2
@ -4335,6 +4422,10 @@ class DeepseekV2Model(Model):
self._set_vocab_gpt2()
def set_gguf_parameters(self):
# note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
self.hparams["num_key_value_heads"] = 1
super().set_gguf_parameters()
hparams = self.hparams
@ -4343,8 +4434,13 @@ class DeepseekV2Model(Model):
if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_value_length(hparams["v_head_dim"])
# note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_value_length(hparams["kv_lora_rank"])
self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
@ -4413,6 +4509,26 @@ class DeepseekV2Model(Model):
else:
return []
# note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
if name.endswith("kv_b_proj.weight"):
name_kb = name.replace("kv_b_proj", "k_b_proj")
name_vb = name.replace("kv_b_proj", "v_b_proj")
n_head_kv = self.hparams["num_key_value_heads"]
v_head_dim = self.hparams["v_head_dim"]
qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
k_b = k_b.transpose(1, 2)
return [
(self.map_tensor_name(name_kb), k_b),
(self.map_tensor_name(name_vb), v_b)
]
return [(self.map_tensor_name(name), data_torch)]
def prepare_tensors(self):
@ -4813,6 +4929,22 @@ class JaisModel(Model):
self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
@Model.register("Glm4ForCausalLM")
class Glm4Model(Model):
model_arch = gguf.MODEL_ARCH.GLM4
def set_vocab(self):
self._set_vocab_gpt2()
def set_gguf_parameters(self):
super().set_gguf_parameters()
if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
if self.hparams["rope_scaling"].get("type") == "yarn":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
@Model.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
class ChatGLMModel(Model):
model_arch = gguf.MODEL_ARCH.CHATGLM
@ -5333,6 +5465,14 @@ class LazyTorchTensor(gguf.LazyBase):
lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
return cast(torch.Tensor, lazy)
@classmethod
def from_remote_tensor(cls, remote_tensor: gguf.utility.RemoteTensor):
dtype = cls._dtype_str_map[remote_tensor.dtype]
shape = remote_tensor.shape
meta = cls.meta_with_dtype_and_shape(dtype, shape)
lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.frombuffer(r.data(), dtype=dtype).reshape(shape))
return cast(torch.Tensor, lazy)
@classmethod
def __torch_function__(cls, func, types, args=(), kwargs=None):
del types # unused
@ -5410,6 +5550,10 @@ def parse_args() -> argparse.Namespace:
"--print-supported-models", action="store_true",
help="Print the supported models"
)
parser.add_argument(
"--remote", action="store_true",
help="(Experimental) Read safetensors file remotely without downloading to disk. Config and tokenizer files will still be downloaded. To use this feature, you need to specify Hugging Face model repo name instead of a local directory. For example: 'HuggingFaceTB/SmolLM2-1.7B-Instruct'. Note: To access gated repo, set HF_TOKEN environment variable to your Hugging Face token.",
)
args = parser.parse_args()
if not args.print_supported_models and args.model is None:
@ -5450,6 +5594,14 @@ def main() -> None:
dir_model = args.model
if args.remote:
from huggingface_hub import snapshot_download
local_dir = snapshot_download(
repo_id=str(dir_model),
allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
dir_model = Path(local_dir)
logger.info(f"Downloaded config and tokenizer to {local_dir}")
if not dir_model.is_dir():
logger.error(f'Error: {args.model} is not a directory')
sys.exit(1)
@ -5471,6 +5623,9 @@ def main() -> None:
if args.outfile is not None:
fname_out = args.outfile
elif args.remote:
# if remote, use the model ID as the output file name
fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
else:
fname_out = dir_model
@ -5481,7 +5636,6 @@ def main() -> None:
with torch.inference_mode():
output_type = ftype_map[args.outtype]
model_architecture = hparams["architectures"][0]
try:
model_class = Model.from_model_architecture(model_architecture)
except NotImplementedError:
@ -5494,7 +5648,8 @@ def main() -> None:
metadata_override=args.metadata, model_name=args.model_name,
split_max_tensors=args.split_max_tensors,
split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
small_first_shard=args.no_tensor_first_split)
small_first_shard=args.no_tensor_first_split,
remote_hf_model_id=str(args.model) if args.remote else None)
if args.vocab_only:
logger.info("Exporting model vocab...")

2
convert_hf_to_gguf_update.py
View File

@ -113,6 +113,8 @@ models = [
{"name": "superbpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
{"name": "trillion", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },
{"name": "bailingmoe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
{"name": "llama4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
{"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
]

10
docs/backend/OPENCL.md
View File

@ -145,8 +145,13 @@ A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the followi
* Clang 19
* Ninja
* Visual Studio 2022
* Powershell 7
Powershell is used for the following instructions.
Visual Studio provides necessary headers and libraries although it is not directly used for building.
Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
Powershell 7 is used for the following commands.
If an older version of Powershell is used, these commands may not work as they are.
### I. Setup Environment
@ -196,10 +201,9 @@ ninja
## Known Issues
- Qwen2.5 0.5B model produces gibberish output with Adreno kernels.
- Currently OpenCL backend does not work on Adreno 6xx GPUs.
## TODO
- Fix Qwen2.5 0.5B
- Optimization for Q6_K
- Support and optimization for Q4_K

95
docs/backend/SYCL.md
View File

@ -425,13 +425,13 @@ Examples:
- Use device 0:
```sh
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
```
- Use multiple devices:
```sh
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
```
*Notes:*
@ -475,6 +475,12 @@ b. Enable oneAPI running environment:
"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
```
- if you are using Powershell, enable the runtime environment with the following:
```
cmd.exe "/K" '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell'
```
c. Verify installation
In the oneAPI command line, run the following to print the available SYCL devices:
@ -505,13 +511,13 @@ You could download the release package for Windows directly, which including bin
Choose one of following methods to build from source code.
1. Script
#### 1. Script
```sh
.\examples\sycl\win-build-sycl.bat
```
2. CMake
#### 2. CMake
On the oneAPI command line window, step into the llama.cpp main directory and run the following:
@ -540,13 +546,84 @@ cmake --preset x64-windows-sycl-debug
cmake --build build-x64-windows-sycl-debug -j --target llama-cli
```
3. Visual Studio
#### 3. Visual Studio
You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
You have two options to use Visual Studio to build llama.cpp:
- As CMake Project using CMake presets.
- Creating a Visual Studio solution to handle the project.
**Note**:
All following commands are executed in PowerShell.
##### - Open as a CMake Project
You can use Visual Studio to open the `llama.cpp` folder directly as a CMake project. Before compiling, select one of the SYCL CMake presets:
- `x64-windows-sycl-release`
- `x64-windows-sycl-debug`
*Notes:*
- For a minimal experimental setup, you can build only the inference executable using:
- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`.
```Powershell
cmake --build build --config Release -j --target llama-cli
```
##### - Generating a Visual Studio Solution
You can use Visual Studio solution to build and work on llama.cpp on Windows. You need to convert the CMake Project into a `.sln` file.
If you want to use the Intel C++ Compiler for the entire `llama.cpp` project, run the following command:
```Powershell
cmake -B build -G "Visual Studio 17 2022" -T "Intel C++ Compiler 2025" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release
```
If you prefer to use the Intel C++ Compiler only for `ggml-sycl`, ensure that `ggml` and its backend libraries are built as shared libraries ( i.e. `-DBUILD_SHARED_LIBRARIES=ON`, this is default behaviour):
```Powershell
cmake -B build -G "Visual Studio 17 2022" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release \
-DSYCL_INCLUDE_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\include" \
-DSYCL_LIBRARY_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\lib"
```
If successful the build files have been written to: *path/to/llama.cpp/build*
Open the project file **build/llama.cpp.sln** with Visual Studio.
Once the Visual Studio solution is created, follow these steps:
1. Open the solution in Visual Studio.
2. Right-click on `ggml-sycl` and select **Properties**.
3. In the left column, expand **C/C++** and select **DPC++**.
4. In the right panel, find **Enable SYCL Offload** and set it to `Yes`.
5. Apply the changes and save.
*Navigation Path:*
```
Properties -> C/C++ -> DPC++ -> Enable SYCL Offload (Yes)
```
Now, you can build `llama.cpp` with the SYCL backend as a Visual Studio project.
To do it from menu: `Build -> Build Solution`.
Once it is completed, final results will be in **build/Release/bin**
*Additional Note*
- You can avoid specifying `SYCL_INCLUDE_DIR` and `SYCL_LIBRARY_DIR` in the CMake command by setting the environment variables:
- `SYCL_INCLUDE_DIR_HINT`
- `SYCL_LIBRARY_DIR_HINT`
- Above instruction has been tested with Visual Studio 17 Community edition and oneAPI 2025.0. We expect them to work also with future version if the instructions are adapted accordingly.
### III. Run the inference
@ -620,13 +697,13 @@ Examples:
- Use device 0:
```
build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
```
- Use multiple devices:
```
build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
```

96
docs/build.md
View File

@ -259,8 +259,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
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`.
However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
@ -296,6 +294,10 @@ 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.
### Unified Memory
On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
## Vulkan
**Windows**
@ -456,6 +458,96 @@ KleidiAI's microkernels implement optimized tensor operations using Arm CPU feat
Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
## OpenCL
This provides GPU acceleration through OpenCL on recent Adreno GPU.
More information about OpenCL backend can be found in [OPENCL.md](./backend/OPENCL.md) for more information.
### Android
Assume NDK is available in `$ANDROID_NDK`. First, install OpenCL headers and ICD loader library if not available,
```sh
mkdir -p ~/dev/llm
cd ~/dev/llm
git clone https://github.com/KhronosGroup/OpenCL-Headers && \
cd OpenCL-Headers && \
cp -r CL $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
cd ~/dev/llm
git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && \
cd OpenCL-ICD-Loader && \
mkdir build_ndk && cd build_ndk && \
cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
-DOPENCL_ICD_LOADER_HEADERS_DIR=$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
-DANDROID_ABI=arm64-v8a \
-DANDROID_PLATFORM=24 \
-DANDROID_STL=c++_shared && \
ninja && \
cp libOpenCL.so $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
```
Then build llama.cpp with OpenCL enabled,
```sh
cd ~/dev/llm
git clone https://github.com/ggml-org/llama.cpp && \
cd llama.cpp && \
mkdir build-android && cd build-android
cmake .. -G Ninja \
-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
-DANDROID_ABI=arm64-v8a \
-DANDROID_PLATFORM=android-28 \
-DBUILD_SHARED_LIBS=OFF \
-DGGML_OPENCL=ON
ninja
```
### Windows Arm64
First, install OpenCL headers and ICD loader library if not available,
```powershell
mkdir -p ~/dev/llm
cd ~/dev/llm
git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers
mkdir build && cd build
cmake .. -G Ninja `
-DBUILD_TESTING=OFF `
-DOPENCL_HEADERS_BUILD_TESTING=OFF `
-DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF `
-DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"
cmake --build . --target install
cd ~/dev/llm
git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader
mkdir build && cd build
cmake .. -G Ninja `
-DCMAKE_BUILD_TYPE=Release `
-DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `
-DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl"
cmake --build . --target install
```
Then build llama.cpp with OpenCL enabled,
```powershell
cmake .. -G Ninja `
-DCMAKE_TOOLCHAIN_FILE="$HOME/dev/llm/llama.cpp/cmake/arm64-windows-llvm.cmake" `
-DCMAKE_BUILD_TYPE=Release `
-DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" `
-DBUILD_SHARED_LIBS=OFF `
-DGGML_OPENCL=ON
ninja
```
## Android
To read documentation for how to build on Android, [click here](./android.md)

31
examples/gguf-split/gguf-split.cpp
View File

@ -408,8 +408,6 @@ static void gguf_merge(const split_params & split_params) {
exit(EXIT_FAILURE);
}
std::ofstream fout(split_params.output.c_str(), std::ios::binary);
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
auto * ctx_out = gguf_init_empty();
@ -453,7 +451,6 @@ static void gguf_merge(const split_params & split_params) {
gguf_free(ctx_gguf);
ggml_free(ctx_meta);
gguf_free(ctx_out);
fout.close();
exit(EXIT_FAILURE);
}
@ -466,7 +463,6 @@ static void gguf_merge(const split_params & split_params) {
gguf_free(ctx_gguf);
ggml_free(ctx_meta);
gguf_free(ctx_out);
fout.close();
exit(EXIT_FAILURE);
}
@ -479,7 +475,6 @@ static void gguf_merge(const split_params & split_params) {
gguf_free(ctx_gguf);
ggml_free(ctx_meta);
gguf_free(ctx_out);
fout.close();
exit(EXIT_FAILURE);
}
@ -500,9 +495,11 @@ static void gguf_merge(const split_params & split_params) {
fprintf(stderr, "\033[3Ddone\n");
}
// placeholder for the meta data
{
std::ofstream fout;
if (!split_params.dry_run) {
fout.open(split_params.output.c_str(), std::ios::binary);
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
// placeholder for the meta data
auto meta_size = gguf_get_meta_size(ctx_out);
::zeros(fout, meta_size);
}
@ -518,7 +515,9 @@ static void gguf_merge(const split_params & split_params) {
ggml_free(ctx_metas[i]);
}
gguf_free(ctx_out);
fout.close();
if (!split_params.dry_run) {
fout.close();
}
exit(EXIT_FAILURE);
}
fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
@ -540,10 +539,11 @@ static void gguf_merge(const split_params & split_params) {
auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
f_input.seekg(offset);
f_input.read((char *)read_data.data(), n_bytes);
// write tensor data + padding
fout.write((const char *)read_data.data(), n_bytes);
zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
if (!split_params.dry_run) {
// write tensor data + padding
fout.write((const char *)read_data.data(), n_bytes);
zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
}
}
gguf_free(ctx_gguf);
@ -552,16 +552,15 @@ static void gguf_merge(const split_params & split_params) {
fprintf(stderr, "\033[3Ddone\n");
}
{
if (!split_params.dry_run) {
// go back to beginning of file and write the updated metadata
fout.seekp(0);
std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
gguf_get_meta_data(ctx_out, data.data());
fout.write((const char *)data.data(), data.size());
fout.close();
gguf_free(ctx_out);
}
gguf_free(ctx_out);
fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
__func__, split_params.output.c_str(), n_split, total_tensors);

1
examples/llama.android/llama/build.gradle.kts
View File

@ -18,6 +18,7 @@ android {
}
externalNativeBuild {
cmake {
arguments += "-DLLAMA_CURL=OFF"
arguments += "-DLLAMA_BUILD_COMMON=ON"
arguments += "-DGGML_LLAMAFILE=OFF"
arguments += "-DCMAKE_BUILD_TYPE=Release"

33
examples/llava/CMakeLists.txt
View File

@ -1,3 +1,5 @@
# llava (legacy)
add_library(llava OBJECT
llava.cpp
llava.h
@ -22,12 +24,41 @@ if (BUILD_SHARED_LIBS)
install(TARGETS llava_shared LIBRARY)
endif()
# mtmd
add_library(mtmd OBJECT
mtmd.cpp
mtmd.h
clip.cpp
clip.h
clip-impl.h
)
target_link_libraries(mtmd PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
target_include_directories(mtmd PUBLIC .)
target_include_directories(mtmd PRIVATE ../..)
target_include_directories(mtmd PRIVATE ../../common) # for stb_image.h
target_compile_features(mtmd PRIVATE cxx_std_17)
add_library(mtmd_static STATIC $<TARGET_OBJECTS:mtmd>)
if (BUILD_SHARED_LIBS)
set_target_properties(mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_definitions(mtmd PRIVATE LLAMA_SHARED LLAMA_BUILD)
add_library(mtmd_shared SHARED $<TARGET_OBJECTS:mtmd>)
target_link_libraries(mtmd_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
install(TARGETS mtmd_shared LIBRARY)
endif()
if (NOT MSVC)
target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
target_compile_options(mtmd PRIVATE -Wno-cast-qual) # stb_image.h
endif()
if(TARGET BUILD_INFO)
add_dependencies(llava BUILD_INFO)
add_dependencies(mtmd BUILD_INFO)
endif()
set(TARGET llama-llava-cli)
@ -55,7 +86,7 @@ set(TARGET llama-gemma3-cli)
add_executable(${TARGET} gemma3-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-llava-clip-quantize-cli)

344
examples/llava/clip-impl.h Normal file
View File

@ -0,0 +1,344 @@
#include "ggml.h"
#include "gguf.h"
#include "clip.h"
#include "clip.h"
#include <climits>
#include <cstdarg>
#include <string>
#include <map>
#include <sstream>
#include <vector>
#include <memory>
// Internal header for clip.cpp
#define KEY_FTYPE "general.file_type"
#define KEY_NAME "general.name"
#define KEY_DESCRIPTION "general.description"
#define KEY_HAS_TEXT_ENC "clip.has_text_encoder"
#define KEY_HAS_VIS_ENC "clip.has_vision_encoder"
#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector"
#define KEY_HAS_GLM_PROJ "clip.has_glm_projector"
#define KEY_MINICPMV_VERSION "clip.minicpmv_version"
#define KEY_HAS_QWEN2VL_MERGER "clip.has_qwen2vl_merger"
#define KEY_USE_GELU "clip.use_gelu"
#define KEY_USE_SILU "clip.use_silu"
#define KEY_N_EMBD "clip.%s.embedding_length"
#define KEY_N_FF "clip.%s.feed_forward_length"
#define KEY_N_BLOCK "clip.%s.block_count"
#define KEY_N_HEAD "clip.%s.attention.head_count"
#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
#define KEY_PROJ_DIM "clip.%s.projection_dim"
#define KEY_TOKENS "tokenizer.ggml.tokens"
#define KEY_N_POSITIONS "clip.text.context_length"
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
#define KEY_IMAGE_MEAN "clip.vision.image_mean"
#define KEY_IMAGE_STD "clip.vision.image_std"
#define KEY_PROJ_TYPE "clip.projector_type"
#define KEY_FEATURE_LAYER "clip.vision.feature_layer"
#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type"
#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints"
#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
//
// tensor name constants
//
#define TN_TOKEN_EMBD "%s.token_embd.weight"
#define TN_POS_EMBD "%s.position_embd.weight"
#define TN_CLASS_EMBD "v.class_embd"
#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backwrad compat
#define TN_PATCH_EMBD_1 "v.patch_embd.weight.1"
#define TN_PATCH_BIAS "v.patch_embd.bias"
#define TN_ATTN_K "%s.blk.%d.attn_k.%s"
#define TN_ATTN_Q "%s.blk.%d.attn_q.%s"
#define TN_ATTN_V "%s.blk.%d.attn_v.%s"
#define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s"
#define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s"
#define TN_FFN_UP "%s.blk.%d.ffn_up.%s"
#define TN_LN_1 "%s.blk.%d.ln1.%s"
#define TN_LN_2 "%s.blk.%d.ln2.%s"
#define TN_LN_PRE "%s.pre_ln.%s"
#define TN_LN_POST "%s.post_ln.%s"
#define TN_TEXT_PROJ "text_projection.weight"
#define TN_VIS_PROJ "visual_projection.weight"
#define TN_LLAVA_PROJ "mm.%d.%s"
#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s"
#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
#define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s"
#define TN_IMAGE_NEWLINE "model.image_newline"
#define TN_MM_INP_PROJ "mm.input_projection.weight" // gemma3
#define TN_MM_SOFT_EMB_N "mm.soft_emb_norm.weight" // gemma3
// mimicpmv
#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
#define TN_MINICPMV_QUERY "resampler.query"
#define TN_MINICPMV_PROJ "resampler.proj.weight"
#define TN_MINICPMV_KV_PROJ "resampler.kv.weight"
#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
#define TN_MINICPMV_LN "resampler.ln_%s.%s"
#define TN_GLM_ADAPER_CONV "adapter.conv.%s"
#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s"
#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s"
#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s"
#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
#define TN_GLM_BOI_W "adapter.boi"
#define TN_GLM_EOI_W "adapter.eoi"
enum projector_type {
PROJECTOR_TYPE_MLP,
PROJECTOR_TYPE_MLP_NORM,
PROJECTOR_TYPE_LDP,
PROJECTOR_TYPE_LDPV2,
PROJECTOR_TYPE_RESAMPLER,
PROJECTOR_TYPE_GLM_EDGE,
PROJECTOR_TYPE_MERGER,
PROJECTOR_TYPE_GEMMA3,
PROJECTOR_TYPE_UNKNOWN,
};
static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_MLP, "mlp" },
{ PROJECTOR_TYPE_LDP, "ldp" },
{ PROJECTOR_TYPE_LDPV2, "ldpv2"},
{ PROJECTOR_TYPE_RESAMPLER, "resampler"},
{ PROJECTOR_TYPE_GLM_EDGE, "adapter"},
{ PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
{ PROJECTOR_TYPE_GEMMA3, "gemma3"},
};
static projector_type clip_projector_type_from_string(const std::string & str) {
for (const auto & pair : PROJECTOR_TYPE_NAMES) {
if (pair.second == str) {
return pair.first;
}
}
return PROJECTOR_TYPE_UNKNOWN;
}
// RGB uint8 image
struct clip_image_u8 {
int nx;
int ny;
std::vector<uint8_t> buf;
};
// RGB float32 image (NHWC)
// Memory layout: RGBRGBRGB...
struct clip_image_f32 {
int nx;
int ny;
std::vector<float> buf;
};
//
// logging
//
static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
fputs(text, stderr);
fflush(stderr);
}
struct clip_logger_state {
ggml_log_level verbosity_thold;
ggml_log_callback log_callback;
void * log_callback_user_data;
};
extern struct clip_logger_state g_logger_state;
static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
if (format == NULL) {
return;
}
va_list args_copy;
va_copy(args_copy, args);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
} else {
char * buffer2 = (char *) calloc(len + 1, sizeof(char));
vsnprintf(buffer2, len + 1, format, args_copy);
buffer2[len] = 0;
g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
free(buffer2);
}
va_end(args_copy);
}
static void clip_log_internal(enum ggml_log_level level, const char * format, ...) {
va_list args;
va_start(args, format);
clip_log_internal_v(level, format, args);
va_end(args);
}
#define LOG_TMPL(level, ...) \
do { \
if ((level) >= g_logger_state.verbosity_thold) { \
clip_log_internal((level), __VA_ARGS__); \
} \
} while (0)
#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT, __VA_ARGS__)
//
// cpp wrappers
//
// wrapper for clip_image_size
struct clip_image_size_deleter {
void operator()(clip_image_size * val) { clip_image_size_free(val); }
};
typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
// wrapper for clip_image_u8
struct clip_image_u8_deleter {
void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
};
typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
// wrapper for clip_image_f32
struct clip_image_f32_deleter {
void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
};
typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
struct clip_image_u8_batch {
std::vector<clip_image_u8_ptr> entries;
};
struct clip_image_f32_batch {
std::vector<clip_image_f32_ptr> entries;
};
//
// common utils
//
static std::string string_format(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::vector<char> buf(size + 1);
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return std::string(buf.data(), buf.size());
}
static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
if (search.empty()) {
return;
}
std::string builder;
builder.reserve(s.length());
size_t pos = 0;
size_t last_pos = 0;
while ((pos = s.find(search, last_pos)) != std::string::npos) {
builder.append(s, last_pos, pos - last_pos);
builder.append(replace);
last_pos = pos + search.length();
}
builder.append(s, last_pos, std::string::npos);
s = std::move(builder);
}
// split string by a `std::string delim` instead of `char delim`
static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
std::vector<std::string> tokens;
size_t pos = 0;
std::string token;
while ((pos = s.find(delimiter)) != std::string::npos) {
token = s.substr(0, pos);
tokens.push_back(token);
s.erase(0, pos + delimiter.length());
}
tokens.push_back(s);
return tokens;
}
//
// gguf utils
//
static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
switch (type) {
case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]);
case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]);
case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]);
case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]);
case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]);
case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]);
case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]);
case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]);
case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]);
case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]);
case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false";
default: return string_format("unknown type %d", type);
}
}
static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
switch (type) {
case GGUF_TYPE_STRING:
return gguf_get_val_str(ctx_gguf, i);
case GGUF_TYPE_ARRAY:
{
const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
int arr_n = gguf_get_arr_n(ctx_gguf, i);
const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
std::stringstream ss;
ss << "[";
for (int j = 0; j < arr_n; j++) {
if (arr_type == GGUF_TYPE_STRING) {
std::string val = gguf_get_arr_str(ctx_gguf, i, j);
// escape quotes
string_replace_all(val, "\\", "\\\\");
string_replace_all(val, "\"", "\\\"");
ss << '"' << val << '"';
} else if (arr_type == GGUF_TYPE_ARRAY) {
ss << "???";
} else {
ss << gguf_data_to_str(arr_type, data, j);
}
if (j < arr_n - 1) {
ss << ", ";
}
}
ss << "]";
return ss.str();
}
default:
return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
}
}
//
// API used internally with mtmd
//
projector_type clip_get_projector_type(const struct clip_ctx * ctx);

1785
examples/llava/clip.cpp
View File

File diff suppressed because it is too large Load Diff

39
examples/llava/clip.h
View File

@ -1,6 +1,7 @@
#ifndef CLIP_H
#define CLIP_H
#include "ggml.h"
#include <stddef.h>
#include <stdint.h>
@ -29,19 +30,12 @@ struct clip_image_size {
int height;
};
struct clip_image_u8_batch {
struct clip_image_u8 * data;
size_t size;
};
struct clip_image_f32_batch {
struct clip_image_f32 * data;
size_t size;
};
struct clip_image_u8_batch;
struct clip_image_f32_batch;
struct clip_context_params {
bool use_gpu;
int verbosity;
ggml_log_level verbosity;
};
// deprecated, use clip_init
@ -54,9 +48,9 @@ CLIP_API void clip_free(struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
// TODO: should be enum, not string
CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
@ -72,15 +66,26 @@ CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
CLIP_API struct clip_image_size * clip_image_size_init();
CLIP_API struct clip_image_u8 * clip_image_u8_init ();
CLIP_API struct clip_image_f32 * clip_image_f32_init();
CLIP_API struct clip_image_size * clip_image_size_init();
CLIP_API struct clip_image_u8 * clip_image_u8_init ();
CLIP_API struct clip_image_f32 * clip_image_f32_init();
CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava
// nx, ny are the output image dimensions
CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch * batch);
CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
// use for accessing underlay data of clip_image_f32_batch
CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
CLIP_API clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
/**
* Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
* The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
@ -105,6 +110,8 @@ CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out
CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);

201
examples/llava/gemma3-cli.cpp
View File

@ -2,15 +2,15 @@
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "stb_image.h"
#include "llama.h"
#include "ggml.h"
#include "console.h"
#include "chat.h"
#include "mtmd.h"
#include <vector>
#include <limits.h>
#include <inttypes.h>
#include <cinttypes>
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
#include <signal.h>
@ -57,13 +57,18 @@ static void sigint_handler(int signo) {
#endif
struct gemma3_context {
struct clip_ctx * ctx_clip = NULL;
common_init_result llama_init;
mtmd_context_ptr ctx_vision;
common_init_result llama_init;
llama_model * model;
llama_context * lctx;
const llama_vocab * vocab;
llama_batch batch;
int n_batch;
// note: we know that gemma3 template is "linear", meaning each turn is completely separated to another
// so here we don't need to keep track of chat history
common_chat_templates_ptr tmpls;
int n_threads = 1;
llama_pos n_past = 0;
@ -74,16 +79,23 @@ struct gemma3_context {
vocab = llama_model_get_vocab(model);
n_threads = params.cpuparams.n_threads;
batch = llama_batch_init(params.n_batch, 0, 1);
init_clip_model(params);
n_batch = params.n_batch;
tmpls = common_chat_templates_init(model, params.chat_template);
init_vision_context(params);
}
void init_clip_model(common_params & params) {
void init_vision_context(common_params & params) {
const char * clip_path = params.mmproj.path.c_str();
ctx_clip = clip_model_load(clip_path, params.verbosity > 1);
}
~gemma3_context() {
clip_free(ctx_clip);
ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{
/* use_gpu */ true,
/* timings */ true,
/* n_threads */ params.cpuparams.n_threads,
/* verbosity */ GGML_LOG_LEVEL_INFO,
}));
if (!ctx_vision.get()) {
LOG_ERR("Failed to load vision model from %s\n", clip_path);
exit(1);
}
}
};
@ -120,77 +132,6 @@ struct decode_embd_batch {
}
};
static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) {
llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true);
common_batch_clear(ctx.batch);
for (llama_token & t : tokens) {
common_batch_add(ctx.batch, t, ctx.n_past++, {0}, false);
}
if (logits_last) {
ctx.batch.logits[ctx.batch.n_tokens - 1] = true;
}
// LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str());
if (llama_decode(ctx.lctx, ctx.batch)) {
LOG_ERR("Failed to decode text\n");
return 1;
}
return 0;
}
static int eval_image(gemma3_context & ctx, std::string & fname) {
std::vector<float> image_embd_v;
int n_embd = llama_model_n_embd(ctx.model);
int n_tokens = 256;
image_embd_v.resize(n_tokens * n_embd);
bool ok;
struct clip_image_u8 * img_u8 = clip_image_u8_init();
ok = clip_image_load_from_file(fname.c_str(), img_u8);
if (!ok) {
LOG_ERR("Unable to load image %s\n", fname.c_str());
clip_image_u8_free(img_u8);
return 2; // non-fatal error
}
clip_image_f32_batch batch_f32;
ok = clip_image_preprocess(ctx.ctx_clip, img_u8, &batch_f32);
if (!ok) {
LOG_ERR("Unable to preprocess image\n");
clip_image_f32_batch_free(&batch_f32);
clip_image_u8_free(img_u8);
return 1;
}
int64_t t0 = ggml_time_ms();
LOG("Encoding image %s\n", fname.c_str());
ok = clip_image_batch_encode(ctx.ctx_clip, ctx.n_threads, &batch_f32, image_embd_v.data());
if (!ok) {
LOG_ERR("Unable to encode image\n");
clip_image_f32_batch_free(&batch_f32);
clip_image_u8_free(img_u8);
return 1;
}
LOG("Image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
clip_image_f32_batch_free(&batch_f32);
clip_image_u8_free(img_u8);
// decode image embeddings
int64_t t1 = ggml_time_ms();
eval_text(ctx, "<start_of_image>");
llama_set_causal_attn(ctx.lctx, false);
decode_embd_batch batch_img(image_embd_v.data(), n_tokens, ctx.n_past, 0);
if (llama_decode(ctx.lctx, batch_img.batch)) {
LOG_ERR("failed to decode image\n");
return 1;
}
ctx.n_past += n_tokens;
llama_set_causal_attn(ctx.lctx, true);
eval_text(ctx, "<end_of_image>");
LOG("Image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
return 0;
}
static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) {
for (int i = 0; i < n_predict; i++) {
if (i > n_predict || !g_is_generating) {
@ -220,6 +161,45 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
return 0;
}
static int eval_message(gemma3_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
std::vector<mtmd_bitmap> bitmaps;
common_chat_templates_inputs tmpl_inputs;
tmpl_inputs.messages = {msg};
tmpl_inputs.add_generation_prompt = true;
tmpl_inputs.use_jinja = false; // jinja is buggy here
auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs);
LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str());
for (auto & fname : images_fname) {
mtmd_bitmap bitmap;
if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
LOG_ERR("Unable to load image %s\n", fname.c_str());
return 2; // image not found
}
bitmaps.push_back(std::move(bitmap));
}
mtmd_input_text text;
text.text = formatted_chat.prompt;
text.add_special = add_bos;
text.parse_special = true;
mtmd_input_chunks_ptr chunks(mtmd_tokenize(ctx.ctx_vision.get(), text, bitmaps));
if (chunks == nullptr) {
LOG_ERR("Unable to tokenize prompt\n");
return 1;
}
if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks.get(), ctx.n_past, 0, ctx.n_batch)) {
LOG_ERR("Unable to eval prompt\n");
return 1;
}
ctx.n_past += mtmd_helper_get_n_tokens(chunks.get());
return 0;
}
int main(int argc, char ** argv) {
ggml_time_init();
@ -261,21 +241,15 @@ int main(int argc, char ** argv) {
#endif
}
if (eval_text(ctx, "<bos>")) {
return 1;
}
if (is_single_turn) {
g_is_generating = true;
if (eval_text(ctx, "<start_of_turn>user\n")) {
return 1;
if (params.prompt.find("<__image__>") == std::string::npos) {
params.prompt += " <__image__>";
}
for (auto & fname : params.image) {
if (eval_image(ctx, fname)) {
return 1;
}
}
if (eval_text(ctx, params.prompt + "<end_of_turn><start_of_turn>model\n", true)) {
common_chat_msg msg;
msg.role = "user";
msg.content = params.prompt;
if (eval_message(ctx, msg, params.image, true)) {
return 1;
}
if (generate_response(ctx, smpl, n_predict)) {
@ -289,9 +263,9 @@ int main(int argc, char ** argv) {
LOG("\n /quit or /exit exit the program");
LOG("\n");
if (eval_text(ctx, "<start_of_turn>user\n")) {
return 1;
}
bool is_first_msg = true;
std::vector<std::string> images_fname;
std::string content;
while (true) {
g_is_generating = false;
@ -316,26 +290,33 @@ int main(int argc, char ** argv) {
g_is_generating = true;
if (line.find("/image") == 0) {
std::string image = line.substr(7);
int res = eval_image(ctx, image);
if (res == 2) {
continue; // image not found
}
if (res) {
return 1;
}
images_fname.push_back(string_strip(image));
content += "<__image__>";
continue;
} else {
content += line;
}
common_chat_msg msg;
msg.role = "user";
msg.content = content;
int ret = eval_message(ctx, msg, images_fname, is_first_msg);
if (ret == 2) {
// non-fatal error
images_fname.clear();
content.clear();
continue;
}
if (eval_text(ctx, line + "<end_of_turn><start_of_turn>model\n", true)) {
if (ret) {
return 1;
}
if (generate_response(ctx, smpl, n_predict)) {
return 1;
}
if (eval_text(ctx, "<end_of_turn><start_of_turn>user\n")) {
return 1;
}
images_fname.clear();
content.clear();
is_first_msg = false;
}
}
llama_perf_context_print(ctx.lctx);
return 0;
}

2
examples/llava/llava-cli.cpp
View File

@ -241,7 +241,7 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings

103
examples/llava/llava.cpp
View File

@ -10,6 +10,7 @@
#include <cstring>
#include <limits>
#include <vector>
#include <memory>
#if defined(LLAVA_LOG_OFF)
# define LOG_INF(...)
@ -45,6 +46,17 @@ struct clip_image_grid_shape {
int second;
};
// convenience cpp wrapper
struct clip_image_f32_batch_deleter {
void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
};
typedef std::unique_ptr<clip_image_f32_batch, clip_image_f32_batch_deleter> clip_image_f32_batch_ptr;
struct clip_image_size_deleter {
void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
};
typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
/**
* Selects the best resolution from a list of possible resolutions based on the original size.
*
@ -105,8 +117,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
struct ggml_context * ctx;
} model;
const int32_t image_size = clip_image_size(ctx_clip);
const int32_t patch_size = clip_patch_size(ctx_clip);
const int32_t image_size = clip_get_image_size(ctx_clip);
const int32_t patch_size = clip_get_patch_size(ctx_clip);
int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
@ -246,12 +258,9 @@ static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size)
static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
// std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
clip_image_f32_batch img_res_v;
img_res_v.size = 0;
img_res_v.data = nullptr;
if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
LOG_ERR("%s: unable to preprocess image\n", __func__);
delete[] img_res_v.data;
return false;
}
@ -259,66 +268,72 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get());
if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
std::vector<float *> image_embd_v;
image_embd_v.resize(img_res_v.size);
struct clip_image_size * load_image_size = clip_image_size_init();
image_embd_v.resize(n_imgs);
clip_image_size load_image_size;
for (size_t i = 0; i < img_res_v.size; i++) {
for (size_t i = 0; i < n_imgs; i++) {
const int64_t t_img_enc_step_start_us = ggml_time_us();
image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
int patch_size=14;
load_image_size->width = img_res_v.data[i].nx;
load_image_size->height = img_res_v.data[i].ny;
clip_add_load_image_size(ctx_clip, load_image_size);
int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny));
int patch_size = 14;
load_image_size.width = nx;
load_image_size.height = ny;
clip_add_load_image_size(ctx_clip, &load_image_size);
bool encoded = false;
clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
if (clip_is_qwen2vl(ctx_clip)) {
encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
}
else {
encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, patch_size), image_embd_v[i]);
}
if (!encoded) {
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
return false;
}
const int64_t t_img_enc_steop_batch_us = ggml_time_us();
LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)n_imgs, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
int n_img_pos_out = 0;
for (size_t i = 0; i < image_embd_v.size(); i++) {
int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
std::memcpy(
image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
image_embd_v[i],
clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
clip_embd_nbytes_by_img(ctx_clip, nx, ny));
n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
}
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {
free(image_embd_v[i]);
}
image_embd_v.clear();
load_image_size->width = img->nx;
load_image_size->height = img->ny;
clip_add_load_image_size(ctx_clip, load_image_size);
LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
delete[] img_res_v.data;
img_res_v.size = 0;
img_res_v.data = nullptr;
load_image_size.width = img->nx;
load_image_size.height = img->ny;
clip_add_load_image_size(ctx_clip, &load_image_size);
LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size.width, load_image_size.height);
}
else if (clip_is_glm(ctx_clip)){
struct clip_image_size * load_image_size = clip_image_size_init();
load_image_size->width = img_res_v.data[0].nx;
load_image_size->height = img_res_v.data[0].ny;
load_image_size->width = clip_image_f32_batch_nx(img_res_v.get(), 0);
load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0);
clip_add_load_image_size(ctx_clip, load_image_size);
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd);
int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2);
clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2);
*n_img_pos = (pos * pos + 2);
if (!encoded){
LOG_ERR("Unable to encode image \n");
@ -328,8 +343,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
*n_img_pos = clip_n_patches(ctx_clip);
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
delete[] img_res_v.data;
clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
if (!encoded) {
LOG_ERR("Unable to encode image\n");
@ -340,17 +355,18 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
// spatial_unpad llava-1.6 type embedding
// TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
std::vector<float *> image_embd_v;
image_embd_v.resize(img_res_v.size);
for (size_t i = 0; i < img_res_v.size; i++) {
image_embd_v.resize(n_imgs);
for (size_t i = 0; i < n_imgs; i++) {
clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
if (!encoded) {
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
return false;
}
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
const int32_t * image_grid = clip_image_grid(ctx_clip);
const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);
@ -360,12 +376,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
}
// free all img_res_v - not needed anymore
delete[] img_res_v.data;
img_res_v.size = 0;
img_res_v.data = nullptr;
const int32_t image_size = clip_image_size(ctx_clip);
const int32_t image_size = clip_get_image_size(ctx_clip);
struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);

2
examples/llava/minicpmv-cli.cpp
View File

@ -88,7 +88,7 @@ static struct clip_ctx * clip_init_context(common_params * params) {
}
struct clip_context_params clip_params = {
/* use_gpu */ params->n_gpu_layers != 0,
/* verbosity */ params->verbosity,
/* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable
};
auto * ctx_clip = clip_init(clip_path, clip_params);
return ctx_clip;

341
examples/llava/mtmd.cpp Normal file
View File

@ -0,0 +1,341 @@
#include "clip.h"
#include "clip-impl.h"
#include "mtmd.h"
#include "llama.h"
#include <algorithm>
#include <cerrno>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <limits>
#include <vector>
struct mtmd_context {
struct clip_ctx * ctx_clip;
const struct llama_model * text_model;
std::vector<float> image_embd_v; // image embedding vector
bool print_timings;
int n_threads;
std::string image_marker;
// TODO @ngxson : add timings
mtmd_context(const char * mmproj_fname,
const llama_model * text_model,
const mtmd_context_params & ctx_params) : print_timings(ctx_params.print_timings), n_threads(ctx_params.n_threads), image_marker(ctx_params.image_marker) {
clip_context_params ctx_clip_params;
ctx_clip_params.use_gpu = ctx_params.use_gpu;
ctx_clip_params.verbosity = ctx_params.verbosity;
ctx_clip = clip_init(mmproj_fname, ctx_clip_params);
if (!ctx_clip) {
throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
}
this->text_model = text_model;
}
~mtmd_context() {
clip_free(ctx_clip);
}
};
struct mtmd_image_tokens_data {
clip_image_f32_batch batch_f32; // preprocessed image patches
};
struct mtmd_image_tokens {
uint32_t nx; // number of tokens in x direction
uint32_t ny; // number of tokens in y direction
uint32_t n_tokens() const { return nx * ny; }
clip_image_f32_batch batch_f32; // preprocessed image patches
};
mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
const struct llama_model * text_model,
const struct mtmd_context_params ctx_params) {
try {
return new mtmd_context(mmproj_fname, text_model, ctx_params);
} catch (const std::exception & e) {
LOG_ERR("%s: error: %s\n", __func__, e.what());
return nullptr;
}
}
void mtmd_free(mtmd_context * ctx) {
if (ctx) {
delete ctx;
}
}
// copied from common_tokenize
static std::vector<llama_token> mtmd_tokenize_text_internal(
const struct llama_vocab * vocab,
const std::string & text,
bool add_special,
bool parse_special) {
// upper limit for the number of tokens
int n_tokens = text.length() + 2 * add_special;
std::vector<llama_token> result(n_tokens);
n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
if (n_tokens < 0) {
result.resize(-n_tokens);
int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
GGML_ASSERT(check == -n_tokens);
} else {
result.resize(n_tokens);
}
return result;
}
mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
const mtmd_input_text & text,
const std::vector<mtmd_bitmap> & bitmaps) {
mtmd_input_chunks * output = new mtmd_input_chunks;
auto vocab = llama_model_get_vocab(ctx->text_model);
std::string prompt_modified(text.text);
std::string marker_modified(ctx->image_marker);
projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
// a bit hacky here, but works for now
// for some models, we need to add prefix and suffix to the image embeddings
if (proj_type == PROJECTOR_TYPE_GEMMA3) {
// <start_of_image> ... (image embeddings) ... <end_of_image>
marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
}
std::vector<std::string> parts = string_split_str(text.text, ctx->image_marker);
output->clear();
output->reserve(parts.size());
size_t i_img = 0;
for (const auto & part : parts) {
//printf("tokenizing part: %s\n", part.c_str());
bool add_bos = &parts.front() == &part;
auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
if (tokens.empty()) {
continue;
}
mtmd_input_chunk chunk{
MTMD_INPUT_CHUNK_TYPE_TEXT,
std::move(tokens),
{},
};
output->emplace_back(std::move(chunk));
if (&parts.back() != &part) {
// add image token to middle of 2 parts
if (i_img >= bitmaps.size()) {
LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size());
return nullptr;
}
// shim layer
clip_image_u8_ptr img_u8(clip_image_u8_init());
img_u8->nx = bitmaps[i_img].nx;
img_u8->ny = bitmaps[i_img].ny;
img_u8->buf.resize(bitmaps[i_img].data.size());
std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
// preprocess image
clip_image_f32_batch batch_f32;
bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), &batch_f32);
if (!ok) {
LOG_ERR("Unable to preprocess image\n");
return nullptr;
}
mtmd_image_tokens * image_tokens = new mtmd_image_tokens;
image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image
image_tokens->ny = 1; // TODO
image_tokens->batch_f32 = std::move(batch_f32);
mtmd_input_chunk chunk{
MTMD_INPUT_CHUNK_TYPE_IMAGE,
{},
image_tokens,
};
output->emplace_back(std::move(chunk));
i_img++;
}
}
return output;
}
void mtmd_input_chunks_free(mtmd_input_chunks * chunks) {
for (auto & chunk : *chunks) {
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE && chunk.tokens_image) {
delete chunk.tokens_image;
}
}
delete chunks;
}
int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
bool ok = clip_image_batch_encode(
ctx->ctx_clip,
ctx->n_threads,
&image_tokens->batch_f32,
ctx->image_embd_v.data());
return ok ? 0 : 1;
}
float * mtmd_get_output_embd(mtmd_context * ctx) {
return ctx->image_embd_v.data();
}
size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks) {
size_t n_tokens = 0;
for (auto & chunk : *chunks) {
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
n_tokens += chunk.tokens_text.size();
} else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
n_tokens += chunk.tokens_image->n_tokens();
} else {
GGML_ASSERT(false && "chunk type not supported");
}
}
return n_tokens;
}
// helper struct to make working with embd batch easier
// note: this will be removed after llama_batch_ext refactoring
struct decode_embd_batch {
std::vector<llama_pos> pos;
std::vector<int32_t> n_seq_id;
std::vector<llama_seq_id> seq_id_0;
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
logits .resize(n_tokens);
seq_id_0.resize(1);
seq_id_0[0] = seq_id;
seq_ids [n_tokens] = nullptr;
batch = {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
/*logits =*/ logits.data(),
};
for (int i = 0; i < n_tokens; i++) {
batch.pos [i] = pos_0 + i;
batch.n_seq_id[i] = 1;
batch.seq_id [i] = seq_id_0.data();
batch.logits [i] = false;
}
}
};
int32_t mtmd_helper_eval(mtmd_context * ctx,
llama_context * lctx,
mtmd_input_chunks * chunks,
llama_pos pos0,
llama_seq_id seq_id,
int32_t n_batch) {
int32_t ret;
llama_pos n_past = pos0;
llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
for (auto & chunk : *chunks) {
bool is_last = &chunk == &chunks->back();
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
// TODO @ngxson : may need to split into smaller batches
text_batch.n_tokens = chunk.tokens_text.size();
for (size_t i = 0; i < chunk.tokens_text.size(); i++) {
text_batch.token [i] = chunk.tokens_text[i];
text_batch.pos [i] = n_past++;
text_batch.n_seq_id[i] = 1;
text_batch.seq_id [i][0] = seq_id;
text_batch.logits [i] = false;
}
if (is_last) {
// always get logits for last input chunk
text_batch.logits[text_batch.n_tokens - 1] = true;
}
ret = llama_decode(lctx, text_batch);
if (ret != 0) {
LOG_ERR("failed to decode text\n");
llama_batch_free(text_batch);
return ret;
}
} else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
GGML_ASSERT(chunk.tokens_image != nullptr);
int64_t t0 = ggml_time_ms();
if (ctx->print_timings) {
LOG_INF("encoding image...\n");
}
ret = mtmd_encode(ctx, chunk.tokens_image);
if (ret != 0) {
LOG_ERR("failed to encode image\n");
llama_batch_free(text_batch);
return ret;
}
if (ctx->print_timings) {
LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
}
int32_t n_tokens = chunk.tokens_image->n_tokens();
float * embd = mtmd_get_output_embd(ctx);
decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
int64_t t1 = ggml_time_ms();
ret = llama_decode(lctx, batch_img.batch);
if (ret != 0) {
LOG_ERR("failed to decode image\n");
llama_batch_free(text_batch);
return ret;
}
if (ctx->print_timings) {
LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
}
n_past += n_tokens;
} else {
GGML_ASSERT(false && "chunk type not supported");
}
}
llama_batch_free(text_batch);
return 0;
}
int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) {
clip_image_u8_ptr img_u8(clip_image_u8_init());
bool ok = clip_image_load_from_bytes(buf, len, img_u8.get());
if (!ok) {
LOG_ERR("Unable to load image from buffer\n");
return 1;
}
unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
output.data.resize(output.nx * output.ny * 3);
std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
return 0;
}
int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) {
clip_image_u8_ptr img_u8(clip_image_u8_init());
bool ok = clip_image_load_from_file(fname, img_u8.get());
if (!ok) {
LOG_ERR("Unable to load image %s\n", fname);
return 1;
}
unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny);
output.data.resize(output.nx * output.ny * 3);
std::memcpy(output.data.data(), data, output.nx * output.ny * 3);
return 0;
}

146
examples/llava/mtmd.h Normal file
View File

@ -0,0 +1,146 @@
#ifndef MTMD_H
#define MTMD_H
#include "ggml.h"
#include "llama.h"
#include "clip.h"
#include <vector>
#include <cinttypes>
#include <memory>
#ifdef LLAMA_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef LLAMA_BUILD
# define MTMD_API __declspec(dllexport)
# else
# define MTMD_API __declspec(dllimport)
# endif
# else
# define MTMD_API __attribute__ ((visibility ("default")))
# endif
#else
# define MTMD_API
#endif
#ifdef __cplusplus
enum mtmd_input_chunk_type {
MTMD_INPUT_CHUNK_TYPE_TEXT,
MTMD_INPUT_CHUNK_TYPE_IMAGE,
};
struct mtmd_context;
struct mtmd_image_tokens;
// represents raw image data, layout is RGBRGBRGB...
// length of data must be nx * ny * 3
struct mtmd_bitmap {
uint32_t nx;
uint32_t ny;
std::vector<unsigned char> data;
};
struct mtmd_input_chunk {
mtmd_input_chunk_type type;
std::vector<llama_token> tokens_text;
mtmd_image_tokens * tokens_image = nullptr;
};
using mtmd_input_chunks = std::vector<mtmd_input_chunk>;
struct mtmd_context_params {
bool use_gpu = true;
bool print_timings = true;
int n_threads = 4;
enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO;
const char * image_marker = "<__image__>";
};
struct mtmd_input_text {
std::string text;
bool add_special;
bool parse_special;
};
// initialize the mtmd context
// return nullptr on failure
MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
const llama_model * text_model,
const mtmd_context_params ctx_params);
MTMD_API void mtmd_free(mtmd_context * ctx);
// tokenize an input text prompt and an image
// the prompt must have the input image marker (default: "<__image__>") in it
// the marker will be replaced with the image tokens
// for example:
// "here is an image: <__image__>\ndescribe it in detail."
// this will gives 3 chunks:
// 1. "here is an image: <start_of_image>"
// 2. (image tokens)
// 3. "<end_of_image>\ndescribe it in detail."
// number of bitmaps must be equal to the number of image markers in the prompt
// this function is thread-safe (shared ctx)
MTMD_API mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx,
const mtmd_input_text & text,
const std::vector<mtmd_bitmap> & bitmaps);
// free image chunk data
MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks);
// returns 0 on success
MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
const mtmd_image_tokens * image_tokens);
// get output embeddings from the last encode pass
MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
//
// helper functions (can be implemented based on other functions)
//
// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks);
// helper function that automatically:
// 1. run llama_decode() on text chunks
// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
// otherwise, returns 0 on success
MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx,
llama_context * lctx,
mtmd_input_chunks * chunks,
llama_pos pos0,
llama_seq_id seq_id,
int32_t n_batch);
// helper function to construct a mtmd_bitmap from a file
// returns 0 on success
// this function is thread-safe
MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output);
// helper function to construct a mtmd_bitmap from a buffer
// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
// returns 0 on success
// this function is thread-safe
MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output);
// convenient unique_ptr wrappers
struct mtmd_context_deleter {
void operator()(mtmd_context * val) { mtmd_free(val); }
};
using mtmd_context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
struct mtmd_input_chunks_deleter {
void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
};
using mtmd_input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
#else
static_assert(false && "C header is not yet supported by this library");
#endif
#endif

2
examples/llava/qwen2vl-cli.cpp
View File

@ -330,7 +330,7 @@ static struct llava_context * llava_init_context(common_params * params, llama_m
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings

BIN
examples/llava/test-1.jpeg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 121 KiB

81
examples/llava/tests.sh Executable file
View File

@ -0,0 +1,81 @@
#!/bin/bash
# make sure we are in the right directory
SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
cd $SCRIPT_DIR
#export LLAMA_CACHE="$SCRIPT_DIR/tmp"
set -eux
mkdir -p $SCRIPT_DIR/output
PROJ_ROOT="$SCRIPT_DIR/../.."
cd $PROJ_ROOT
###############
arr_bin=()
arr_hf=()
add_test() {
local bin=$1
local hf=$2
arr_bin+=("$bin")
arr_hf+=("$hf")
}
add_test "llama-gemma3-cli" "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
add_test "llama-llava-cli" "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
add_test "llama-llava-cli" "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"
add_test "llama-llava-cli" "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
add_test "llama-llava-cli" "second-state/Llava-v1.5-7B-GGUF:Q2_K"
add_test "llama-llava-cli" "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"
add_test "llama-llava-cli" "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
add_test "llama-qwen2vl-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
###############
cmake --build build -j --target "${arr_bin[@]}"
arr_res=()
for i in "${!arr_bin[@]}"; do
bin="${arr_bin[$i]}"
hf="${arr_hf[$i]}"
echo "Running test with binary: $bin and HF model: $hf"
echo ""
echo ""
output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty)
echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
if echo "$output" | grep -iq "new york"; then
result="\033[32mOK\033[0m: $bin $hf"
else
result="\033[31mFAIL\033[0m: $bin $hf"
fi
echo -e "$result"
arr_res+=("$result")
echo ""
echo ""
echo ""
echo "#################################################"
echo "#################################################"
echo ""
echo ""
done
set +x
for i in "${!arr_res[@]}"; do
echo -e "${arr_res[$i]}"
done
echo ""
echo "Output logs are saved in $SCRIPT_DIR/output"

20
examples/perplexity/perplexity.cpp
View File

@ -851,7 +851,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
LOG_INF("%s : calculating hellaswag score over selected tasks.\n", __func__);
LOG("\ntask\tacc_norm\n");
LOG("\ntask\tacc_norm\t95%% confidence interval\n");
double acc = 0.0f;
@ -985,8 +985,22 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
acc += 1.0;
}
// Print the accumulated accuracy mean x 100
LOG("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0);
double freq = acc / double(i + 1);
const double za = 1.95996398454;
// // Wald normal approx
// double conf =za*sqrt(freq*(1-freq)/double(i + 1));
// LOG("%zu\t%.8lf +/- %.8lf\n", i + 1, freq*100.0, conf*100.0);
// Wilson score interval, more accurate
double z = za * za / double(i + 1);
double cnf = z * sqrt(double(i + 1) * (4.0 * freq * (1 - freq) + z)) / (za + za);
double a = (freq + z * 0.5 - cnf) / (1.0 + z);
double b = (freq + z * 0.5 + cnf) / (1.0 + z);
// Print the accumulated accuracy mean x 100 and confidence interval
LOG("%zu\t%3.8lf%%\t[%3.4lf%%, %3.4lf%%]\n", i + 1, freq * 100.0, a * 100.0, b * 100.0);
}
i0 = i1 - 1;

117
examples/quantize/quantize.cpp
View File

@ -9,6 +9,7 @@
#include <fstream>
#include <cmath>
#include <cctype>
#include <algorithm>
struct quant_option {
std::string name;
@ -16,7 +17,7 @@ struct quant_option {
std::string desc;
};
static const std::vector<struct quant_option> QUANT_OPTIONS = {
static const std::vector<quant_option> QUANT_OPTIONS = {
{ "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", },
{ "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", },
@ -105,7 +106,8 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
//
[[noreturn]]
static void usage(const char * executable) {
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable);
printf(" [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
printf(" --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
@ -114,6 +116,8 @@ static void usage(const char * executable) {
printf(" --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
printf(" --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
printf(" --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
printf(" --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
printf(" Advanced option to selectively quantize tensors. May be specified multiple times.\n");
printf(" --keep-split: will generate quantized model in the same shards as input\n");
printf(" --override-kv KEY=TYPE:VALUE\n");
printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
@ -244,6 +248,107 @@ static ggml_type parse_ggml_type(const char * arg) {
return GGML_TYPE_COUNT;
}
// Allowed tensors for arbitrary quantization with --tensor-type option
static const std::vector<std::string> ALLOWED_TENSOR_TYPE = {
"attn_k",
"attn_kv_a_mqa",
"attn_kv_b",
"attn_o",
"attn_output",
"attn_q",
"attn_q_a",
"attn_q_b",
"attn_qkv",
"attn_v",
"channel_mix_key",
"channel_mix_receptance",
"channel_mix_value",
"cls",
"cls.output",
"cross_attn_k",
"cross_attn_o",
"cross_attn_q",
"cross_attn_v",
"ffn_act",
"ffn_down",
"ffn_down_exps",
"ffn_down_shexp",
"ffn_gate",
"ffn_gate_exps",
"ffn_gate_shexp",
"ffn_up",
"ffn_up_exps",
"ffn_up_shexp",
"ssm_in",
"ssm_out",
"time_mix_gate",
"time_mix_key",
"time_mix_output",
"time_mix_receptance",
"time_mix_value",
};
// changes to this struct must be replicated in llama-quant.cpp
struct tensor_quantization {
std::string name;
ggml_type quant = GGML_TYPE_COUNT;
};
static bool parse_tensor_type(const char * data, std::vector<tensor_quantization> & tensor_type) {
const char * sep = strchr(data, '=');
if (sep == nullptr) {
printf("\n%s: malformed tensor type '%s'\n\n", __func__, data);
return false;
}
const size_t tn_len = sep - data;
if (tn_len == 0) {
printf("\n%s: missing tensor name\n\n", __func__);
return false;
}
if (const size_t qt_len = strlen(sep); qt_len == 1) {
printf("\n%s: missing quantization type\n\n", __func__);
return false;
}
std::string tn(data, tn_len);
std::transform(tn.begin(), tn.end(), tn.begin(), tolower);
sep++;
const std::string qt(sep);
bool found = false;
for (const auto & allowed : ALLOWED_TENSOR_TYPE) {
std::string tensor;
tensor = tn.rfind('.') != std::string::npos ? tn.substr(tn.rfind('.') + 1) : tn;
// handle special case of cls.output
std::string cls_output = "cls.output";
if (tn.find(cls_output) != std::string::npos) {
tensor = "cls.output";
}
// check if an allowed tensor exists and it's at the end of the kv string
if (tensor == allowed) {
found = true;
break;
}
}
if (!found) {
printf("\n%s: invalid tensor name '%s'\n\n", __func__, tn.c_str());
return false;
}
if (parse_ggml_type(qt.c_str()) == GGML_TYPE_COUNT) {
printf("\n%s: invalid quantization type '%s'\n\n", __func__, qt.c_str());
return false;
}
tensor_quantization tqz;
tqz.name = tn;
tqz.quant = parse_ggml_type(qt.c_str());
tensor_type.emplace_back(std::move(tqz));
return true;
}
int main(int argc, char ** argv) {
if (argc < 3) {
usage(argv[0]);
@ -255,6 +360,7 @@ int main(int argc, char ** argv) {
std::string imatrix_file;
std::vector<std::string> included_weights, excluded_weights;
std::vector<llama_model_kv_override> kv_overrides;
std::vector<tensor_quantization> tensor_types;
for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@ -277,6 +383,10 @@ int main(int argc, char ** argv) {
} else {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--tensor-type") == 0) {
if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
usage(argv[0]);
@ -361,6 +471,9 @@ int main(int argc, char ** argv) {
kv_overrides.back().key[0] = 0;
params.kv_overrides = &kv_overrides;
}
if (!tensor_types.empty()) {
params.tensor_types = &tensor_types;
}
llama_backend_init();

6
examples/rpc/rpc-server.cpp
View File

@ -126,7 +126,7 @@ static std::string fs_get_cache_directory() {
if (getenv("LLAMA_CACHE")) {
cache_directory = std::getenv("LLAMA_CACHE");
} else {
#ifdef __linux__
#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
if (std::getenv("XDG_CACHE_HOME")) {
cache_directory = std::getenv("XDG_CACHE_HOME");
} else {
@ -136,7 +136,9 @@ static std::string fs_get_cache_directory() {
cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
#elif defined(_WIN32)
cache_directory = std::getenv("LOCALAPPDATA");
#endif // __linux__
#else
# error Unknown architecture
#endif
cache_directory = ensure_trailing_slash(cache_directory);
cache_directory += "llama.cpp";
}

13
examples/run/CMakeLists.txt
View File

@ -1,5 +1,16 @@
set(TARGET llama-run)
add_executable(${TARGET} run.cpp linenoise.cpp/linenoise.cpp)
# TODO: avoid copying this code block from common/CMakeLists.txt
set(LLAMA_RUN_EXTRA_LIBS "")
if (LLAMA_CURL)
find_package(CURL REQUIRED)
target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
include_directories(${CURL_INCLUDE_DIRS})
find_library(CURL_LIBRARY curl REQUIRED)
set(LLAMA_RUN_EXTRA_LIBS ${LLAMA_RUN_EXTRA_LIBS} ${CURL_LIBRARY})
endif ()
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_RUN_EXTRA_LIBS})
target_compile_features(${TARGET} PRIVATE cxx_std_17)

6
examples/run/run.cpp
View File

@ -697,8 +697,10 @@ class LlamaData {
std::vector<std::string> headers = { "User-Agent: llama-cpp", "Accept: application/json" };
std::string url;
std::string model_endpoint = get_model_endpoint();
if (pos == std::string::npos) {
auto [model_name, manifest_url] = extract_model_and_tag(model, "https://huggingface.co/v2/");
auto [model_name, manifest_url] = extract_model_and_tag(model, model_endpoint + "v2/");
hfr = model_name;
nlohmann::json manifest;
@ -713,7 +715,7 @@ class LlamaData {
hff = model.substr(pos + 1);
}
url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
url = model_endpoint + hfr + "/resolve/main/" + hff;
return download(url, bn, true, headers);
}

BIN
examples/server/public/index.html.gz
View File

Binary file not shown.

39
examples/server/server.cpp
View File

@ -1705,6 +1705,8 @@ private:
};
struct server_response {
bool running = true;
// for keeping track of all tasks waiting for the result
std::unordered_set<int> waiting_task_ids;
@ -1759,6 +1761,10 @@ struct server_response {
while (true) {
std::unique_lock<std::mutex> lock(mutex_results);
condition_results.wait(lock, [&]{
if (!running) {
SRV_DBG("%s : queue result stop\n", __func__);
std::terminate(); // we cannot return here since the caller is HTTP code
}
return !queue_results.empty();
});
@ -1789,6 +1795,10 @@ struct server_response {
}
std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
if (!running) {
SRV_DBG("%s : queue result stop\n", __func__);
std::terminate(); // we cannot return here since the caller is HTTP code
}
if (cr_res == std::cv_status::timeout) {
return nullptr;
}
@ -1818,6 +1828,12 @@ struct server_response {
}
}
}
// terminate the waiting loop
void terminate() {
running = false;
condition_results.notify_all();
}
};
struct server_context {
@ -3891,6 +3907,21 @@ int main(int argc, char ** argv) {
res_ok(res, {{ "success", true }});
};
const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
json data = {
{
"template", common_chat_templates_source(ctx_server.chat_templates.get()),
},
{
"model_info", {
{ "llama.context_length", ctx_server.slots.back().n_ctx, },
}
},
};
res_ok(res, data);
};
// handle completion-like requests (completion, chat, infill)
// we can optionally provide a custom format for partial results and final results
const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
@ -4455,6 +4486,7 @@ int main(int argc, char ** argv) {
svr->Get ("/metrics", handle_metrics);
svr->Get ("/props", handle_props);
svr->Post("/props", handle_props_change);
svr->Post("/api/show", handle_api_show);
svr->Get ("/models", handle_models); // public endpoint (no API key check)
svr->Get ("/v1/models", handle_models); // public endpoint (no API key check)
svr->Post("/completion", handle_completions); // legacy
@ -4491,9 +4523,10 @@ int main(int argc, char ** argv) {
svr->new_task_queue = [&params] { return new httplib::ThreadPool(params.n_threads_http); };
// clean up function, to be called before exit
auto clean_up = [&svr]() {
auto clean_up = [&svr, &ctx_server]() {
SRV_INF("%s: cleaning up before exit...\n", __func__);
svr->stop();
ctx_server.queue_results.terminate();
llama_backend_free();
};
@ -4534,7 +4567,7 @@ int main(int argc, char ** argv) {
if (!ctx_server.load_model(params)) {
clean_up();
// t.join(); // FIXME: see below
t.join();
LOG_ERR("%s: exiting due to model loading error\n", __func__);
return 1;
}
@ -4582,7 +4615,7 @@ int main(int argc, char ** argv) {
ctx_server.queue_tasks.start_loop();
clean_up();
// t.join(); // FIXME: http thread may stuck if there is an on-going request. we don't need to care about this for now as the HTTP connection will already be closed at this point, but it's better to fix this
t.join();
return 0;
}

2
examples/server/tests/README.md
View File

@ -17,7 +17,7 @@ To mitigate it, you can increase values in `n_predict`, `kv_size`.
```shell
cd ../../..
cmake -B build -DLLAMA_CURL=ON
cmake -B build
cmake --build build --target llama-server
```

20
examples/server/tests/unit/test_embedding.py
View File

@ -49,6 +49,26 @@ def test_embedding_multiple():
assert len(d['embedding']) > 1
def test_embedding_multiple_with_fa():
server = ServerPreset.bert_bge_small_with_fa()
server.pooling = 'last'
server.start()
# one of these should trigger the FA branch (i.e. context size % 256 == 0)
res = server.make_request("POST", "/v1/embeddings", data={
"input": [
"a "*253,
"b "*254,
"c "*255,
"d "*256,
],
})
assert res.status_code == 200
assert len(res.body['data']) == 4
for d in res.body['data']:
assert 'embedding' in d
assert len(d['embedding']) > 1
@pytest.mark.parametrize(
"input,is_multi_prompt",
[

15
examples/server/tests/utils.py
View File

@ -323,6 +323,21 @@ class ServerPreset:
server.server_embeddings = True
return server
@staticmethod
def bert_bge_small_with_fa() -> ServerProcess:
server = ServerProcess()
server.model_hf_repo = "ggml-org/models"
server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf"
server.model_alias = "bert-bge-small"
server.n_ctx = 1024
server.n_batch = 300
server.n_ubatch = 300
server.n_slots = 2
server.fa = True
server.seed = 42
server.server_embeddings = True
return server
@staticmethod
def tinyllama_infill() -> ServerProcess:
server = ServerProcess()

2
examples/server/utils.hpp
View File

@ -3,7 +3,7 @@
#include "common.h"
#include "log.h"
#include "llama.h"
#include "common/base64.hpp"
#include "base64.hpp"
// increase max payload length to allow use of larger context size
#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576

1436
examples/server/webui/package-lock.json generated
View File

File diff suppressed because it is too large Load Diff

6
examples/server/webui/package.json
View File

@ -13,9 +13,11 @@
"dependencies": {
"@heroicons/react": "^2.2.0",
"@sec-ant/readable-stream": "^0.6.0",
"@tailwindcss/postcss": "^4.1.1",
"@tailwindcss/vite": "^4.1.1",
"@vscode/markdown-it-katex": "^1.1.1",
"autoprefixer": "^10.4.20",
"daisyui": "^4.12.14",
"daisyui": "^5.0.12",
"dexie": "^4.0.11",
"highlight.js": "^11.10.0",
"katex": "^0.16.15",
@ -29,7 +31,7 @@
"remark-breaks": "^4.0.0",
"remark-gfm": "^4.0.0",
"remark-math": "^6.0.0",
"tailwindcss": "^3.4.15",
"tailwindcss": "^4.1.1",
"textlinestream": "^1.1.1",
"vite-plugin-singlefile": "^2.0.3"
},

3
examples/server/webui/postcss.config.js
View File

@ -1,6 +1,5 @@
export default {
plugins: {
tailwindcss: {},
autoprefixer: {},
"@tailwindcss/postcss": {},
},
}

2
examples/server/webui/src/App.tsx
View File

@ -28,7 +28,7 @@ function AppLayout() {
<>
<Sidebar />
<div
className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto"
className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto bg-base-100"
id="main-scroll"
>
<Header />

2
examples/server/webui/src/Config.ts
View File

@ -1,4 +1,4 @@
import daisyuiThemes from 'daisyui/src/theming/themes';
import daisyuiThemes from 'daisyui/theme/object';
import { isNumeric } from './utils/misc';
export const isDev = import.meta.env.MODE === 'development';

58
examples/server/webui/src/components/ChatScreen.tsx
View File

@ -1,4 +1,4 @@
import { useEffect, useMemo, useRef, useState } from 'react';
import { useEffect, useMemo, useState } from 'react';
import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context';
import ChatMessage from './ChatMessage';
import { CanvasType, Message, PendingMessage } from '../utils/types';
@ -6,6 +6,7 @@ import { classNames, cleanCurrentUrl, throttle } from '../utils/misc';
import CanvasPyInterpreter from './CanvasPyInterpreter';
import StorageUtils from '../utils/storage';
import { useVSCodeContext } from '../utils/llama-vscode';
import { useChatTextarea, ChatTextareaApi } from './useChatTextarea.ts';
/**
* A message display is a message node with additional information for rendering.
@ -99,7 +100,8 @@ export default function ChatScreen() {
canvasData,
replaceMessageAndGenerate,
} = useAppContext();
const textarea = useOptimizedTextarea(prefilledMsg.content());
const textarea: ChatTextareaApi = useChatTextarea(prefilledMsg.content());
const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
// TODO: improve this when we have "upload file" feature
@ -248,14 +250,16 @@ export default function ChatScreen() {
</div>
{/* chat input */}
<div className="flex flex-row items-center pt-8 pb-6 sticky bottom-0 bg-base-100">
<div className="flex flex-row items-end pt-8 pb-6 sticky bottom-0 bg-base-100">
<textarea
className="textarea textarea-bordered w-full"
// Default (mobile): Enable vertical resize, overflow auto for scrolling if needed
// Large screens (lg:): Disable manual resize, apply max-height for autosize limit
className="textarea textarea-bordered w-full resize-vertical lg:resize-none lg:max-h-48 lg:overflow-y-auto" // Adjust lg:max-h-48 as needed (e.g., lg:max-h-60)
placeholder="Type a message (Shift+Enter to add a new line)"
ref={textarea.ref}
onInput={textarea.onInput} // Hook's input handler (will only resize height on lg+ screens)
onKeyDown={(e) => {
if (e.nativeEvent.isComposing || e.keyCode === 229) return;
if (e.key === 'Enter' && e.shiftKey) return;
if (e.key === 'Enter' && !e.shiftKey) {
e.preventDefault();
sendNewMessage();
@ -263,7 +267,11 @@ export default function ChatScreen() {
}}
id="msg-input"
dir="auto"
// Set a base height of 2 rows for mobile views
// On lg+ screens, the hook will calculate and set the initial height anyway
rows={2}
></textarea>
{isGenerating(currConvId ?? '') ? (
<button
className="btn btn-neutral ml-2"
@ -286,43 +294,3 @@ export default function ChatScreen() {
</div>
);
}
export interface OptimizedTextareaValue {
value: () => string;
setValue: (value: string) => void;
focus: () => void;
ref: React.RefObject<HTMLTextAreaElement>;
}
// This is a workaround to prevent the textarea from re-rendering when the inner content changes
// See https://github.com/ggml-org/llama.cpp/pull/12299
function useOptimizedTextarea(initValue: string): OptimizedTextareaValue {
const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
const textareaRef = useRef<HTMLTextAreaElement>(null);
useEffect(() => {
if (textareaRef.current && savedInitValue) {
textareaRef.current.value = savedInitValue;
setSavedInitValue('');
}
}, [textareaRef, savedInitValue, setSavedInitValue]);
return {
value: () => {
return textareaRef.current?.value ?? savedInitValue;
},
setValue: (value: string) => {
if (textareaRef.current) {
textareaRef.current.value = value;
}
},
focus: () => {
if (textareaRef.current) {
// focus and move the cursor to the end
textareaRef.current.focus();
textareaRef.current.selectionStart = textareaRef.current.value.length;
}
},
ref: textareaRef,
};
}

3
examples/server/webui/src/components/Header.tsx
View File

@ -2,7 +2,7 @@ import { useEffect, useState } from 'react';
import StorageUtils from '../utils/storage';
import { useAppContext } from '../utils/app.context';
import { classNames } from '../utils/misc';
import daisyuiThemes from 'daisyui/src/theming/themes';
import daisyuiThemes from 'daisyui/theme/object';
import { THEMES } from '../Config';
import { useNavigate } from 'react-router';
@ -20,7 +20,6 @@ export default function Header() {
document.body.setAttribute('data-theme', selectedTheme);
document.body.setAttribute(
'data-color-scheme',
// @ts-expect-error daisyuiThemes complains about index type, but it should work
daisyuiThemes[selectedTheme]?.['color-scheme'] ?? 'auto'
);
}, [selectedTheme]);

96
examples/server/webui/src/components/useChatTextarea.ts Normal file
View File

@ -0,0 +1,96 @@
import { useEffect, useRef, useState, useCallback } from 'react';
// Media Query for detecting "large" screens (matching Tailwind's lg: breakpoint)
const LARGE_SCREEN_MQ = '(min-width: 1024px)';
// Calculates and sets the textarea height based on its scrollHeight
const adjustTextareaHeight = (textarea: HTMLTextAreaElement | null) => {
if (!textarea) return;
// Only perform auto-sizing on large screens
if (!window.matchMedia(LARGE_SCREEN_MQ).matches) {
// On small screens, reset inline height and max-height styles.
// This allows CSS (e.g., `rows` attribute or classes) to control the height,
// and enables manual resizing if `resize-vertical` is set.
textarea.style.height = ''; // Use 'auto' or '' to reset
textarea.style.maxHeight = '';
return; // Do not adjust height programmatically on small screens
}
const computedStyle = window.getComputedStyle(textarea);
// Get the max-height specified by CSS (e.g., from `lg:max-h-48`)
const currentMaxHeight = computedStyle.maxHeight;
// Temporarily remove max-height to allow scrollHeight to be calculated correctly
textarea.style.maxHeight = 'none';
// Reset height to 'auto' to measure the actual scrollHeight needed
textarea.style.height = 'auto';
// Set the height to the calculated scrollHeight
textarea.style.height = `${textarea.scrollHeight}px`;
// Re-apply the original max-height from CSS to enforce the limit
textarea.style.maxHeight = currentMaxHeight;
};
// Interface describing the API returned by the hook
export interface ChatTextareaApi {
value: () => string;
setValue: (value: string) => void;
focus: () => void;
ref: React.RefObject<HTMLTextAreaElement>;
onInput: (event: React.FormEvent<HTMLTextAreaElement>) => void; // Input handler
}
// This is a workaround to prevent the textarea from re-rendering when the inner content changes
// See https://github.com/ggml-org/llama.cpp/pull/12299
// combined now with auto-sizing logic.
export function useChatTextarea(initValue: string): ChatTextareaApi {
const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
const textareaRef = useRef<HTMLTextAreaElement>(null);
// Effect to set initial value and height on mount or when initValue changes
useEffect(() => {
const textarea = textareaRef.current;
if (textarea) {
if (typeof savedInitValue === 'string' && savedInitValue.length > 0) {
textarea.value = savedInitValue;
// Call adjustTextareaHeight - it will check screen size internally
setTimeout(() => adjustTextareaHeight(textarea), 0);
setSavedInitValue(''); // Reset after applying
} else {
// Adjust height even if there's no initial value (for initial render)
setTimeout(() => adjustTextareaHeight(textarea), 0);
}
}
}, [textareaRef, savedInitValue]); // Depend on ref and savedInitValue
const handleInput = useCallback(
(event: React.FormEvent<HTMLTextAreaElement>) => {
// Call adjustTextareaHeight on every input - it will decide whether to act
adjustTextareaHeight(event.currentTarget);
},
[]
);
return {
// Method to get the current value directly from the textarea
value: () => {
return textareaRef.current?.value ?? '';
},
// Method to programmatically set the value and trigger height adjustment
setValue: (value: string) => {
const textarea = textareaRef.current;
if (textarea) {
textarea.value = value;
// Call adjustTextareaHeight - it will check screen size internally
setTimeout(() => adjustTextareaHeight(textarea), 0);
}
},
focus: () => {
if (textareaRef.current) {
textareaRef.current.focus();
}
},
ref: textareaRef,
onInput: handleInput,
};
}

11
examples/server/webui/src/index.scss
View File

@ -1,8 +1,13 @@
@use 'sass:meta';
@use 'tailwindcss';
@tailwind base;
@tailwind components;
@tailwind utilities;
@plugin 'daisyui' {
themes: all;
}
html {
scrollbar-gutter: auto;
}
.markdown {
h1,

4
examples/server/webui/src/utils/llama-vscode.ts
View File

@ -1,6 +1,6 @@
import { useEffect, useState } from 'react';
import { MessageExtraContext } from './types';
import { OptimizedTextareaValue } from '../components/ChatScreen';
import { ChatTextareaApi } from '../components/useChatTextarea.ts';
// Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
// Ref: https://github.com/ggml-org/llama.cpp/pull/11940
@ -15,7 +15,7 @@ interface SetTextEvData {
* window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n return 123' }, '*');
*/
export const useVSCodeContext = (textarea: OptimizedTextareaValue) => {
export const useVSCodeContext = (textarea: ChatTextareaApi) => {
const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
null
);

2
examples/server_embd.py
View File

@ -15,7 +15,7 @@ async def main():
model_url = "http://127.0.0.1:6900"
responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
url= f"{model_url}/embedding",
json= {"content": str(0)*1024}
json= {"content": "a "*1022}
) for i in range(n)])
for response in responses:

4
examples/sycl/build.sh
View File

@ -8,10 +8,10 @@ cd build
source /opt/intel/oneapi/setvars.sh
#for FP16
#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON # faster for long-prompt inference
#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DLLAMA_CURL=OFF # faster for long-prompt inference
#for FP32
cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=OFF
#build example/main
#cmake --build . --config Release --target main

4
examples/sycl/win-build-sycl.bat
View File

@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR
:: for FP16
:: faster for long-prompt inference
:: cmake -G "MinGW Makefiles" .. -DGGML_SYCL=ON -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON
:: cmake -G "MinGW Makefiles" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON
:: for FP32
cmake -G "Ninja" .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
cmake -G "Ninja" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
if %errorlevel% neq 0 goto ERROR
:: build example/main only
:: make main

1
ggml/CMakeLists.txt
View File

@ -170,7 +170,6 @@ option(GGML_HIP "ggml: use HIP"
option(GGML_HIP_GRAPHS "ggml: use HIP graph, experimental, slow" OFF)
option(GGML_HIP_NO_VMM "ggml: do not try to use HIP VMM" ON)
option(GGML_HIP_ROCWMMA_FATTN "ggml: enable rocWMMA for FlashAttention" OFF)
option(GGML_HIP_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)
option(GGML_VULKAN_DEBUG "ggml: enable Vulkan debug output" OFF)

129
ggml/include/ggml.h
View File

@ -507,17 +507,12 @@ extern "C" {
GGML_OP_UNARY,
GGML_OP_MAP_UNARY,
GGML_OP_MAP_BINARY,
GGML_OP_MAP_CUSTOM1_F32,
GGML_OP_MAP_CUSTOM2_F32,
GGML_OP_MAP_CUSTOM3_F32,
GGML_OP_MAP_CUSTOM1,
GGML_OP_MAP_CUSTOM2,
GGML_OP_MAP_CUSTOM3,
GGML_OP_CUSTOM,
GGML_OP_CROSS_ENTROPY_LOSS,
GGML_OP_CROSS_ENTROPY_LOSS_BACK,
GGML_OP_OPT_STEP_ADAMW,
@ -1722,24 +1717,29 @@ extern "C" {
float p0,
float p1);
// nearest interpolate
enum ggml_scale_mode {
GGML_SCALE_MODE_NEAREST = 0,
GGML_SCALE_MODE_BILINEAR = 1,
};
// interpolate
// multiplies ne0 and ne1 by scale factor
// used in stable-diffusion
GGML_API struct ggml_tensor * ggml_upscale(
struct ggml_context * ctx,
struct ggml_tensor * a,
int scale_factor);
int scale_factor,
enum ggml_scale_mode mode);
// nearest interpolate
// nearest interpolate to specified dimensions
// used in tortoise.cpp
// interpolate
// interpolate scale to specified dimensions
GGML_API struct ggml_tensor * ggml_upscale_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2,
int ne3);
int ne3,
enum ggml_scale_mode mode);
// pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
GGML_API struct ggml_tensor * ggml_pad(
@ -1916,83 +1916,6 @@ extern "C" {
// custom operators
typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_unary_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_unary_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_binary_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_binary_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_f32_t fun),
"use ggml_map_custom1 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
ggml_custom1_op_f32_t fun),
"use ggml_map_custom1_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_f32_t fun),
"use ggml_map_custom2 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
ggml_custom2_op_f32_t fun),
"use ggml_map_custom2_inplace instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_f32_t fun),
"use ggml_map_custom3 instead");
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
ggml_custom3_op_f32_t fun),
"use ggml_map_custom3_inplace instead");
// custom operators v2
typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
@ -2048,6 +1971,30 @@ extern "C" {
int n_tasks,
void * userdata);
typedef void (*ggml_custom_op_t)(struct ggml_tensor * dst , int ith, int nth, void * userdata);
GGML_API struct ggml_tensor * ggml_custom_4d(
struct ggml_context * ctx,
enum ggml_type type,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
struct ggml_tensor ** args,
int n_args,
ggml_custom_op_t fun,
int n_tasks,
void * userdata);
GGML_API struct ggml_tensor * ggml_custom_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor ** args,
int n_args,
ggml_custom_op_t fun,
int n_tasks,
void * userdata);
// loss function
GGML_API struct ggml_tensor * ggml_cross_entropy_loss(

2
ggml/src/ggml-cann/acl_tensor.cpp
View File

@ -41,6 +41,8 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
return ACL_INT4;
case GGML_TYPE_Q8_0:
return ACL_INT8;
case GGML_TYPE_I64:
return ACL_INT64;
default:
return ACL_DT_UNDEFINED;
}

1601
ggml/src/ggml-cann/aclnn_ops.cpp
View File

File diff suppressed because it is too large Load Diff

496
ggml/src/ggml-cann/aclnn_ops.h
View File

@ -1,15 +1,4 @@
#ifndef CANN_ACLNN_OPS
#define CANN_ACLNN_OPS
/**
* @file acl_tensor
* @brief This file contains related functions of ggml_tensor and acl_tensor.
* Contains conversion from ggml_tensor to acl_tensor, broadcast and other
* functions.
* @author hipudding <huafengchun@gmail.com>
* @author wangshuai09 <391746016@qq.com>
* @date July 15, 2024
*
* Copyright (c) 2023-2024 The ggml authors
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
@ -31,20 +20,30 @@
* IN THE SOFTWARE.
*/
#include <aclnnop/aclnn_add.h>
#ifndef CANN_ACLNN_OPS
#define CANN_ACLNN_OPS
#include <aclnnop/aclnn_abs.h>
#include <aclnnop/aclnn_neg.h>
#include <aclnnop/aclnn_exp.h>
#include <aclnnop/aclnn_arange.h>
#include <aclnnop/aclnn_argsort.h>
#include <aclnnop/aclnn_cat.h>
#include <aclnnop/aclnn_clamp.h>
#include <aclnnop/aclnn_div.h>
#include <aclnnop/aclnn_gelu.h>
#include <aclnnop/aclnn_gelu_v2.h>
#include <aclnnop/aclnn_sigmoid.h>
#include <aclnnop/aclnn_hardsigmoid.h>
#include <aclnnop/aclnn_hardswish.h>
#include <aclnnop/aclnn_leaky_relu.h>
#include <aclnnop/aclnn_mul.h>
#include <aclnnop/aclnn_relu.h>
#include <aclnnop/aclnn_silu.h>
#include <aclnnop/aclnn_tanh.h>
#include <aclnnop/aclnn_sqrt.h>
#include <aclnnop/aclnn_sin.h>
#include <aclnnop/aclnn_cos.h>
#include <aclnnop/aclnn_log.h>
#include <aclnnop/aclnn_sign.h>
#include "acl_tensor.h"
#include "common.h"
@ -63,23 +62,6 @@
*/
void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Adds two ggml tensors using the CANN backend.
*
* @details This function performs an element-wise addition of two tensors. In
* case the tensors do not have the same shape, one or both tensors
* will be broadcasted to match the shape of the other before the
* addition is performed.The formula for the operation is given by:
* \f[
* \text{dst} = \text{acl_src0} + \alpha \cdot \text{acl_src1}
* \f]
*
* @param ctx The CANN context used for operations.
* @param dst The ggml tensor representing the destination, result of the
* addition is stored at dst->data, and dst->op is `GGML_OP_ADD`
*/
void ggml_cann_add(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies the Leaky ReLU activation function to a tensor using the CANN
* backend.
@ -131,19 +113,6 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
*/
void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Computes the square of the elements of a ggml tensor using the CANN
* backend.
* @details The function sets the second source tensor of the destination
* tensor `dst` to be equal to the first source tensor. This is
* effectively squaring the elements since the multiplication becomes
* `element * element`.
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the squared values will be stored，
* which dst->op is `GGML_OP_SQR`.
*/
void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies a clamp operation to the elements of a ggml tensor using the
* CANN backend.
@ -275,6 +244,20 @@ void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
*/
void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Computes the sum of elements in a ggml tensor.
*
* @details This function performs a reduction sum operation along the last
* dimension of the input tensor `src`. The result of the sum is stored
* in the destination tensor `dst`.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the reduced values will be stored。
*
*/
void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Upsamples a ggml tensor using nearest neighbor interpolation using
* the CANN backend.
@ -494,134 +477,377 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
* operation is executed using the CANN backend for optimized performance.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the indices of the maximum values will be stored.
* dst->op is `GGML_OP_ARGMAX`.
* @param dst The destination tensor where the indices of the maximum values will
* be stored. dst->op is `GGML_OP_ARGMAX`.
*/
void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Computes the cosine of each element in a ggml tensor using the CANN backend.
* @brief Adds two tensors element-wise and stores the result in a destination
* tensor.
*
* @details This function applies the cosine function element-wise to the input tensor.
* The computed cosine values are stored in the destination tensor `dst`.
* The operation is optimized using the CANN backend for improved performance.
* This function performs the operation:
* \f[
* dst = acl\_src0 + alpha \times acl\_src1
* \f]
* where alpha is a scalar value and defaults to 1.0f.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the cosine values will be stored.
* dst->op is `GGML_OP_COS`.
* @param ctx The context for the CANN backend operations.
* @param acl_src0 The first source tensor.
* @param acl_src1 The second source tensor.
* @param acl_dst The destination tensor where the result will be stored.
*/
void ggml_cann_cos(ggml_backend_cann_context& ctx, ggml_tensor* dst);
void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
/**
* @brief Computes the sine of each element in a ggml tensor using the CANN backend.
* @brief Sub two tensors element-wise and stores the result in a destination
* tensor.
*
* @details This function applies the sine function element-wise to the input tensor.
* The computed sine values are stored in the destination tensor `dst`.
* The operation is optimized using the CANN backend for improved performance.
* This function performs the operation:
* \f[
* dst = acl\_src0 - alpha \times acl\_src1
* \f]
* where alpha is a scalar value and defaults to 1.0f.
*
* @param ctx The context for the CANN backend operations.
* @param acl_src0 The first source tensor.
* @param acl_src1 The second source tensor.
* @param acl_dst The destination tensor where the result will be stored.
*/
void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
/**
* @brief Performs element-wise multiplication of two tensors and stores the
* result in a destination tensor.
*
* This function performs element-wise multiplication of the tensors `acl_src`
* and `acl_other` and stores the result in the destination tensor `acl_dst`.
* The operation is defined as:
* \f[
* \text {acl_dst }_i=\text {acl_src }_i \times \text {acl_other }_i
* \f]
*
* @param ctx The context for the CANN backend operations.
* @param acl_src The first tensor for element-wise multiplication.
* @param acl_other The second tensor for element-wise multiplication.
* @param acl_dst The destination tensor where the result will be stored.
*/
void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_other, aclTensor* acl_dst = nullptr);
/**
* @brief Matrix division, optionally in-place.
*
* This function division each element of the source tensor `acl_src` by the
* tensor `acl_other` and stores the result in the destination tensor `acl_dst`.
* If `inplace` is true, `acl_dst` will not be used and the operation is
* performed in-place on `acl_src`. The operation is defined as: \f[
* \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i}
* \f]
*
* @param ctx The context for the CANN backend operations.
* @param acl_src Numerator tensor..
* @param acl_other Denominator tensor.
* @param acl_dst The destination tensor where the result will be stored if
* `inplace` is false.
* @param inplace Flag indicating whether to perform the operation in-place on
* `acl_src`.
*/
void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_other, aclTensor* acl_dst = nullptr);
/**
* @brief Applies element-wise cosine function to the elements of a tensor.
*
* This function computes the cosine of each element in the source tensor
* `acl_src` and stores the result in the destination tensor `acl_dst`. The
* operation is defined as: \f[ \text {acl_dst }_i=\cos \left(\text {acl_src
* }_i\right) \f]
*
* @param ctx The context for the CANN backend operations.
* @param acl_src The source tensor on which the cosine function will be
* applied.
* @param acl_dst The destination tensor where the cosine results will be
* stored.
*/
void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_dst);
/**
* @brief Applies element-wise sine function to the elements of a tensor.
*
* This function computes the sine of each element in the source tensor
`acl_src`
* and stores the result in the destination tensor `acl_dst`.
* The operation is defined as:
* \f[
* \text {acl_dst }_i=\sin \left(\text {acl_src }_i\right)
* \f]
* @param ctx The context for the CANN backend operations.
* @param acl_src The source tensor on which the sine function will be applied.
* @param acl_dst The destination tensor where the sine results will be stored.
*/
void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_dst);
/**
* @brief Prepares broadcast-compatible ACL tensors for two input tensors and one
* output tensor.
*
* This function checks whether broadcasting is needed between `src0` and `src1`.
* If broadcasting is required, it calculates the proper shapes and creates
* ACL tensors with broadcast parameters. Otherwise, it directly creates ACL tensors
* based on the original tensor shapes.
*
* @param src0 The first input tensor (reference shape).
* @param src1 The second input tensor (possibly broadcasted).
* @param dst The destination/output tensor.
* @param acl_src0 Output pointer to the created ACL tensor corresponding to src0.
* @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
* @param acl_dst Output pointer to the created ACL tensor corresponding to dst.
*/
void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst);
/**
* @brief Computes the 1D transposed convolution (deconvolution) of a ggml
* tensor using the CANN backend.
*
* @details This function performs a 1D transposed convolution (also known as
* deconvolution) operation on the input tensor. The computed result is stored
* in the destination tensor `dst`. The operation is optimized using the CANN
* backend for improved performance.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the sine values will be stored.
* dst->op is `GGML_OP_SIN`.
* @param dst The destination tensor where the transposed convolution result
* will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`.
*/
void ggml_cann_sin(ggml_backend_cann_context& ctx, ggml_tensor* dst);
void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
aclTensor*, uint64_t*, aclOpExecutor**),
aclnnStatus execute(void*, uint64_t, aclOpExecutor*, aclrtStream)>
void ggml_cann_mul_div(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
/**
* @brief Applies the ELU (Exponential Linear Unit) activation to a ggml tensor
* using the CANN backend.
*
* @details This function performs an element-wise ELU activation on the input
* tensor.
* The result is written to the destination tensor `dst` in-place.
* The ELU function is defined as:
*
* \text{ELU}(x) =
* \begin{cases}
* x, & \text{if } x > 0 \\
* \alpha \left( \exp(x) - 1 \right), & \text{if } x \leq 0
* \end{cases}
*
* where α (alpha) is a hyperparameter, typically set to 1.0.
* This operation is optimized using the CANN backend for high-performance
* inference or training.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the ELU-activated result will be stored.
* dst->op is expected to be `GGML_OP_ELU`.
*/
void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Computes the mean of a ggml tensor element-wise using the CANN backend.
*
* @details This function calculates the element-wise mean of the input tensor.
* The result is written to the destination tensor `dst`.
* The mean is computed by averaging the values across the entire tensor.
*
* This operation is optimized using the CANN backend for high-performance inference or training.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the mean result will be stored.
* dst->op is expected to be `GGML_OP_MEAN`.
*/
void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies 1D reflect padding to a ggml tensor using the CANN backend.
*
* @details This function performs 1D reflect padding on the input tensor.
* The amount of padding on each side is specified by parameters stored in `dst->op_params`.
* The operation reflects the values at the borders of the tensor to generate the padded output.
*
* This operation is optimized using the CANN backend for high-performance inference or training.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the padded result will be stored.
* dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
*/
void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Counts the number of equal elements in two ggml tensors using the CANN backend.
*
* @details This function performs an element-wise comparison between two input tensors,
* and counts the number of positions where the elements are equal. The result is
* stored in the destination tensor `dst` as a scalar.
*
* The operation is optimized using the CANN backend, making it suitable for
* high-performance inference or training scenarios.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
*/
void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies the Step activation function to a ggml tensor using the CANN backend.
*
* @details This function applies a step function element-wise to the input tensor, where
* each element is transformed to 1.0 if it is greater than 0, and 0.0 otherwise.
* The result is stored in the destination tensor `dst`.
*
* This operation is accelerated using the CANN backend to improve runtime performance.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* dst->op is expected to be `GGML_OP_STEP`.
*/
void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies a element-wise operation to two input tensors using the CANN
* backend.
*
* This templated function takes a binary operator and applies it to two source
* tensors
* associated with the destination tensor. The function handles broadcasting as
* needed.
*
* @tparam binary_op A callable object (e.g., lambda or function pointer) representing
* the binary operation to be performed. It must take three arguments:
* (ggml_backend_cann_context&, aclTensor*, aclTensor*, aclTensor*).
*
* @param ctx The CANN backend context used to manage execution and resources.
* @param dst The destination tensor.
*/
template <auto binary_op>
void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src0 = dst->src[0];
ggml_tensor* src1 = dst->src[1];
GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
aclTensor* acl_src0;
aclTensor* acl_src1;
aclTensor* acl_dst;
// Need bcast
if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) {
BCAST_SHAPE(src0, src1)
acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0));
acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1));
acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0));
} else {
acl_src0 = ggml_cann_create_tensor(src0);
acl_src1 = ggml_cann_create_tensor(src1);
acl_dst = ggml_cann_create_tensor(dst);
}
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
ACL_CHECK(getWorkspaceSize(acl_src0, acl_src1, acl_dst, &workspaceSize,
&executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
aclrtStream main_stream = ctx.stream();
ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
binary_op(ctx, acl_src0, acl_src1, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src0));
ACL_CHECK(aclDestroyTensor(acl_src1));
ACL_CHECK(aclDestroyTensor(acl_dst));
}
// Activation functions template.
template <aclnnStatus getWorkspaceSize(const aclTensor*, aclTensor*, uint64_t*,
aclOpExecutor**),
aclnnStatus execute(void*, uint64_t, aclOpExecutor*,
const aclrtStream)>
void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
/**
* @brief Launches an asynchronous task using the memory allocator.
*
* This macro submit an asynchronous task on the specified stream.
* The task uses memory allocated by the allocator. It is guaranteed
* that the memory will not be accessed by other tasks until this task
* completes, due to the sequential execution order within the same stream.
*
* @param OP_NAME aclnn operator name.
* @param args Additional arguments required by the task.
*
* @note
* Memory from the allocator will be "freed" immediately and can be
* reallocated to other pointers. However, it won't be accessed by any
* other task before this asynchronous task ends, because all tasks in the
* same stream are executed in queue order.
*/
#define GGML_CANN_CALL_ACLNN_OP(OP_NAME, ...) \
do { \
uint64_t workspaceSize = 0; \
aclOpExecutor * executor; \
void * workspaceAddr = nullptr; \
\
ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
\
if (workspaceSize > 0) { \
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); \
workspaceAddr = workspace_allocator.get(); \
} \
ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, ctx.stream())); \
} while (0)
/**
* @brief Applies a unary operation to an input tensor using the CANN backend.
*
* This templated function applies a unary operator to the source tensor of `dst`
* and stores the result in the destination tensor.
*
* @tparam unary_op A callable with the signature:
* void(ggml_backend_cann_context&, aclTensor*, aclTensor*)
* where the first aclTensor is the source and the second is the destination.
* @param ctx The CANN backend context for managing resources and execution.
* @param dst The destination tensor. Its src[0] is treated as the input tensor.
*/
template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
void ggml_cann_unary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
aclrtStream main_stream = ctx.stream();
ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
unary_op(ctx, acl_src, acl_dst);
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
}
// Activation functions template for const aclTensors.
template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
uint64_t*, aclOpExecutor**),
aclnnStatus execute(void*, uint64_t, aclOpExecutor*,
const aclrtStream)>
void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
aclrtStream main_stream = ctx.stream();
ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
ACL_CHECK(aclDestroyTensor(acl_src));
ACL_CHECK(aclDestroyTensor(acl_dst));
}
/**
* @brief Applies a unary operation to a ggml tensor using the CANN backend.
*
* @details This function performs a unary operation on the input tensor using
* a user-provided lambda or callable object `unary_op`, which accepts the CANN
* context and two ACL tensors (source and destination). Internally, this function
* creates ACL representations of the ggml tensors and invokes the unary operation.
* The result is stored in the destination tensor `dst`. This utility abstracts the
* common boilerplate of tensor conversion and cleanup when implementing unary ops.
*
* @param unary_op A callable that performs the unary operation using CANN APIs.
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* The source tensor is retrieved from `dst->src[0]`.
*/
void ggml_cann_unary_op(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op.
*
* This macro defines an inline lambda wrapping a specific ACL operation name,
* and passes it to the templated ggml_cann_unary_op function. It simplifies
* calling unary ops by hiding the lambda boilerplate.
*
* Internally, the lambda will call:
* @code
* GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);
* @endcode
*
* @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
*
* @see ggml_cann_unary_op
* @see GGML_CANN_CALL_ACLNN_OP
*/
#define GGML_CANN_CALL_UNARY_OP(OP_NAME) \
do { \
auto lambda = [](ggml_backend_cann_context& ctx, \
aclTensor* acl_src, \
aclTensor* acl_dst) { \
GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst); \
}; \
ggml_cann_unary_op(lambda, ctx, dst); \
} \
while (0)
#endif // CANN_ACLNN_OPS

514
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -29,6 +29,8 @@
#include <cstdio>
#include <cstring>
#include <mutex>
#include <queue>
#include <chrono>
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
@ -119,9 +121,10 @@ static ggml_cann_device_info ggml_cann_init() {
prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
prop.location.id = id;
prop.reserve = 0;
ACL_CHECK(aclrtMemGetAllocationGranularity(
err = aclrtMemGetAllocationGranularity(
&prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
&info.devices[id].vmm_granularity));
&info.devices[id].vmm_granularity);
info.devices[id].vmm = err == ACL_SUCCESS;
size_t free, total;
ggml_backend_cann_get_device_memory(id, &free, &total);
@ -148,11 +151,223 @@ const ggml_cann_device_info& ggml_cann_info() {
//#define DEBUG_CANN_MALLOC
/**
* @brief A pool of CANN buffers(legacy).
* @brief A pool of CANN buffers(priority segment buffer).
*
* This class manages a pool of CANN buffers for a specific device.
*/
struct ggml_cann_pool_leg : public ggml_cann_pool {
struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
/**
* @brief The maximum reuse margin for a buffer.
*/
static const size_t max_reuse_margin = 1ull << 22; // 4MB
/**
* @brief The minimum free margin for a buffer.
*/
static const size_t min_free_margin = 1ull << 20; // 1MB
/**
* @brief The alignment for buffer allocation.
*/
static const size_t alignment = 128;
/**
* @brief The device ID associated with this buffer pool.
*/
int device;
/**
* @brief Whether to disable clean during buffer allocation.
*/
bool disable_clean = false;
/**
* @brief Structure representing a CANN buffer.
*/
struct ggml_cann_buffer {
void* ptr = nullptr; ///< Pointer to the buffer.
size_t size = 0; ///< Size of the buffer.
std::chrono::steady_clock::time_point last_used; ///< Last used time.
bool operator>(const ggml_cann_buffer& other) const {
return size > other.size;
}
};
/**
* @brief Array of CANN buffers in the pool.
*/
std::unordered_map<void*, size_t> buffer_pool;
std::priority_queue<ggml_cann_buffer,
std::vector<ggml_cann_buffer>,
std::greater<>> free_buffers ;
/**
* @brief Total size of all buffers in the pool.
*/
size_t pool_size = 0;
/**
* @brief Constructor to initialize the buffer pool for a specific device.
*
* @param device The device ID to associate with this buffer pool.
*/
explicit ggml_cann_pool_buf_prio(int device) : device(device) {
disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
}
/**
* @brief Destructor to free all buffers in the pool.
*/
~ggml_cann_pool_buf_prio() {
ggml_cann_set_device(device);
for (auto& [b_ptr, b_size] : buffer_pool) {
aclrtFree(b_ptr);
pool_size -= b_size;
}
buffer_pool.clear();
GGML_ASSERT(pool_size == 0);
}
/**
* @brief Allocate a buffer of the given size.
*
* @param size The size of the buffer to allocate.
* @param actual_size A pointer to a variable to receive the actual size of
* the allocated buffer.
* @return A pointer to the allocated buffer.
*/
void* alloc(size_t size, size_t* actual_size) override {
size = GGML_PAD(size, alignment);
if (size == 0) {
size = alignment;
}
void* ptr = nullptr;
auto now = std::chrono::steady_clock::now();
std::vector<ggml_cann_buffer> free_buffers_rest;
free_buffers_rest.reserve(free_buffers.size());
while (!free_buffers.empty()) {
auto b = free_buffers.top();
free_buffers.pop();
if (b.size >= size) {
// reuse the buffer if the size is enough
const size_t margin = b.size - size;
if (margin <= max_reuse_margin) {
*actual_size = b.size;
ptr = b.ptr;
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"cann pool[%d]: reused %p, "
"pool_size = %5u MB, "
"size = %5u MB, "
"margin = %5u MB\n",
device, b.ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
#endif
break;
}
}
bool should_clean = !disable_clean &&
b.size > min_free_margin &&
std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
if (should_clean) {
// free the buffer if the size is needed to be freed
ACL_CHECK(aclrtFree(b.ptr));
pool_size -= b.size;
buffer_pool.erase(b.ptr);
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"cann pool[%d]: clean %p, "
"pool_size = %5u MB, "
"size = %5u MB\n",
device, b.ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
#endif
continue;
}
free_buffers_rest.push_back(b);
}
for (ggml_cann_buffer &b : free_buffers_rest) {
free_buffers.push(std::move(b));
}
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
#endif
if (ptr != nullptr) {
return ptr;
}
// allocate a new buffer if no buffer can be reused
ggml_cann_set_device(device);
ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
*actual_size = size;
pool_size += size;
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"cann pool[%d]: allocate %p, "
"pool_size = %5u MB, "
"size = %5u MB\n",
device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(size, 1048576) / 1048576));
#endif
buffer_pool.emplace(ptr, size);
return ptr;
}
/**
* @brief Free a buffer and return it to the pool.
*
* @param ptr Pointer to the buffer to free.
* @param size Size of the buffer to free.
*/
void free(void* ptr, size_t size) override {
GGML_UNUSED(size);
auto it = buffer_pool.find(ptr);
if (it == buffer_pool.end()) {
GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
}
auto now = std::chrono::steady_clock::now();
free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"cann pool[%d]: return %p, "
"pool_size = %5u MB\n",
device, ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
#endif
}
};
/**
* @brief A pool of CANN buffers(segment buffer).
*
* This class manages a pool of CANN buffers for a specific device.
*/
struct ggml_cann_pool_buf : public ggml_cann_pool {
/**
* @brief The maximum reuse margin for a buffer.
*/
static const size_t max_reuse_margin = 1ull << 22; // 4MB
/**
* @brief The minimum free margin for a buffer.
*/
static const size_t min_free_margin = 1ull << 20; // 1MB
/**
* @brief The alignment for buffer allocation.
*/
static const size_t alignment = 128;
/**
* @brief The maximum number of buffers in the pool.
*/
@ -163,12 +378,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
*/
int device;
/**
* @brief Whether to disable clean during buffer allocation.
*/
bool disable_clean = false;
/**
* @brief Structure representing a CANN buffer.
*/
struct ggml_cann_buffer {
void* ptr = nullptr; ///< Pointer to the buffer memory.
size_t size = 0; ///< Size of the buffer.
bool used = false; ///< Whether the buffer is currently in use.
std::chrono::steady_clock::time_point last_used; ///< Last used time.
};
/**
@ -186,17 +408,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
*
* @param device The device ID to associate with this buffer pool.
*/
explicit ggml_cann_pool_leg(int device) : device(device) {}
explicit ggml_cann_pool_buf(int device) : device(device) {
disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
}
/**
* @brief Destructor to free all buffers in the pool.
*/
~ggml_cann_pool_leg() {
~ggml_cann_pool_buf() {
ggml_cann_set_device(device);
for (int i = 0; i < MAX_BUFFERS; ++i) {
ggml_cann_buffer& b = buffer_pool[i];
if (b.ptr != nullptr) {
ACL_CHECK(aclrtFree(b.ptr));
aclrtFree(b.ptr);
pool_size -= b.size;
}
}
@ -212,63 +436,93 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
* @return A pointer to the allocated buffer.
*/
void* alloc(size_t size, size_t* actual_size) override {
const size_t alignment = 128;
size = GGML_PAD(size, alignment);
if (size == 0) {
size = alignment;
}
#ifdef DEBUG_CANN_MALLOC
int nnz = 0;
size_t max_size = 0;
#endif
size_t best_diff = 1ull << 36;
int ibest = -1;
for (int i = 0; i < MAX_BUFFERS; ++i) {
void* ptr = nullptr;
auto now = std::chrono::steady_clock::now();
int i = 0;
for (; i < MAX_BUFFERS; ++i) {
ggml_cann_buffer& b = buffer_pool[i];
if (b.ptr != nullptr) {
if (b.ptr == nullptr) {
break;
}
if (b.used) {
continue;
}
if (b.size >= size) {
// reuse the buffer if the size is enough
const size_t margin = b.size - size;
if (margin <= max_reuse_margin) {
*actual_size = b.size;
b.used = true;
ptr = b.ptr;
#ifdef DEBUG_CANN_MALLOC
++nnz;
if (b.size > max_size) max_size = b.size;
GGML_LOG_INFO(
"cann pool[%d]: reused %p, "
"pool_size = %5u MB, "
"size = %5u MB, "
"margin = %5u MB\n",
device, b.ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
#endif
if (b.size >= size) {
size_t diff = b.size - size;
if (diff < best_diff) {
best_diff = diff;
ibest = i;
if (!best_diff) {
void* ptr = b.ptr;
*actual_size = b.size;
b.ptr = nullptr;
b.size = 0;
return ptr;
}
}
break;
}
}
bool should_clean = !disable_clean &&
b.size > min_free_margin &&
std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
if (should_clean) {
// free the buffer if the size is needed to be freed
ACL_CHECK(aclrtFree(b.ptr));
pool_size -= b.size;
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"cann pool[%d]: clean %p, "
"pool_size = %5u MB, "
"size = %5u MB\n",
device, b.ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
#endif
b.ptr = nullptr;
}
}
if (ibest >= 0) {
ggml_cann_buffer& b = buffer_pool[ibest];
void* ptr = b.ptr;
*actual_size = b.size;
b.ptr = nullptr;
b.size = 0;
if (ptr != nullptr) {
return ptr;
}
void* ptr;
ggml_cann_set_device(device);
ACL_CHECK(
aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
*actual_size = size;
pool_size += size;
if (i < MAX_BUFFERS) {
// allocate a new buffer if no buffer can be reused
ggml_cann_buffer& b = buffer_pool[i];
ggml_cann_set_device(device);
ACL_CHECK(aclrtMalloc(&b.ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
pool_size += size;
*actual_size = size;
b.size = size;
b.used = true;
if (i >= MAX_BUFFERS - 8) {
GGML_LOG_WARN("cann pool[%d]: slots almost full\n", device);
}
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
"requested %u MB\n",
__func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
(uint32_t)(pool_size / 1024 / 1024),
(uint32_t)(size / 1024 / 1024));
GGML_LOG_INFO(
"cann pool[%d]: allocate %p, "
"pool_size = %5u MB, "
"size = %5u MB\n",
device, b.ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
(uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
#endif
return ptr;
return b.ptr;
}
GGML_ABORT("cann pool[%d]: slots full\n", device);
}
/**
@ -278,18 +532,24 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
* @param size Size of the buffer to free.
*/
void free(void* ptr, size_t size) override {
GGML_UNUSED(size);
for (int i = 0; i < MAX_BUFFERS; ++i) {
ggml_cann_buffer& b = buffer_pool[i];
if (b.ptr == nullptr) {
b.ptr = ptr;
b.size = size;
return;
if (b.ptr != ptr) {
continue;
}
b.used = false;
b.last_used = std::chrono::steady_clock::now();
#ifdef DEBUG_CANN_MALLOC
GGML_LOG_INFO(
"cann pool[%d]: return %p, "
"pool_size = %5u MB\n",
device, b.ptr,
(uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
#endif
return;
}
// memory should always buffered. these memory may still needed by
// tasks in stream.
// TODO, fix me.
GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
GGML_ABORT("cann pool[%d]: slots full\n", device);
}
};
@ -347,8 +607,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
* @param device The device ID to associate with this buffer pool.
*/
explicit ggml_cann_pool_vmm(int device)
: device(device),
granularity(ggml_cann_info().devices[device].vmm_granularity) {
: device(device) {
auto dev = ggml_cann_info().devices[device];
granularity = dev.vmm_granularity;
max_size = dev.total_vram;
@ -471,7 +730,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
*/
std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
int device) {
return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
}
bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
if (enable_buf_prio) {
GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
}
GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
}
// cann buffer
@ -803,7 +1073,7 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
return GGML_STATUS_SUCCESS;
}
// TODO: can backend doesn't support quantized yet. Just leave the code
// TODO: cann backend doesn't support quantized yet. Just leave the code
// here.
if (ggml_is_quantized(tensor->type)) {
// Initialize padding to 0 to avoid possible NaN values
@ -1020,8 +1290,11 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
ggml_cann_set_device(buft_ctx->device);
size = std::max(size, (size_t)1);
const size_t alignment = 128;
size = GGML_PAD(size, alignment);
if (size == 0) {
size = alignment;
}
void* dev_ptr;
aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
if (err != ACL_SUCCESS) {
@ -1300,47 +1573,69 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
ggml_cann_dup(ctx, dst);
break;
case GGML_OP_ADD:
ggml_cann_add(ctx, dst);
case GGML_OP_ADD1:
ggml_cann_binary_op<aclnn_add>(ctx, dst);
break;
case GGML_OP_SUB:
ggml_cann_binary_op<aclnn_sub>(ctx, dst);
break;
case GGML_OP_ACC:
ggml_cann_acc(ctx, dst);
break;
case GGML_OP_MUL:
ggml_cann_mul_div<aclnnMulGetWorkspaceSize, aclnnMul>(ctx, dst);
ggml_cann_binary_op<aclnn_mul>(ctx, dst);
break;
case GGML_OP_DIV:
ggml_cann_mul_div<aclnnDivGetWorkspaceSize, aclnnDiv>(ctx, dst);
ggml_cann_binary_op<aclnn_div>(ctx, dst);
break;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(dst)) {
case GGML_UNARY_OP_ABS:
GGML_CANN_CALL_UNARY_OP(Abs);
break;
case GGML_UNARY_OP_NEG:
GGML_CANN_CALL_UNARY_OP(Neg);
break;
case GGML_UNARY_OP_GELU:
ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
ctx, dst);
GGML_CANN_CALL_UNARY_OP(Gelu);
break;
case GGML_UNARY_OP_SILU:
ggml_cann_activation<aclnnSiluGetWorkspaceSize, aclnnSilu>(
ctx, dst);
break;
// TODO: Use faster gelu??
case GGML_UNARY_OP_GELU_QUICK:
ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
ctx, dst);
GGML_CANN_CALL_UNARY_OP(Silu);
break;
case GGML_UNARY_OP_GELU_QUICK: {
auto lambda = [](ggml_backend_cann_context& ctx,
aclTensor* acl_src,
aclTensor* acl_dst) {
GGML_CANN_CALL_ACLNN_OP(GeluV2, acl_src, 0, acl_dst);
};
ggml_cann_unary_op(lambda, ctx, dst);
} break;
case GGML_UNARY_OP_TANH:
ggml_cann_activation<aclnnTanhGetWorkspaceSize, aclnnTanh>(
ctx, dst);
GGML_CANN_CALL_UNARY_OP(Tanh);
break;
case GGML_UNARY_OP_RELU:
ggml_cann_activation<aclnnReluGetWorkspaceSize, aclnnRelu>(
ctx, dst);
GGML_CANN_CALL_UNARY_OP(Relu);
break;
case GGML_UNARY_OP_SIGMOID:
GGML_CANN_CALL_UNARY_OP(Sigmoid);
break;
case GGML_UNARY_OP_HARDSIGMOID:
ggml_cann_activation<aclnnHardsigmoidGetWorkspaceSize,
aclnnHardsigmoid>(ctx, dst);
GGML_CANN_CALL_UNARY_OP(Hardsigmoid);
break;
case GGML_UNARY_OP_HARDSWISH:
ggml_cann_activation<aclnnHardswishGetWorkspaceSize,
aclnnHardswish>(ctx, dst);
GGML_CANN_CALL_UNARY_OP(Hardswish);
break;
case GGML_UNARY_OP_EXP:
GGML_CANN_CALL_UNARY_OP(Exp);
break;
case GGML_UNARY_OP_ELU:
ggml_cann_elu(ctx, dst);
break;
case GGML_UNARY_OP_SGN:
GGML_CANN_CALL_UNARY_OP(Sign);
break;
case GGML_UNARY_OP_STEP:
ggml_cann_step(ctx, dst);
break;
default:
return false;
@ -1382,7 +1677,12 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
ggml_cann_scale(ctx, dst);
break;
case GGML_OP_SQR:
ggml_cann_sqr(ctx, dst);
GGML_ASSERT(dst->src[1] == nullptr);
dst->src[1] = dst->src[0];
ggml_cann_binary_op<aclnn_mul>(ctx, dst);
break;
case GGML_OP_SQRT:
GGML_CANN_CALL_UNARY_OP(Sqrt);
break;
case GGML_OP_CLAMP:
ggml_cann_clamp(ctx, dst);
@ -1414,6 +1714,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
case GGML_OP_POOL_2D:
ggml_cann_pool2d(ctx, dst);
break;
case GGML_OP_SUM:
ggml_cann_sum(ctx, dst);
break;
case GGML_OP_SUM_ROWS:
ggml_cann_sum_rows(ctx, dst);
break;
@ -1424,10 +1727,25 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
ggml_cann_argmax(ctx, dst);
break;
case GGML_OP_COS:
ggml_cann_cos(ctx, dst);
ggml_cann_unary_op<aclnn_cos>(ctx, dst);
break;
case GGML_OP_SIN:
ggml_cann_sin(ctx, dst);
ggml_cann_unary_op<aclnn_sin>(ctx, dst);
break;
case GGML_OP_CONV_TRANSPOSE_1D:
ggml_cann_conv_transpose_1d(ctx, dst);
break;
case GGML_OP_LOG:
GGML_CANN_CALL_UNARY_OP(Log);
break;
case GGML_OP_MEAN:
ggml_cann_mean(ctx, dst);
break;
case GGML_OP_PAD_REFLECT_1D:
ggml_cann_pad_reflect_1d(ctx, dst);
break;
case GGML_OP_COUNT_EQUAL:
ggml_cann_count_equal(ctx, dst);
break;
default:
return false;
@ -1679,13 +1997,20 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_ABS:
case GGML_UNARY_OP_NEG:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_HARDSIGMOID:
case GGML_UNARY_OP_HARDSWISH:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_TANH:
case GGML_UNARY_OP_EXP:
case GGML_UNARY_OP_ELU:
case GGML_UNARY_OP_SGN:
case GGML_UNARY_OP_STEP:
return true;
default:
return false;
@ -1764,6 +2089,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
return false;
}
if(!ggml_is_contiguous(op->src[0])){
return false;
}
return true;
}
case GGML_OP_UPSCALE: {
@ -1772,6 +2100,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
return false;
}
if (op->op_params[0] != GGML_SCALE_MODE_NEAREST) {
return false;
}
return true;
}
case GGML_OP_POOL_2D: {
@ -1784,6 +2115,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
// value of paddingW should be at most half of kernelW
return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
}
case GGML_OP_SUM:
case GGML_OP_DUP:
case GGML_OP_IM2COL:
case GGML_OP_CONCAT:
@ -1795,11 +2127,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_TRANSPOSE:
case GGML_OP_NORM:
case GGML_OP_ADD:
case GGML_OP_ADD1:
case GGML_OP_SUB:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_RMS_NORM:
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_SQRT:
case GGML_OP_CLAMP:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
@ -1814,6 +2149,11 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_ARGMAX:
case GGML_OP_COS:
case GGML_OP_SIN:
case GGML_OP_CONV_TRANSPOSE_1D:
case GGML_OP_LOG:
case GGML_OP_MEAN:
case GGML_OP_PAD_REFLECT_1D:
case GGML_OP_COUNT_EQUAL:
return true;
default:
return false;

5
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -28,6 +28,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
ggml-cpu/binary-ops.cpp
ggml-cpu/unary-ops.h
ggml-cpu/unary-ops.cpp
ggml-cpu/simd-mappings.h
ggml-cpu/vec.h
ggml-cpu/vec.cpp
ggml-cpu/ops.h
ggml-cpu/ops.cpp
)
target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)

910
ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
View File

File diff suppressed because it is too large Load Diff

23
ggml/src/ggml-cpu/ggml-cpu-impl.h
View File

@ -4,13 +4,13 @@
#include "ggml.h"
#include "ggml-impl.h"
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
//#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h> // fabsf
#ifdef __cplusplus
extern "C" {
#endif
@ -69,33 +69,16 @@ struct ggml_compute_params {
#endif
#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#include <sys/prctl.h>
#endif
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#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>
// ref: https://github.com/ggml-org/llama.cpp/pull/5404
#ifdef _MSC_VER
typedef uint16_t ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
typedef __fp16 ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
@ -340,8 +323,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>

10683
ggml/src/ggml-cpu/ggml-cpu.c
View File

File diff suppressed because it is too large Load Diff

2
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -425,6 +425,8 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
}
case GGML_OP_IM2COL_BACK:
return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
case GGML_OP_GET_ROWS_BACK:
return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
case GGML_OP_OUT_PROD:
return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;

8637
ggml/src/ggml-cpu/ops.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

109
ggml/src/ggml-cpu/ops.h Normal file
View File

@ -0,0 +1,109 @@
#pragma once
#include "ggml.h"
//
// cache line
//
#if defined(__cpp_lib_hardware_interference_size)
#define CACHE_LINE_SIZE std::hardware_destructive_interference_size
#else
#if defined(__POWER9_VECTOR__)
#define CACHE_LINE_SIZE 128
#elif defined(__VXE__) || defined(__VXE2__)
#define CACHE_LINE_SIZE 256
#else
#define CACHE_LINE_SIZE 64
#endif
#endif
static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_repeat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_repeat_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_concat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_silu_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rms_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rms_norm_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_group_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_l2_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_out_prod(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_scale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_soft_max(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_soft_max_ext_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rope(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rope_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(
const struct ggml_compute_params * params,
const struct ggml_tensor * q,
const struct ggml_tensor * k,
const struct ggml_tensor * v,
const struct ggml_tensor * mask,
struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_back(
const struct ggml_compute_params * params,
const bool masked,
struct ggml_tensor * dst);
void ggml_compute_forward_ssm_conv(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif

892
ggml/src/ggml-cpu/simd-mappings.h Normal file
View File

@ -0,0 +1,892 @@
#pragma once
#include "ggml-cpu-impl.h"
//
// simd mappings
//
// we define a common set of C macros which map to specific intrinsics based on the current architecture
// we then implement the fundamental computation operations below using only these macros
// adding support for new architectures requires to define the corresponding SIMD macros
//
// GGML_F32_STEP / GGML_F16_STEP
// number of elements to process in a single step
//
// GGML_F32_EPR / GGML_F16_EPR
// number of elements to fit in a single register
//
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
#define GGML_SIMD
// F32 NEON
#define GGML_F32_STEP 16
#define GGML_F32_EPR 4
#define GGML_F32x4 float32x4_t
#define GGML_F32x4_ZERO vdupq_n_f32(0.0f)
#define GGML_F32x4_SET1(x) vdupq_n_f32(x)
#define GGML_F32x4_LOAD vld1q_f32
#define GGML_F32x4_STORE vst1q_f32
#define GGML_F32x4_FMA(a, b, c) vfmaq_f32(a, b, c)
#define GGML_F32x4_ADD vaddq_f32
#define GGML_F32x4_MUL vmulq_f32
#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
(res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 NEON
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
#define GGML_F16_STEP 32
#define GGML_F16_EPR 8
#define GGML_F16x8 float16x8_t
#define GGML_F16x8_ZERO vdupq_n_f16(0.0f)
#define GGML_F16x8_SET1(x) vdupq_n_f16(x)
#define GGML_F16x8_LOAD(x) vld1q_f16((const __fp16 *)(x))
#define GGML_F16x8_STORE vst1q_f16
#define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
#define GGML_F16x8_ADD vaddq_f16
#define GGML_F16x8_MUL vmulq_f16
#define GGML_F16x8_REDUCE(res, x) \
do { \
int offset = GGML_F16_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
(res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
} while (0)
#define GGML_F16_VEC GGML_F16x8
#define GGML_F16_VEC_ZERO GGML_F16x8_ZERO
#define GGML_F16_VEC_SET1 GGML_F16x8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
#define GGML_F16_VEC_FMA GGML_F16x8_FMA
#define GGML_F16_VEC_ADD GGML_F16x8_ADD
#define GGML_F16_VEC_MUL GGML_F16x8_MUL
#define GGML_F16_VEC_REDUCE GGML_F16x8_REDUCE
#else
// if FP16 vector arithmetic is not supported, we use FP32 instead
// and take advantage of the vcvt_ functions to convert to/from FP16
#define GGML_F16_STEP 16
#define GGML_F16_EPR 4
#define GGML_F32Cx4 float32x4_t
#define GGML_F32Cx4_ZERO vdupq_n_f32(0.0f)
#define GGML_F32Cx4_SET1(x) vdupq_n_f32(x)
#define GGML_F32Cx4_LOAD(x) vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
#define GGML_F32Cx4_STORE(x, y) vst1_f16(x, vcvt_f16_f32(y))
#define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
#define GGML_F32Cx4_ADD vaddq_f32
#define GGML_F32Cx4_MUL vmulq_f32
#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
#define GGML_F16_VEC GGML_F32Cx4
#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
#endif
#elif defined(__AVX512F__)
#define GGML_SIMD
// F32 AVX512
#define GGML_F32_STEP 64
#define GGML_F32_EPR 16
#define GGML_F32x16 __m512
#define GGML_F32x16_ZERO _mm512_setzero_ps()
#define GGML_F32x16_SET1(x) _mm512_set1_ps(x)
#define GGML_F32x16_LOAD _mm512_loadu_ps
#define GGML_F32x16_STORE _mm512_storeu_ps
// _mm512_fmadd_ps is defined in AVX512F so no guard is required
#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
#define GGML_F32x16_ADD _mm512_add_ps
#define GGML_F32x16_MUL _mm512_mul_ps
#define GGML_F32x16_REDUCE(res, x) \
do { \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
res = (ggml_float) _mm512_reduce_add_ps(x[0]); \
} while (0)
// TODO: is this optimal ?
#define GGML_F32_VEC GGML_F32x16
#define GGML_F32_VEC_ZERO GGML_F32x16_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x16_SET1
#define GGML_F32_VEC_LOAD GGML_F32x16_LOAD
#define GGML_F32_VEC_STORE GGML_F32x16_STORE
#define GGML_F32_VEC_FMA GGML_F32x16_FMA
#define GGML_F32_VEC_ADD GGML_F32x16_ADD
#define GGML_F32_VEC_MUL GGML_F32x16_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE
// F16 AVX512
// F16 AVX
#define GGML_F16_STEP 64
#define GGML_F16_EPR 16
// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead
#define GGML_F32Cx16 __m512
#define GGML_F32Cx16_ZERO _mm512_setzero_ps()
#define GGML_F32Cx16_SET1(x) _mm512_set1_ps(x)
// unlike _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F
// so F16C guard isn't required
#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x)))
#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0))
#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
#define GGML_F32Cx16_ADD _mm512_add_ps
#define GGML_F32Cx16_MUL _mm512_mul_ps
#define GGML_F32Cx16_REDUCE(res, x) \
do { \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ps(x[i], x[offset+i]); \
} \
res = (ggml_float) _mm512_reduce_add_ps(x[0]); \
} while (0)
#define GGML_F16_VEC GGML_F32Cx16
#define GGML_F16_VEC_ZERO GGML_F32Cx16_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx16_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx16_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx16_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx16_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx16_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE
#elif defined(__AVX__)
#define GGML_SIMD
// F32 AVX
#define GGML_F32_STEP 32
#define GGML_F32_EPR 8
#define GGML_F32x8 __m256
#define GGML_F32x8_ZERO _mm256_setzero_ps()
#define GGML_F32x8_SET1(x) _mm256_set1_ps(x)
#define GGML_F32x8_LOAD _mm256_loadu_ps
#define GGML_F32x8_STORE _mm256_storeu_ps
#if defined(__FMA__)
#define GGML_F32x8_FMA(a, b, c) _mm256_fmadd_ps(b, c, a)
#else
#define GGML_F32x8_FMA(a, b, c) _mm256_add_ps(_mm256_mul_ps(b, c), a)
#endif
#define GGML_F32x8_ADD _mm256_add_ps
#define GGML_F32x8_MUL _mm256_mul_ps
#define GGML_F32x8_REDUCE(res, x) \
do { \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm256_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm256_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm256_add_ps(x[i], x[offset+i]); \
} \
const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]), \
_mm256_extractf128_ps(x[0], 1)); \
const __m128 t1 = _mm_hadd_ps(t0, t0); \
res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \
} while (0)
// TODO: is this optimal ?
#define GGML_F32_VEC GGML_F32x8
#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x8_SET1
#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD
#define GGML_F32_VEC_STORE GGML_F32x8_STORE
#define GGML_F32_VEC_FMA GGML_F32x8_FMA
#define GGML_F32_VEC_ADD GGML_F32x8_ADD
#define GGML_F32_VEC_MUL GGML_F32x8_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE
// F16 AVX
#define GGML_F16_STEP 32
#define GGML_F16_EPR 8
// F16 arithmetic is not supported by AVX, so we use F32 instead
#define GGML_F32Cx8 __m256
#define GGML_F32Cx8_ZERO _mm256_setzero_ps()
#define GGML_F32Cx8_SET1(x) _mm256_set1_ps(x)
#if defined(__F16C__)
// the _mm256_cvt intrinsics require F16C
#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
#else
static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
float tmp[8];
for (int i = 0; i < 8; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]);
}
return _mm256_loadu_ps(tmp);
}
static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
float arr[8];
_mm256_storeu_ps(arr, y);
for (int i = 0; i < 8; i++)
x[i] = GGML_FP32_TO_FP16(arr[i]);
}
#define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x)
#define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y)
#endif
#define GGML_F32Cx8_FMA GGML_F32x8_FMA
#define GGML_F32Cx8_ADD _mm256_add_ps
#define GGML_F32Cx8_MUL _mm256_mul_ps
#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE
#define GGML_F16_VEC GGML_F32Cx8
#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE
#elif defined(__POWER9_VECTOR__)
#define GGML_SIMD
// F32 POWER9
#define GGML_F32_STEP 32
#define GGML_F32_EPR 4
#define GGML_F32x4 vector float
#define GGML_F32x4_ZERO 0.0f
#define GGML_F32x4_SET1 vec_splats
#define GGML_F32x4_LOAD(p) vec_xl(0, p)
#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p)
#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
#define GGML_F32x4_ADD vec_add
#define GGML_F32x4_MUL vec_mul
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset+i]); \
} \
res = vec_extract(x[0], 0) + \
vec_extract(x[0], 1) + \
vec_extract(x[0], 2) + \
vec_extract(x[0], 3); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 POWER9
#define GGML_F16_STEP GGML_F32_STEP
#define GGML_F16_EPR GGML_F32_EPR
#define GGML_F16_VEC GGML_F32x4
#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32x4_SET1
#define GGML_F16_VEC_FMA GGML_F32x4_FMA
#define GGML_F16_VEC_ADD GGML_F32x4_ADD
#define GGML_F16_VEC_MUL GGML_F32x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
// Use vec_xl, not vec_ld, in case the load address is not aligned.
#define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ? \
vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \
vec_extract_fp32_from_shortl(vec_xl(0, p))
static inline unsigned char ggml_endian_byte(int i) {
uint16_t tmp_val = 1;
return ((unsigned char *)&tmp_val)[i];
}
#define GGML_ENDIAN_BYTE(i) ggml_endian_byte(i)
#define GGML_F16_VEC_STORE(p, r, i) \
if (i & 0x1) \
vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)], \
r[i - GGML_ENDIAN_BYTE(0)]), \
0, p - GGML_F16_EPR)
#elif defined(__wasm_simd128__)
#define GGML_SIMD
// F32 WASM
#define GGML_F32_STEP 16
#define GGML_F32_EPR 4
#define GGML_F32x4 v128_t
#define GGML_F32x4_ZERO wasm_f32x4_splat(0.0f)
#define GGML_F32x4_SET1(x) wasm_f32x4_splat(x)
#define GGML_F32x4_LOAD wasm_v128_load
#define GGML_F32x4_STORE wasm_v128_store
#define GGML_F32x4_FMA(a, b, c) wasm_f32x4_add(wasm_f32x4_mul(b, c), a)
#define GGML_F32x4_ADD wasm_f32x4_add
#define GGML_F32x4_MUL wasm_f32x4_mul
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
} \
res = wasm_f32x4_extract_lane(x[0], 0) + \
wasm_f32x4_extract_lane(x[0], 1) + \
wasm_f32x4_extract_lane(x[0], 2) + \
wasm_f32x4_extract_lane(x[0], 3); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 WASM
#define GGML_F16_STEP 16
#define GGML_F16_EPR 4
inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) {
float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(p[0]);
tmp[1] = GGML_FP16_TO_FP32(p[1]);
tmp[2] = GGML_FP16_TO_FP32(p[2]);
tmp[3] = GGML_FP16_TO_FP32(p[3]);
return wasm_v128_load(tmp);
}
inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
float tmp[4];
wasm_v128_store(tmp, x);
p[0] = GGML_FP32_TO_FP16(tmp[0]);
p[1] = GGML_FP32_TO_FP16(tmp[1]);
p[2] = GGML_FP32_TO_FP16(tmp[2]);
p[3] = GGML_FP32_TO_FP16(tmp[3]);
}
#define GGML_F16x4 v128_t
#define GGML_F16x4_ZERO wasm_f32x4_splat(0.0f)
#define GGML_F16x4_SET1(x) wasm_f32x4_splat(x)
#define GGML_F16x4_LOAD(x) __wasm_f16x4_load(x)
#define GGML_F16x4_STORE(x, y) __wasm_f16x4_store(x, y)
#define GGML_F16x4_FMA GGML_F32x4_FMA
#define GGML_F16x4_ADD wasm_f32x4_add
#define GGML_F16x4_MUL wasm_f32x4_mul
#define GGML_F16x4_REDUCE(res, x) \
{ \
int offset = GGML_F16_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
} \
res = (ggml_float) (wasm_f32x4_extract_lane(x[0], 0) + \
wasm_f32x4_extract_lane(x[0], 1) + \
wasm_f32x4_extract_lane(x[0], 2) + \
wasm_f32x4_extract_lane(x[0], 3)); \
}
#define GGML_F16_VEC GGML_F16x4
#define GGML_F16_VEC_ZERO GGML_F16x4_ZERO
#define GGML_F16_VEC_SET1 GGML_F16x4_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F16x4_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x4_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F16x4_FMA
#define GGML_F16_VEC_ADD GGML_F16x4_ADD
#define GGML_F16_VEC_MUL GGML_F16x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F16x4_REDUCE
#elif defined(__SSE3__)
#define GGML_SIMD
// F32 SSE
#define GGML_F32_STEP 32
#define GGML_F32_EPR 4
#define GGML_F32x4 __m128
#define GGML_F32x4_ZERO _mm_setzero_ps()
#define GGML_F32x4_SET1(x) _mm_set1_ps(x)
#define GGML_F32x4_LOAD _mm_loadu_ps
#define GGML_F32x4_STORE _mm_storeu_ps
#if defined(__FMA__)
// TODO: Does this work?
#define GGML_F32x4_FMA(a, b, c) _mm_fmadd_ps(b, c, a)
#else
#define GGML_F32x4_FMA(a, b, c) _mm_add_ps(_mm_mul_ps(b, c), a)
#endif
#define GGML_F32x4_ADD _mm_add_ps
#define GGML_F32x4_MUL _mm_mul_ps
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm_add_ps(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm_add_ps(x[i], x[offset+i]); \
} \
const __m128 t0 = _mm_hadd_ps(x[0], x[0]); \
res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t0, t0)); \
}
// TODO: is this optimal ?
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 SSE
#define GGML_F16_STEP 32
#define GGML_F16_EPR 4
static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(x[0]);
tmp[1] = GGML_FP16_TO_FP32(x[1]);
tmp[2] = GGML_FP16_TO_FP32(x[2]);
tmp[3] = GGML_FP16_TO_FP32(x[3]);
return _mm_loadu_ps(tmp);
}
static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
float arr[4];
_mm_storeu_ps(arr, y);
x[0] = GGML_FP32_TO_FP16(arr[0]);
x[1] = GGML_FP32_TO_FP16(arr[1]);
x[2] = GGML_FP32_TO_FP16(arr[2]);
x[3] = GGML_FP32_TO_FP16(arr[3]);
}
#define GGML_F32Cx4 __m128
#define GGML_F32Cx4_ZERO _mm_setzero_ps()
#define GGML_F32Cx4_SET1(x) _mm_set1_ps(x)
#define GGML_F32Cx4_LOAD(x) __sse_f16x4_load(x)
#define GGML_F32Cx4_STORE(x, y) __sse_f16x4_store(x, y)
#define GGML_F32Cx4_FMA GGML_F32x4_FMA
#define GGML_F32Cx4_ADD _mm_add_ps
#define GGML_F32Cx4_MUL _mm_mul_ps
#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
#define GGML_F16_VEC GGML_F32Cx4
#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
#elif defined(__loongarch_asx)
#define GGML_SIMD
// F32 LASX
#define GGML_F32_STEP 32
#define GGML_F32_EPR 8
#define GGML_F32x8 __m256
#define GGML_F32x8_ZERO (__m256)__lasx_xvldi(0)
#define GGML_F32x8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
#define GGML_F32x8_LOAD(x) (__m256)__lasx_xvld((x), 0)
#define GGML_F32x8_STORE(x,y) __lasx_xvst((y), (x), 0)
#define GGML_F32x8_FMA(a, b, c) __lasx_xvfmadd_s(b, c, a)
#define GGML_F32x8_ADD __lasx_xvfadd_s
#define GGML_F32x8_MUL __lasx_xvfmul_s
#define GGML_F32x8_REDUCE(res, x) \
do { \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \
} \
float *tmp_p = (float *)&x[0]; \
res = tmp_p[0] + tmp_p[1] + tmp_p[2] + tmp_p[3] + tmp_p[4] + tmp_p[5] + tmp_p[6] + tmp_p[7]; \
} while (0)
// TODO: is this optimal ?
#define GGML_F32_VEC GGML_F32x8
#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x8_SET1
#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD
#define GGML_F32_VEC_STORE GGML_F32x8_STORE
#define GGML_F32_VEC_FMA GGML_F32x8_FMA
#define GGML_F32_VEC_ADD GGML_F32x8_ADD
#define GGML_F32_VEC_MUL GGML_F32x8_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE
// F16 LASX
#define GGML_F16_STEP 32
#define GGML_F16_EPR 8
// F16 arithmetic is not supported by LASX, so we use F32 instead
#define GGML_F32Cx8 __m256
#define GGML_F32Cx8_ZERO (__m256)__lasx_xvldi(0)
#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
__m256i a;
memcpy(&a, x, sizeof(ggml_fp16_t) * 8);
a = __lasx_xvpermi_d(a, 0 | (1 << 4));
return __lasx_xvfcvtl_s_h(a);
}
static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
__m256i a = __lasx_xvfcvt_h_s(y, y);
a = __lasx_xvpermi_d(a, 0 | (2 << 2));
memcpy(x, &a, sizeof(ggml_fp16_t) * 8);
}
#define GGML_F32Cx8_LOAD(x) __lasx_f32cx8_load(x)
#define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y)
#define GGML_F32Cx8_FMA GGML_F32x8_FMA
#define GGML_F32Cx8_ADD __lasx_xvfadd_s
#define GGML_F32Cx8_MUL __lasx_xvfmul_s
#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE
#define GGML_F16_VEC GGML_F32Cx8
#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE
#elif defined(__loongarch_sx)
#define GGML_SIMD
// F32 LSX
#define GGML_F32_STEP 32
#define GGML_F32_EPR 4
#define GGML_F32x4 __m128
#define GGML_F32x4_ZERO __lsx_vldi(0)
#define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
#define GGML_F32x4_LOAD(x) __lsx_vld((x), 0)
#define GGML_F32x4_STORE((x),(y)) __lsx_vst((y), (x), 0)
#define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
#define GGML_F32x4_ADD __lsx_vfadd_s
#define GGML_F32x4_MUL __lsx_vfmul_s
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \
} \
__m128i tmp = __lsx_vsrli_d((__m128i) x[0], 32); \
tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]); \
tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \
tmp = __lsx_vsrli_d((__m128i) t0, 32); \
tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, t0); \
tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 LSX
#define GGML_F16_STEP 32
#define GGML_F16_EPR 4
static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(x[0]);
tmp[1] = GGML_FP16_TO_FP32(x[1]);
tmp[2] = GGML_FP16_TO_FP32(x[2]);
tmp[3] = GGML_FP16_TO_FP32(x[3]);
return __lsx_vld(tmp, 0);
}
static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
float arr[4];
__lsx_vst(y, arr, 0);
x[0] = GGML_FP32_TO_FP16(arr[0]);
x[1] = GGML_FP32_TO_FP16(arr[1]);
x[2] = GGML_FP32_TO_FP16(arr[2]);
x[3] = GGML_FP32_TO_FP16(arr[3]);
}
#define GGML_F32Cx4 __m128
#define GGML_F32Cx4_ZERO __lsx_vldi(0)
#define GGML_F32Cx4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
#define GGML_F32Cx4_LOAD(x) __lsx_f16x4_load(x)
#define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
#define GGML_F32Cx4_FMA GGML_F32x4_FMA
#define GGML_F32Cx4_ADD __lsx_vfadd_s
#define GGML_F32Cx4_MUL __lsx_vfmul_s
#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
#define GGML_F16_VEC GGML_F32Cx4
#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
#elif defined(__VXE__) || defined(__VXE2__)
#define GGML_SIMD
// F32 s390x
#define GGML_F32_STEP 32
#define GGML_F32_EPR 4
#define GGML_F32x4 __vector float
#define GGML_F32x4_ZERO vec_splats(0.0f)
#define GGML_F32x4_SET1 vec_splats
#define GGML_F32x4_LOAD(p) vec_xl(0, p)
#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p)
#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
#define GGML_F32x4_ADD vec_add
#define GGML_F32x4_MUL vec_mul
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset + i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset + i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset + i]); \
} \
res = vec_extract(x[0], 0) + \
vec_extract(x[0], 1) + \
vec_extract(x[0], 2) + \
vec_extract(x[0], 3); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 s390x
#define GGML_F16_STEP GGML_F32_STEP
#define GGML_F16_EPR GGML_F32_EPR
static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
float tmp[4];
for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]);
}
// note: keep type-cast here to prevent compiler bugs
// see: https://github.com/ggml-org/llama.cpp/issues/12846
return vec_xl(0, (const float *)(tmp));
}
static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
float arr[4];
// note: keep type-cast here to prevent compiler bugs
// see: https://github.com/ggml-org/llama.cpp/issues/12846
vec_xst(y, 0, (float *)(arr));
for (int i = 0; i < 4; i++) {
x[i] = GGML_FP32_TO_FP16(arr[i]);
}
}
#define GGML_F16_VEC GGML_F32x4
#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32x4_SET1
#define GGML_F16_VEC_LOAD(p, i) __lzs_f16cx4_load(p)
#define GGML_F16_VEC_STORE(p, r, i) __lzs_f16cx4_store(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32x4_FMA
#define GGML_F16_VEC_ADD GGML_F32x4_ADD
#define GGML_F16_VEC_MUL GGML_F32x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
#endif
// GGML_F32_ARR / GGML_F16_ARR
// number of registers to use per step
#ifdef GGML_SIMD
#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
#endif

258
ggml/src/ggml-cpu/vec.cpp Normal file
View File

@ -0,0 +1,258 @@
#include "vec.h"
#include <cassert>
#if defined(_MSC_VER)
// disable "possible loss of data" to avoid hundreds of casts
// we should just be careful :)
#pragma warning(disable: 4244 4267)
#endif
// precomputed gelu table for f16 (128 KB)
ggml_fp16_t ggml_table_gelu_f16[1 << 16];
// precomputed quick gelu table for f16 (128 KB)
ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
#if defined(GGML_SIMD)
float sumf = 0.0f;
const int np = (n & ~(GGML_F32_STEP - 1));
GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
}
}
// reduce sum0..sum3 to sum0
GGML_F32_VEC_REDUCE(sumf, sum);
// leftovers
for (int i = np; i < n; ++i) {
sumf += x[i]*y[i];
}
#else
// scalar
ggml_float sumf = 0.0;
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(x[i]*y[i]);
}
#endif
*s = sumf;
}
void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
int i = 0;
ggml_float sumf = 0;
#if defined(__AVX512BF16__)
__m512 c1 = _mm512_setzero_ps();
__m512 c2 = _mm512_setzero_ps();
for (; i + 64 <= n; i += 64) {
c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))),
m512bh(_mm512_loadu_si512((y + i))));
c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))),
m512bh(_mm512_loadu_si512((y + i + 32))));
}
sumf += (ggml_float)_mm512_reduce_add_ps(c1);
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#elif defined(__AVX512F__)
#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16))
__m512 c1 = _mm512_setzero_ps();
__m512 c2 = _mm512_setzero_ps();
for (; i + 32 <= n; i += 32) {
c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2);
}
sumf += (ggml_float)_mm512_reduce_add_ps(c1);
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#undef LOAD
#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();
__m256 c4 = _mm256_setzero_ps();
for (; i + 32 <= n; i += 32) {
c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2);
c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3);
c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4);
}
__m128 g;
c1 = _mm256_add_ps(_mm256_add_ps(c1, c3),
_mm256_add_ps(c2, c4));
g = _mm_add_ps(_mm256_extractf128_ps(c1, 1),
_mm256_castps256_ps128(c1));
g = _mm_add_ps(g, _mm_movehl_ps(g, g));
g = _mm_add_ss(g, _mm_movehdup_ps(g));
sumf += (ggml_float)_mm_cvtss_f32(g);
#undef LOAD
#endif
for (; i < n; ++i) {
sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
GGML_BF16_TO_FP32(y[i]));
}
*s = sumf;
}
void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
ggml_float sumf = 0.0;
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
GGML_F16_VEC ax[GGML_F16_ARR];
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
}
}
// reduce sum0..sum3 to sum0
GGML_F16_VEC_REDUCE(sumf, sum);
// leftovers
for (int i = np; i < n; ++i) {
sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
}
#else
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
}
#endif
*s = sumf;
}
void ggml_vec_silu_f32(const int n, float * y, const float * x) {
int i = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
_mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i)));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
_mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i)));
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
_mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i)));
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]);
}
}
ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
int i = 0;
ggml_float sum = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
__m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i),
_mm512_set1_ps(max)));
_mm512_storeu_ps(y + i, val);
sum += (ggml_float)_mm512_reduce_add_ps(val);
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
__m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i),
_mm256_set1_ps(max)));
_mm256_storeu_ps(y + i, val);
__m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
_mm256_castps256_ps128(val));
val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
sum += (ggml_float)_mm_cvtss_f32(val2);
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
__m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i),
_mm_set1_ps(max)));
_mm_storeu_ps(y + i, val);
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
val = _mm_add_ps(val, _mm_movehl_ps(val, val));
val = _mm_add_ss(val, _mm_movehdup_ps(val));
#else
__m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
val = _mm_add_ps(val, tmp);
tmp = _mm_movehl_ps(tmp, val);
val = _mm_add_ss(val, tmp);
#endif
sum += (ggml_float)_mm_cvtss_f32(val);
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i),
vdupq_n_f32(max)));
vst1q_f32(y + i, val);
sum += (ggml_float)vaddvq_f32(val);
}
#endif
for (; i < n; ++i) {
float val = expf(x[i] - max);
sum += (ggml_float)val;
y[i] = val;
}
return sum;
}
ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) {
// log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i)
int i = 0;
ggml_float sum = 0;
for (; i < n; ++i) {
float val = x[i] - max;
y[i] = val;
sum += (ggml_float)expf(val);
}
return sum = (ggml_float)logf(sum);
}

802
ggml/src/ggml-cpu/vec.h Normal file
View File

@ -0,0 +1,802 @@
// Vectorized functions for fundamental operations
#pragma once
#include "ggml-impl.h"
#include "simd-mappings.h"
#include "ggml.h"
#if defined(GGML_USE_ACCELERATE)
#include <Accelerate/Accelerate.h>
#endif
// floating point type used to accumulate sums
typedef double ggml_float;
#define GGML_GELU_FP16
#define GGML_GELU_QUICK_FP16
#define GGML_SOFT_MAX_UNROLL 4
#define GGML_VEC_DOT_UNROLL 2
#define GGML_VEC_MAD_UNROLL 32
#ifdef __cplusplus
extern "C" {
#endif
//
// global data
//
// precomputed gelu table for f16 (128 KB)
extern ggml_fp16_t ggml_table_gelu_f16[1 << 16];
// precomputed quick gelu table for f16 (128 KB)
extern ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
//
// fundamental operations
//
void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc);
void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc);
void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
void ggml_vec_silu_f32(const int n, float * y, const float * x);
ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max);
ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max);
inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const ggml_fp16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; }
inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i]));
}
}
inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; }
inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; }
inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; }
inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; }
inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i]));
}
}
inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; }
inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; }
inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i]));
}
}
inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; }
inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i]));
}
}
inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; }
inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i]));
}
}
// compute GGML_VEC_DOT_UNROLL dot products at once
// xs - x row stride in bytes
inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GGML_RESTRICT s, void * GGML_RESTRICT xv, ggml_fp16_t * GGML_RESTRICT y) {
ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
ggml_fp16_t * GGML_RESTRICT x[GGML_VEC_DOT_UNROLL];
for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
}
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
GGML_F16_VEC ax[GGML_F16_ARR];
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
}
}
}
// reduce sum0..sum3 to sum0
for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
}
// leftovers
for (int i = np; i < n; ++i) {
for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
}
}
#else
for (int i = 0; i < n; ++i) {
for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
}
}
#endif
for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
s[i] = (float)sumf[i];
}
}
inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));
GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx);
GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
}
}
// leftovers
for (int i = np; i < n; ++i) {
y[i] += x[i]*v;
}
#else
// scalar
for (int i = 0; i < n; ++i) {
y[i] += x[i]*v;
}
#endif
}
inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
GGML_F16_VEC ax[GGML_F16_ARR];
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
}
}
// leftovers
for (int i = np; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
}
#else
// scalar
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
}
#endif
}
// xs and vs are byte strides of x and v
inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * GGML_RESTRICT y, const float * GGML_RESTRICT xv, const float * GGML_RESTRICT vv) {
const float * GGML_RESTRICT x[GGML_VEC_MAD_UNROLL];
const float * GGML_RESTRICT v[GGML_VEC_MAD_UNROLL];
for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) {
x[i] = (const float *) ((const char *) xv + i*xs);
v[i] = (const float *) ((const char *) vv + i*vs);
}
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));
GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
vx[k] = GGML_F32_VEC_SET1(v[k][0]);
}
GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
}
GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
}
}
// leftovers
for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
for (int i = np; i < n; ++i) {
y[i] += x[k][i]*v[k][0];
}
}
#else
// scalar
for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
for (int i = 0; i < n; ++i) {
y[i] += x[k][i]*v[k][0];
}
}
#endif
}
//inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] *= v; }
inline static void ggml_vec_scale_f32(const int n, float * y, const float v) {
#if defined(GGML_USE_ACCELERATE)
vDSP_vsmul(y, 1, &v, y, 1, n);
#elif defined(GGML_SIMD)
const int np = (n & ~(GGML_F32_STEP - 1));
GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
GGML_F32_VEC ay[GGML_F32_ARR];
for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_MUL(ay[j], vx);
GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
}
}
// leftovers
for (int i = np; i < n; ++i) {
y[i] *= v;
}
#else
// scalar
for (int i = 0; i < n; ++i) {
y[i] *= v;
}
#endif
}
inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
#if defined(GGML_SIMD)
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
}
}
// leftovers
for (int i = np; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
}
#else
// scalar
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
}
#endif
}
inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s); }
inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; }
inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(v*v);
}
}
inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); }
inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); }
inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); }
inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
}
}
inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
}
}
inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); }
inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i])));
}
}
inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f);
}
}
inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) {
for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
}
}
inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i]))));
}
}
// TODO: optimize performance
inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
}
}
inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
}
}
inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i])));
}
}
static const float GELU_COEF_A = 0.044715f;
static const float GELU_QUICK_COEF = -1.702f;
static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
inline static float ggml_gelu_f32(float x) {
return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
}
inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
const uint16_t * i16 = (const uint16_t *) x;
for (int i = 0; i < n; ++i) {
y[i] = ggml_table_gelu_f16[i16[i]];
}
}
#ifdef GGML_GELU_FP16
inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
uint16_t t;
for (int i = 0; i < n; ++i) {
if (x[i] <= -10.0f) {
y[i] = 0.0f;
} else if (x[i] >= 10.0f) {
y[i] = x[i];
} else {
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
memcpy(&t, &fp16, sizeof(uint16_t));
y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
}
}
}
#else
inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
for (int i = 0; i < n; ++i) {
y[i] = ggml_gelu_f32(x[i]);
}
}
#endif
inline static float ggml_gelu_quick_f32(float x) {
return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x)));
}
//inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
// const uint16_t * i16 = (const uint16_t *) x;
// for (int i = 0; i < n; ++i) {
// y[i] = ggml_table_gelu_quick_f16[i16[i]];
// }
//}
#ifdef GGML_GELU_QUICK_FP16
inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
uint16_t t;
for (int i = 0; i < n; ++i) {
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
memcpy(&t, &fp16, sizeof(uint16_t));
y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
}
}
#else
inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
for (int i = 0; i < n; ++i) {
y[i] = ggml_gelu_quick_f32(x[i]);
}
}
#endif
inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
}
}
// Sigmoid Linear Unit (SiLU) function
inline static float ggml_silu_f32(float x) {
return x/(1.0f + expf(-x));
}
inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
float v = GGML_FP16_TO_FP32(x);
return GGML_FP32_TO_FP16(v/(1.0f + expf(-v)));
}
#if __FINITE_MATH_ONLY__
#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
#error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461"
#endif
#if defined(__ARM_NEON) && defined(__aarch64__)
// adapted from arm limited optimized routine
// the maximum error is 1.45358 plus 0.5 ulps
// numbers above 88.38 will flush to infinity
// numbers beneath -103.97 will flush to zero
inline static float32x4_t ggml_v_expf(float32x4_t x) {
const float32x4_t r = vdupq_n_f32(0x1.8p23f);
const float32x4_t z = vfmaq_f32(r, x, vdupq_n_f32(0x1.715476p+0f));
const float32x4_t n = vsubq_f32(z, r);
const float32x4_t b = vfmsq_f32(vfmsq_f32(x, n, vdupq_n_f32(0x1.62e4p-1f)), n,
vdupq_n_f32(0x1.7f7d1cp-20f));
const uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_f32(z), 23);
const float32x4_t k = vreinterpretq_f32_u32(vaddq_u32(e, vreinterpretq_u32_f32(vdupq_n_f32(1))));
const uint32x4_t c = vcagtq_f32(n, vdupq_n_f32(126));
const float32x4_t u = vmulq_f32(b, b);
const float32x4_t j = vfmaq_f32(
vmulq_f32(vdupq_n_f32(0x1.ffffecp-1f), b),
vfmaq_f32(vfmaq_f32(vdupq_n_f32(0x1.fffdb6p-2f), vdupq_n_f32(0x1.555e66p-3f), b),
vfmaq_f32(vdupq_n_f32(0x1.573e2ep-5f), vdupq_n_f32(0x1.0e4020p-7f), b), u), u);
if (!vpaddd_u64(vreinterpretq_u64_u32(c)))
return vfmaq_f32(k, j, k);
const uint32x4_t d = vandq_u32(vclezq_f32(n), vdupq_n_u32(0x82000000));
const float32x4_t s1 = vreinterpretq_f32_u32(vaddq_u32(d, vdupq_n_u32(0x7f000000)));
const float32x4_t s2 = vreinterpretq_f32_u32(vsubq_u32(e, d));
return vbslq_f32(vcagtq_f32(n, vdupq_n_f32(192)), vmulq_f32(s1, s1),
vbslq_f32(c, vmulq_f32(vfmaq_f32(s2, s2, j), s1), vfmaq_f32(k, k, j)));
}
// computes silu x/(1+exp(-x)) in single precision vector
inline static float32x4_t ggml_v_silu(float32x4_t x) {
const float32x4_t one = vdupq_n_f32(1.0f);
const float32x4_t zero = vdupq_n_f32(0.0f);
const float32x4_t neg_x = vsubq_f32(zero, x);
const float32x4_t exp_neg_x = ggml_v_expf(neg_x);
const float32x4_t one_plus_exp_neg_x = vaddq_f32(one, exp_neg_x);
return vdivq_f32(x, one_plus_exp_neg_x);
}
#elif defined(__AVX512F__) && defined(__AVX512DQ__)
// adapted from arm limited optimized routine
// the maximum error is 1.45358 plus 0.5 ulps
// numbers above 88.38 will flush to infinity
// numbers beneath -103.97 will flush to zero
inline static __m512 ggml_v_expf(__m512 x) {
const __m512 r = _mm512_set1_ps(0x1.8p23f);
const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r);
const __m512 n = _mm512_sub_ps(z, r);
const __m512 b =
_mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f),
_mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x));
const __mmask16 d =
_mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ);
const __m512 u = _mm512_mul_ps(b, b);
const __m512 j = _mm512_fmadd_ps(
_mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b,
_mm512_set1_ps(0x1.573e2ep-5f)),
u,
_mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b,
_mm512_set1_ps(0x1.fffdb6p-2f))),
u,
_mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F)));
const __m512 res = _mm512_scalef_ps(j, n);
if (_mm512_kortestz(d, d))
return res;
const __m512 zero = _mm512_setzero_ps();
const __m512 alt = _mm512_mask_blend_ps(
_mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero);
return _mm512_mask_blend_ps(d, res, alt);
}
// computes silu x/(1+exp(-x)) in single precision vector
inline static __m512 ggml_v_silu(__m512 x) {
const __m512 one = _mm512_set1_ps(1);
const __m512 zero = _mm512_setzero_ps();
const __m512 neg_x = _mm512_sub_ps(zero, x);
const __m512 exp_neg_x = ggml_v_expf(neg_x);
const __m512 one_plus_exp_neg_x = _mm512_add_ps(one, exp_neg_x);
return _mm512_div_ps(x, one_plus_exp_neg_x);
}
#elif defined(__AVX2__) && defined(__FMA__)
// adapted from arm limited optimized routine
// the maximum error is 1.45358 plus 0.5 ulps
// numbers above 88.38 will flush to infinity
// numbers beneath -103.97 will flush to zero
inline static __m256 ggml_v_expf(__m256 x) {
const __m256 r = _mm256_set1_ps(0x1.8p23f);
const __m256 z = _mm256_fmadd_ps(x, _mm256_set1_ps(0x1.715476p+0f), r);
const __m256 n = _mm256_sub_ps(z, r);
const __m256 b = _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.7f7d1cp-20f),
_mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.62e4p-1f), x));
const __m256i e = _mm256_slli_epi32(_mm256_castps_si256(z), 23);
const __m256 k = _mm256_castsi256_ps(
_mm256_add_epi32(e, _mm256_castps_si256(_mm256_set1_ps(1))));
const __m256i c = _mm256_castps_si256(
_mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n),
_mm256_set1_ps(126), _CMP_GT_OQ));
const __m256 u = _mm256_mul_ps(b, b);
const __m256 j = _mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_set1_ps(0x1.0e4020p-7f), b,
_mm256_set1_ps(0x1.573e2ep-5f)), u,
_mm256_fmadd_ps(_mm256_set1_ps(0x1.555e66p-3f), b,
_mm256_set1_ps(0x1.fffdb6p-2f))),
u, _mm256_mul_ps(_mm256_set1_ps(0x1.ffffecp-1f), b));
if (!_mm256_movemask_ps(_mm256_castsi256_ps(c)))
return _mm256_fmadd_ps(j, k, k);
const __m256i g = _mm256_and_si256(
_mm256_castps_si256(_mm256_cmp_ps(n, _mm256_setzero_ps(), _CMP_LE_OQ)),
_mm256_set1_epi32(0x82000000u));
const __m256 s1 =
_mm256_castsi256_ps(_mm256_add_epi32(g, _mm256_set1_epi32(0x7f000000u)));
const __m256 s2 = _mm256_castsi256_ps(_mm256_sub_epi32(e, g));
const __m256i d = _mm256_castps_si256(
_mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n),
_mm256_set1_ps(192), _CMP_GT_OQ));
return _mm256_or_ps(
_mm256_and_ps(_mm256_castsi256_ps(d), _mm256_mul_ps(s1, s1)),
_mm256_andnot_ps(
_mm256_castsi256_ps(d),
_mm256_or_ps(
_mm256_and_ps(_mm256_castsi256_ps(c),
_mm256_mul_ps(_mm256_fmadd_ps(s2, j, s2), s1)),
_mm256_andnot_ps(_mm256_castsi256_ps(c), _mm256_fmadd_ps(k, j, k)))));
}
// computes silu x/(1+exp(-x)) in single precision vector
inline static __m256 ggml_v_silu(__m256 x) {
const __m256 one = _mm256_set1_ps(1);
const __m256 zero = _mm256_setzero_ps();
const __m256 neg_x = _mm256_sub_ps(zero, x);
const __m256 exp_neg_x = ggml_v_expf(neg_x);
const __m256 one_plus_exp_neg_x = _mm256_add_ps(one, exp_neg_x);
return _mm256_div_ps(x, one_plus_exp_neg_x);
}
#elif defined(__SSE2__) // __AVX2__ / __ARM_NEON
#if defined(__FMA__)
#define MADD128(x, y, z) _mm_fmadd_ps(x, y, z)
#define NMADD128(x, y, z) _mm_fnmadd_ps(x, y, z)
#else
#define MADD128(x, y, z) _mm_add_ps(_mm_mul_ps(x, y), z)
#define NMADD128(x, y, z) _mm_sub_ps(z, _mm_mul_ps(x, y))
#endif
// adapted from arm limited optimized routine
// the maximum error is 1.45358 plus 0.5 ulps
// numbers above 88.38 will flush to infinity
// numbers beneath -103.97 will flush to zero
inline static __m128 ggml_v_expf(__m128 x) {
const __m128 r = _mm_set1_ps(0x1.8p23f);
const __m128 z = MADD128(x, _mm_set1_ps(0x1.715476p+0f), r);
const __m128 n = _mm_sub_ps(z, r);
const __m128 b =
NMADD128(n, _mm_set1_ps(0x1.7f7d1cp-20f), NMADD128(n, _mm_set1_ps(0x1.62e4p-1f), x));
const __m128i e = _mm_slli_epi32(_mm_castps_si128(z), 23);
const __m128 k = _mm_castsi128_ps(_mm_add_epi32(e, _mm_castps_si128(_mm_set1_ps(1))));
const __m128i c =
_mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(126)));
const __m128 u = _mm_mul_ps(b, b);
const __m128 j =
MADD128(MADD128(MADD128(_mm_set1_ps(0x1.0e4020p-7f), b, _mm_set1_ps(0x1.573e2ep-5f)), u,
MADD128(_mm_set1_ps(0x1.555e66p-3f), b, _mm_set1_ps(0x1.fffdb6p-2f))),
u, _mm_mul_ps(_mm_set1_ps(0x1.ffffecp-1f), b));
if (!_mm_movemask_epi8(c))
return MADD128(j, k, k);
const __m128i g = _mm_and_si128(_mm_castps_si128(_mm_cmple_ps(n, _mm_setzero_ps())),
_mm_set1_epi32(0x82000000u));
const __m128 s1 = _mm_castsi128_ps(_mm_add_epi32(g, _mm_set1_epi32(0x7f000000u)));
const __m128 s2 = _mm_castsi128_ps(_mm_sub_epi32(e, g));
const __m128i d =
_mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(192)));
return _mm_or_ps(
_mm_and_ps(_mm_castsi128_ps(d), _mm_mul_ps(s1, s1)),
_mm_andnot_ps(_mm_castsi128_ps(d),
_mm_or_ps(_mm_and_ps(_mm_castsi128_ps(c), _mm_mul_ps(MADD128(s2, j, s2), s1)),
_mm_andnot_ps(_mm_castsi128_ps(c), MADD128(k, j, k)))));
}
// computes silu x/(1+exp(-x)) in single precision vector
inline static __m128 ggml_v_silu(__m128 x) {
const __m128 one = _mm_set1_ps(1);
const __m128 zero = _mm_setzero_ps();
const __m128 neg_x = _mm_sub_ps(zero, x);
const __m128 exp_neg_x = ggml_v_expf(neg_x);
const __m128 one_plus_exp_neg_x = _mm_add_ps(one, exp_neg_x);
return _mm_div_ps(x, one_plus_exp_neg_x);
}
#endif // __ARM_NEON / __AVX2__ / __SSE2__
inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) {
y[i] = ggml_silu_f16(x[i]);
}
}
inline static float ggml_silu_backward_f32(float x, float dy) {
const float s = 1.0f/(1.0f + expf(-x));
return dy*s*(1.0f + x*(1.0f - s));
}
inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) {
const float v = GGML_FP16_TO_FP32(x);
const float s = 1.0f/(1.0f + expf(-v));
return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
}
inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
for (int i = 0; i < n; ++i) {
dx[i] = ggml_silu_backward_f32(x[i], dy[i]);
}
}
inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, const ggml_fp16_t * x, const ggml_fp16_t * dy) {
for (int i = 0; i < n; ++i) {
dx[i] = ggml_silu_backward_f16(x[i], dy[i]);
}
}
inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
#ifndef GGML_USE_ACCELERATE
ggml_float sum = 0.0;
for (int i = 0; i < n; ++i) {
sum += (ggml_float)x[i];
}
*s = (float)sum;
#else
vDSP_sve(x, 1, s, n);
#endif
}
inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) {
ggml_float sum = 0.0;
for (int i = 0; i < n; ++i) {
sum += (ggml_float)x[i];
}
*s = sum;
}
inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) {
float sum = 0.0f;
for (int i = 0; i < n; ++i) {
sum += GGML_FP16_TO_FP32(x[i]);
}
*s = sum;
}
inline static void ggml_vec_sum_bf16_ggf(const int n, float * s, const ggml_bf16_t * x) {
float sum = 0.0f;
for (int i = 0; i < n; ++i) {
sum += GGML_BF16_TO_FP32(x[i]);
}
*s = sum;
}
inline static void ggml_vec_max_f32(const int n, float * s, const float * x) {
#ifndef GGML_USE_ACCELERATE
float max = -INFINITY;
for (int i = 0; i < n; ++i) {
max = MAX(max, x[i]);
}
*s = max;
#else
vDSP_maxv(x, 1, s, n);
#endif
}
inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x) {
ggml_vec_norm_f32(n, s, x);
*s = 1.f/(*s);
}
inline static void ggml_vec_argmax_f32(const int n, int * s, const float * x) {
float max = -INFINITY;
int idx = 0;
for (int i = 0; i < n; ++i) {
max = MAX(max, x[i]);
if (max == x[i]) { idx = i; }
}
*s = idx;
}
#ifdef __cplusplus
}
#endif

15
ggml/src/ggml-cuda/convert.cu
View File

@ -579,7 +579,7 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
const src_t * x = (const src_t *) vx;
y[i] = x[i];
y[i] = float(x[i]);
}
template <typename src_t, typename dst_t>
@ -588,6 +588,17 @@ static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict_
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_F32:
return convert_unary_cuda<float>;
case GGML_TYPE_F16:
return convert_unary_cuda<half>;
default:
return nullptr;
}
}
to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
@ -633,6 +644,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
return dequantize_row_iq3_s_cuda;
case GGML_TYPE_F32:
return convert_unary_cuda<float>;
case GGML_TYPE_BF16:
return convert_unary_cuda<nv_bfloat16>;
default:
return nullptr;
}

3
ggml/src/ggml-cuda/convert.cuh
View File

@ -7,7 +7,10 @@ using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, in
typedef to_t_cuda_t<float> to_fp32_cuda_t;
typedef to_t_cuda_t<half> to_fp16_cuda_t;
typedef to_t_cuda_t<nv_bfloat16> to_bf16_cuda_t;
to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);

24
ggml/src/ggml-cuda/cpy.cu
View File

@ -10,6 +10,13 @@ static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
*dsti = *xi;
}
static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
*dsti = *xi;
}
static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
half * dsti = (half *) cdsti;
@ -360,6 +367,9 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
// copy destination pointers to GPU
CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
cuda_graph->graph_cpynode_index = 0; // reset index
#else
GGML_UNUSED(cuda_graph); GGML_UNUSED(host_dest_ptrs);
GGML_UNUSED(host_dest_ptrs_size); GGML_UNUSED(stream);
#endif
}
@ -383,6 +393,16 @@ static void ggml_cpy_f32_f32_cuda(
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
}
static void ggml_cpy_f32_bf16_cuda(
const char * cx, char * cdst, const int ne,
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
}
static void ggml_cpy_f32_f16_cuda(
const char * cx, char * cdst, const int ne,
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@ -578,6 +598,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
@ -631,6 +653,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
return nullptr;
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
return (void*) cpy_f32_f16<cpy_1_f32_f32>;
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
return (void*) cpy_f32_f16<cpy_1_f32_f16>;
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {

10
ggml/src/ggml-cuda/fattn-common.cuh
View File

@ -62,7 +62,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0(
T sum = 0.0f;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
const int k_KQ = k_KQ_0 + threadIdx.x;
const int ib = k_KQ / QI8_1;
@ -102,7 +102,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1(
T sum = 0.0f;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
const int k_KQ = k_KQ_0 + threadIdx.x;
const int ib = k_KQ / QI8_1;
@ -146,7 +146,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0(
T sum = 0.0f;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
const int k_KQ = k_KQ_0 + threadIdx.x;
const int ib = k_KQ / QI8_1;
@ -193,7 +193,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1(
T sum = 0.0f;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
const int k_KQ = k_KQ_0 + threadIdx.x;
const int ib = k_KQ / QI8_1;
@ -244,7 +244,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
T sum = 0.0f;
#pragma unroll
for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
const int k_KQ = k_KQ_0 + threadIdx.x;
const int ib = k_KQ / QI8_0;

12
ggml/src/ggml-cuda/fattn-tile-f32.cu
View File

@ -52,6 +52,18 @@ static __global__ void flash_attn_tile_ext_f32(
return;
#endif // FP16_MMA_AVAILABLE
if (use_logit_softcap && !(D == 128 || D == 256)) {
GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
GGML_UNUSED(ne2); GGML_UNUSED(ne3);
NO_DEVICE_CODE;
return;
}

18
ggml/src/ggml-cuda/fattn-vec-f32.cuh
View File

@ -45,6 +45,18 @@ static __global__ void flash_attn_vec_ext_f32(
// Skip unused kernel variants for faster compilation:
if (use_logit_softcap && !(D == 128 || D == 256)) {
GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
GGML_UNUSED(ne2); GGML_UNUSED(ne3);
NO_DEVICE_CODE;
return;
}
@ -114,7 +126,7 @@ static __global__ void flash_attn_vec_ext_f32(
// Set memory to zero if out of bounds:
if (ncols > 2 && ic0 + j >= ne01) {
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
tmp_q_i32[i] = 0;
@ -127,7 +139,7 @@ static __global__ void flash_attn_vec_ext_f32(
const float * Q_f = (const float *) (Q + j*nb01);
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
quantize_q8_1_to_shared<float2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
}
}
@ -140,7 +152,7 @@ static __global__ void flash_attn_vec_ext_f32(
float2 * tmp_q_ds = (float2 *) (tmp_q_i32 + D/sizeof(int));
#pragma unroll
for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
const int i = i0 + threadIdx.x;
Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];

2
ggml/src/ggml-cuda/fattn.cu
View File

@ -299,7 +299,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
const bool can_use_vector_kernel = (Q->ne[0] % (2*warp_size) == 0) && (prec == GGML_PREC_DEFAULT || Q->ne[0] <= 128);
const bool can_use_vector_kernel = Q->ne[0] % (2*warp_size) == 0;
if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
if (prec == GGML_PREC_DEFAULT) {
ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);

69
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -96,31 +96,32 @@ 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_HIP) && defined(GGML_HIP_UMA)
auto res = hipMallocManaged(ptr, size);
if (res == hipSuccess) {
// if error we "need" to know why...
CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
}
return res;
#else
#if !defined(GGML_USE_HIP)
cudaError_t err;
if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
{
err = cudaMallocManaged(ptr, size);
#if defined(GGML_USE_HIP)
if (err == hipSuccess) {
CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
}
// fall back to cudaMalloc if not supported (e.g. on Windows)
if (err == hipErrorNotSupported) {
static bool warned_unsupported = false;
if (!warned_unsupported) {
GGML_LOG_WARN("hipMallocManaged unsupported, falling back to hipMalloc.\n");
warned_unsupported = true;
}
err = cudaMalloc(ptr, size);
}
#endif // defined(GGML_USE_HIP)
}
else
{
err = cudaMalloc(ptr, size);
}
return err;
#else
return cudaMalloc(ptr, size);
#endif // !defined(GGML_USE_HIP)
#endif
}
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@ -1194,7 +1195,35 @@ static void ggml_cuda_op_mul_mat_cublas(
const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;
if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
ggml_cuda_pool_alloc<nv_bfloat16> src1_as_bf16(ctx.pool(id));
if (src1->type != GGML_TYPE_BF16) {
const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type);
GGML_ASSERT(to_bf16_cuda != nullptr);
size_t ne = src1_ncols*ne10;
src1_as_bf16.alloc(ne);
to_bf16_cuda(src1_ddf_i, src1_as_bf16.get(), ne, stream);
}
const nv_bfloat16 * src1_ptr = src1->type == GGML_TYPE_BF16 ? (const nv_bfloat16 *) src1_ddf_i : src1_as_bf16.get();
const nv_bfloat16 * src0_ptr = (const nv_bfloat16 *)src0_dd_i;
ggml_cuda_pool_alloc<nv_bfloat16> dst_bf16(ctx.pool(id), row_diff*src1_ncols);
const float alpha_f32 = 1.0f;
const float beta_f32 = 0.0f;
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
CUBLAS_CHECK(
cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
row_diff, src1_ncols, ne10,
&alpha_f32, src0_ptr, CUDA_R_16BF, ne00,
src1_ptr, CUDA_R_16BF, ne10,
&beta_f32, dst_bf16.get(), CUDA_R_16BF, ldc,
CUBLAS_COMPUTE_32F,
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream);
} else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
// convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
if (src0->type != GGML_TYPE_F16) {
@ -2460,10 +2489,10 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
#endif
}
if (node->op == GGML_OP_MUL_MAT_ID) {
if (node->op == GGML_OP_MUL_MAT_ID || node->op == GGML_OP_CONT || node->op == GGML_OP_DUP) {
use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
#endif
}
@ -3051,6 +3080,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) {
return true;
}
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
return true;
}
@ -3185,6 +3217,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_GROUP_NORM:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_UPSCALE:
return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
case GGML_OP_PAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:

3
ggml/src/ggml-cuda/vendors/hip.h vendored
View File

@ -20,6 +20,7 @@
#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
#define CUBLAS_TF32_TENSOR_OP_MATH 0
#define CUDA_R_16F HIPBLAS_R_16F
#define CUDA_R_16BF HIPBLAS_R_16B
#define CUDA_R_32F HIPBLAS_R_32F
#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED hipDeviceAttributeVirtualMemoryManagementSupported
#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED hipMemAllocationGranularityRecommended
@ -70,6 +71,8 @@
#define cudaLaunchHostFunc hipLaunchHostFunc
#define cudaMalloc hipMalloc
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
#define cudaMallocManaged hipMallocManaged
#define cudaMemAdvise hipMemAdvise
#define cudaMemcpy hipMemcpy
#define cudaMemcpyAsync hipMemcpyAsync
#define cudaMemcpyPeerAsync hipMemcpyPeerAsync

1
ggml/src/ggml-cuda/vendors/musa.h vendored
View File

@ -15,6 +15,7 @@
#define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
#define CUDA_R_16F MUSA_R_16F
#define CUDA_R_16BF MUSA_R_16BF
#define CUDA_R_32F MUSA_R_32F
#define cublasComputeType_t cudaDataType_t
#define cublasCreate mublasCreate

4
ggml/src/ggml-hip/CMakeLists.txt
View File

@ -89,10 +89,6 @@ endif()
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()

41
ggml/src/ggml-impl.h
View File

@ -148,8 +148,14 @@ struct ggml_map_custom2_op_params {
struct ggml_map_custom3_op_params {
ggml_custom3_op_t fun;
int n_tasks;
void * userdata;
int n_tasks;
void * userdata;
};
struct ggml_custom_op_params {
ggml_custom_op_t fun;
int n_tasks;
void * userdata;
};
// bitset
@ -311,29 +317,28 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
// FP16 to FP32 conversion
#if defined(__ARM_NEON)
#if defined(_MSC_VER) || (defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11)
typedef uint16_t ggml_fp16_internal_t;
#else
typedef __fp16 ggml_fp16_internal_t;
#endif
#endif
#if defined(__ARM_NEON) && !defined(_MSC_VER) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11)
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
//
// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
//
#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
__fp16 tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
__fp16 tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
@ -357,8 +362,8 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
#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;
float f;
double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
@ -370,8 +375,8 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
double d;
ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
@ -485,7 +490,7 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
#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)
#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()

42
ggml/src/ggml-metal/ggml-metal.m
View File

@ -402,6 +402,13 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,
@ -1059,6 +1066,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192, flash_attn_ext_q8_0_h192, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96, flash_attn_ext_vec_f16_h96, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96, flash_attn_ext_vec_bf16_h96, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96, flash_attn_ext_vec_q4_0_h96, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96, flash_attn_ext_vec_q4_1_h96, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96, flash_attn_ext_vec_q5_0_h96, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96, flash_attn_ext_vec_q5_1_h96, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96, flash_attn_ext_vec_q8_0_h96, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128, flash_attn_ext_vec_bf16_h128, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128, flash_attn_ext_vec_q4_0_h128, has_simdgroup_reduction);
@ -1334,8 +1348,9 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
return op->src[0]->type == GGML_TYPE_F16;
case GGML_OP_POOL_1D:
return false;
case GGML_OP_POOL_2D:
case GGML_OP_UPSCALE:
return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
case GGML_OP_POOL_2D:
case GGML_OP_PAD:
case GGML_OP_PAD_REFLECT_1D:
case GGML_OP_TIMESTEP_EMBEDDING:
@ -1345,6 +1360,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_ARANGE:
return true;
case GGML_OP_FLASH_ATTN_EXT:
if (op->src[0]->ne[0] == 32) {
// head size == 32 (e.g. bert-bge-small)
// TODO: not sure if it is worth adding kernels for this size
return false;
}
if (op->src[1]->type != op->src[2]->type) {
return false;
}
@ -3837,7 +3857,7 @@ static void ggml_metal_encode_node(
// TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
// for now avoiding mainly to keep the number of templates/kernels a bit lower
// these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 192)) {
if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192)) {
switch (src1->type) {
case GGML_TYPE_F16:
{
@ -4004,6 +4024,24 @@ static void ggml_metal_encode_node(
use_vec_kernel = true;
switch (ne00) {
case 96:
{
switch (src1->type) {
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96].pipeline; break;
case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96].pipeline; break;
default:
{
GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
GGML_LOG_ERROR("add template specialization for this type\n");
GGML_ABORT("add template specialization for this type");
}
}
} break;
case 128:
{
switch (src1->type) {

10
ggml/src/ggml-metal/ggml-metal.metal
View File

@ -3959,6 +3959,16 @@ kernel void kernel_flash_attn_ext_vec(
typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 96, 96, 4>;
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 96, 96, 4>;
#endif
template [[host_name("kernel_flash_attn_ext_vec_q4_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0, 8, dequantize_q4_0_t4, 96, 96, 4>;
template [[host_name("kernel_flash_attn_ext_vec_q4_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1, 8, dequantize_q4_1_t4, 96, 96, 4>;
template [[host_name("kernel_flash_attn_ext_vec_q5_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0, 8, dequantize_q5_0_t4, 96, 96, 4>;
template [[host_name("kernel_flash_attn_ext_vec_q5_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1, 8, dequantize_q5_1_t4, 96, 96, 4>;
template [[host_name("kernel_flash_attn_ext_vec_q8_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0, 8, dequantize_q8_0_t4, 96, 96, 4>;
template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>;
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 128, 128, 4>;

14
ggml/src/ggml-opencl/ggml-opencl.cpp
View File

@ -415,6 +415,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
unsigned number;
cl_device_type type;
char name[128];
char version[128];
};
enum { NPLAT = 16, NDEV = 16 };
@ -455,6 +456,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
d->platform = p;
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_VERSION, sizeof(d->version), &d->version, NULL));
if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
p->default_device = d;
@ -547,7 +549,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
}
GGML_LOG_INFO("ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
GGML_LOG_INFO("ggml_opencl: selecting device: '%s'\n", default_device->name);
GGML_LOG_INFO("ggml_opencl: selecting device: '%s (%s)'\n", default_device->name, default_device->version);
if (default_device->type != CL_DEVICE_TYPE_GPU) {
GGML_LOG_WARN("ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
}
@ -556,9 +558,15 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
dev_ctx->device = default_device->id;
backend_ctx->device = default_device->id;
if (strstr(default_device->name, "Adreno")) {
if (strstr(default_device->name, "Adreno") ||
strstr(default_device->name, "Qualcomm") ||
strstr(default_device->version, "Adreno")) {
backend_ctx->gpu_family = GPU_FAMILY::ADRENO;
backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->name);
// Usually device version contains the detailed device name
backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->version);
if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::ADRENO_UNKNOWN) {
backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->name);
}
// Use wave size of 64 for all Adreno GPUs.
backend_ctx->adreno_wave_size = 64;

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