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' of https://github.com/ggerganov/llama.cpp into cli_output

This commit is contained in:
David Baker 2026-03-18 11:32:38 +00:00
parent c522288ab6 312cf03328
commit e4cfe15e69
No known key found for this signature in database
GPG Key ID: 89298D31E5B7B548
203 changed files with 14152 additions and 2797 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

138
.devops/openvino.Dockerfile Normal file
View File

@ -0,0 +1,138 @@
ARG OPENVINO_VERSION_MAJOR=2026.0
ARG OPENVINO_VERSION_FULL=2026.0.0.20965.c6d6a13a886
ARG UBUNTU_VERSION=24.04
# Optional proxy build arguments - empty by default
ARG http_proxy=
ARG https_proxy=
## Build Image
FROM ubuntu:${UBUNTU_VERSION} AS build
# Pass proxy args to build stage
ARG http_proxy
ARG https_proxy
RUN apt-get update && \
apt-get install -y --no-install-recommends \
ca-certificates \
gnupg \
wget \
git \
cmake \
ninja-build \
build-essential \
libtbb12 \
libssl-dev \
ocl-icd-opencl-dev \
opencl-headers \
opencl-clhpp-headers \
intel-opencl-icd && \
rm -rf /var/lib/apt/lists/*
# Install OpenVINO for Ubuntu 24.04
ARG OPENVINO_VERSION_MAJOR
ARG OPENVINO_VERSION_FULL
RUN mkdir -p /opt/intel && \
wget https://storage.openvinotoolkit.org/repositories/openvino/packages/${OPENVINO_VERSION_MAJOR}/linux/openvino_toolkit_ubuntu24_${OPENVINO_VERSION_FULL}_x86_64.tgz && \
tar -xf openvino_toolkit_ubuntu24_${OPENVINO_VERSION_FULL}_x86_64.tgz && \
mv openvino_toolkit_ubuntu24_${OPENVINO_VERSION_FULL}_x86_64 /opt/intel/openvino_${OPENVINO_VERSION_MAJOR} && \
cd /opt/intel/openvino_${OPENVINO_VERSION_MAJOR} && \
echo "Y" | ./install_dependencies/install_openvino_dependencies.sh && \
cd - && \
ln -s /opt/intel/openvino_${OPENVINO_VERSION_MAJOR} /opt/intel/openvino
ENV OpenVINO_DIR=/opt/intel/openvino
WORKDIR /app
COPY . .
# Build Stage
RUN bash -c "source ${OpenVINO_DIR}/setupvars.sh && \
cmake -B build/ReleaseOV -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_OPENVINO=ON && \
cmake --build build/ReleaseOV -j$(nproc)"
# Copy all necessary libraries
RUN mkdir -p /app/lib && \
find build/ReleaseOV -name '*.so*' -exec cp {} /app/lib \; && \
find ${OpenVINO_DIR}/runtime/lib/intel64 -name '*.so*' -exec cp -P {} /app/lib \; 2>/dev/null || \
find ${OpenVINO_DIR}/lib/intel64 -name '*.so*' -exec cp -P {} /app/lib \;
# Create runtime directories and copy binaries
RUN mkdir -p /app/full \
&& cp build/ReleaseOV/bin/* /app/full/ \
&& cp *.py /app/full \
&& cp -r gguf-py /app/full \
&& cp -r requirements /app/full \
&& cp requirements.txt /app/full \
&& cp .devops/tools.sh /app/full/tools.sh
## Base Runtime Image
FROM ubuntu:${UBUNTU_VERSION} AS base
# Pass proxy args to runtime stage
ARG http_proxy
ARG https_proxy
RUN apt-get update \
&& apt-get install -y libgomp1 libtbb12 curl\
&& apt autoremove -y \
&& apt clean -y \
&& rm -rf /tmp/* /var/tmp/* \
&& find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
&& find /var/cache -type f -delete
COPY --from=build /app/lib/ /app/
### Full (all binaries)
FROM base AS full
ARG http_proxy
ARG https_proxy
COPY --from=build /app/full /app/
WORKDIR /app
RUN apt-get update && \
apt-get install -y --no-install-recommends \
git \
python3 \
python3-venv \
python3-pip && \
python3 -m venv /ov-venv && \
/ov-venv/bin/pip install --no-cache-dir --upgrade pip setuptools wheel && \
/ov-venv/bin/pip install --no-cache-dir -r requirements.txt && \
apt-get autoremove -y && \
apt-get clean && \
rm -rf /tmp/* /var/tmp/* && \
find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
find /var/cache -type f -delete
ENTRYPOINT ["/bin/bash", "-c", "source /ov-venv/bin/activate && exec /app/tools.sh \"$@\"", "--"]
### Light, CLI only
FROM base AS light
COPY --from=build /app/full/llama-cli /app/
WORKDIR /app
ENTRYPOINT [ "/app/llama-cli" ]
### Server, Server only
FROM base AS server
ENV LLAMA_ARG_HOST=0.0.0.0
COPY --from=build /app/full/llama-server /app/
WORKDIR /app
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/app/llama-server" ]

5
.devops/vulkan.Dockerfile
View File

@ -53,10 +53,11 @@ RUN apt-get update \
&& apt-get install -y \
build-essential \
git \
python3 \
python3-dev \
python3.13 \
python3.13-dev \
python3-pip \
python3-wheel \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.13 100 \
&& pip install --break-system-packages --upgrade setuptools \
&& pip install --break-system-packages -r requirements.txt \
&& apt autoremove -y \

25
.github/actions/linux-setup-openvino/action.yml vendored Normal file
View File

@ -0,0 +1,25 @@
name: "Linux - Setup OpenVINO Toolkit"
description: "Setup OpenVINO Toolkit for Linux"
inputs:
path:
description: "Installation path"
required: true
version_major:
description: "OpenVINO major version (e.g., 2025.3)"
required: true
version_full:
description: "OpenVINO full version (e.g., 2025.3.0.19807.44526285f24)"
required: true
runs:
using: "composite"
steps:
- name: Setup OpenVINO Toolkit
id: setup
uses: ./.github/actions/unarchive-tar
with:
url: https://storage.openvinotoolkit.org/repositories/openvino/packages/${{ inputs.version_major }}/linux/openvino_toolkit_ubuntu24_${{ inputs.version_full }}_x86_64.tgz
path: ${{ inputs.path }}
type: z
strip: 1

17
.github/labeler.yml vendored
View File

@ -104,3 +104,20 @@ OpenCL:
- any-glob-to-any-file:
- ggml/include/ggml-opencl.h
- ggml/src/ggml-opencl/**
- docs/backend/OPENCL.md
Hexagon:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-hexagon.h
- ggml/src/ggml-hexagon/**
WebGPU:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-webgpu.h
- ggml/src/ggml-webgpu/**
OpenVINO:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-openvino.h
- ggml/src/ggml-openvino/**
- docs/backend/OPENVINO.md

57
.github/workflows/build-3rd-party.yml vendored Normal file
View File

@ -0,0 +1,57 @@
name: CI (3rd-party)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-3rd-party.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
ubuntu-24-llguidance:
runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libssl-dev
- name: Build
id: cmake_build
run: |
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_LLGUIDANCE=ON
cmake --build build --config Release -j $(nproc)
- name: Test
id: cmake_test
run: |
cd build
ctest -L main --verbose --timeout 900

140
.github/workflows/build-android.yml vendored Normal file
View File

@ -0,0 +1,140 @@
name: CI (android)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-android.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp'
]
pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-android.yml',
'examples/llama.android/**'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
android:
runs-on: ubuntu-latest
steps:
- name: Clone
uses: actions/checkout@v6
# Disabled due to size (400MB) and always 0 cache hits
# - name: ccache
# uses: ggml-org/ccache-action@v1.2.16
# with:
# key: android-build
# evict-old-files: 1d
- name: Set up JDK
uses: actions/setup-java@v5
with:
java-version: 17
distribution: zulu
- name: Setup Android SDK
uses: android-actions/setup-android@v3
with:
log-accepted-android-sdk-licenses: false
- name: Build
run: |
cd examples/llama.android
./gradlew build --no-daemon
android-ndk:
runs-on: ubuntu-latest
env:
OPENCL_VERSION: 2025.07.22
strategy:
matrix:
include:
- build: 'arm64-cpu'
defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_OPENSSL=OFF -D GGML_OPENMP=OFF'
- build: 'arm64-snapdragon'
defines: '--preset arm64-android-snapdragon-release'
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Install OpenCL Headers and Libs
id: install_opencl
if: ${{ matrix.build == 'arm64-snapdragon' }}
run: |
mkdir opencl
curl -L -o opencl/clhpp.tar.gz https://github.com/KhronosGroup/OpenCL-CLHPP/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
curl -L -o opencl/headers.tar.gz https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
curl -L -o opencl/icd-loader.tar.gz https://github.com/KhronosGroup/OpenCL-ICD-Loader/archive/refs/tags/v${OPENCL_VERSION}.tar.gz
tar -xaf opencl/headers.tar.gz -C opencl
tar -xaf opencl/clhpp.tar.gz -C opencl
tar -xaf opencl/icd-loader.tar.gz -C opencl
sudo cp -r opencl/OpenCL-Headers-${OPENCL_VERSION}/CL ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
sudo cp -r opencl/OpenCL-CLHPP-${OPENCL_VERSION}/include/CL/* ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/CL
cd opencl/OpenCL-ICD-Loader-${OPENCL_VERSION}
cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -DOPENCL_ICD_LOADER_HEADERS_DIR=${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=31 -DANDROID_STL=c++_shared
cmake --build build
sudo cp build/libOpenCL.so ${ANDROID_NDK_ROOT}/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
rm -rf opencl
- name: Install Hexagon SDK
id: install_hexsdk
if: ${{ matrix.build == 'arm64-snapdragon' }}
env:
HEXSDK_VER: 6.4.0.2
HEXTLS_VER: 19.0.04
run: |
curl -L -o hex-sdk.tar.gz https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v$HEXSDK_VER/hexagon-sdk-v$HEXSDK_VER-amd64-lnx.tar.xz
mkdir hex-sdk
tar -xaf hex-sdk.tar.gz -C hex-sdk
ls -l hex-sdk
sudo mv hex-sdk /opt/hexagon
echo "HEXAGON_SDK_ROOT=/opt/hexagon/$HEXSDK_VER" >> "$GITHUB_ENV"
echo "HEXAGON_TOOLS_ROOT=/opt/hexagon/$HEXSDK_VER/tools/HEXAGON_Tools/$HEXTLS_VER" >> "$GITHUB_ENV"
echo "DEFAULT_HLOS_ARCH=64" >> "$GITHUB_ENV"
echo "DEFAULT_TOOLS_VARIANT=toolv19" >> "$GITHUB_ENV"
echo "DEFAULT_NO_QURT_INC=0" >> "$GITHUB_ENV"
echo "DEFAULT_DSP_ARCH=v73" >> "$GITHUB_ENV"
- name: Update CMake presets
id: update_presets
if: ${{ matrix.build == 'arm64-snapdragon' }}
run: |
cp docs/backend/snapdragon/CMakeUserPresets.json .
- name: Build
id: ndk_build
run: |
cmake ${{ matrix.defines }} -B build
cmake --build build
cmake --install build --prefix pkg-adb/llama.cpp
- name: Test
id: cmake_test
run: |
echo "FIXME: test on devices"

214
.github/workflows/build-apple.yml vendored Normal file
View File

@ -0,0 +1,214 @@
name: CI (apple)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-apple.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.swift',
'**/*.m',
'**/*.metal'
]
pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-apple.yml',
'ggml/src/ggml-metal/**'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
macOS-latest-ios:
runs-on: macos-latest
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.21
with:
key: macOS-latest-ios
evict-old-files: 1d
save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
- name: Build
id: cmake_build
run: |
sysctl -a
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_TOOLS=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_SYSTEM_NAME=iOS \
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
macos-latest-ios-xcode:
runs-on: macos-latest
steps:
- name: Checkout code
uses: actions/checkout@v6
- name: Setup Xcode
uses: ggml-org/setup-xcode@v1
with:
xcode-version: latest-stable
- name: Build
id: cmake_build
run: |
sysctl -a
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_OPENSSL=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TOOLS=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_SYSTEM_NAME=iOS \
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
- name: xcodebuild for swift package
id: xcodebuild
run: |
./build-xcframework.sh
- name: Upload xcframework artifact
uses: actions/upload-artifact@v6
with:
name: llama-xcframework
path: build-apple/llama.xcframework/
retention-days: 1
- name: Build Xcode project
run: |
xcodebuild -downloadPlatform iOS
xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
macOS-latest-tvos:
runs-on: macos-latest
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.21
with:
key: macOS-latest-tvos
evict-old-files: 1d
save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
- name: Build
id: cmake_build
run: |
sysctl -a
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_TOOLS=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_SYSTEM_NAME=tvOS \
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
macOS-latest-visionos:
runs-on: macos-latest
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Build
id: cmake_build
run: |
sysctl -a
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_TOOLS=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_SYSTEM_NAME=visionOS \
-DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
macOS-latest-swift:
runs-on: macos-latest
needs: macos-latest-ios-xcode
strategy:
matrix:
destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.21
with:
key: macOS-latest-swift
evict-old-files: 1d
save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
- name: Download xcframework artifact
uses: actions/download-artifact@v7
with:
name: llama-xcframework
path: build-apple/llama.xcframework/
- name: Build llama.cpp with CMake
id: cmake_build
run: |
sysctl -a
cmake -B build -G Xcode \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_OPENSSL=OFF \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TOOLS=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

50
.github/workflows/build-cache.yml vendored
View File

@ -37,12 +37,39 @@ jobs:
path: ./vulkan_sdk
version: ${{ env.VULKAN_SDK_VERSION }}
ubuntu-24-spacemit-cache:
#ubuntu-24-spacemit-cache:
# runs-on: ubuntu-24.04
# env:
# # Make sure this is in sync with build-linux-cross.yml
# SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
# steps:
# - name: Clone
# id: checkout
# uses: actions/checkout@v6
# - name: Setup Cache
# uses: actions/cache@v5
# id: cache-toolchain
# with:
# path: ./spacemit_toolchain
# key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
# - name: Setup SpacemiT Toolchain
# if: steps.cache-toolchain.outputs.cache-hit != 'true'
# uses: ./.github/actions/linux-setup-spacemit
# with:
# path: ./spacemit_toolchain
# version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
ubuntu-24-openvino-cache:
runs-on: ubuntu-24.04
env:
# Make sure this is in sync with build-linux-cross.yml
SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
# Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
OPENVINO_VERSION_MAJOR: "2026.0"
OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
steps:
- name: Clone
@ -51,17 +78,18 @@ jobs:
- name: Setup Cache
uses: actions/cache@v5
id: cache-toolchain
id: cache-openvino
with:
path: ./spacemit_toolchain
key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
path: ./openvino_toolkit
key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
- name: Setup SpacemiT Toolchain
if: steps.cache-toolchain.outputs.cache-hit != 'true'
uses: ./.github/actions/linux-setup-spacemit
- name: Setup OpenVINO Toolkit
if: steps.cache-openvino.outputs.cache-hit != 'true'
uses: ./.github/actions/linux-setup-openvino
with:
path: ./spacemit_toolchain
version: ${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
path: ./openvino_toolkit
version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
version_full: ${{ env.OPENVINO_VERSION_FULL }}
windows-2022-rocm-cache:
runs-on: windows-2022

102
.github/workflows/build-cann.yml vendored Normal file
View File

@ -0,0 +1,102 @@
name: CI (cann)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-cann.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp'
]
pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-cann.yml',
'ggml/src/ggml-cann/**'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
openEuler-latest-cann:
defaults:
run:
shell: bash -el {0}
strategy:
matrix:
arch: [x86, aarch64]
chip_type: ['910b', '310p']
build: ['Release']
use_acl_graph: ['on', 'off']
exclude:
# 310P does not support USE_ACL_GRAPH=on
- chip_type: '310p'
use_acl_graph: 'on'
runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
steps:
- name: Checkout
uses: actions/checkout@v6
with:
fetch-depth: 0
- name: Free up disk space
uses: ggml-org/free-disk-space@v1.3.1
with:
tool-cache: true
- name: Set container image
id: cann-image
run: |
image="ascendai/cann:${{ matrix.chip_type == '910b' && '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-openeuler24.03-py3.11' }}"
echo "image=${image}" >> "${GITHUB_OUTPUT}"
- name: Pull container image
run: docker pull "${{ steps.cann-image.outputs.image }}"
- name: Build
env:
BUILD_TYPE: ${{ matrix.build }}
SOC_TYPE: ascend${{ matrix.chip_type }}
USE_ACL_GRAPH: ${{ matrix.use_acl_graph }}
run: |
HOST_UID=$(id -u)
HOST_GID=$(id -g)
docker run --rm \
-v "${PWD}:/workspace" \
-w /workspace \
-e SOC_TYPE=${SOC_TYPE} \
-e BUILD_TYPE=${BUILD_TYPE} \
-e USE_ACL_GRAPH=${USE_ACL_GRAPH} \
"${{ steps.cann-image.outputs.image }}" \
bash -lc '
set -e
yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
yum clean all && rm -rf /var/cache/yum
git config --global --add safe.directory "/workspace"
export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
cmake -S . -B build \
-DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
-DGGML_CANN=on \
-DSOC_TYPE=${SOC_TYPE} \
-DUSE_ACL_GRAPH=${USE_ACL_GRAPH}
cmake --build build -j $(nproc)
chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
'

4
.github/workflows/build-cmake-pkg.yml vendored
View File

@ -5,7 +5,7 @@ on:
jobs:
linux:
runs-on: ubuntu-24.04
runs-on: ubuntu-slim
steps:
- uses: actions/checkout@v6
with:
@ -14,7 +14,7 @@ jobs:
- name: Install dependencies
run: |
sudo apt update
sudo apt install -y build-essential tcl
sudo apt install -y build-essential tcl cmake
- name: Build
run: |

39
.github/workflows/build-linux-cross.yml → .github/workflows/build-cross.yml vendored
View File

@ -1,7 +1,24 @@
name: Build on Linux using cross-compiler
name: CI (cross)
on:
# only manual triggers due to low-importance of the workflows
# TODO: for regular runs, provision dedicated self-hosted runners
workflow_dispatch:
workflow_call:
push:
branches:
- master
paths: [
'.github/workflows/build-cross.yml',
'ggml/src/spacemit/*',
'ggml/src/arch/loongarch/*'
]
# run once every week
schedule:
- cron: '0 0 * * 0'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
# ubuntu-24-riscv64-cpu-cross:
@ -142,7 +159,7 @@ jobs:
# cmake --build build --config Release -j $(nproc)
debian-13-loongarch64-cpu-cross:
runs-on: ubuntu-24.04
runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
steps:
@ -197,7 +214,7 @@ jobs:
cmake --build build --config Release -j $(nproc)
debian-13-loongarch64-vulkan-cross:
runs-on: ubuntu-24.04
runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
steps:
@ -264,15 +281,15 @@ jobs:
steps:
- uses: actions/checkout@v6
- name: Use SpacemiT Toolchain Cache
uses: actions/cache@v5
id: cache-toolchain
with:
path: ./spacemit_toolchain
key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
#- name: Use SpacemiT Toolchain Cache
# uses: actions/cache@v5
# id: cache-toolchain
# with:
# path: ./spacemit_toolchain
# key: spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}-${{ runner.os }}
- name: Setup SpacemiT Toolchain
if: steps.cache-toolchain.outputs.cache-hit != 'true'
#if: steps.cache-toolchain.outputs.cache-hit != 'true'
uses: ./.github/actions/linux-setup-spacemit
with:
path: ./spacemit_toolchain

72
.github/workflows/build-msys.yml vendored Normal file
View File

@ -0,0 +1,72 @@
name: CI (msys)
on:
# only manual triggers due to low-importance of the workflows
# TODO: for regular runs, provision dedicated self-hosted runners
workflow_dispatch:
# run once every week
schedule:
- cron: '0 0 * * 0'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
windows-msys2:
runs-on: windows-2025
strategy:
fail-fast: false
matrix:
include:
- { sys: UCRT64, env: ucrt-x86_64, build: Release }
- { sys: CLANG64, env: clang-x86_64, build: Release }
steps:
- name: Clone
uses: actions/checkout@v6
#- name: ccache
# uses: ggml-org/ccache-action@v1.2.16
# with:
# key: windows-msys2
# variant: ccache
# evict-old-files: 1d
# save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
- name: Setup ${{ matrix.sys }}
uses: msys2/setup-msys2@v2
with:
update: true
msystem: ${{matrix.sys}}
install: >-
base-devel
git
mingw-w64-${{matrix.env}}-toolchain
mingw-w64-${{matrix.env}}-cmake
mingw-w64-${{matrix.env}}-openblas
- name: Build using CMake
shell: msys2 {0}
run: |
cmake -B build
cmake --build build --config ${{ matrix.build }} -j $(nproc)
- name: Clean after building using CMake
shell: msys2 {0}
run: |
rm -rf build
- name: Build using CMake w/ OpenBLAS
shell: msys2 {0}
run: |
cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
cmake --build build --config ${{ matrix.build }} -j $(nproc)

136
.github/workflows/build-riscv.yml vendored Normal file
View File

@ -0,0 +1,136 @@
name: CI (riscv)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-riscv.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp'
]
pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-riscv.yml',
'ggml/src/ggml-cpu/arch/riscv/**'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
ubuntu-riscv64-native-sanitizer:
runs-on: RISCV64
continue-on-error: true
strategy:
matrix:
sanitizer: [ADDRESS, THREAD, UNDEFINED]
build_type: [Debug]
steps:
- name: Install dependencies
run: |
sudo apt-get update
# Install necessary packages
sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 rustup cmake build-essential wget ccache git-lfs
# Set gcc-14 and g++-14 as the default compilers
sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
sudo ln -sf /usr/bin/gcc-14 /usr/bin/gcc
sudo ln -sf /usr/bin/g++-14 /usr/bin/g++
# Install Rust stable version
rustup install stable
rustup default stable
git lfs install
- name: GCC version check
run: |
gcc --version
g++ --version
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Setup ccache
run: |
# Unique cache directory per matrix combination
export CCACHE_DIR="$HOME/.ccache/sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}"
mkdir -p "$CCACHE_DIR"
# Configure ccache
ccache --set-config=max_size=5G
ccache --set-config=compression=true
ccache --set-config=compression_level=6
ccache --set-config=cache_dir="$CCACHE_DIR"
ccache --set-config=sloppiness=file_macro,time_macros,include_file_mtime,include_file_ctime
ccache --set-config=hash_dir=false
# Export for subsequent steps
echo "CCACHE_DIR=$CCACHE_DIR" >> $GITHUB_ENV
echo "PATH=/usr/lib/ccache:$PATH" >> $GITHUB_ENV
- name: Build
id: cmake_build
if: ${{ matrix.sanitizer != 'THREAD' }}
run: |
cmake -B build \
-DLLAMA_OPENSSL=OFF \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DGGML_OPENMP=ON \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TOOLS=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_C_COMPILER_LAUNCHER=ccache \
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
- name: Build (no OpenMP)
id: cmake_build_no_openmp
if: ${{ matrix.sanitizer == 'THREAD' }}
run: |
cmake -B build \
-DLLAMA_OPENSSL=OFF \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TOOLS=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_C_COMPILER_LAUNCHER=ccache \
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
- name: Test
id: cmake_test
run: |
cd build
ctest -L main --verbose --timeout 900

87
.github/workflows/build-sanitize.yml vendored Normal file
View File

@ -0,0 +1,87 @@
name: CI (sanitize)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-sanitize.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
ubuntu-latest-sanitizer:
runs-on: ubuntu-latest
continue-on-error: true
strategy:
matrix:
sanitizer: [ADDRESS, THREAD, UNDEFINED]
build_type: [Debug]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.21
with:
key: ubuntu-latest-sanitizer-${{ matrix.sanitizer }}
evict-old-files: 1d
save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libssl-dev
- name: Build
id: cmake_build
if: ${{ matrix.sanitizer != 'THREAD' }}
run: |
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
- name: Build (no OpenMP)
id: cmake_build_no_openmp
if: ${{ matrix.sanitizer == 'THREAD' }}
run: |
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DGGML_OPENMP=OFF
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
- name: Test
id: cmake_test
run: |
cd build
ctest -L main --verbose --timeout 900

245
.github/workflows/build-self-hosted.yml vendored Normal file
View File

@ -0,0 +1,245 @@
name: CI (self-hosted)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.cuh',
'**/*.swift',
'**/*.m',
'**/*.metal',
'**/*.comp',
'**/*.glsl',
'**/*.wgsl'
]
pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-self-hosted.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.cuh',
'**/*.swift',
'**/*.m',
'**/*.metal',
'**/*.comp',
'**/*.glsl',
'**/*.wgsl'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
ggml-ci-nvidia-cuda:
runs-on: [self-hosted, Linux, NVIDIA]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
nvidia-smi
GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
ggml-ci-nvidia-vulkan-cm:
runs-on: [self-hosted, Linux, NVIDIA]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
ggml-ci-nvidia-vulkan-cm2:
runs-on: [self-hosted, Linux, NVIDIA, COOPMAT2]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
# TODO: provision AMX-compatible machine
#ggml-ci-cpu-amx:
# runs-on: [self-hosted, Linux, CPU, AMX]
# steps:
# - name: Clone
# id: checkout
# uses: actions/checkout@v6
# - name: Test
# id: ggml-ci
# run: |
# bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
# TODO: provision AMD GPU machine
# ggml-ci-amd-vulkan:
# runs-on: [self-hosted, Linux, AMD]
# steps:
# - name: Clone
# id: checkout
# uses: actions/checkout@v6
# - name: Test
# id: ggml-ci
# run: |
# vulkaninfo --summary
# GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
# TODO: provision AMD GPU machine
# ggml-ci-amd-rocm:
# runs-on: [self-hosted, Linux, AMD]
# steps:
# - name: Clone
# id: checkout
# uses: actions/checkout@v6
# - name: Test
# id: ggml-ci
# run: |
# amd-smi static
# GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
ggml-ci-mac-metal:
runs-on: [self-hosted, macOS, ARM64]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
ggml-ci-mac-webgpu:
runs-on: [self-hosted, macOS, ARM64]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Dawn Dependency
id: dawn-depends
run: |
DAWN_VERSION="v2.0.0"
DAWN_OWNER="reeselevine"
DAWN_REPO="dawn"
DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
curl -L -o artifact.zip \
"https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
mkdir dawn
unzip artifact.zip
tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
- name: Test
id: ggml-ci
run: |
GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
ggml-ci-mac-vulkan:
runs-on: [self-hosted, macOS, ARM64]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
ggml-ci-linux-intel-vulkan:
runs-on: [self-hosted, Linux, Intel]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
with:
persist-credentials: false
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
ggml-ci-intel-openvino-gpu-low-perf:
runs-on: [self-hosted, Linux, Intel, OpenVINO]
env:
# Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
OPENVINO_VERSION_MAJOR: "2026.0"
OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: Setup OpenVINO Toolkit
uses: ./.github/actions/linux-setup-openvino
with:
path: ./openvino_toolkit
version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
version_full: ${{ env.OPENVINO_VERSION_FULL }}
- name: Install OpenVINO dependencies
run: |
cd ./openvino_toolkit
chmod +x ./install_dependencies/install_openvino_dependencies.sh
echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
- name: Test
id: ggml-ci
run: |
source ./openvino_toolkit/setupvars.sh
GG_BUILD_OPENVINO=1 GGML_OPENVINO_DEVICE=GPU GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

96
.github/workflows/build-vulkan.yml vendored Normal file
View File

@ -0,0 +1,96 @@
name: CI (vulkan)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-vulkan.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.comp',
'**/*.glsl'
]
pull_request:
types: [opened, synchronize, reopened]
paths: [
'.github/workflows/build-vulkan.yml',
'ggml/src/ggml-vulkan/**'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
ubuntu-24-vulkan-llvmpipe:
runs-on: ubuntu-24.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.21
with:
key: ubuntu-24-vulkan-llvmpipe
evict-old-files: 1d
save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
- name: Dependencies
id: depends
run: |
sudo add-apt-repository -y ppa:kisak/kisak-mesa
sudo apt-get update -y
sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
- name: Get latest Vulkan SDK version
id: vulkan_sdk_version
run: |
echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
- name: Use Vulkan SDK Cache
uses: actions/cache@v5
id: cache-sdk
with:
path: ./vulkan_sdk
key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
- name: Setup Vulkan SDK
if: steps.cache-sdk.outputs.cache-hit != 'true'
uses: ./.github/actions/linux-setup-vulkan-llvmpipe
with:
path: ./vulkan_sdk
version: ${{ env.VULKAN_SDK_VERSION }}
- name: Build
id: cmake_build
run: |
source ./vulkan_sdk/setup-env.sh
cmake -B build \
-DGGML_VULKAN=ON
cmake --build build --config Release -j $(nproc)
- name: Test
id: cmake_test
run: |
cd build
export GGML_VK_VISIBLE_DEVICES=0
export GGML_VK_DISABLE_F16=1
export GGML_VK_DISABLE_COOPMAT=1
# This is using llvmpipe and runs slower than other backends
ctest -L main --verbose --timeout 4800

1388
.github/workflows/build.yml vendored
View File

File diff suppressed because it is too large Load Diff

4
.github/workflows/copilot-setup-steps.yml vendored
View File

@ -29,7 +29,7 @@ jobs:
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: copilot-setup-steps
evict-old-files: 1d
@ -52,6 +52,6 @@ jobs:
- name: Install Python dependencies
run: |
python3 -m venv .venv
.venv/bin/activate
source .venv/bin/activate
pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
pip install flake8 pyright pre-commit

1
.github/workflows/docker.yml vendored
View File

@ -47,6 +47,7 @@ jobs:
- { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
- { tag: "s390x", dockerfile: ".devops/s390x.Dockerfile", platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04-s390x" }
- { tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true, runs_on: "ubuntu-22.04" }
- { tag: "openvino", dockerfile: ".devops/openvino.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
steps:
- name: Check out the repo
uses: actions/checkout@v6

10
.github/workflows/python-lint.yml vendored
View File

@ -4,10 +4,16 @@ on:
push:
branches:
- master
paths: ['.github/workflows/python-lint.yml', '**/*.py']
paths: [
'.github/workflows/python-lint.yml',
'**/*.py'
]
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/python-lint.yml', '**/*.py']
paths: [
'.github/workflows/python-lint.yml',
'**/*.py'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

140
.github/workflows/release.yml vendored
View File

@ -10,7 +10,22 @@ on:
push:
branches:
- master
paths: ['.github/workflows/release.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
paths: [
'.github/workflows/release.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.cuh',
'**/*.swift',
'**/*.m',
'**/*.metal',
'**/*.comp',
'**/*.glsl'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@ -32,9 +47,9 @@ jobs:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: macOS-latest-cmake-arm64
key: macOS-latest-arm64
evict-old-files: 1d
- name: Build
@ -79,9 +94,9 @@ jobs:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: macOS-latest-cmake-x64
key: macOS-latest-x64
evict-old-files: 1d
- name: Build
@ -138,9 +153,10 @@ jobs:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
if: ${{ matrix.build != 's390x' }}
uses: ggml-org/ccache-action@v1.2.21
with:
key: ubuntu-cpu-cmake-${{ matrix.build }}
key: ubuntu-cpu-${{ matrix.build }}
evict-old-files: 1d
- name: Dependencies
@ -189,9 +205,9 @@ jobs:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: ubuntu-22-cmake-vulkan
key: ubuntu-22-vulkan
evict-old-files: 1d
- name: Dependencies
@ -231,6 +247,86 @@ jobs:
path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz
name: llama-bin-ubuntu-vulkan-x64.tar.gz
ubuntu-24-openvino:
runs-on: ubuntu-24.04
outputs:
openvino_version: ${{ steps.openvino_version.outputs.value }}
env:
# Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
OPENVINO_VERSION_MAJOR: "2026.0"
OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
steps:
- name: Set OpenVINO version output
id: openvino_version
run: echo "value=${{ env.OPENVINO_VERSION_MAJOR }}" >> $GITHUB_OUTPUT
- name: Clone
id: checkout
uses: actions/checkout@v6
with:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.21
with:
key: ubuntu-24-openvino-release-no-preset-v1
evict-old-files: 1d
- name: Dependencies
run: |
sudo apt-get update
sudo apt-get install -y build-essential libssl-dev libtbb12 cmake ninja-build python3-pip
sudo apt install ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd
- name: Use OpenVINO Toolkit Cache
uses: actions/cache@v5
id: cache-openvino
with:
path: ./openvino_toolkit
key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
- name: Setup OpenVINO Toolkit
if: steps.cache-openvino.outputs.cache-hit != 'true'
uses: ./.github/actions/linux-setup-openvino
with:
path: ./openvino_toolkit
version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
version_full: ${{ env.OPENVINO_VERSION_FULL }}
- name: Install OpenVINO dependencies
run: |
cd ./openvino_toolkit
chmod +x ./install_dependencies/install_openvino_dependencies.sh
echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
- name: Build
id: cmake_build
run: |
source ./openvino_toolkit/setupvars.sh
cmake -B build/ReleaseOV -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_OPENVINO=ON
cmake --build build/ReleaseOV --config Release -j $(nproc)
- name: Determine tag name
id: tag
uses: ./.github/actions/get-tag-name
- name: Pack artifacts
id: pack_artifacts
run: |
cp LICENSE ./build/ReleaseOV/bin/
tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/ReleaseOV/bin .
- name: Upload artifacts
uses: actions/upload-artifact@v6
with:
path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
windows-cpu:
runs-on: windows-2025
@ -247,9 +343,9 @@ jobs:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: windows-latest-cmake-cpu-${{ matrix.arch }}
key: windows-latest-cpu-${{ matrix.arch }}
variant: ccache
evict-old-files: 1d
@ -308,9 +404,9 @@ jobs:
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
key: windows-latest-${{ matrix.backend }}-${{ matrix.arch }}
variant: ccache
evict-old-files: 1d
@ -378,7 +474,7 @@ jobs:
uses: actions/checkout@v6
- name: Install ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: windows-cuda-${{ matrix.cuda }}
variant: ccache
@ -454,9 +550,9 @@ jobs:
uses: actions/checkout@v6
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: windows-latest-cmake-sycl
key: windows-latest-sycl
variant: ccache
evict-old-files: 1d
@ -534,9 +630,9 @@ jobs:
fetch-depth: 0
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
key: ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
evict-old-files: 1d
- name: Dependencies
@ -644,9 +740,9 @@ jobs:
key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
uses: ggml-org/ccache-action@v1.2.21
with:
key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}-x64
key: windows-latest-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}-x64
evict-old-files: 1d
- name: Install ROCm
@ -872,7 +968,7 @@ jobs:
permissions:
contents: write # for creating release
runs-on: ubuntu-latest
runs-on: ubuntu-slim
needs:
- windows
@ -883,6 +979,7 @@ jobs:
- ubuntu-22-rocm
- ubuntu-22-cpu
- ubuntu-22-vulkan
- ubuntu-24-openvino
- macOS-arm64
- macOS-x64
- ios-xcode-build
@ -967,6 +1064,7 @@ jobs:
- [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
- [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
- [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
- [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
**Windows:**
- [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)

105
.github/workflows/server-sanitize.yml vendored Normal file
View File

@ -0,0 +1,105 @@
name: Server (sanitize)
on:
workflow_dispatch: # allows manual triggering
inputs:
sha:
description: 'Commit SHA1 to build'
required: false
type: string
slow_tests:
description: 'Run slow tests'
required: true
type: boolean
push:
branches:
- master
paths: [
'.github/workflows/server-sanitize.yml',
'**/CMakeLists.txt',
'**/Makefile',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'tools/server/**.*'
]
env:
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
LLAMA_LOG_VERBOSITY: 10
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
cancel-in-progress: true
jobs:
server:
runs-on: ubuntu-latest
strategy:
matrix:
sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
build_type: [RelWithDebInfo]
fail-fast: false
steps:
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get -y install \
build-essential \
xxd \
git \
cmake \
curl \
wget \
language-pack-en \
libssl-dev
- name: Clone
id: checkout
uses: actions/checkout@v6
with:
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: Build
id: cmake_build
run: |
cmake -B build \
-DLLAMA_BUILD_BORINGSSL=ON \
-DGGML_SCHED_NO_REALLOC=ON \
-DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
-DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
- name: Python setup
id: setup_python
uses: actions/setup-python@v6
with:
python-version: '3.11'
pip-install: -r tools/server/tests/requirements.txt
- name: Tests
id: server_integration_tests
if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
run: |
cd tools/server/tests
export ${{ matrix.extra_args }}
pytest -v -x -m "not slow"
- name: Slow tests
id: server_integration_tests_slow
if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
run: |
cd tools/server/tests
export ${{ matrix.extra_args }}
SLOW_TESTS=1 pytest -v -x

57
.github/workflows/server-metal.yml → .github/workflows/server-self-hosted.yml vendored
View File

@ -1,4 +1,4 @@
name: Server-Metal
name: Server (self-hosted)
on:
workflow_dispatch: # allows manual triggering
@ -14,7 +14,19 @@ on:
push:
branches:
- master
paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
paths: [
'.github/workflows/server-self-hosted.yml',
'**/CMakeLists.txt',
'**/Makefile',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.swift',
'**/*.m',
'tools/server/**.*'
]
env:
LLAMA_LOG_COLORS: 1
@ -28,7 +40,7 @@ concurrency:
jobs:
server-metal:
runs-on: [self-hosted, macOS, ARM64]
runs-on: [self-hosted, llama-server, macOS, ARM64]
name: server-metal (${{ matrix.wf_name }})
strategy:
@ -71,3 +83,42 @@ jobs:
pip install -r requirements.txt
export ${{ matrix.extra_args }}
pytest -v -x -m "not slow"
server-cuda:
runs-on: [self-hosted, llama-server, Linux, NVIDIA]
name: server-cuda (${{ matrix.wf_name }})
strategy:
matrix:
build_type: [Release]
wf_name: ["GPUx1"]
include:
- build_type: Release
extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
wf_name: "GPUx1, backend-sampling"
fail-fast: false
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
with:
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: Build
id: cmake_build
run: |
cmake -B build -DGGML_SCHED_NO_REALLOC=ON
cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
- name: Tests
id: server_integration_tests
if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
run: |
cd tools/server/tests
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
export ${{ matrix.extra_args }}
pytest -v -x -m "not slow"

17
.github/workflows/server-webui.yml vendored
View File

@ -1,4 +1,3 @@
# Server WebUI build and tests
name: Server WebUI
on:
@ -11,10 +10,20 @@ on:
push:
branches:
- master
paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
paths: [
'.github/workflows/server-webui.yml',
'tools/server/webui/**.*',
'tools/server/tests/**.*',
'tools/server/public/**'
]
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
paths: [
'.github/workflows/server-webui.yml',
'tools/server/webui/**.*',
'tools/server/tests/**.*',
'tools/server/public/**'
]
env:
LLAMA_LOG_COLORS: 1
@ -29,7 +38,7 @@ concurrency:
jobs:
webui-check:
name: WebUI Checks
runs-on: ubuntu-latest
runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
continue-on-error: true
steps:
- name: Checkout code

46
.github/workflows/server.yml vendored
View File

@ -1,4 +1,3 @@
# Server build and tests
name: Server
on:
@ -15,10 +14,34 @@ on:
push:
branches:
- master
paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
paths: [
'.github/workflows/server.yml',
'**/CMakeLists.txt',
'**/Makefile',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.swift',
'**/*.m',
'tools/server/**.*'
]
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
paths: [
'.github/workflows/server.yml',
'**/CMakeLists.txt',
'**/Makefile',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.swift',
'**/*.m',
'tools/server/**.*'
]
env:
LLAMA_LOG_COLORS: 1
@ -34,17 +57,18 @@ jobs:
server:
runs-on: ubuntu-latest
name: server (${{ matrix.wf_name }})
strategy:
matrix:
sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
build_type: [RelWithDebInfo]
build_type: [Release]
wf_name: ["default"]
include:
- build_type: Release
sanitizer: ""
extra_args: ""
wf_name: "default"
- build_type: Release
sanitizer: ""
extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
wf_name: "backend-sampling"
fail-fast: false
steps:
@ -74,13 +98,7 @@ jobs:
run: |
cmake -B build \
-DLLAMA_BUILD_BORINGSSL=ON \
-DGGML_SCHED_NO_REALLOC=ON \
-DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
-DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
-DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
-DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
-DGGML_SCHED_NO_REALLOC=ON
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
- name: Python setup

5
.gitignore vendored
View File

@ -124,6 +124,11 @@ poetry.toml
# Scripts
!/scripts/install-oneapi.bat
# Generated by scripts
/hellaswag_val_full.txt
/winogrande-debiased-eval.csv
/wikitext-2-raw/
# Test models for lora adapters
/lora-tests

57
CODEOWNERS
View File

@ -2,29 +2,13 @@
# multiplie collaborators per item can be specified
/.devops/*.Dockerfile @ngxson
/.github/actions/ @CISC
/.github/workflows/ @CISC
/.github/actions/ @ggml-org/ci
/.github/workflows/ @ggml-org/ci
/ci/ @ggerganov
/cmake/ @ggerganov
/common/CMakeLists.txt @ggerganov
/common/arg.* @ggerganov
/common/base64.hpp.* @ggerganov
/common/build-info.* @ggerganov
/common/chat.* @pwilkin
/common/chat-auto*.* @pwilkin
/common/chat-diff-analyzer.* @pwilkin
/common/chat-peg-parser.* @aldehir
/common/common.* @ggerganov
/common/console.* @ggerganov
/common/http.* @angt
/common/jinja/ @ngxson @CISC @aldehir
/common/llguidance.* @ggerganov
/common/log.* @ggerganov
/common/ @ggml-org/llama-common
/common/jinja/ @CISC
/common/ngram-map.* @srogmann
/common/peg-parser.* @aldehir
/common/sampling.* @ggerganov
/common/speculative.* @ggerganov
/common/unicode.* @aldehir
/convert_*.py @CISC
/examples/batched.swift/ @ggerganov
/examples/batched/ @ggerganov
@ -51,29 +35,28 @@
/examples/speculative/ @ggerganov
/ggml/cmake/ @ggerganov
/ggml/include/ @ggerganov
/ggml/src/ggml-cann/ @ggml-org/ggml-cann
/ggml/src/ggml-common.h @ggerganov
/ggml/src/ggml-cpu/ @ggerganov
/ggml/src/ggml-cpu/spacemit/ @alex-spacemit
/ggml/src/ggml-cuda/fattn* @JohannesGaessler
/ggml/src/ggml-cuda/mmf.* @JohannesGaessler @am17an
/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
/ggml/src/ggml-cuda/mmvf.* @JohannesGaessler
/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
/ggml/src/ggml-cuda/ @ggml-org/ggml-cuda
/ggml/src/ggml-cuda/fattn-wmma* @IMbackK
/ggml/src/ggml-hip/ @IMbackK
/ggml/src/ggml-cuda/vendors/hip.h @IMbackK
/ggml/src/ggml-impl.h @ggerganov
/ggml/src/ggml-metal/ @ggerganov
/ggml/src/ggml-opencl/ @lhez @max-krasnyansky
/ggml/src/ggml-hexagon/ @max-krasnyansky @lhez
/ggml/src/ggml-metal/ @ggml-org/ggml-metal
/ggml/src/ggml-opencl/ @ggml-org/ggml-opencl
/ggml/src/ggml-hexagon/ @ggml-org/ggml-hexagon
/ggml/src/ggml-opt.cpp @JohannesGaessler
/ggml/src/ggml-quants.* @ggerganov
/ggml/src/ggml-rpc/ @rgerganov
/ggml/src/ggml-rpc/ @ggml-org/ggml-rpc
/ggml/src/ggml-sycl/ @ggml-org/ggml-sycl
/ggml/src/ggml-threading.* @ggerganov
/ggml/src/ggml-vulkan/ @0cc4m
/ggml/src/ggml-vulkan/ @ggml-org/ggml-vulkan
/ggml/src/ggml-virtgpu/ @kpouget
/ggml/src/ggml-webgpu/ @reeselevine
/ggml/src/ggml-zdnn/ @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
/ggml/src/ggml-webgpu/ @ggml-org/ggml-webgpu
/ggml/src/ggml-zdnn/ @ggml-org/ggml-zdnn @Andreas-Krebbel @AlekseiNikiforovIBM
/ggml/src/ggml-openvino/ @cavusmustafa @wine99
/ggml/src/ggml.c @ggerganov
/ggml/src/ggml.cpp @ggerganov
/ggml/src/gguf.cpp @JohannesGaessler @Green-Sky
@ -92,16 +75,18 @@
/src/models/ @CISC
/tests/ @ggerganov
/tests/test-chat.* @pwilkin
/tests/test-llama-archs.cpp @JohannesGaessler
/tools/batched-bench/ @ggerganov
/tools/cli/ @ngxson
/tools/completion/ @ggerganov
/tools/mtmd/ @ngxson
/tools/mtmd/ @ggml-org/llama-mtmd
/tools/perplexity/ @ggerganov
/tools/parser/ @pwilkin
/tools/quantize/ @ggerganov
/tools/rpc/ @rgerganov
/tools/server/* @ngxson @ggerganov # no subdir
/tools/server/webui/ @allozaur
/tools/rpc/ @ggml-org/ggml-rpc
/tools/server/* @ggml-org/llama-server # no subdir
/tools/server/tests/ @ggml-org/llama-server
/tools/server/webui/ @ggml-org/llama-webui
/tools/tokenize/ @ggerganov
/tools/tts/ @ggerganov
/vendor/ @ggerganov

1
README.md
View File

@ -279,6 +279,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
| [BLAS](docs/build.md#blas-build) | All |
| [BLIS](docs/backend/BLIS.md) | All |
| [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU |
| [OpenVINO [In Progress]](docs/backend/OPENVINO.md) | Intel CPUs, GPUs, and NPUs |
| [MUSA](docs/build.md#musa) | Moore Threads GPU |
| [CUDA](docs/build.md#cuda) | Nvidia GPU |
| [HIP](docs/build.md#hip) | AMD GPU |

51
benches/nemotron/nemotron-dgx-spark.md
View File

@ -24,9 +24,9 @@ Fri Mar 6 11:39:45 2026
+-----------------------------------------+------------------------+----------------------+
```
## ggml-org/nemotron-3-super-120b-GGUF
## ggml-org/Nemotron-3-Super-120B-GGUF
Model: https://huggingface.co/ggml-org/nemotron-3-super-120b-GGUF
Model: https://huggingface.co/ggml-org/Nemotron-3-Super-120B-GGUF
- `llama-batched-bench`
@ -53,7 +53,6 @@ main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_
| 8192 | 32 | 16 | 131584 | 171.066 | 766.21 | 10.774 | 47.52 | 181.840 | 723.62 |
| 8192 | 32 | 32 | 263168 | 342.140 | 766.19 | 18.969 | 53.98 | 361.109 | 728.78 |
- `llama-bench`
| model | size | params | backend | n_ubatch | fa | test | t/s |
@ -70,3 +69,49 @@ main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_
| nemotron 120B.A12B Q4_K | 65.10 GiB | 120.67 B | CUDA | 2048 | 1 | tg32 @ d32768 | 19.45 ± 0.18 |
build: 04a65daab (8268)
## ggml-org/Nemotron-3-Nano-4B-GGUF
Model: https://huggingface.co/ggml-org/Nemotron-3-Nano-4B-GGUF
- `llama-batched-bench`
main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = 99, n_threads = 20, n_threads_batch = 20
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
| 512 | 32 | 1 | 544 | 0.152 | 3371.61 | 0.597 | 53.64 | 0.748 | 726.90 |
| 512 | 32 | 2 | 1088 | 0.319 | 3208.68 | 0.857 | 74.66 | 1.176 | 924.89 |
| 512 | 32 | 4 | 2176 | 0.720 | 2843.56 | 1.323 | 96.78 | 2.043 | 1065.18 |
| 512 | 32 | 8 | 4352 | 1.428 | 2867.96 | 2.311 | 110.76 | 3.739 | 1163.82 |
| 512 | 32 | 16 | 8704 | 2.857 | 2866.94 | 4.203 | 121.82 | 7.060 | 1232.82 |
| 512 | 32 | 32 | 17408 | 5.709 | 2869.76 | 7.964 | 128.58 | 13.673 | 1273.14 |
| 4096 | 32 | 1 | 4128 | 1.458 | 2809.76 | 0.605 | 52.92 | 2.062 | 2001.52 |
| 4096 | 32 | 2 | 8256 | 2.905 | 2819.95 | 0.875 | 73.12 | 3.780 | 2183.95 |
| 4096 | 32 | 4 | 16512 | 5.790 | 2829.74 | 1.361 | 94.07 | 7.151 | 2309.17 |
| 4096 | 32 | 8 | 33024 | 11.598 | 2825.32 | 2.378 | 107.65 | 13.976 | 2362.89 |
| 4096 | 32 | 16 | 66048 | 23.208 | 2823.88 | 4.348 | 117.76 | 27.556 | 2396.89 |
| 4096 | 32 | 32 | 132096 | 46.515 | 2817.85 | 8.279 | 123.69 | 54.794 | 2410.79 |
| 8192 | 32 | 1 | 8224 | 2.950 | 2776.95 | 0.617 | 51.89 | 3.567 | 2305.75 |
| 8192 | 32 | 2 | 16448 | 5.921 | 2767.32 | 0.896 | 71.45 | 6.816 | 2413.05 |
| 8192 | 32 | 4 | 32896 | 11.842 | 2767.21 | 1.401 | 91.34 | 13.243 | 2484.03 |
| 8192 | 32 | 8 | 65792 | 23.726 | 2762.17 | 2.461 | 104.03 | 26.187 | 2512.38 |
| 8192 | 32 | 16 | 131584 | 47.777 | 2743.43 | 4.577 | 111.86 | 52.354 | 2513.36 |
| 8192 | 32 | 32 | 263168 | 96.691 | 2711.16 | 8.772 | 116.73 | 105.463 | 2495.36 |
- `llama-bench`
| model | size | params | backend | n_ubatch | fa | test | t/s |
| ----------------------- | ---------: | ---------: | ---------- | -------: | -: | --------------: | -------------------: |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | pp2048 | 2761.90 ± 19.31 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | tg32 | 52.85 ± 0.12 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | pp2048 @ d4096 | 2687.07 ± 21.84 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | tg32 @ d4096 | 52.32 ± 0.23 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | pp2048 @ d8192 | 2564.52 ± 57.69 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | tg32 @ d8192 | 51.27 ± 0.34 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | pp2048 @ d16384 | 2334.02 ± 37.83 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | tg32 @ d16384 | 49.71 ± 0.14 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | pp2048 @ d32768 | 2041.46 ± 40.45 |
| nemotron 4B Q8_0 | 3.94 GiB | 3.97 B | CUDA | 2048 | 1 | tg32 @ d32768 | 46.71 ± 0.13 |
build: 1bbec6a75 (8382)

32
ci/run.sh
View File

@ -25,6 +25,15 @@
# # with KLEIDIAI support
# GG_BUILD_KLEIDIAI=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
# # with BLAS support
# GG_BUILD_BLAS=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
# with BLAS support (custom vendor)
# GG_BUILD_BLAS=1 GG_BUILD_BLAS_VENDOR=Intel10_64lp bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
# with OPENVINO support
# GG_BUILD_OPENVINO=1 GG_BUILD_LOW_PERF=1 GGML_OPENVINO_DEVICE=CPU bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
if [ -z "$2" ]; then
echo "usage: $0 <output-dir> <mnt-dir>"
@ -46,6 +55,7 @@ cd $sd/../
SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=${LLAMA_FATAL_WARNINGS:-ON} -DLLAMA_OPENSSL=OFF -DGGML_SCHED_NO_REALLOC=ON"
CTEST_EXTRA=""
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
@ -165,6 +175,22 @@ if [ -n "${GG_BUILD_KLEIDIAI}" ]; then
-DBUILD_SHARED_LIBS=OFF"
fi
if [ ! -z ${GG_BUILD_BLAS} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=${GG_BUILD_BLAS_VENDOR:-OpenBLAS}"
fi
if [ ! -z ${GG_BUILD_OPENVINO} ]; then
if [ -z ${OpenVINO_DIR} ]; then
echo "OpenVINO_DIR not found, please install OpenVINO via archives and enable it by:"
echo "source /opt/intel/openvino/setupvars.sh"
exit 1
fi
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"
# TODO: fix and re-enable the `test-llama-archs` test below
CTEST_EXTRA="-E test-llama-archs"
fi
## helpers
# download a file if it does not exist or if it is outdated
@ -222,7 +248,7 @@ function gg_run_ctest_debug {
(time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
(time ctest --output-on-failure -L main -E "test-opt|test-backend-ops" ) 2>&1 | tee -a $OUT/${ci}-ctest.log
(time ctest --output-on-failure -L main -E "test-opt|test-backend-ops" ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
set +e
}
@ -254,9 +280,9 @@ function gg_run_ctest_release {
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
if [ -z ${GG_BUILD_LOW_PERF} ]; then
(time ctest --output-on-failure -L 'main|python' ) 2>&1 | tee -a $OUT/${ci}-ctest.log
(time ctest --output-on-failure -L 'main|python' ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
else
(time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
(time ctest --output-on-failure -L main -E test-opt ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
fi
set +e

11
common/arg.cpp
View File

@ -3116,6 +3116,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.chat_template = read_file(value);
}
).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
add_opt(common_arg(
{"--skip-chat-parsing"},
{"--no-skip-chat-parsing"},
string_format(
"force a pure content parser, even if a Jinja template is specified; model will output everything "
"in the content section, including any reasoning and/or tool calls (default: disabled)"
),
[](common_params & params, bool value) {
params.force_pure_content_parser = value;
}
).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SKIP_CHAT_PARSING"));
add_opt(common_arg(
{"--prefill-assistant"},
{"--no-prefill-assistant"},

6
common/chat-auto-parser-generator.cpp
View File

@ -3,6 +3,7 @@
#include "chat.h"
#include "common.h"
#include "json-schema-to-grammar.h"
#include "log.h"
#include "nlohmann/json.hpp"
#include <stdexcept>
@ -182,7 +183,10 @@ common_peg_parser analyze_tools::build_parser(parser_build_context & ctx) const
case tool_format::TAG_WITH_TAGGED:
return build_tool_parser_tag_tagged(ctx);
default:
GGML_ABORT("Unable to create tool parser");
LOG_ERR("[ERROR] Template seems to support tool calls, but failed to determine tool format. Tool calling will not work properly. "
"Check for a fixed template for your model in the models/templates directory of your llama.cpp installation or "
"report an issue at https://github.com/ggml-org/llama.cpp/issues\n");
return ctx.p.eps();
}
}

7
common/chat-diff-analyzer.cpp
View File

@ -479,6 +479,7 @@ analyze_content::analyze_content(const common_chat_template & tmpl, const analyz
if (!comparison_with_tools || !comparison_with_reasoning) {
LOG_DBG(ANSI_ORANGE "%s: Template application failed\n" ANSI_RESET, __func__);
return;
}
const auto & diff_tools = comparison_with_tools->diff;
@ -911,8 +912,10 @@ void analyze_tools::extract_function_markers() {
// we'll have to rely on an extra diff with no-calls version
auto notool_comp = compare_variants(
*tmpl, params, [&](template_params & p) { p.messages = json::array({ user_msg, assistant_nocall }); });
auto nt_diff = notool_comp->diff;
closer_suffix = nt_diff.left.substr(nt_diff.left.find("YYYY") + 4);
if (notool_comp) {
auto nt_diff = notool_comp->diff;
closer_suffix = nt_diff.left.substr(nt_diff.left.find("YYYY") + 4);
}
} else {
closer_suffix = diff.suffix.substr(0, diff.suffix.find(suffix_marker));
}

129
common/chat.cpp
View File

@ -933,17 +933,12 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
// Copy reasoning to the "thinking" field as expected by the gpt-oss template
auto adjusted_messages = json::array();
for (const auto & msg : inputs.messages) {
auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string();
auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array();
if (has_reasoning_content && has_tool_calls) {
auto adjusted_message = msg;
adjusted_message["thinking"] = msg.at("reasoning_content");
adjusted_messages.push_back(adjusted_message);
} else {
adjusted_messages.push_back(msg);
for (auto msg : inputs.messages) {
if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
msg["thinking"] = msg.at("reasoning_content");
msg.erase("content");
}
adjusted_messages.push_back(msg);
}
auto prompt = common_chat_template_direct_apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
@ -969,45 +964,31 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
"<|channel|>", "<|constrain|>", "<|message|>", "<|start|>", "<|end|>",
};
auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
auto include_grammar = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && has_tools;
auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
auto include_grammar = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
const std::string END = "<|end|>";
const std::string START = "<|start|>";
const std::string MESSAGE = "<|message|>";
const std::string CHANNEL = "<|channel|>";
const std::string CONSTRAIN = "<|constrain|>";
const std::string START_ASSISTANT = START + "assistant";
const std::string CHANNEL_ANALYSIS = CHANNEL + "analysis";
const std::string CHANNEL_COMMENTARY = CHANNEL + "commentary";
const std::string CHANNEL_FINAL = CHANNEL + "final";
auto start = p.rule("start", p.literal("<|start|>assistant"));
auto end = p.rule("end", p.literal("<|end|>"));
auto content = p.rule("message-content", p.until("<|end|>"));
auto channel = p.literal("<|channel|>") + (p.literal("commentary") | p.literal("analysis"));
auto constrain_type = p.chars("[A-Za-z0-9_-]", 1, -1);
auto the_end = END | p.end();
auto analysis = p.rule("analysis", p.literal("<|channel|>analysis<|message|>") + p.reasoning(content) + end);
auto preamble = p.rule("preamble", p.literal("<|channel|>commentary<|message|>") + p.content(content) + end);
auto final_msg = p.rule("final", p.literal("<|channel|>final<|message|>") + p.content(content));
auto any = p.rule("any", preamble | analysis);
const std::string analysis_header = CHANNEL_ANALYSIS + MESSAGE;
auto segment_content = p.until(END);
auto analysis_segment = extract_reasoning ?
p.literal(analysis_header) + p.reasoning(segment_content) + p.until(END) + the_end :
p.content(analysis_header + p.until(END) + the_end);
if (has_response_format) {
auto constraint = p.optional(p.space() + p.literal("<|constrain|>") + constrain_type);
auto response_format = p.rule("response-format",
p.literal("<|channel|>final") + constraint + p.literal("<|message|>") +
p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
auto channel_header_content = p.until_one_of({ " to=functions.", MESSAGE });
auto content_header = p.choice({ p.literal(CHANNEL_COMMENTARY), p.literal(CHANNEL_FINAL) });
auto content_segment = p.rule("content-segment", content_header + channel_header_content + MESSAGE +
p.content(segment_content) + the_end);
if (!inputs.json_schema.is_null()) {
auto final_header = p.literal(CHANNEL_FINAL);
auto constraint = p.optional(p.space() + p.literal(CONSTRAIN) + channel_header_content);
return p.optional(analysis_segment) + final_header + constraint + MESSAGE +
p.content(p.schema(p.json(), "response-format", inputs.json_schema));
return response_format | (analysis + p.zero_or_more(start + analysis) + start + response_format);
}
auto segment = p.optional(START_ASSISTANT + p.space()) + p.choice({ content_segment, analysis_segment });
auto contents = p.optional(segment + p.repeat(p.optional(p.space()) + segment, 0, -1)) + p.end();
// Tool call parser
if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
auto tool_choice = p.choice();
@ -1016,42 +997,37 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
std::string name = function.at("name");
const auto & params = function.at("parameters");
// Tool call can appear as:
// 1. In role header: " to=functions.NAME<|channel|>..."
// 2. In channel: "<|channel|>(analysis|commentary) to=functions.NAME..."
auto func_name = p.literal(" to=functions.") + p.tool_name(p.literal(name));
auto channel = p.literal(CHANNEL_COMMENTARY) | p.literal(CHANNEL_ANALYSIS);
auto constraint = p.space() + p.optional(p.literal(CONSTRAIN) + channel_header_content);
auto func_name = p.literal(" to=functions.") + p.tool_name(p.literal(name));
auto constraint = p.optional(p.space() + p.literal("<|constrain|>") + constrain_type);
auto args = p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", params));
// Pattern 1: recipient in role header
// " to=functions.NAME<|channel|>(analysis|commentary)[constraint]<|message|>ARGS"
auto tool_in_role = p.tool(p.tool_open(func_name + channel) + constraint + MESSAGE + args);
// recipient in role header
// <|start|>assistant to=functions.NAME<|channel|>(commentary|analysis)[constraint]<|message|>ARGS
auto tool_in_role = p.tool(p.tool_open(func_name + channel + constraint + p.literal("<|message|>")) + args);
// Pattern 2: recipient in channel header
// "<|channel|>(analysis|commentary) to=functions.NAME[constraint]<|message|>ARGS"
auto tool_in_channel = p.tool(channel + p.tool_open(func_name + constraint + MESSAGE) + args);
// recipient in channel header
// <|channel|>(commentary|analysis) to=functions.NAME[constraint]<|message|>ARGS
auto tool_in_channel = p.tool(p.tool_open(channel + func_name + constraint + p.literal("<|message|>")) + args);
tool_choice |= tool_in_role | tool_in_channel;
tool_choice |= p.rule("tool-" + name, tool_in_role | tool_in_channel);
});
auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
auto max_calls = inputs.parallel_tool_calls ? -1 : 1;
auto tool_call = p.trigger_rule("tool-call", tool_choice);
auto role_start = p.optional(p.space() + p.literal(START_ASSISTANT));
auto tool_call = p.rule("tool-call", p.repeat(role_start + tool_choice, min_calls, max_calls) + p.end());
if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
return tool_call | ( any + p.zero_or_more(start + any) + start + tool_call);
}
return p.choice({ p.trigger_rule("single-tool", tool_call), p.trigger_rule("tools", p.one_or_more(segment) + tool_call) });
return tool_call | final_msg | (any + p.zero_or_more(start + any) + start + (tool_call | final_msg));
}
return contents;
return final_msg | (any + p.zero_or_more(start + any) + start + final_msg);
});
data.parser = parser.save();
if (include_grammar) {
data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool.at("function");
@ -1062,10 +1038,9 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
});
data.grammar_triggers = {
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^(?:<\\|start\\|>assistant\\s*)?(\\s+to=functions)" },
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "(?:<\\|end\\|>)(?:<\\|start\\|>assistant\\s*)?(\\s+to=functions)" },
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
"(?:<\\|start\\|>assistant\\s*)?(<\\|channel\\|>(?:commentary|analysis)\\s+to=functions)" }
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^\\s+to$" },
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>assistant(\\s+to)" },
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>assistant(<\\|channel\\|>(?:commentary|analysis)\\s+to)" }
};
}
@ -1519,7 +1494,6 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
// map developer to system for all models except for GPT-OSS
workaround::map_developer_role_to_system(params.messages);
}
workaround::func_args_not_string(params.messages);
if (!tmpl.original_caps().supports_system_role) {
workaround::system_message_not_supported(params.messages);
@ -1532,6 +1506,10 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
workaround::requires_non_null_content(params.messages);
}
if (tmpl.original_caps().supports_object_arguments) {
workaround::func_args_not_string(params.messages);
}
params.extra_context = common_chat_extra_context();
for (auto el : inputs.chat_template_kwargs) {
params.extra_context[el.first] = json::parse(el.second);
@ -1559,6 +1537,21 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
}
}
if (inputs.force_pure_content) {
LOG_WRN("Forcing pure content template, will not render reasoning or tools separately.");
// Create the result structure
common_chat_params data;
auto params_copy = params;
params_copy.reasoning_format = COMMON_REASONING_FORMAT_NONE;
data.prompt = common_chat_template_direct_apply(tmpl, params_copy);
data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
auto parser = build_chat_peg_parser([](common_chat_peg_builder &p) {
return p.content(p.rest());
});
data.parser = parser.save();
return data;
}
// Ministral/Mistral Large 3 - uses special reasoning structure fixes, can't use autoparser
// Note: Mistral Small 3.2 uses [CALL_ID] which Ministral doesn't have, so we can distinguish them
if (src.find("[SYSTEM_PROMPT]") != std::string::npos && src.find("[TOOL_CALLS]") != std::string::npos &&

1
common/chat.h
View File

@ -204,6 +204,7 @@ struct common_chat_templates_inputs {
std::map<std::string, std::string> chat_template_kwargs;
bool add_bos = false;
bool add_eos = false;
bool force_pure_content = false;
};
struct common_chat_params {

2
common/common.cpp
View File

@ -1067,7 +1067,7 @@ common_init_result::common_init_result(common_params & params) :
const llama_vocab * vocab = llama_model_get_vocab(model);
// load and optionally apply lora adapters (must be loaded before context creation)
// load and optionally apply lora adapters
for (auto & la : params.lora_adapters) {
llama_adapter_lora_ptr lora;
lora.reset(llama_adapter_lora_init(model, la.path.c_str()));

1
common/common.h
View File

@ -544,6 +544,7 @@ struct common_params {
std::string chat_template = ""; // NOLINT
bool use_jinja = true; // NOLINT
bool enable_chat_template = true;
bool force_pure_content_parser = false;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
int enable_reasoning = -1; // -1 = auto, 0 = disable, 1 = enable
int reasoning_budget = -1;

120
common/jinja/caps.cpp
View File

@ -75,6 +75,7 @@ std::map<std::string, bool> caps::to_map() const {
{"supports_parallel_tool_calls", supports_parallel_tool_calls},
{"supports_system_role", supports_system_role},
{"supports_preserve_reasoning", supports_preserve_reasoning},
{"supports_object_arguments", supports_object_arguments},
};
}
@ -158,9 +159,9 @@ caps caps_get(jinja::program & prog) {
}
);
JJ_DEBUG("%s\n", ">>> Running capability check: single tool support");
JJ_DEBUG("%s\n", ">>> Running capability check: single tool with object arguments support");
// case: tools support: single call
// case: tools support: single call with object arguments
caps_try_execute(
prog,
[&]() {
@ -226,9 +227,7 @@ caps caps_get(jinja::program & prog) {
},
[&](bool success, value & messages, value & tools) {
if (!success) {
result.supports_tool_calls = false;
result.supports_tools = false;
return;
return; // Nothing can be inferred
}
auto & tool_name = tools->at(0)->at("function")->at("name");
@ -242,16 +241,117 @@ caps caps_get(jinja::program & prog) {
caps_print_stats(tool_calls, "messages[1].tool_calls");
if (!tool_calls->stats.used) {
result.supports_tool_calls = false;
return;
}
auto & tool_arg = tool_calls->at(0)->at("function")->at("arguments")->at("arg");
caps_print_stats(tool_arg, "messages[1].tool_calls[0].function.arguments.arg");
if (tool_arg->stats.used) {
result.supports_object_arguments = true;
}
}
);
if (!result.supports_object_arguments) {
JJ_DEBUG("%s\n", ">>> Running capability check: single tool with string arguments support");
// case: tools support: single call with string arguments
caps_try_execute(
prog,
[&]() {
// messages
return json::array({
{
{"role", "user"},
{"content", "User message"},
},
{
{"role", "assistant"},
{"content", ""}, // Some templates expect content to be empty with tool calls
{"tool_calls", json::array({
{
{"id", "call00001"},
{"type", "function"},
{"function", {
{"name", "tool1"},
{"arguments", R"({"arg": "value"})"}
}}
}
})}
},
{
{"role", "tool"},
{"content", "Tool response"},
{"tool_call_id", "call00001"}
},
{
{"role", "assistant"},
{"content", "The tool response was 'tool response'"}
},
{
{"role", "user"},
{"content", "User message"},
},
});
},
[&]() {
// tools
return json::array({
{
{"name", "tool"},
{"type", "function"},
{"function", {
{"name", "tool1"},
{"description", "Tool description"},
{"parameters", {
{"type", "object"},
{"properties", {
{"arg", {
{"type", "string"},
{"description", "Arg description"},
}},
}},
{"required", json::array({ "arg" })},
}},
}},
},
});
},
[&](bool success, value & messages, value & tools) {
if (!success) {
result.supports_tool_calls = false;
result.supports_tools = false;
return;
}
auto & tool_name = tools->at(0)->at("function")->at("name");
caps_print_stats(tool_name, "tools[0].function.name");
caps_print_stats(tools, "tools");
if (!tool_name->stats.used) {
result.supports_tools = false;
}
auto & tool_calls = messages->at(1)->at("tool_calls");
caps_print_stats(tool_calls, "messages[1].tool_calls");
if (!tool_calls->stats.used) {
result.supports_tool_calls = false;
return;
}
}
);
}
JJ_DEBUG("%s\n", ">>> Running capability check: parallel tool support");
// case: tools support: parallel calls
caps_try_execute(
prog,
[&]() {
json args = json(R"({"arg": "value"})");
if (result.supports_object_arguments) {
args = json{{"arg", "value"}};
}
// messages
return json::array({
{
@ -267,9 +367,7 @@ caps caps_get(jinja::program & prog) {
{"type", "function"},
{"function", {
{"name", "tool1"},
{"arguments", {
{"arg", "value"}
}}
{"arguments", args}
}}
},
{
@ -277,9 +375,7 @@ caps caps_get(jinja::program & prog) {
{"type", "function"},
{"function", {
{"name", "tool1"},
{"arguments", {
{"arg", "value"}
}}
{"arguments", args}
}}
}
})}
@ -328,7 +424,7 @@ caps caps_get(jinja::program & prog) {
return;
}
auto & tool_calls = messages->at(1)->at("tool_calls");;
auto & tool_calls = messages->at(1)->at("tool_calls");
caps_print_stats(tool_calls, "messages[1].tool_calls");
// check for second tool call usage

2
common/jinja/caps.h
View File

@ -18,6 +18,8 @@ struct caps {
bool supports_string_content = true;
bool supports_typed_content = false;
bool supports_object_arguments = false;
// for reporting on server
std::map<std::string, bool> to_map() const;

2
common/regex-partial.cpp
View File

@ -102,7 +102,7 @@ std::string regex_to_reversed_partial_regex(const std::string & pattern) {
auto is_star = *it == '*';
++it;
if (is_star) {
if (*it == '?') {
if (it != end && *it == '?') {
++it;
}
}

189
convert_hf_to_gguf.py
View File

@ -272,8 +272,9 @@ class ModelBase:
return tensors
def dequant_model(self):
if self._is_nvfp4:
return # NVFP4 weights are repacked in _generate_nvfp4_tensors
# If all quantized tensors were already handled (e.g. pure NVFP4), skip
if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv")) for k in self.model_tensors):
return
tensors_to_remove: list[str] = []
new_tensors: dict[str, Callable[[], Tensor]] = {}
@ -297,11 +298,16 @@ class ModelBase:
scale = scale.float()
if block_size is not None:
dim_offset = scale.ndim - len(block_size)
for i, size in enumerate(block_size):
scale = scale.repeat_interleave(size, i)
scale = scale.repeat_interleave(size, dim_offset + i)
# unpad the scale (e.g. when the tensor size isn't a multiple of the block size)
scale = scale[tuple(slice(0, size) for size in weight.shape)]
# align scale dims to weight for correct broadcasting (e.g. [128] -> [128, 1, 1])
while scale.ndim < weight.ndim:
scale = scale.unsqueeze(-1)
return weight.float() * scale
# ref: https://github.com/ModelCloud/GPTQModel/blob/037c5c0f6c9e33c500d975b038d02e7ca437546d/gptqmodel/nn_modules/qlinear/__init__.py#L437-L476
@ -392,7 +398,7 @@ class ModelBase:
elif quant_method == "fp8":
block_size = quant_config.get("weight_block_size")
for name in self.model_tensors.keys():
if name.endswith(".weight_scale_inv"):
if name.endswith("_scale_inv"):
weight_name = name.removesuffix("_scale_inv")
w = self.model_tensors[weight_name]
s = self.model_tensors[name]
@ -400,6 +406,8 @@ class ModelBase:
tensors_to_remove.append(name)
if name.endswith(".activation_scale"): # unused
tensors_to_remove.append(name)
if name.endswith("_activation_scale"): # Mistral-Small-4-119B-2602, unused
tensors_to_remove.append(name)
# mistral format
if name.endswith(".qscale_weight"):
weight_name = name.removesuffix("qscale_weight") + "weight"
@ -474,7 +482,20 @@ class ModelBase:
tensors_to_remove.append(base_name + "_zero_point")
else:
raise NotImplementedError(f"Quant format {quant_format!r} for method {quant_method!r} is not yet supported")
else:
elif quant_method == "modelopt":
# Mixed-precision ModelOpt models: NVFP4 tensors are handled by
# _generate_nvfp4_tensors; FP8 tensors have 1D weight_scale and
# are dequantized here. input_scale tensors are unused.
for name in self.model_tensors.keys():
if name.endswith(".weight_scale"):
weight_name = name.removesuffix("_scale")
w = self.model_tensors[weight_name]
s = self.model_tensors[name]
self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
tensors_to_remove.append(name)
if name.endswith((".input_scale", ".k_scale", ".v_scale")):
tensors_to_remove.append(name)
elif quant_method is not None:
raise NotImplementedError(f"Quant method is not yet supported: {quant_method!r}")
for name in tensors_to_remove:
@ -520,12 +541,6 @@ class ModelBase:
raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# skip NVFP4 auxiliary tensors (handled in _generate_nvfp4_tensors)
if self._is_nvfp4:
if name.endswith((".weight_scale", ".weight_scale_2", ".input_scale", ".k_scale", ".v_scale")):
return []
if name.endswith(".weight") and name.replace(".weight", ".weight_scale") in self.model_tensors:
return []
new_name = self.map_tensor_name(name)
@ -609,6 +624,7 @@ class ModelBase:
expert_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
expert_shapes: dict[tuple[int, str], list[int]] = {}
n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=True) or 0
consumed: list[str] = []
for name in list(self.model_tensors.keys()):
if not name.endswith(".weight"):
@ -620,8 +636,18 @@ class ModelBase:
# Force eager materialization of lazy tensors
weight = LazyTorchTensor.to_eager(self.model_tensors[name]())
scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
# Skip non-NVFP4 tensors (e.g. FP8 with per-channel 1D scales)
if scale.ndim < 2:
continue
scale2 = LazyTorchTensor.to_eager(self.model_tensors.get(scale2_name, lambda: torch.tensor(1.0))())
# Mark tensors for removal from model_tensors (already written to gguf)
consumed.extend([name, scale_name])
if scale2_name in self.model_tensors:
consumed.append(scale2_name)
# Check if this is a per-expert tensor
m = re.search(r'\.experts\.(\d+)\.(gate_proj|up_proj|down_proj)\.weight$', name)
if m:
@ -652,6 +678,15 @@ class ModelBase:
for (bid, proj_type) in list(expert_blocks.keys()):
self._flush_nvfp4_experts((bid, proj_type), expert_blocks, expert_scales, expert_shapes, bid, proj_type)
# Remove consumed tensors so get_tensors/modify_tensors won't see them
for name in consumed:
self.model_tensors.pop(name, None)
# Remove unused auxiliary tensors (input_scale, k_scale, v_scale)
for name in list(self.model_tensors.keys()):
if name.endswith((".input_scale", ".k_scale", ".v_scale")):
del self.model_tensors[name]
def _flush_nvfp4_experts(self, key, expert_blocks, expert_scales, expert_shapes, bid, proj_type):
experts = expert_blocks.pop(key)
scales = expert_scales.pop(key)
@ -677,20 +712,31 @@ class ModelBase:
def prepare_tensors(self):
# detect NVFP4 quantization (ModelOpt format)
quant_algo = (self.hparams.get("quantization_config") or {}).get("quant_algo")
quant_layers = (self.hparams.get("quantization_config") or {}).get("quantized_layers") or {}
quant_config_file = self.dir_model / "hf_quant_config.json"
if not quant_algo and quant_config_file.is_file():
if (not quant_algo or not quant_layers) and quant_config_file.is_file():
with open(quant_config_file, "r", encoding="utf-8") as f:
quant_algo = (json.load(f).get("quantization") or {}).get("quant_algo")
quant_config = json.load(f).get("quantization") or {}
quant_algo = quant_config.get("quant_algo", quant_algo)
quant_layers = quant_config.get("quantized_layers", quant_layers) or {}
# Some models use per-tensor quant_algo (e.g. "MIXED_PRECISION" with
# per-layer NVFP4/FP8) instead of a single global "NVFP4" value.
if quant_algo != "NVFP4":
if any(v.get("quant_algo") == "NVFP4" for v in quant_layers.values() if isinstance(v, dict)):
quant_algo = "NVFP4"
self._is_nvfp4 = quant_algo == "NVFP4"
self.dequant_model()
# NVFP4 weights are repacked and written directly to gguf_writer
# NVFP4 weights are repacked and written directly to gguf_writer.
# This must run before dequant_model so NVFP4 tensors are removed
# from model_tensors, leaving only non-NVFP4 (e.g. FP8) for dequant.
if self._is_nvfp4:
self._generate_nvfp4_tensors()
self.dequant_model()
# Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
if self.tensor_map.mapping:
max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@ -2992,10 +3038,16 @@ class LlavaVisionModel(MmprojModel):
def get_token_id(self, token: str) -> int:
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
added_tokens_decoder = json.load(f)['added_tokens_decoder']
added_tokens_decoder = json.load(f).get('added_tokens_decoder') or {}
for id_, token_data in added_tokens_decoder.items():
if token_data["content"] == token:
if token_data.get("content") == token:
return int(id_)
# fallthrough to tokenizer.json
with open(self.dir_model / "tokenizer.json", "r", encoding="utf-8") as f:
tokenizer_json = json.load(f)
for token_data in tokenizer_json["added_tokens"]:
if token_data["content"] == token:
return int(token_data["id"])
raise ValueError(f"Token '{token}' not found in tokenizer config.")
def set_gguf_parameters(self):
@ -3159,40 +3211,6 @@ class Llama4VisionModel(MmprojModel):
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register(
"Mistral3ForConditionalGeneration",
"Ministral3ForCausalLM",
)
class Mistral3Model(LlamaModel):
model_arch = gguf.MODEL_ARCH.MISTRAL3
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# for compatibility, we use LLAMA arch for older models
# TODO: remove this once everyone has migrated to newer version of llama.cpp
if self.hparams.get("model_type") != "ministral3":
self.model_arch = gguf.MODEL_ARCH.LLAMA
self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
self.gguf_writer.add_architecture()
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
def set_gguf_parameters(self):
super().set_gguf_parameters()
rope_params = self.rope_parameters
if self.hparams.get("model_type") == "ministral3":
assert rope_params, "ministral3 must have 'rope_parameters' config"
assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
name = name.replace("language_model.", "")
if "multi_modal_projector" in name or "vision_tower" in name:
return
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("DeciLMForCausalLM")
class DeciModel(TextModel):
model_arch = gguf.MODEL_ARCH.DECI
@ -8232,6 +8250,8 @@ class DeepseekV2Model(TextModel):
# TODO @ngxson : remove this when we support MTP for deepseek models
skip_mtp = True
merge_expert = True
def set_vocab(self):
try:
self._set_vocab_gpt2()
@ -8370,7 +8390,7 @@ class DeepseekV2Model(TextModel):
return
# process the experts separately
if name.find("mlp.experts") != -1:
if self.merge_expert and name.find("mlp.experts") != -1:
n_experts = self.hparams["n_routed_experts"]
assert bid is not None
@ -8429,6 +8449,69 @@ class DeepseekV2Model(TextModel):
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register(
"Mistral3ForConditionalGeneration",
"Ministral3ForCausalLM",
)
class Mistral3Model(TextModel):
class Ministral3Model(LlamaModel):
model_arch = gguf.MODEL_ARCH.MISTRAL3
def set_gguf_parameters(self):
super().set_gguf_parameters()
rope_params = self.rope_parameters
if self.hparams.get("model_type") == "ministral3":
assert rope_params, "ministral3 must have 'rope_parameters' config"
assert rope_params["rope_type"] == "yarn", "ministral3 rope_type must be 'yarn'"
self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
name = name.replace("language_model.", "")
if "multi_modal_projector" in name or "vision_tower" in name:
return
yield from super().modify_tensors(data_torch, name, bid)
class Mistral4Model(DeepseekV2Model):
model_arch = gguf.MODEL_ARCH.MISTRAL4
skip_mtp = False # model contains no MTP layers, so no need to skip
merge_expert = False # experts are already stacked as 3D
def modify_tensors(self, data_torch, name, bid):
if name.endswith(".down_proj") or name.endswith(".gate_up_proj"):
name = name + ".weight"
yield from super().modify_tensors(data_torch, name, bid)
model_arch = gguf.MODEL_ARCH.MISTRAL3 # unused
impl: TextModel
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
if self.hparams.get("model_type") == "mistral4":
self.impl = Mistral3Model.Mistral4Model(*args, **kwargs)
else:
self.impl = Mistral3Model.Ministral3Model(*args, **kwargs)
def set_vocab(self):
self.impl.set_vocab()
def set_gguf_parameters(self):
self.impl.set_gguf_parameters()
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
yield from self.impl.modify_tensors(data_torch, name, bid)
def prepare_tensors(self):
self.impl.prepare_tensors()
def write_vocab(self):
self.impl.write_vocab()
def write(self):
self.impl.write()
@ModelBase.register("MiniMaxM2ForCausalLM")
class MiniMaxM2Model(TextModel):
model_arch = gguf.MODEL_ARCH.MINIMAXM2

6
convert_lora_to_gguf.py
View File

@ -128,6 +128,12 @@ class LoraTorchTensor:
assert dim is None
return self.shape
def contiguous(self) -> LoraTorchTensor:
return LoraTorchTensor(
self._lora_A.contiguous(),
self._lora_B.contiguous(),
)
def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
if isinstance(shape[0], tuple):
new_shape: tuple[int, ...] = shape[0]

343
docs/backend/OPENVINO.md Normal file
View File

@ -0,0 +1,343 @@
# OpenVINO Backend for llama.cpp
[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge.
This document describes the [OpenVINO backend for llama.cpp](../../src/ggml-openvino), which enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a translation layer for core GGML operations. The OpenVINO backend replaces the standard GGML graph execution path with Intel's OpenVINO inference engine. This approach allows the same GGUF model file to run on Intel CPUs, Intel GPUs (integrated and discrete), and Intel NPUs without changes to the model or the rest of the llama.cpp stack. When a `ggml_cgraph` is dispatched to OpenVINO backend, it:
- Walks the GGML graph and identifies inputs, outputs, weights, and KV cache tensors.
- Translates the GGML operations into an `ov::Model` using OpenVINO's frontend API.
- Compiles and caches the model for the target device.
- Binds GGML tensor memory to OpenVINO inference tensors and runs inference.
## Supported Devices
OpenVINO backend supports the following hardware:
- Intel CPUs
- Intel GPUs (integrated and discrete)
- Intel NPUs
Although OpenVINO supports a wide range of [Intel hardware](https://docs.openvino.ai/2026/about-openvino/release-notes-openvino/system-requirements.html), the llama.cpp OpenVINO backend has been validated specifically on AI PCs such as the Intel® Core™ Ultra Series 1 and Series 2.
## Supported Model Precisions
- `FP16`
- `BF16` (on Intel Xeon)
- `Q8_0`
- `Q4_0`
- `Q4_1`
- `Q4_K`
- `Q4_K_M`
- `Q5_K` (converted to Q8_0_C at runtime)
- `Q6_K` (converted to Q8_0_C at runtime)
> [!NOTE]
> Accuracy validation and performance optimizations for quantized models are a work in progress.
## Quantization Support Details
### CPU and GPU
- **`Q4_0`, `Q4_1`, `Q4_K_M`, `Q6_K` models are supported**
- `Q5_K` and `Q6_K` tensors are converted to `Q8_0_C`
### NPU
- **Primary supported quantization scheme is `Q4_0`**
- `Q6_K` tensors are requantized to `Q4_0_128` in general. For embedding weights, `Q6_K` tensors are requantized to `Q8_0_C` except for the token embedding matrix which is dequantized to fp16
### Additional Notes
- Both `Q4_0` and `Q4_1` models use `Q6_K` for the token embedding tensor and the final matmul weight tensor (often the same tensor)
- `Q4_0` models may produce some `Q4_1` tensors if an imatrix is provided during quantization using `llama-quantize`
- `Q4_K_M` models may include both `Q6_K` and `Q5_K` tensors (observed in Phi-3)
## Validated Models
The following models have been validated for functionality on Intel® Core™ Ultra Series 1 and Series 2:
- [Llama-3.2-1B-Instruct-GGUF](https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/)
- [Llama-3.1-8B-Instruct](https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF)
- [microsoft/Phi-3-mini-4k-instruct-gguf](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf)
- [Qwen/Qwen2.5-1.5B-Instruct-GGUF](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF)
- [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B-GGUF)
- [openbmb/MiniCPM-1B-sft-bf16](https://huggingface.co/openbmb/MiniCPM-S-1B-sft-gguf)
- [tencent/Hunyuan-7B-Instruct](https://huggingface.co/bartowski/tencent_Hunyuan-7B-Instruct-GGUF)
- [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/bartowski/Mistral-7B-Instruct-v0.3-GGUF)
- [bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF)
## Build Instructions
### Prerequisites
- Linux or Windows system with Intel hardware (CPU, GPU, or NPU)
- **For Intel GPU or NPU Usage**: Install the appropriate hardware drivers for your Intel GPU or NPU. For detailed instructions, see: [Additional Configurations for Hardware Acceleration](https://docs.openvino.ai/2025/get-started/install-openvino/configurations.html).
- **Linux:**
- Git, CMake, and Ninja software tools are needed for building.
```bash
sudo apt-get update
sudo apt-get install -y build-essential libcurl4-openssl-dev libtbb12 cmake ninja-build python3-pip curl wget tar
```
- OpenCL
```bash
sudo apt install ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd
```
- **Windows:**
- Download and install [Microsoft Visual Studio 2022 Build Tools](https://aka.ms/vs/17/release/vs_BuildTools.exe). During installation, select the **"Desktop development with C++"** workload.
- Install required tools:
```powershell
# Windows PowerShell
winget install Git.Git
winget install GNU.Wget
winget install Ninja-build.Ninja
```
- Install **OpenCL** using **vcpkg**:
```powershell
# Windows PowerShell
cd C:\
git clone https://github.com/microsoft/vcpkg
cd vcpkg
.\bootstrap-vcpkg.bat
.\vcpkg install opencl
# Optional but recommended: Integrate vcpkg with Visual Studio / CMake:
.\vcpkg integrate install
```
### 1. Install OpenVINO Runtime
- Follow the guide to install OpenVINO Runtime from an archive file: [Linux](https://docs.openvino.ai/2026/get-started/install-openvino/install-openvino-archive-linux.html) | [Windows](https://docs.openvino.ai/2026/get-started/install-openvino/install-openvino-archive-windows.html)
- **Linux:**
<details>
<summary>📦 Click to expand OpenVINO installation from an archive file on Ubuntu</summary>
<br>
```bash
wget https://raw.githubusercontent.com/ravi9/misc-scripts/main/openvino/ov-archive-install/install-openvino-from-archive.sh
chmod +x install-openvino-from-archive.sh
./install-openvino-from-archive.sh
```
Verify OpenVINO is initialized properly:
```bash
echo $OpenVINO_DIR
```
</details>
### 2. Build llama.cpp with OpenVINO Backend
Clone the OpenVINO-enabled llama.cpp fork and build it:
```bash
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
```
- **Linux:**
```bash
source /opt/intel/openvino/setupvars.sh
cmake -B build/ReleaseOV -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_OPENVINO=ON
cmake --build build/ReleaseOV --parallel
```
- **Windows:**
```cmd
# x64 Native Tools Command Prompt for VS 2022
"C:\Program Files (x86)\Intel\openvino_2026.0\setupvars.bat"
cmake -B build\ReleaseOV -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_OPENVINO=ON -DLLAMA_CURL=OFF -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake
cmake --build build\ReleaseOV --parallel
```
> [!NOTE]
> Use `x64 Native Tools Command Prompt` for Windows build. After building, you could use either `cmd` or `PowerShell` to run the OpenVINO backend.
### 3. Download Sample Model
Download models for testing:
```bash
# Linux
mkdir -p ~/models/
wget https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf \
-O ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf
# Windows PowerShell
mkdir C:\models
Invoke-WebRequest -Uri https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf -OutFile C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf
# Windows Command Line
mkdir C:\models
curl -L https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf -o C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf
```
### 4. Run Inference with OpenVINO Backend
When using the OpenVINO backend, the first inference token may have slightly higher latency due to on-the-fly conversion to the OpenVINO graph. Subsequent tokens and runs will be faster.
```bash
# If device is unset or unavailable, defaults to CPU.
# If the system has multiple GPUs, use GPU.0 or GPU.1 to explicitly target a specific GPU.
# Linux
export GGML_OPENVINO_DEVICE=GPU
# To run llama-simple:
./build/ReleaseOV/bin/llama-simple -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -n 50 "The story of AI is "
# To run in chat mode:
./build/ReleaseOV/bin/llama-cli -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf
# Windows Command Line
set GGML_OPENVINO_DEVICE=GPU
# Windows PowerShell
$env:GGML_OPENVINO_DEVICE = "GPU"
# To run llama-simple
build\ReleaseOV\bin\llama-simple.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -n 50 "The story of AI is "
# To run in chat mode:
build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf"
```
> [!NOTE]
> On systems with multiple GPUs, use `GPU.0` or `GPU.1` to explicitly target specific GPU. See [OpenVINO GPU Device](https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/gpu-device.html) for more details.
### Docker Build
You can build and run llama.cpp with OpenVINO backend using Docker.
```bash
# Build the base runtime image with compiled shared libraries and minimal dependencies.
docker build -t llama-openvino:base -f .devops/openvino.Dockerfile .
# Build the complete image with all binaries, Python tools, gguf-py library, and model conversion utilities.
docker build --target=full -t llama-openvino:full -f .devops/openvino.Dockerfile .
# Build a minimal CLI-only image containing just the llama-cli executable.
docker build --target=light -t llama-openvino:light -f .devops/openvino.Dockerfile .
# Builds a server-only image with llama-server executable, health check endpoint, and REST API support.
docker build --target=server -t llama-openvino:server -f .devops/openvino.Dockerfile .
# If you are behind a proxy:
docker build --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_proxy --target=light -t llama-openvino:light -f .devops/openvino.Dockerfile .
```
Run llama.cpp with OpenVINO backend Docker container.
Save sample models in `~/models` as [shown above](#3-download-sample-model). It will be mounted to the container in the examples below.
```bash
# Run Docker container
docker run --rm -it -v ~/models:/models llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
# With Intel GPU access (iGPU or dGPU)
docker run --rm -it -v ~/models:/models \
--device=/dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -u $(id -u):$(id -g) \
llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
# With Intel NPU access
docker run --rm -it --env GGML_OPENVINO_DEVICE=NPU -v ~/models:/models \
--device=/dev/accel --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) -u $(id -u):$(id -g) \
llama-openvino:light --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
```
Run Llama.cpp Server with OpenVINO Backend:
```bash
# Run the Server Docker container
docker run --rm -it -p 8080:8080 -v ~/models:/models llama-openvino:server --no-warmup -m /models/Llama-3.2-1B-Instruct-Q4_0.gguf
# In a NEW terminal, test the server with curl
# If you are behind a proxy, make sure to set NO_PROXY to avoid proxy for localhost
export NO_PROXY=localhost,127.0.0.1
# Test health endpoint
curl -f http://localhost:8080/health
# Test with a simple prompt
curl -X POST "http://localhost:8080/v1/chat/completions" -H "Content-Type: application/json" \
-d '{"messages":[{"role":"user","content":"Write a poem about OpenVINO"}],"max_tokens":100}' | jq .
```
## Runtime Configuration
The OpenVINO backend can be configured using the following environment variables at runtime to control device selection, caching, debugging, and profiling behavior.
### Configuration Options
| Variable | Default | Description |
|-----------------------------------|------------|-------------------------------------------------------------------------------------------------------------|
| `GGML_OPENVINO_DEVICE` | `CPU` | Specify the target device (CPU, GPU, NPU). On systems with multiple GPUs, use `GPU.0` or `GPU.1` to explicitly target specific GPU. See [OpenVINO GPU Device](https://docs.openvino.ai/2026/openvino-workflow/running-inference/inference-devices-and-modes/gpu-device.html). When set to **NPU**, static compilation mode is enabled for optimal performance. |
| `GGML_OPENVINO_CACHE_DIR` | `not set` | Directory for OpenVINO model caching (recommended: `/tmp/ov_cache`). Enables model caching when set. **Not supported on NPU devices.** |
| `GGML_OPENVINO_PREFILL_CHUNK_SIZE`| `256` | Token chunk size for **NPU** prefill. |
| `GGML_OPENVINO_STATEFUL_EXECUTION`| `0` | Enable stateful KV cache on for better performance. Recommended on CPU, GPU. |
| `GGML_OPENVINO_PROFILING` | `0` | Enable execution-time profiling. |
| `GGML_OPENVINO_DUMP_CGRAPH` | `0` | Dump the GGML compute graph to `cgraph_ov.txt`. |
| `GGML_OPENVINO_DUMP_IR` | `0` | Serialize OpenVINO IR files with timestamps. |
| `GGML_OPENVINO_DEBUG_INPUT` | `0` | Enable input debugging and print input tensor info. |
| `GGML_OPENVINO_DEBUG_OUTPUT` | `0` | Enable output debugging and print output tensor info. |
| `GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS` | `0` | Print tensor address map once. |
> [!NOTE]
>`GGML_OPENVINO_STATEFUL_EXECUTION` is an **Experimental** feature to allow stateful execution for managing the KV cache internally inside the OpenVINO model, improving performance on CPUs and GPUs. Stateful execution is not effective on NPUs, and not all models currently support this feature. This feature is experimental and has been validated only with the llama-simple, llama-cli, llama-bench, and llama-run applications and is recommended to enable for the best performance. Other applications, such as llama-server and llama-perplexity, are not yet supported.
### Example Usage
#### GPU Inference with Profiling
```bash
# If the system has multiple GPUs, use GPU.0 or GPU.1 to explicitly target a specific GPU.
# Linux
export GGML_OPENVINO_CACHE_DIR=/tmp/ov_cache
export GGML_OPENVINO_PROFILING=1
export GGML_OPENVINO_DEVICE=GPU
./build/ReleaseOV/bin/llama-simple -m ~/models/Llama-3.2-1B-Instruct-Q4_0.gguf -n 50 "The story of AI is "
# Windows Command Line
set GGML_OPENVINO_CACHE_DIR=C:\tmp\ov_cache
set GGML_OPENVINO_PROFILING=1
set GGML_OPENVINO_DEVICE=GPU
# Windows PowerShell
$env:GGML_OPENVINO_CACHE_DIR = "C:\tmp\ov_cache"
$env:GGML_OPENVINO_PROFILING = "1"
$env:GGML_OPENVINO_DEVICE = "GPU"
build\ReleaseOV\bin\llama-simple.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf" -n 50 "The story of AI is "
```
#### llama-bench
```bash
# -fa 1 is required when running llama-bench with the OpenVINO backend.
GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1
```
### NPU Notes
- Model caching is not yet supported
- Does not support llama-server -np > 1 (multiple parallel sequences)
- Only supports llama-perplexity -b 512 or smaller
## Llama.cpp Tools
The following tools work with the OpenVINO backend on CPU, GPU, NPU:
- llama-simple
- llama-run
- llama-cli
- llama-server
- llama-bench
- llama-perplexity
## Work in Progress
- Performance and memory optimizations
- Accuracy validation
- Broader quantization coverage
- Support for additional model architectures

33
docs/build.md
View File

@ -13,6 +13,21 @@ cd llama.cpp
The following sections describe how to build with different backends and options.
* [CPU Build](#cpu-build)
* [BLAS Build](#blas-build)
* [Metal Build](#metal-build)
* [SYCL](#sycl)
* [CUDA](#cuda)
* [MUSA](#musa)
* [HIP](#hip)
* [Vulkan](#vulkan)
* [CANN](#cann)
* [Arm® KleidiAI™](#arm-kleidiai)
* [OpenCL](#opencl)
* [Android](#android-1)
* [OpenVINO](#openvino)
* [Notes about GPU-accelerated backends](#notes-about-gpu-accelerated-backends)
## CPU Build
Build llama.cpp using `CMake`:
@ -254,6 +269,14 @@ The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cu
Consider setting `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
#### GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F
Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F` environment variable to use FP32 compute type on all GPUs in FP16 cuBLAS for preventing possible numerical overflows in exchange for slower prompt processing (small impact on RTX PRO/Datacenter products and significant on GeForce products).
#### GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F
Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F` environment variable to force use FP16 compute type (instead of default FP32) in FP16 cuBLAS for V100, CDNA and RDNA4.
### Unified Memory
The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
@ -265,7 +288,7 @@ The following compilation options are also available to tweak performance:
| Option | Legal values | Default | Description |
|-------------------------------|------------------------|---------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for CDNA and RDNA4) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000). |
| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for V100, CDNA and RDNA4 which use FP32 compute type by default) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000). |
| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
| GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. |
@ -724,6 +747,14 @@ Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/m
To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)
## OpenVINO
[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware (CPUs, GPUs, and NPUs).
For build instructions and usage examples, refer to [OPENVINO.md](backend/OPENVINO.md).
---
## Notes about GPU-accelerated backends
The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.

6
docs/ops.md
View File

@ -15,7 +15,7 @@ Legend:
| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
|-----------|------|------|------|------|------|------|------|------|------|------|------|
| ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
| ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
| ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@ -47,7 +47,7 @@ Legend:
| FILL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
| FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
| FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
| GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
| GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
| GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
| GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@ -117,5 +117,5 @@ Legend:
| TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
| TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
| UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
| XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |

345
docs/ops/SYCL.csv
View File

@ -5937,6 +5937,20 @@
"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.100000","support","1","yes","SYCL"
"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=0.100000,v=0","support","1","yes","SYCL"
"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=0.100000,v=1","support","1","yes","SYCL"
"SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=10.000000","support","1","yes","SYCL"
"SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=10.000000,inplace=0","support","1","yes","SYCL"
"SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=10.000000","support","1","yes","SYCL"
"SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=10.000000,inplace=0","support","1","yes","SYCL"
"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=10.000000","support","1","yes","SYCL"
"SYCL0","L2_NORM","type=f32,ne=[64,5,4,3],eps=10.000000,v=0","support","1","yes","SYCL"
"SYCL0","L2_NORM","type=f32,ne=[64,5,4,3],eps=10.000000,v=1","support","1","yes","SYCL"
"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=10.000000","support","1","yes","SYCL"
"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=10.000000,inplace=0","support","1","yes","SYCL"
"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=10.000000","support","1","yes","SYCL"
"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=10.000000,inplace=0","support","1","yes","SYCL"
"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=10.000000","support","1","yes","SYCL"
"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=10.000000,v=0","support","1","yes","SYCL"
"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3],eps=10.000000,v=1","support","1","yes","SYCL"
"SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","1","yes","SYCL"
"SYCL0","SSM_CONV","type=f32,ne_a=[3,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
"SYCL0","SSM_CONV","type=f32,ne_a=[6,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
@ -6841,10 +6855,6 @@
"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=193,bs=[1,1],nr=[4,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=67,bs=[1,1],nr=[4,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f32,type_b=f32,m=64,n=77,k=77,bs=[12,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=2,n=1,k=3,bs=[128,1024],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=2,n=3,k=4,bs=[128,1024],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=2,n=1,k=3,bs=[131072,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=2,n=1,k=3,bs=[131072,1],nr=[1,1],per=[0,1,2,3],k_v=64,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=576,n=512,k=576,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=2048,k=8192,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
"SYCL0","MUL_MAT","type_a=f32,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
@ -10213,24 +10223,24 @@
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","0","no","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","1","yes","SYCL"
"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","1","yes","SYCL"
"SYCL0","SUM","type=f32,ne=[10,5,4,3]","support","1","yes","SYCL"
"SYCL0","SUM","type=f32,ne=[11,5,6,3],permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","SUM","type=f32,ne=[11,5,6,3],permute=[0,3,2,1]","support","0","no","SYCL"
@ -10261,8 +10271,8 @@
"SYCL0","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1],stride_dim=-1","support","1","yes","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[256,16,2,3],stride_dim=-1","support","1","yes","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[128,16,2,3],stride_dim=-1","support","1","yes","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[256,16,2,3],stride_dim=1","support","1","yes","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[128,16,2,3],stride_dim=2","support","1","yes","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[256,16,2,3],stride_dim=1","support","0","no","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[128,16,2,3],stride_dim=2","support","0","no","SYCL"
"SYCL0","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[64,16,2,3],stride_dim=3","support","1","yes","SYCL"
"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1,circular=0","support","1","yes","SYCL"
"SYCL0","PAD","type=f32,ne_a=[33,17,2,1],pad_0=4,pad_1=3,circular=1","support","0","no","SYCL"
@ -13329,6 +13339,262 @@
"SYCL0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=113,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[4,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=1,nr23=[32,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=113,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=320,hsv=256,nh=4,nr23=[4,1],kv=512,nb=75,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=1,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=1,nr23=[1,1],kv=113,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
"SYCL0","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=1,nr23=[1,1],kv=113,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
@ -13591,16 +13857,21 @@
"SYCL0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","SYCL"
"SYCL0","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","0","no","SYCL"
"SYCL0","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=32,head_size=128,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=16,head_size=64,n_seq_tokens=1,n_seqs=2,v_repeat=1,permuted=0,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=1,v_repeat=1,permuted=0,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=0,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=8,head_size=32,n_seq_tokens=4,n_seqs=2,v_repeat=2,permuted=0,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=1,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=1,v_repeat=1,permuted=1,kda=0","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=1","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=1,n_seqs=2,v_repeat=1,permuted=0,kda=1","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=32,n_seq_tokens=4,n_seqs=1,v_repeat=1,permuted=0,kda=1","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=0,kda=1","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=8,head_size=32,n_seq_tokens=4,n_seqs=2,v_repeat=2,permuted=0,kda=1","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=1,kda=1","support","0","no","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=32,head_size=128,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=32,head_size=16,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=32,head_size=16,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=1,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=32,head_size=16,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=16,head_size=64,n_seq_tokens=1,n_seqs=2,v_repeat=1,permuted=0,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=1,v_repeat=1,permuted=0,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=0,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=8,head_size=32,n_seq_tokens=4,n_seqs=2,v_repeat=2,permuted=0,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=1,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=1,v_repeat=1,permuted=1,kda=0","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=1,n_seqs=2,v_repeat=1,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=16,n_seq_tokens=1,n_seqs=2,v_repeat=1,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=32,n_seq_tokens=4,n_seqs=1,v_repeat=1,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=8,head_size=32,n_seq_tokens=4,n_seqs=2,v_repeat=2,permuted=0,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=64,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=1,kda=1","support","1","yes","SYCL"
"SYCL0","GATED_DELTA_NET","type=f32,head_count=4,head_size=16,n_seq_tokens=4,n_seqs=2,v_repeat=1,permuted=1,kda=1","support","1","yes","SYCL"

Can't render this file because it is too large.

3
ggml/CMakeLists.txt
View File

@ -248,6 +248,8 @@ set (GGML_SYCL_TARGET "INTEL" CACHE STRING
set (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
"ggml: sycl device architecture")
option(GGML_OPENVINO "ggml: use OPENVINO" OFF)
option(GGML_OPENCL "ggml: use OpenCL" OFF)
option(GGML_OPENCL_PROFILING "ggml: use OpenCL profiling (increases overhead)" OFF)
option(GGML_OPENCL_EMBED_KERNELS "ggml: embed kernels" ON)
@ -327,6 +329,7 @@ set(GGML_PUBLIC_HEADERS
include/ggml-vulkan.h
include/ggml-webgpu.h
include/ggml-zendnn.h
include/ggml-openvino.h
include/gguf.h)
set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

37
ggml/include/ggml-openvino.h Normal file
View File

@ -0,0 +1,37 @@
#pragma once
#include "ggml-backend.h"
#include <cstring>
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_OPENVINO_NAME "OPENVINO"
// backend API
GGML_BACKEND_API ggml_backend_t ggml_backend_openvino_init(int device);
GGML_BACKEND_API bool ggml_backend_is_openvino(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_buffer_is_openvino(ggml_backend_buffer_t buffer);
GGML_BACKEND_API bool ggml_backend_buft_is_openvino(ggml_backend_buffer_type_t buft);
GGML_BACKEND_API bool ggml_backend_buft_is_openvino_host(ggml_backend_buffer_type_t buft);
GGML_BACKEND_API size_t ggml_backend_openvino_buffer_get_ctx_id(ggml_backend_buffer_t buffer);
// device buffer
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_openvino_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_openvino_host_buffer_type(int device);
GGML_BACKEND_API int ggml_backend_openvino_get_device_count(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_openvino_reg(void);
#ifdef __cplusplus
}
#endif

1
ggml/src/CMakeLists.txt
View File

@ -460,6 +460,7 @@ ggml_add_backend(zDNN)
ggml_add_backend(OpenCL)
ggml_add_backend(Hexagon)
ggml_add_backend(ZenDNN)
ggml_add_backend(OPENVINO)
foreach (target ggml-base ggml)
target_include_directories(${target} PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

8
ggml/src/ggml-backend-reg.cpp
View File

@ -82,6 +82,10 @@
#include "ggml-zendnn.h"
#endif
#ifdef GGML_USE_OPENVINO
#include "ggml-openvino.h"
#endif
namespace fs = std::filesystem;
static std::string path_str(const fs::path & path) {
@ -154,6 +158,9 @@ struct ggml_backend_registry {
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_OPENVINO
register_backend(ggml_backend_openvino_reg());
#endif
#ifdef GGML_USE_CPU
register_backend(ggml_backend_cpu_reg());
#endif
@ -557,6 +564,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
ggml_backend_load_best("opencl", silent, dir_path);
ggml_backend_load_best("hexagon", silent, dir_path);
ggml_backend_load_best("musa", silent, dir_path);
ggml_backend_load_best("openvino", silent, dir_path);
ggml_backend_load_best("cpu", silent, dir_path);
// check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend
const char * backend_path = std::getenv("GGML_BACKEND_PATH");

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

@ -121,6 +121,8 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
bli_thread_set_num_threads(ctx->n_threads);
#elif defined(GGML_BLAS_USE_NVPL)
nvpl_blas_set_num_threads(ctx->n_threads);
#elif defined(GGML_BLAS_USE_MKL)
mkl_set_num_threads(ctx->n_threads);
#endif
for (int64_t i13 = 0; i13 < ne13; i13++) {

2
ggml/src/ggml-cpu/arch/arm/quants.c
View File

@ -666,7 +666,7 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0;
#if defined __ARM_NEON
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
const int8x16_t values = vld1q_s8(kvalues_mxfp4);
const uint8x16_t m4b = vdupq_n_u8(0x0f);
float32x4_t acc = vdupq_n_f32(0.0f);

40
ggml/src/ggml-cpu/arch/riscv/quants.c
View File

@ -115,10 +115,10 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
assert(k % QK_K == 0);
block_q8_K * y_blocks = (block_q8_K *)y;
size_t nb = k / QK_K;
#if defined(__riscv_v_intrinsic)
block_q8_K * y_blocks = (block_q8_K *)y;
const size_t vlmax_f32m8 = __riscv_vsetvlmax_e32m8();
for (size_t i = 0; i < nb; i++) {
@ -2052,6 +2052,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq1_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -2147,6 +2148,7 @@ static void ggml_vec_dot_iq1_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
*s = sumf;
}
#endif
void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2163,6 +2165,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq1_m_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -2269,6 +2272,7 @@ static void ggml_vec_dot_iq1_m_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
*s = sumf;
}
#endif
void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2285,6 +2289,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static const uint8_t sign_gather_indices_arr[64] = {
0,0,0,0,0,0,0,0, 1,1,1,1,1,1,1,1, 2,2,2,2,2,2,2,2, 3,3,3,3,3,3,3,3,
4,4,4,4,4,4,4,4, 5,5,5,5,5,5,5,5, 6,6,6,6,6,6,6,6, 7,7,7,7,7,7,7,7
@ -2488,6 +2493,7 @@ static void ggml_vec_dot_iq2_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
}
*s = 0.125f * sumf;
}
#endif
void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2507,7 +2513,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined(__riscv_v)
#if defined(__riscv_v_intrinsic)
static const int8_t keven_signs_q2xs[1024] = {
1, 1, 1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, 1,
1, 1, -1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, 1, 1, -1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, -1,
@ -2542,7 +2548,6 @@ static const int8_t keven_signs_q2xs[1024] = {
1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, -1, 1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, 1,
1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, -1,
};
#endif
static void ggml_vec_dot_iq2_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
@ -2618,6 +2623,7 @@ static void ggml_vec_dot_iq2_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_
}
*s = 0.125f * sumf;
}
#endif
void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2634,6 +2640,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq2_xxs_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -2818,6 +2825,7 @@ static void ggml_vec_dot_iq2_xxs_q8_K_vl256(int n, float * GGML_RESTRICT s, size
}
*s = 0.125f * sumf;
}
#endif
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2830,10 +2838,11 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
break;
}
#else
ggml_vec_dot_iq2_xxs_q8_K(n, s, bs, vx, bx, vy, by, nrc);
ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq3_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
UNUSED(nrc);
@ -2928,6 +2937,7 @@ static void ggml_vec_dot_iq3_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
}
*s = sumf;
}
#endif
void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2944,6 +2954,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq3_xxs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -3036,6 +3047,7 @@ static void ggml_vec_dot_iq3_xxs_q8_K_vl256(int n, float * GGML_RESTRICT s, size
}
*s = 0.25f * sumf;
}
#endif
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3052,6 +3064,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq4_nl_q8_0_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -3161,6 +3174,7 @@ static void ggml_vec_dot_iq4_nl_q8_0_vl256(int n, float * GGML_RESTRICT s, size_
*s = sumf;
}
#endif
void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3177,6 +3191,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -3190,7 +3205,6 @@ static void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_
const int nb = n / QK_K;
#if defined __riscv_v_intrinsic
const vint8m4_t values = __riscv_vle8_v_i8m4(kvalues_iq4nl, 16);
float sumf = 0;
int acc[4];
@ -3252,14 +3266,8 @@ static void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_
}
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
#endif
void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3276,6 +3284,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_tq1_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -3381,6 +3390,7 @@ static void ggml_vec_dot_tq1_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
*s = sumf;
}
#endif
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3397,6 +3407,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_tq2_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -3467,6 +3478,7 @@ static void ggml_vec_dot_tq2_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
*s = sumf;
}
#endif
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3483,6 +3495,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static void ggml_vec_dot_mxfp4_q8_0_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -3592,6 +3605,7 @@ static void ggml_vec_dot_mxfp4_q8_0_vl256(int n, float * GGML_RESTRICT s, size_t
*s = sumf;
}
#endif
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3604,6 +3618,6 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
break;
}
#else
return ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}

40
ggml/src/ggml-cpu/arch/riscv/repack.cpp
View File

@ -107,8 +107,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
}
#else
UNUSED(nb);
UNUSED(y);
ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
#endif
}
@ -203,6 +202,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
#if defined __riscv_zvfh
void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
@ -222,7 +222,6 @@ void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb);
@ -256,9 +255,6 @@ void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
__riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
}
return;
#endif
ggml_gemv_q4_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -280,7 +276,6 @@ void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
const block_q8_K * a_ptr = (const block_q8_K *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
@ -392,9 +387,6 @@ void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
__riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
}
return;
#endif
ggml_gemv_q4_K_16x1_q8_K_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -416,7 +408,6 @@ void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16);
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
@ -451,9 +442,6 @@ void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
__riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
}
return;
#endif
ggml_gemv_iq4_nl_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -476,7 +464,6 @@ void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
UNUSED(blocklen);
UNUSED(bs);
#if defined __riscv_v_intrinsic
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb);
@ -505,9 +492,6 @@ void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
__riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
}
return;
#endif
ggml_gemv_q8_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -679,9 +663,9 @@ void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
} // End K-Block
__riscv_vse32_v_f32m2(s + col_tile, v_sumf, vl);
}
}
#endif
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
@ -909,6 +893,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
#if defined __riscv_zvfh
void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
@ -929,7 +914,6 @@ void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
@ -994,9 +978,6 @@ void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
__riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
}
}
return;
#endif
ggml_gemm_q4_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -1019,7 +1000,6 @@ void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
for (int y = 0; y < nr / 4; y++) {
const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
@ -1267,9 +1247,6 @@ void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
__riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
}
}
return;
#endif
ggml_gemm_q4_K_16x1_q8_K_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -1292,7 +1269,6 @@ void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16);
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
@ -1355,9 +1331,6 @@ void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
__riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
}
}
return;
#endif
ggml_gemm_iq4_nl_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -1380,7 +1353,6 @@ void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v_intrinsic
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
@ -1429,9 +1401,6 @@ void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
__riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
}
}
return;
#endif
ggml_gemm_q8_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@ -1731,3 +1700,4 @@ void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
}
}
}
#endif

8
ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
View File

@ -1461,7 +1461,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
return false;
}
if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) &&
ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) {
ggml_ne(op->src[1], 3) == 1) {
return true;
}
}
@ -1473,10 +1473,12 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
} else {
if (op->src[0]->type != GGML_TYPE_F16) {
return nullptr;
}
std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
const int slot_total = kleidiai_collect_kernel_chain(op, kernel_chain);
const bool has_kernel = slot_total > 0;
if (has_kernel && op->src[1]->ne[1] > 1) {
if (slot_total > 0 && op->src[1]->ne[1] > 1) {
if ((op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
(op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
return nullptr;

53
ggml/src/ggml-cpu/ops.cpp
View File

@ -6205,7 +6205,7 @@ static void ggml_compute_forward_im2col_f16(
const ggml_tensor * src1 = dst->src[1];
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16);
GGML_TENSOR_BINARY_OP_LOCALS;
@ -6236,7 +6236,7 @@ static void ggml_compute_forward_im2col_f16(
int ofs1 = is_2D ? nb12 : nb11;
GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
GGML_ASSERT(nb10 == sizeof(float));
GGML_ASSERT(nb10 == ggml_type_size(src1->type));
// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
{
@ -6249,7 +6249,12 @@ static void ggml_compute_forward_im2col_f16(
// micro kernel
ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW]
const float * const src_data_f32 = src1->type == GGML_TYPE_F32
? (const float *)((const char *) src1->data + in*ofs0 + iic*ofs1)
: nullptr; // [IH, IW]
const ggml_fp16_t * const src_data_f16 = src1->type == GGML_TYPE_F16
? (const ggml_fp16_t *)((const char *) src1->data + in*ofs0 + iic*ofs1)
: nullptr; // [IH, IW]
for (int64_t ikh = 0; ikh < KH; ikh++) { // 1
for (int64_t ikw = 0; ikw < KW; ikw++) {
@ -6259,7 +6264,11 @@ static void ggml_compute_forward_im2col_f16(
if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
} else {
dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]);
if (src_data_f32 != nullptr) {
dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data_f32[iih*IW + iiw]);
} else {
dst_data[iic*(KH*KW) + ikh*KW + ikw] = src_data_f16[iih*IW + iiw];
}
}
}
}
@ -10477,34 +10486,40 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
const float beta_val = *(const float *)((const char *)src_beta->data + iv3 * nbb3 + t * nbb2 + iv1 * nbb1);
const float * g_d = (const float *)((const char *)src_g->data + iv3 * nbg3 + t * nbg2 + iv1 * nbg1);
// state is stored transposed: s_out[j*S_v + i] = S[i][j]
// so row j of s_out = column j of S (contiguous access)
if (kda) {
// precompute exp(g) into delta scratch (reused below)
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_scale_f32(S_v, &s_out[i * S_v], expf(g_d[i]));
delta[i] = expf(g_d[i]);
}
// S[i][:] *= exp(g[i]) => for each row j of M: M[j][i] *= exp(g[i])
for (int64_t j = 0; j < S_v; ++j) {
ggml_vec_mul_f32(S_v, &s_out[j * S_v], &s_out[j * S_v], delta);
}
} else {
ggml_vec_scale_f32(S_v * S_v, s_out, expf(g_d[0]));
}
// delta[j] = sum_i S[j][i] * k[i]
memset(delta, 0, S_v * sizeof(float));
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_mad_f32(S_v, delta, &s_out[i * S_v], k_d[i]);
}
// delta[j] = sum_i S[i][j] * k[i] = dot(row j of M, k)
for (int64_t j = 0; j < S_v; ++j) {
delta[j] = (v_d[j] - delta[j]) * beta_val;
float sum = 0.0f;
ggml_vec_dot_f32(S_v, &sum, 0, &s_out[j * S_v], 0, k_d, 0, 1);
delta[j] = (v_d[j] - sum) * beta_val;
}
// outer product: S[j][i] += k[i] * delta[j]
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_mad_f32(S_v, &s_out[i * S_v], delta, k_d[i]);
// outer product: S[i][j] += k[i] * delta[j] => M[j][i] += delta[j] * k[i]
for (int64_t j = 0; j < S_v; ++j) {
ggml_vec_mad_f32(S_v, &s_out[j * S_v], k_d, delta[j]);
}
// attn_out[j] = sum_i S[j][i] * q[i]
memset(attn_data, 0, S_v * sizeof(float));
for (int64_t i = 0; i < S_v; ++i) {
ggml_vec_mad_f32(S_v, attn_data, &s_out[i * S_v], q_d[i]);
// attn_out[j] = sum_i S[i][j] * q[i] = dot(row j of M, q)
for (int64_t j = 0; j < S_v; ++j) {
float sum = 0.0f;
ggml_vec_dot_f32(S_v, &sum, 0, &s_out[j * S_v], 0, q_d, 0, 1);
attn_data[j] = sum * scale;
}
ggml_vec_scale_f32(S_v, attn_data, scale);
attn_data += S_v * H; // advance to next token
}

3
ggml/src/ggml-cpu/repack.cpp
View File

@ -1365,6 +1365,7 @@ void ggml_gemv_q8_0_4x8_q8_0_generic(int n,
}
}
// Only enable these for RISC-V.
#if defined __riscv_zvfh
void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
@ -1568,6 +1569,7 @@ void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
assert(nc % 16 == 0);
UNUSED(bs);
UNUSED(nr);
const int nb = n / QK_K;
const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx;
@ -2381,6 +2383,7 @@ void ggml_gemm_q8_0_4x8_q8_0_generic(int n,
}
}
// Only enable these for RISC-V.
#if defined __riscv_zvfh
void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;

46
ggml/src/ggml-cpu/simd-mappings.h
View File

@ -479,13 +479,51 @@ do { \
// F16 AVX512
// F16 AVX
#if defined(__AVX512FP16__)
#define GGML_F16_STEP 128
#define GGML_F16_EPR 32
#define GGML_F16x32 __m512h
#define GGML_F16x32_ZERO _mm512_setzero_ph()
#define GGML_F16x32_SET1(x) _mm512_set1_ph(__extension__(_Float16)(x))
#define GGML_F16x32_LOAD(x) _mm512_loadu_ph(x)
#define GGML_F16x32_STORE(x, y) _mm512_storeu_ph(x, y)
#define GGML_F16x32_FMA(a, b, c) _mm512_fmadd_ph(b, c, a)
#define GGML_F16x32_ADD _mm512_add_ph
#define GGML_F16x32_MUL _mm512_mul_ph
#define GGML_F16x32_REDUCE(res, x) \
do { \
int offset = GGML_F16_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ph(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ph(x[i], x[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = _mm512_add_ph(x[i], x[offset+i]); \
} \
res = (ggml_float) _mm512_reduce_add_ph(x[0]); \
} while (0)
#define GGML_F16_VEC GGML_F16x32
#define GGML_F16_VEC_ZERO GGML_F16x32_ZERO
#define GGML_F16_VEC_SET1 GGML_F16x32_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F16x32_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x32_STORE(p, r[i])
#define GGML_F16_VEC_FMA GGML_F16x32_FMA
#define GGML_F16_VEC_ADD GGML_F16x32_ADD
#define GGML_F16_VEC_MUL GGML_F16x32_MUL
#define GGML_F16_VEC_REDUCE GGML_F16x32_REDUCE
#else // Fallback FP16 <-> FP32
#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)
@ -525,6 +563,8 @@ do { \
#define GGML_F16_VEC_MUL GGML_F32Cx16_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE
#endif // __AVX512FP16__
#elif defined(__AVX__)
#define GGML_SIMD

9
ggml/src/ggml-cuda/cpy.cu
View File

@ -56,7 +56,8 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x; // transpose block offset
const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
__shared__ float tile[CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
__shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
int cur_tile_buf = 0;
#pragma unroll
for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
@ -70,7 +71,7 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
if(x < ne01 && y + j < ne00){
const int row = threadIdx.y+j;
const int col = threadIdx.x * sizeof(float)/sizeof(T);
T *tile2 = reinterpret_cast<T*>(tile[row]);
T *tile2 = reinterpret_cast<T*>(tile[cur_tile_buf][row]);
tile2[col] = src[imat*n + (y+j)*ne01 + x];
}
}
@ -81,10 +82,12 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
if (ty + j < ne01 && tx < ne00) {
const int col = (threadIdx.y+j)*sizeof(float)/sizeof(T);
const T *tile2 = reinterpret_cast<const T*>(tile[threadIdx.x]);
const T *tile2 = reinterpret_cast<const T*>(tile[cur_tile_buf][threadIdx.x]);
dst[imat*n + (ty+j)*ne00 + tx] = tile2[col];
}
}
cur_tile_buf = (cur_tile_buf + 1) % 2;
}
GGML_UNUSED_VARS(ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11,

24
ggml/src/ggml-cuda/fattn-common.cuh
View File

@ -892,7 +892,7 @@ void launch_fattn(
const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
const int gqa_ratio = Q->ne[2] / K->ne[2];
const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2);
const int ntiles_total = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];
const int ntiles_dst = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];
// Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
// Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
@ -919,37 +919,37 @@ void launch_fattn(
GGML_ASSERT(max_blocks_per_sm > 0);
int parallel_blocks = max_blocks_per_sm;
const int ntiles_KV = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by KV cache length.
dim3 blocks_num;
if (stream_k) {
// For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
const int max_blocks = max_blocks_per_sm*nsm;
const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
const int tiles_efficiency_percent = 100 * ntiles_total / (max_blocks*tiles_nwaves);
const int tiles_nwaves = (ntiles_dst + max_blocks - 1) / max_blocks;
const int tiles_efficiency_percent = 100 * ntiles_dst / (max_blocks*tiles_nwaves);
const int nblocks_stream_k = max_blocks;
const int nblocks_stream_k = std::min(max_blocks, ntiles_KV*ntiles_dst);
const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || amd_wmma_available(cc) || tiles_efficiency_percent < 75;
blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total;
blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_dst;
blocks_num.y = 1;
blocks_num.z = 1;
if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
if (ntiles_dst % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
dst_tmp_meta.alloc((size_t(blocks_num.x) * ncols * (2 + DV/2)));
}
} else {
const int ntiles_KQ = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by tensor size.
// parallel_blocks must not be larger than what the tensor size allows:
parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
parallel_blocks = std::min(parallel_blocks, ntiles_KV);
// If ntiles_total % blocks_per_wave != 0 then some efficiency is lost due to tail effects.
// Test whether parallel_blocks can be set to a higher value for better efficiency.
const int blocks_per_wave = nsm * max_blocks_per_sm;
int nwaves_best = 0;
int efficiency_percent_best = 0;
for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KQ; ++parallel_blocks_test) {
const int nblocks_total = ntiles_total * parallel_blocks_test;
for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KV; ++parallel_blocks_test) {
const int nblocks_total = ntiles_dst * parallel_blocks_test;
const int nwaves = (nblocks_total + blocks_per_wave - 1) / blocks_per_wave;
const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave);
@ -1015,7 +1015,7 @@ void launch_fattn(
CUDA_CHECK(cudaGetLastError());
if (stream_k) {
if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
if (ntiles_dst % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
const dim3 block_dim_combine(DV, 1, 1);
const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};

54
ggml/src/ggml-cuda/gated_delta_net.cu
View File

@ -1,7 +1,8 @@
#include "gated_delta_net.cuh"
template <int S_v, bool KDA>
__global__ void gated_delta_net_cuda(const float * q,
__global__ void __launch_bounds__((ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v) * 4, 2)
gated_delta_net_cuda(const float * q,
const float * k,
const float * v,
const float * g,
@ -38,17 +39,19 @@ __global__ void gated_delta_net_cuda(const float * q,
const int64_t state_offset = (sequence * H + h_idx) * S_v * S_v;
state += state_offset;
curr_state += state_offset;
curr_state += state_offset + col * S_v;
attn_data += (sequence * n_tokens * H + h_idx) * S_v;
constexpr int warp_size = ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v;
static_assert(S_v % warp_size == 0, "S_v must be a multiple of warp_size");
constexpr int rows_per_lane = (S_v + warp_size - 1) / warp_size;
float s_shard[rows_per_lane];
// state is stored transposed: M[col][i] = S[i][col], row col is contiguous
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
s_shard[r] = curr_state[i * S_v + col];
s_shard[r] = curr_state[i];
}
for (int t = 0; t < n_tokens; t++) {
@ -62,6 +65,16 @@ __global__ void gated_delta_net_cuda(const float * q,
const float beta_val = *beta_t;
// Cache k and q in registers
float k_reg[rows_per_lane];
float q_reg[rows_per_lane];
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
k_reg[r] = k_t[i];
q_reg[r] = q_t[i];
}
if constexpr (!KDA) {
const float g_val = expf(*g_t);
@ -69,8 +82,7 @@ __global__ void gated_delta_net_cuda(const float * q,
float kv_shard = 0.0f;
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
kv_shard += s_shard[r] * k_t[i];
kv_shard += s_shard[r] * k_reg[r];
}
float kv_col = warp_reduce_sum<warp_size>(kv_shard);
@ -82,9 +94,8 @@ __global__ void gated_delta_net_cuda(const float * q,
float attn_partial = 0.0f;
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
s_shard[r] = g_val * s_shard[r] + k_t[i] * delta_col;
attn_partial += s_shard[r] * q_t[i];
s_shard[r] = g_val * s_shard[r] + k_reg[r] * delta_col;
attn_partial += s_shard[r] * q_reg[r];
}
float attn_col = warp_reduce_sum<warp_size>(attn_partial);
@ -98,7 +109,7 @@ __global__ void gated_delta_net_cuda(const float * q,
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
kv_shard += expf(g_t[i]) * s_shard[r] * k_t[i];
kv_shard += expf(g_t[i]) * s_shard[r] * k_reg[r];
}
float kv_col = warp_reduce_sum<warp_size>(kv_shard);
@ -112,8 +123,8 @@ __global__ void gated_delta_net_cuda(const float * q,
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
s_shard[r] = expf(g_t[i]) * s_shard[r] + k_t[i] * delta_col;
attn_partial += s_shard[r] * q_t[i];
s_shard[r] = expf(g_t[i]) * s_shard[r] + k_reg[r] * delta_col;
attn_partial += s_shard[r] * q_reg[r];
}
float attn_col = warp_reduce_sum<warp_size>(attn_partial);
@ -126,23 +137,14 @@ __global__ void gated_delta_net_cuda(const float * q,
attn_data += S_v * H;
}
// Write state back to global memory
// Write state back to global memory (transposed layout)
#pragma unroll
for (int r = 0; r < rows_per_lane; r++) {
const int i = r * warp_size + lane;
state[i * S_v + col] = s_shard[r];
state[col * S_v + i] = s_shard[r];
}
}
static size_t calculate_smem(const int sv, int cc)
{
size_t smem = 0;
if ((GGML_CUDA_CC_IS_AMD(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_RDNA4(cc)) || GGML_CUDA_CC_IS_MTHREADS(cc)) {
smem = sv * sv * sizeof(float);
}
return smem;
}
template <bool KDA>
static void launch_gated_delta_net(
const float * q_d, const float * k_d, const float * v_d,
@ -179,18 +181,14 @@ static void launch_gated_delta_net(
sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
break;
case 64: {
constexpr int sv = 64;
size_t smem = calculate_smem(sv, cc);
gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
gated_delta_net_cuda<64, KDA><<<grid_dims, block_dims, 0, stream>>>(
q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
break;
}
case 128: {
constexpr int sv = 128;
size_t smem = calculate_smem(sv, cc);
gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
gated_delta_net_cuda<128, KDA><<<grid_dims, block_dims, 0, stream>>>(
q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);

94
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -124,7 +124,10 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
err = cudaMallocManaged(ptr, size);
#if defined(GGML_USE_HIP)
if (err == hipSuccess) {
CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
// hipMemAdviseSetCoarseGrain is an optional performance hint;
// ignore errors (e.g. hipErrorInvalidValue on some APU/iGPU configs).
(void)cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device);
(void)hipGetLastError(); // clear any error
}
// fall back to cudaMalloc if not supported (e.g. on Windows)
@ -251,11 +254,6 @@ static ggml_cuda_device_info ggml_cuda_init() {
info.devices[id].supports_cooperative_launch = false;
#endif // !(GGML_USE_MUSA)
// cudaMemGetInfo returns info for the current device
size_t free_mem;
CUDA_CHECK(cudaSetDevice(id));
CUDA_CHECK(cudaMemGetInfo(&free_mem, NULL));
#if defined(GGML_USE_HIP)
info.devices[id].smpbo = prop.sharedMemPerBlock;
@ -270,25 +268,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
info.devices[id].cc += prop.minor * 0x10;
}
}
GGML_LOG_INFO(" Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d, VRAM: %zu MiB (%zu MiB free)\n",
GGML_LOG_INFO(" Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d, VRAM: %zu MiB\n",
id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
device_vmm ? "yes" : "no", prop.warpSize,
(size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
(size_t)(prop.totalGlobalMem / (1024 * 1024)));
#elif defined(GGML_USE_MUSA)
// FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
info.devices[id].warp_size = 32;
info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100;
info.devices[id].cc += prop.minor * 0x10;
GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB (%zu MiB free)\n",
GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB\n",
id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
(size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
(size_t)(prop.totalGlobalMem / (1024 * 1024)));
#else
info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
info.devices[id].cc = 100*prop.major + 10*prop.minor;
GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB (%zu MiB free)\n",
GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s, VRAM: %zu MiB\n",
id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
(size_t)(prop.totalGlobalMem / (1024 * 1024)), free_mem / (1024 * 1024));
(size_t)(prop.totalGlobalMem / (1024 * 1024)));
std::string device_name(prop.name);
if (device_name == "NVIDIA GeForce MX450") {
turing_devices_without_mma.push_back({ id, device_name });
@ -303,6 +301,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
// TODO: Check for future drivers the default scheduling strategy and
// remove this call again when cudaDeviceScheduleSpin is default.
if (prop.major == 12 && prop.minor == 1) {
CUDA_CHECK(cudaSetDevice(id));
CUDA_CHECK(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
}
@ -1242,6 +1241,34 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
}
}
struct cublas_force_compute_type {
bool fp32 = false;
bool fp16 = false;
};
static const cublas_force_compute_type & ggml_cuda_cublas_get_force_compute_type() {
static const cublas_force_compute_type compute_type = [] {
cublas_force_compute_type result;
const bool ggml_cuda_force_cublas_compute_32f_env = getenv("GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F") != nullptr;
const bool ggml_cuda_force_cublas_compute_16f_env = getenv("GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F") != nullptr;
GGML_ASSERT(ggml_cuda_force_cublas_compute_16f_env == false || ggml_cuda_force_cublas_compute_32f_env == false);
if (ggml_cuda_force_cublas_compute_32f_env) {
GGML_LOG_INFO("Detected GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F\n");
result.fp32 = true;
} else if (ggml_cuda_force_cublas_compute_16f_env) {
GGML_LOG_INFO("Detected GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F\n");
result.fp16 = true;
}
return result;
}();
return compute_type;
}
static void ggml_cuda_op_mul_mat_cublas(
ggml_backend_cuda_context & ctx,
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
@ -1324,7 +1351,13 @@ static void ggml_cuda_op_mul_mat_cublas(
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
const auto & force_compute_type = ggml_cuda_cublas_get_force_compute_type();
if (!force_compute_type.fp16 && (GGML_CUDA_CC_IS_CDNA(cc)
|| GGML_CUDA_CC_IS_RDNA4(cc)
|| cc == GGML_CUDA_CC_VOLTA
|| force_compute_type.fp32))
{
const float alpha = 1.0f;
const float beta = 0.0f;
CUBLAS_CHECK(
@ -1923,10 +1956,23 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
cudaDataType_t cu_data_type_b = traits::data_type;
const void * alpha = traits::get_alpha();
const void * beta = traits::get_beta();
const float alpha_f32 = 1.0f;
const float beta_f32 = 0.0f;
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
const auto & force_compute_type = ggml_cuda_cublas_get_force_compute_type();
int id = ggml_cuda_get_device();
const int cc = ggml_cuda_info().devices[id].cc;
static constexpr bool is_src0_type_f16 = src0_type == GGML_TYPE_F16;
// bf16 and fp32 are already being computed in fp32 (ensure it using static_assert),
// so checking necessity of forced fp32 only for fp16 src0_type
static_assert(is_src0_type_f16 || traits::compute_type == CUBLAS_COMPUTE_32F);
const bool need_compute_32f = is_src0_type_f16 && !force_compute_type.fp16 && (GGML_CUDA_CC_IS_CDNA(cc)
|| GGML_CUDA_CC_IS_RDNA4(cc)
|| cc == GGML_CUDA_CC_VOLTA
|| force_compute_type.fp32);
if (dst->op_params[0] == GGML_PREC_DEFAULT && !need_compute_32f) {
if constexpr (src0_type == GGML_TYPE_F32) {
dst_t = (char *) dst_ddf; // Direct F32 output
} else {
@ -1936,18 +1982,10 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
}
} else {
dst_t = (char *) dst_ddf;
cu_compute_type = CUBLAS_COMPUTE_32F;
cu_data_type = CUDA_R_32F;
alpha = &alpha_f32;
beta = &beta_f32;
}
int id = ggml_cuda_get_device();
const int cc = ggml_cuda_info().devices[id].cc;
if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
cu_compute_type = CUBLAS_COMPUTE_32F;
alpha = &alpha_f32;
beta = &beta_f32;
cu_compute_type = batched_mul_mat_traits<GGML_TYPE_F32>::compute_type;
cu_data_type = batched_mul_mat_traits<GGML_TYPE_F32>::data_type;
alpha = batched_mul_mat_traits<GGML_TYPE_F32>::get_alpha();
beta = batched_mul_mat_traits<GGML_TYPE_F32>::get_beta();
}
GGML_ASSERT(ne12 % ne02 == 0);

89
ggml/src/ggml-cuda/mmvq.cu
View File

@ -60,11 +60,17 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
enum mmvq_parameter_table_id {
MMVQ_PARAMETERS_GENERIC = 0,
MMVQ_PARAMETERS_GCN,
MMVQ_PARAMETERS_RDNA2
MMVQ_PARAMETERS_RDNA2,
MMVQ_PARAMETERS_RDNA3_0,
MMVQ_PARAMETERS_RDNA4
};
static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
#if defined(RDNA2) || defined(RDNA3) || defined(RDNA4)
#if defined(RDNA4)
return MMVQ_PARAMETERS_RDNA4;
#elif defined(RDNA3_0)
return MMVQ_PARAMETERS_RDNA3_0;
#elif defined(RDNA2) || defined(RDNA3_5)
return MMVQ_PARAMETERS_RDNA2;
#elif defined(GCN) || defined(CDNA)
return MMVQ_PARAMETERS_GCN;
@ -74,7 +80,13 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
}
static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
if (GGML_CUDA_CC_IS_RDNA4(cc)) {
return MMVQ_PARAMETERS_RDNA4;
}
if (GGML_CUDA_CC_IS_RDNA3_0(cc)) {
return MMVQ_PARAMETERS_RDNA3_0;
}
if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc)) {
return MMVQ_PARAMETERS_RDNA2;
}
if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
@ -83,7 +95,7 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
return MMVQ_PARAMETERS_GENERIC;
}
static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
if (table_id == MMVQ_PARAMETERS_GENERIC) {
switch (ncols_dst) {
case 1:
@ -114,6 +126,50 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_paramet
return 1;
}
}
if (table_id == MMVQ_PARAMETERS_RDNA4) {
// nwarps=8 benefits types with simple vec_dot on RDNA4 (ncols_dst=1).
// Types with complex vec_dot (Q3_K, IQ2_*, IQ3_*) regress due to register
// pressure and lookup table contention at higher thread counts.
if (ncols_dst == 1) {
switch (type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
return 8;
default:
return 1;
}
}
return 1;
}
if (table_id == MMVQ_PARAMETERS_RDNA3_0) {
// RDNA3 (W7900): stricter whitelist than RDNA4.
// Q2_K / Q5_K / IQ4_XS regress in full quant sweeps.
if (ncols_dst == 1) {
switch (type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_IQ4_NL:
return 8;
default:
return 1;
}
}
return 1;
}
return 1;
}
@ -138,7 +194,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
}
template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false>
__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
static __global__ void mul_mat_vec_q(
const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@ -151,7 +207,7 @@ static __global__ void mul_mat_vec_q(
constexpr int qi = ggml_cuda_type_traits<type>::qi;
constexpr int vdr = get_vdr_mmvq(type);
constexpr mmvq_parameter_table_id table_id = get_device_table_id();
constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
@ -355,12 +411,13 @@ static __global__ void mul_mat_vec_q(
}
}
template<ggml_type type>
static std::pair<dim3, dim3> calc_launch_params(
const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
const int warp_size, const mmvq_parameter_table_id table_id) {
const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
const dim3 block_dims(warp_size, calc_nwarps(type, ncols_dst, table_id), 1);
return {block_nums, block_dims};
}
@ -420,7 +477,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
if (has_ids && ncols_dst > 1) {
// Multi-token MUL_MAT_ID path only - single-token goes through regular path below
constexpr int c_ncols_dst = 1;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -431,7 +488,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
switch (ncols_dst) {
case 1: {
constexpr int c_ncols_dst = 1;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -439,7 +496,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 2: {
constexpr int c_ncols_dst = 2;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -447,7 +504,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 3: {
constexpr int c_ncols_dst = 3;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -455,7 +512,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 4: {
constexpr int c_ncols_dst = 4;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -463,7 +520,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 5: {
constexpr int c_ncols_dst = 5;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -471,7 +528,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 6: {
constexpr int c_ncols_dst = 6;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -479,7 +536,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 7: {
constexpr int c_ncols_dst = 7;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
@ -487,7 +544,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
} break;
case 8: {
constexpr int c_ncols_dst = 8;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,

8
ggml/src/ggml-cuda/vendors/hip.h vendored
View File

@ -207,6 +207,14 @@
#define RDNA3
#endif // defined(__GFX11__)
#if defined(__gfx1150__) || defined(__gfx1151__)
#define RDNA3_5
#endif // defined(__gfx1150__) || defined(__gfx1151__)
#if defined(RDNA3) && !defined(RDNA3_5)
#define RDNA3_0
#endif // defined(RDNA3) && !defined(RDNA3_5)
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
#define RDNA2

195
ggml/src/ggml-hexagon/ggml-hexagon.cpp
View File

@ -402,6 +402,7 @@ static void pack_q4_0_quants(block_q4_0 * x, const uint8_t * qs, unsigned int bi
static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
static const int qk = QK_Q4_0x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)
const int nloe = k % qk; // leftovers
const int dblk_size = 8 * 2; // 8x __fp16
const int qblk_size = qk / 2; // int4
@ -435,9 +436,11 @@ static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
unpack_q4_0_quants(qs, &x[i * 8 + 6], 6);
unpack_q4_0_quants(qs, &x[i * 8 + 7], 7);
bool partial = (nloe && i == nb-1);
uint8_t * q = y_q + (i * qblk_size);
for (int j = 0; j < qk / 2; j++) {
q[j] = (qs[j + 128] << 4) | qs[j];
q[j] = partial ? (qs[j*2+1] << 4) | qs[j*2+0] : (qs[j+128] << 4) | qs[j+000];
}
}
@ -467,6 +470,7 @@ static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
static void unpack_row_q4x4x2(block_q4_0 * x, const uint8_t * y, int64_t k) {
static const int qk = QK_Q4_0x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)
const int nloe = k % qk; // leftovers
const int dblk_size = 8 * 2; // 8x __fp16
const int qblk_size = qk / 2; // int4
@ -485,10 +489,17 @@ static void unpack_row_q4x4x2(block_q4_0 * x, const uint8_t * y, int64_t k) {
for (int i = 0; i < nb; i++) {
uint8_t qs[QK_Q4_0x4x2]; // unpacked quants
bool partial = (nloe && i == nb-1);
const uint8_t * q = y_q + (i * qblk_size);
for (int j = 0; j < qk / 2; j++) {
qs[j] = q[j] & 0xf;
qs[j + 128] = q[j] >> 4;
if (partial) {
qs[j*2+0] = q[j] & 0xf;
qs[j*2+1] = q[j] >> 4;
} else {
qs[j+000] = q[j] & 0xf;
qs[j+128] = q[j] >> 4;
}
}
pack_q4_0_quants(&x[i * 8 + 0], qs, 0);
@ -1078,6 +1089,7 @@ static void pack_mxfp4_quants(block_mxfp4 * x, const uint8_t * qs, unsigned int
static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k) {
static const int qk = QK_MXFP4x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)
const int nloe = k % qk; // leftovers
const int eblk_size = 8 * 1; // 8x E8M0
const int qblk_size = qk / 2; // int4
@ -1112,9 +1124,11 @@ static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k)
unpack_mxfp4_quants(qs, &x[i * 8 + 6], 6);
unpack_mxfp4_quants(qs, &x[i * 8 + 7], 7);
bool partial = (nloe && i == nb-1);
uint8_t * q = y_q + (i * qblk_size);
for (int j = 0; j < qk / 2; j++) {
q[j] = (qs[j + 128] << 4) | qs[j];
q[j] = partial ? (qs[j*2+1] << 4) | qs[j*2+0] : (qs[j+128] << 4) | qs[j+000];
}
}
@ -1144,6 +1158,7 @@ static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k)
static void unpack_row_mxfp4x4x2(block_mxfp4 * x, const uint8_t * y, int64_t k) {
static const int qk = QK_MXFP4x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)
const int nloe = k % qk; // leftovers
const int eblk_size = 8 * 1; // 8x E8M0
const int qblk_size = qk / 2; // int4
@ -1162,10 +1177,17 @@ static void unpack_row_mxfp4x4x2(block_mxfp4 * x, const uint8_t * y, int64_t k)
for (int i = 0; i < nb; i++) {
uint8_t qs[QK_MXFP4x4x2]; // unpacked quants
bool partial = (nloe && i == nb-1);
const uint8_t * q = y_q + (i * qblk_size);
for (int j = 0; j < qk / 2; j++) {
qs[j] = q[j] & 0xf;
qs[j + 128] = q[j] >> 4;
if (partial) {
qs[j*2+0] = q[j] & 0xf;
qs[j*2+1] = q[j] >> 4;
} else {
qs[j+000] = q[j] & 0xf;
qs[j+128] = q[j] >> 4;
}
}
pack_mxfp4_quants(&x[i * 8 + 0], qs, 0);
@ -1801,12 +1823,12 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
return false;
}
if (src0->ne[1] > 16 * 1024) {
if (ggml_nrows(src0) > 16 * 1024) {
return false; // typically the lm-head which would be too large for VTCM
}
if ((src1->ne[2] != 1 || src1->ne[3] != 1)) {
return false;
if (ggml_nrows(src1) > 1024 || src1->ne[2] != 1 || src1->ne[3] != 1) {
return false; // no huge batches or broadcasting (for now)
}
// src0 (weights) must be repacked
@ -1820,6 +1842,9 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
GGML_LOG_DEBUG("ggml_hexagon_supported_mul_mat: permuted F16 src0 not supported\n");
return false;
}
if (ggml_nrows(src1) > 1024) {
return false; // no huge batches (for now)
}
break;
default:
@ -2337,6 +2362,27 @@ static inline size_t init_cpy_req(htp_general_req * req, dspqueue_buffer * bufs,
return n_bufs;
}
static inline size_t init_cont_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
// CONT is just a contiguous copy — reuse CPY op
req->op = HTP_OP_CPY;
size_t n_bufs = 0;
n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
n_bufs += htp_req_buff_init(&req->dst, &bufs[n_bufs], t, DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
return n_bufs;
}
static inline size_t init_repeat_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
req->op = HTP_OP_REPEAT;
size_t n_bufs = 0;
n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
n_bufs += htp_req_buff_init(&req->dst, &bufs[n_bufs], t, DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
return n_bufs;
}
static inline size_t init_get_rows_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
req->op = HTP_OP_GET_ROWS;
@ -2424,12 +2470,33 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
break;
case GGML_OP_UNARY:
if (ggml_get_unary_op(t) == GGML_UNARY_OP_SILU) {
switch (ggml_get_unary_op(t)) {
case GGML_UNARY_OP_SILU:
req->op = HTP_OP_UNARY_SILU;
supported = true;
} else if (ggml_get_unary_op(t) == GGML_UNARY_OP_GELU) {
break;
case GGML_UNARY_OP_GELU:
req->op = HTP_OP_UNARY_GELU;
supported = true;
break;
case GGML_UNARY_OP_SIGMOID:
req->op = HTP_OP_UNARY_SIGMOID;
supported = true;
break;
case GGML_UNARY_OP_NEG:
req->op = HTP_OP_UNARY_NEG;
supported = true;
break;
case GGML_UNARY_OP_EXP:
req->op = HTP_OP_UNARY_EXP;
supported = true;
break;
case GGML_UNARY_OP_SOFTPLUS:
req->op = HTP_OP_UNARY_SOFTPLUS;
supported = true;
break;
default:
break;
}
break;
@ -2615,16 +2682,28 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
ggml_hexagon_dispatch_op<init_sum_rows_req>(sess, node, flags);
break;
case GGML_OP_UNARY:
if ((ggml_get_unary_op(node) == GGML_UNARY_OP_SILU) ||
(ggml_get_unary_op(node) == GGML_UNARY_OP_GELU)) {
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
switch (ggml_get_unary_op(node)) {
case GGML_UNARY_OP_NEG:
case GGML_UNARY_OP_EXP:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_SOFTPLUS:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_GELU:
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
break;
default:
break;
}
break;
case GGML_OP_GLU:
if ((ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU) ||
(ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI) ||
(ggml_get_glu_op(node) == GGML_GLU_OP_GEGLU)) {
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
switch (ggml_get_glu_op(node)) {
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU:
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
break;
default:
break;
}
break;
case GGML_OP_SOFT_MAX:
@ -2651,6 +2730,14 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
ggml_hexagon_dispatch_op<init_cpy_req>(sess, node, flags);
break;
case GGML_OP_CONT:
ggml_hexagon_dispatch_op<init_cont_req>(sess, node, flags);
break;
case GGML_OP_REPEAT:
ggml_hexagon_dispatch_op<init_repeat_req>(sess, node, flags);
break;
case GGML_OP_ARGSORT:
ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
break;
@ -2981,6 +3068,39 @@ static bool ggml_hexagon_supported_cpy(const struct ggml_hexagon_session * sess,
return true;
}
static bool ggml_hexagon_supported_cont(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
GGML_UNUSED(sess);
const struct ggml_tensor * src0 = op->src[0];
// CONT is same-type only, supports f32 and f16
if (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) return false;
return true;
}
static bool ggml_hexagon_supported_repeat(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
GGML_UNUSED(sess);
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * dst = op;
// Support f32 and f16
if (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) return false;
// src and dst must be the same type
if (src0->type != dst->type) return false;
// dst dims must be multiples of src dims
if (dst->ne[0] % src0->ne[0] != 0) return false;
if (dst->ne[1] % src0->ne[1] != 0) return false;
if (dst->ne[2] % src0->ne[2] != 0) return false;
if (dst->ne[3] % src0->ne[3] != 0) return false;
// require contiguous tensors (no transposition)
if (ggml_is_transposed(src0) || ggml_is_transposed(dst)) return false;
return true;
}
static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
auto sess = static_cast<ggml_hexagon_session *>(dev->context);
@ -3038,21 +3158,32 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
break;
case GGML_OP_UNARY:
{
const auto unary_op = ggml_get_unary_op(op);
if (unary_op == GGML_UNARY_OP_SILU || unary_op == GGML_UNARY_OP_GELU) {
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_NEG:
case GGML_UNARY_OP_EXP:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_SOFTPLUS:
supp = ggml_hexagon_supported_unary(sess, op);
break;
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_GELU:
supp = ggml_hexagon_supported_activations(sess, op);
}
break;
break;
default:
break;
}
break;
case GGML_OP_GLU:
{
const auto glu_op = ggml_get_glu_op(op);
if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI) || (glu_op == GGML_GLU_OP_GEGLU)) {
switch (ggml_get_glu_op(op)) {
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU:
supp = ggml_hexagon_supported_activations(sess, op);
}
break;
break;
default:
break;
}
break;
case GGML_OP_ROPE:
supp = ggml_hexagon_supported_rope(sess, op);
break;
@ -3073,6 +3204,14 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
supp = ggml_hexagon_supported_cpy(sess, op);
break;
case GGML_OP_CONT:
supp = ggml_hexagon_supported_cont(sess, op);
break;
case GGML_OP_REPEAT:
supp = ggml_hexagon_supported_repeat(sess, op);
break;
case GGML_OP_ARGSORT:
supp = ggml_hexagon_supported_argsort(sess, op);
break;

1
ggml/src/ggml-hexagon/htp/CMakeLists.txt
View File

@ -30,6 +30,7 @@ add_library(${HTP_LIB} SHARED
set-rows-ops.c
get-rows-ops.c
cpy-ops.c
repeat-ops.c
argsort-ops.c
ssm-conv.c
)

5
ggml/src/ggml-hexagon/htp/htp-msg.h
View File

@ -53,6 +53,10 @@ enum htp_op {
HTP_OP_RMS_NORM,
HTP_OP_UNARY_SILU,
HTP_OP_UNARY_GELU,
HTP_OP_UNARY_SIGMOID,
HTP_OP_UNARY_EXP,
HTP_OP_UNARY_NEG,
HTP_OP_UNARY_SOFTPLUS,
HTP_OP_GLU_SWIGLU,
HTP_OP_GLU_SWIGLU_OAI,
HTP_OP_GLU_GEGLU,
@ -69,6 +73,7 @@ enum htp_op {
HTP_OP_SQRT,
HTP_OP_SUM_ROWS,
HTP_OP_SSM_CONV,
HTP_OP_REPEAT,
INVALID
};

1
ggml/src/ggml-hexagon/htp/htp-ops.h
View File

@ -57,6 +57,7 @@ int op_flash_attn_ext(struct htp_ops_context * octx);
int op_set_rows(struct htp_ops_context * octx);
int op_get_rows(struct htp_ops_context * octx);
int op_cpy(struct htp_ops_context * octx);
int op_repeat(struct htp_ops_context * octx);
int op_argsort(struct htp_ops_context * octx);
int op_ssm_conv(struct htp_ops_context * octx);

2
ggml/src/ggml-hexagon/htp/hvx-base.h
View File

@ -3,6 +3,8 @@
#include <stdbool.h>
#include <stdint.h>
#include <math.h>
#include <assert.h>
#include "hex-utils.h"
#include "hvx-types.h"

17
ggml/src/ggml-hexagon/htp/hvx-exp.h
View File

@ -3,6 +3,7 @@
#include <stdbool.h>
#include <stdint.h>
#include <math.h>
#include "hvx-base.h"
#include "hvx-floor.h"
@ -16,8 +17,8 @@
#define EXP_LOGN2 (0x3F317218) // ln(2) = 0.6931471805
#define EXP_LOG2E (0x3FB8AA3B) // log2(e) = 1/ln(2) = 1.4426950408
#define EXP_ONE (0x3f800000) // 1.0
#define EXP_RANGE_R (0x41a00000) // 20.0
#define EXP_RANGE_L (0xc1a00000) // -20.0
#define EXP_RANGE_R (0x42B16666) // 88.7
#define EXP_RANGE_L (0xC2B00000) // -88.0 (approx log(FLT_MIN))
static inline HVX_Vector hvx_vec_exp_f32(HVX_Vector in_vec) {
HVX_Vector z_qf32_v;
@ -47,12 +48,12 @@ static inline HVX_Vector hvx_vec_exp_f32(HVX_Vector in_vec) {
HVX_Vector temp_v = in_vec;
// Clamp inputs to (-20.0, 20.0)
// Clamp inputs to (-88.0, 88.0) to avoid overflow/underflow
HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VsfVsf(in_vec, Q6_V_vsplat_R(EXP_RANGE_R));
HVX_VectorPred pred_cap_left = Q6_Q_vcmp_gt_VsfVsf(Q6_V_vsplat_R(EXP_RANGE_L), in_vec);
in_vec = Q6_V_vmux_QVV(pred_cap_right, Q6_V_vsplat_R(EXP_RANGE_R), temp_v);
in_vec = Q6_V_vmux_QVV(pred_cap_left, Q6_V_vsplat_R(EXP_RANGE_L), temp_v);
in_vec = Q6_V_vmux_QVV(pred_cap_left, Q6_V_vsplat_R(EXP_RANGE_L), in_vec);
epsilon_v = Q6_Vqf32_vmpy_VsfVsf(log2e, in_vec);
epsilon_v = Q6_Vsf_equals_Vqf32(epsilon_v);
@ -69,12 +70,12 @@ static inline HVX_Vector hvx_vec_exp_f32(HVX_Vector in_vec) {
// normalize before every QFloat's vmpy
x_qf32_v = Q6_Vqf32_vadd_Vqf32Vsf(x_qf32_v, zero_v);
x_v = Q6_Vsf_equals_Vqf32(x_qf32_v);
// z = x * x;
z_qf32_v = Q6_Vqf32_vmpy_Vqf32Vqf32(x_qf32_v, x_qf32_v);
z_qf32_v = Q6_Vqf32_vadd_Vqf32Vsf(z_qf32_v, zero_v);
x_v = Q6_Vsf_equals_Vqf32(x_qf32_v);
// y = E4 + E5 * x;
E_const = Q6_V_vsplat_R(EXP_COEFF_5);
y_v = Q6_Vqf32_vmpy_VsfVsf(E_const, x_v);
@ -145,7 +146,7 @@ static inline HVX_Vector hvx_vec_exp_f32_guard(HVX_Vector in_vec, HVX_Vector max
return Q6_V_vmux_QVV(pred0, inf, out);
}
static inline void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems, bool negate) {
static inline void hvx_exp_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems, bool negate) {
int left_over = num_elems & (VLEN_FP32 - 1);
int num_elems_whole = num_elems - left_over;
@ -162,7 +163,7 @@ static inline void hvx_exp_f32(const uint8_t * restrict src, uint8_t * restrict
HVX_Vector vec_out = Q6_V_vzero();
static const float kInf = INFINITY;
static const float kMaxExp = 88.02f; // log(INF)
static const float kMaxExp = 88.7f;
const HVX_Vector max_exp = hvx_vec_splat_f32(kMaxExp);
const HVX_Vector inf = hvx_vec_splat_f32(kInf);

1
ggml/src/ggml-hexagon/htp/hvx-sigmoid.h
View File

@ -2,6 +2,7 @@
#define HVX_SIGMOID_H
#include "hvx-base.h"
#include "hvx-inverse.h"
#define FAST_SIGMOID_LOG2F (0x3fb8aa3b) // 1.442695022
#define FAST_SIGMOID_C1 (0x3d009076) // 0.03138777

45
ggml/src/ggml-hexagon/htp/main.c
View File

@ -516,6 +516,39 @@ static void proc_cpy_req(struct htp_context * ctx, struct htp_general_req * req,
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
}
static void proc_repeat_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
struct dspqueue_buffer rsp_bufs[1];
// We had written to the output buffer, we'd also need to flush it
rsp_bufs[0].fd = bufs[1].fd;
rsp_bufs[0].ptr = bufs[1].ptr;
rsp_bufs[0].offset = bufs[1].offset;
rsp_bufs[0].size = bufs[1].size;
rsp_bufs[0].flags = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | // Flush HTP
DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT); // Invalidate CPU
// Setup Op context
struct htp_ops_context octx = { 0 };
octx.ctx = ctx;
octx.src0 = req->src0;
octx.dst = req->dst;
octx.flags = req->flags;
octx.op = req->op;
// Update data pointers
octx.src0.data = (uint32_t) bufs[0].ptr;
octx.dst.data = (uint32_t) bufs[1].ptr;
octx.n_threads = ctx->n_threads;
struct profile_data prof;
profile_start(&prof);
uint32_t rsp_status = op_repeat(&octx);
profile_stop(&prof);
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
}
static void proc_get_rows_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
struct dspqueue_buffer rsp_bufs[1];
@ -1090,6 +1123,10 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
case HTP_OP_SQR:
case HTP_OP_SQRT:
case HTP_OP_UNARY_NEG:
case HTP_OP_UNARY_EXP:
case HTP_OP_UNARY_SIGMOID:
case HTP_OP_UNARY_SOFTPLUS:
if (n_bufs != 2) {
FARF(ERROR, "Bad unary-req buffer list");
continue;
@ -1175,6 +1212,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
proc_cpy_req(ctx, &req, bufs);
break;
case HTP_OP_REPEAT:
if (n_bufs != 2) {
FARF(ERROR, "Bad repeat-req buffer list");
continue;
}
proc_repeat_req(ctx, &req, bufs);
break;
case HTP_OP_ARGSORT:
if (n_bufs != 2) {
FARF(ERROR, "Bad argsort-req buffer list");

407
ggml/src/ggml-hexagon/htp/matmul-ops.c
View File

@ -77,7 +77,7 @@ static inline size_t q8x4x2_row_size(uint32_t ne) {
return hex_round_up(ne + nb * 8 * sizeof(__fp16), 128);
}
static inline HVX_Vector_x8 hvx_vec_load_q4x4x8(const uint8_t * restrict ptr) {
static inline HVX_Vector_x8 hvx_vec_load_q4x4x8_full(const uint8_t * restrict ptr) {
const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
HVX_Vector v0_1 = vptr[0]; // first 256 elements (128 bytes)
@ -88,9 +88,9 @@ static inline HVX_Vector_x8 hvx_vec_load_q4x4x8(const uint8_t * restrict ptr) {
const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
const HVX_Vector i8 = Q6_Vb_vsplat_R(8);
HVX_Vector v0 = Q6_V_vand_VV(v0_1, mask_h4); // & 0x0F
HVX_Vector v1 = Q6_Vub_vlsr_VubR(v0_1, 4); // >> 4
HVX_Vector v2 = Q6_V_vand_VV(v2_3, mask_h4); // & 0x0F
HVX_Vector v0 = Q6_V_vand_VV(v0_1, mask_h4); // & 0x0F : first 128 elements
HVX_Vector v1 = Q6_Vub_vlsr_VubR(v0_1, 4); // >> 4 : second 128 elements
HVX_Vector v2 = Q6_V_vand_VV(v2_3, mask_h4); // & 0x0F ...
HVX_Vector v3 = Q6_Vub_vlsr_VubR(v2_3, 4); // >> 4
HVX_Vector v4 = Q6_V_vand_VV(v4_5, mask_h4); // & 0x0F
HVX_Vector v5 = Q6_Vub_vlsr_VubR(v4_5, 4); // >> 4
@ -111,7 +111,41 @@ static inline HVX_Vector_x8 hvx_vec_load_q4x4x8(const uint8_t * restrict ptr) {
return r;
}
static inline HVX_Vector_x8 hvx_vec_load_mxfp4x4x8(const uint8_t * restrict ptr) {
static HVX_Vector_x8 hvx_vec_load_q4x4x8_partial(const uint8_t * restrict ptr, uint32_t n) {
const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
const uint32_t qk = QK_Q4_0x4x2; // 256
const uint32_t nb = n / qk;
const uint32_t nloe = n % qk;
const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
const HVX_Vector i8 = Q6_Vb_vsplat_R(8);
HVX_Vector_x8 r;
uint32_t i = 0;
#pragma unroll(2)
for (i=0; i < nb; i++) {
HVX_Vector v = vptr[i]; // 256 elements (128 bytes)
HVX_Vector v0 = Q6_V_vand_VV(v, mask_h4); // & 0x0F : first 128 elements
HVX_Vector v1 = Q6_Vub_vlsr_VubR(v, 4); // >> 4 : second 128 elements
r.v[i*2+0] = Q6_Vb_vsub_VbVb(v0, i8);
r.v[i*2+1] = Q6_Vb_vsub_VbVb(v1, i8);
}
if (nloe) {
HVX_Vector v = vptr[i]; // 256 elements (128 bytes)
HVX_Vector v0 = Q6_V_vand_VV(v, mask_h4); // & 0x0F : even 128 elements
HVX_Vector v1 = Q6_Vub_vlsr_VubR(v, 4); // >> 4 : odd 128 elements
HVX_VectorPair v0_1_p = Q6_W_vshuff_VVR(v1, v0, -1); // zip even:odd:...
r.v[i*2+0] = Q6_Vb_vsub_VbVb(Q6_V_lo_W(v0_1_p), i8);
r.v[i*2+1] = Q6_Vb_vsub_VbVb(Q6_V_hi_W(v0_1_p), i8);
}
return r;
}
static inline HVX_Vector_x8 hvx_vec_load_mxfp4x4x8_full(const uint8_t * restrict ptr) {
const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
HVX_Vector v0_1 = vptr[0]; // first 256 elements (128 bytes)
@ -144,7 +178,41 @@ static inline HVX_Vector_x8 hvx_vec_load_mxfp4x4x8(const uint8_t * restrict ptr)
return r;
}
static inline HVX_Vector_x8 hvx_vec_load_q8x4x8(const uint8_t * restrict ptr) {
static inline HVX_Vector_x8 hvx_vec_load_mxfp4x4x8_partial(const uint8_t * restrict ptr, uint32_t n) {
const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
const uint32_t qk = QK_Q4_0x4x2; // 256
const uint32_t nb = n / qk;
const uint32_t nloe = n % qk;
const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
const HVX_Vector lut = *(const HVX_Vector *) kvalues_mxfp4_lut;
HVX_Vector_x8 r;
uint32_t i = 0;
#pragma unroll(2)
for (i=0; i < nb; i++) {
HVX_Vector v = vptr[i]; // 256 elements (128 bytes)
HVX_Vector v0 = Q6_V_vand_VV(v, mask_h4); // & 0x0F : first 128 elements
HVX_Vector v1 = Q6_Vub_vlsr_VubR(v, 4); // >> 4 : second 128 elements
r.v[i*2+0] = Q6_Vb_vlut32_VbVbI(v0, lut, 0);
r.v[i*2+1] = Q6_Vb_vlut32_VbVbI(v1, lut, 0);
}
if (nloe) {
HVX_Vector v = vptr[i]; // 256 elements (128 bytes)
HVX_Vector v0 = Q6_V_vand_VV(v, mask_h4); // & 0x0F : even 128 elements
HVX_Vector v1 = Q6_Vub_vlsr_VubR(v, 4); // >> 4 : odd 128 elements
HVX_VectorPair v0_1_p = Q6_W_vshuff_VVR(v1, v0, -1); // zip even:odd:...
r.v[i*2+0] = Q6_Vb_vlut32_VbVbI(Q6_V_lo_W(v0_1_p), lut, 0);
r.v[i*2+1] = Q6_Vb_vlut32_VbVbI(Q6_V_hi_W(v0_1_p), lut, 0);
}
return r;
}
static inline HVX_Vector_x8 hvx_vec_load_q8x4x8_full(const uint8_t * restrict ptr) {
const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
HVX_Vector v0 = vptr[0]; // first 128 vals
@ -160,6 +228,10 @@ static inline HVX_Vector_x8 hvx_vec_load_q8x4x8(const uint8_t * restrict ptr) {
return r;
}
static inline HVX_Vector_x8 hvx_vec_load_q8x4x8_partial(const uint8_t * restrict ptr, uint32_t nloe) {
return hvx_vec_load_q8x4x8_full(ptr);
}
// Reduce multiply 1024 x 1024 int8 elements (32x q4/8 blocks in 8x HVX vectors).
// Accumulate each block into a single int32 value.
// Return a single HVX vector with 32x int32 accumulators.
@ -167,14 +239,14 @@ static inline HVX_Vector_x8 hvx_vec_load_q8x4x8(const uint8_t * restrict ptr) {
// if() checks are optimized out at compile time -- make sure to pass N as a constexpr.
static inline HVX_Vector hvx_vec_rmpy_x8_n(HVX_Vector_x8 x, HVX_Vector_x8 y, unsigned int n) {
HVX_Vector r0 = Q6_V_vsplat_R(0);
HVX_Vector r1 = Q6_V_vsplat_R(0);
HVX_Vector r2 = Q6_V_vsplat_R(0);
HVX_Vector r3 = Q6_V_vsplat_R(0);
HVX_Vector r4 = Q6_V_vsplat_R(0);
HVX_Vector r5 = Q6_V_vsplat_R(0);
HVX_Vector r6 = Q6_V_vsplat_R(0);
HVX_Vector r7 = Q6_V_vsplat_R(0);
HVX_Vector r0 = Q6_V_vzero();
HVX_Vector r1 = Q6_V_vzero();
HVX_Vector r2 = Q6_V_vzero();
HVX_Vector r3 = Q6_V_vzero();
HVX_Vector r4 = Q6_V_vzero();
HVX_Vector r5 = Q6_V_vzero();
HVX_Vector r6 = Q6_V_vzero();
HVX_Vector r7 = Q6_V_vzero();
HVX_VectorPair p3;
HVX_VectorPair p2;
@ -213,15 +285,42 @@ static inline HVX_Vector hvx_vec_rmpy_x8_n(HVX_Vector_x8 x, HVX_Vector_x8 y, uns
}
static inline HVX_Vector hvx_vec_rmpy_x8_full(HVX_Vector_x8 x, HVX_Vector_x8 y) {
return hvx_vec_rmpy_x8_n(x, y, 1024);
HVX_Vector r0 = Q6_Vw_vrmpy_VbVb(x.v[0], y.v[0]);
HVX_Vector r1 = Q6_Vw_vrmpy_VbVb(x.v[1], y.v[1]);
HVX_Vector r2 = Q6_Vw_vrmpy_VbVb(x.v[2], y.v[2]);
HVX_Vector r3 = Q6_Vw_vrmpy_VbVb(x.v[3], y.v[3]);
HVX_Vector r4 = Q6_Vw_vrmpy_VbVb(x.v[4], y.v[4]);
HVX_Vector r5 = Q6_Vw_vrmpy_VbVb(x.v[5], y.v[5]);
HVX_Vector r6 = Q6_Vw_vrmpy_VbVb(x.v[6], y.v[6]);
HVX_Vector r7 = Q6_Vw_vrmpy_VbVb(x.v[7], y.v[7]);
HVX_VectorPair p0 = Q6_W_vdeal_VVR(r1, r0, -4);
HVX_VectorPair p1 = Q6_W_vdeal_VVR(r3, r2, -4);
HVX_VectorPair p2 = Q6_W_vdeal_VVR(r5, r4, -4);
HVX_VectorPair p3 = Q6_W_vdeal_VVR(r7, r6, -4);
r0 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p0), Q6_V_hi_W(p0));
r1 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p1), Q6_V_hi_W(p1));
r2 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p2), Q6_V_hi_W(p2));
r3 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p3), Q6_V_hi_W(p3));
p0 = Q6_W_vdeal_VVR(r1, r0, -4);
p1 = Q6_W_vdeal_VVR(r3, r2, -4);
r0 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p0), Q6_V_hi_W(p0));
r1 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p1), Q6_V_hi_W(p1));
p0 = Q6_W_vdeal_VVR(r1, r0, -4);
r0 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p0), Q6_V_hi_W(p0));
return r0;
}
// Handle most common cases of tensors not multiple of 1024.
static inline HVX_Vector hvx_vec_rmpy_x8_nloe(HVX_Vector_x8 x, HVX_Vector_x8 y, unsigned int n) {
if (n <= 256) { return hvx_vec_rmpy_x8_n(x, y, 256); };
if (n <= 512) { return hvx_vec_rmpy_x8_n(x, y, 512); };
if (n <= 768) { return hvx_vec_rmpy_x8_n(x, y, 768); };
return hvx_vec_rmpy_x8_n(x, y, 1024);
static inline HVX_Vector hvx_vec_rmpy_x8_partial(HVX_Vector_x8 x, HVX_Vector_x8 y, unsigned int n) {
if (n >= 512)
return hvx_vec_rmpy_x8_full(x, y);
return hvx_vec_rmpy_x8_partial(x, y, 512);
}
static void vec_dot_q4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const void * restrict vx0, const void * restrict vy0) {
@ -246,7 +345,7 @@ static void vec_dot_q4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const vo
const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size); // then scales
// Row sum (sf)
HVX_Vector r0_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_sum = Q6_V_vzero();
// Multiply and accumulate into int32.
// Compute combined scale (fp32).
@ -257,12 +356,12 @@ static void vec_dot_q4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const vo
uint32_t i = 0;
for (; i < nb; i++) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
@ -272,19 +371,19 @@ static void vec_dot_q4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const vo
r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
}
// Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial(y_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy_q, nloe));
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
// Zero out unused scales
// Zero out unused elements
HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
r0_dd = Q6_V_vand_QV(bmask, r0_dd);
r0_ia = Q6_V_vand_QV(bmask, r0_ia);
@ -326,8 +425,8 @@ static void vec_dot_q4x4x2_q8x4x2_2x1(const int n, float * restrict s0,
const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size); // then scales
// Row sum (sf)
HVX_Vector r0_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_sum = Q6_V_vzero();
HVX_Vector r1_sum = Q6_V_vzero();
// Multiply and accumulate into int32.
// Compute combined scale (fp32).
@ -338,14 +437,14 @@ static void vec_dot_q4x4x2_q8x4x2_2x1(const int n, float * restrict s0,
uint32_t i = 0;
for (; i < nb; i++) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8_full(r1_x_q + i * x_qblk_size);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy_q));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
@ -359,23 +458,23 @@ static void vec_dot_q4x4x2_q8x4x2_2x1(const int n, float * restrict s0,
r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
}
// Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial(y_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8_partial(r1_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy_q, nloe));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy_q, nloe));
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy_q, nloe));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
HVX_Vector r1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy_d)));
// Zero out unused scales
// Zero out unused elements
HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
r0_dd = Q6_V_vand_QV(bmask, r0_dd);
r1_dd = Q6_V_vand_QV(bmask, r1_dd);
@ -423,10 +522,10 @@ static void vec_dot_q4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
const uint8_t * restrict y1_d = ((const uint8_t *) vy1) + y_qrow_size; // then scales
// Row sums (sf) - 4 accumulators for 2×2 tile
HVX_Vector r0_c0_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_c1_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_c0_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_c1_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_c0_sum = Q6_V_vzero();
HVX_Vector r0_c1_sum = Q6_V_vzero();
HVX_Vector r1_c0_sum = Q6_V_vzero();
HVX_Vector r1_c1_sum = Q6_V_vzero();
const uint32_t nb = n / qk; // num full blocks
const uint32_t nloe = n % qk; // num leftover elements
@ -434,12 +533,12 @@ static void vec_dot_q4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
uint32_t i = 0;
for (; i < nb; i++) {
// Load src1 columns (reused across both src0 rows)
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8(y1_q + i * y_qblk_size);
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_full(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_full(y1_q + i * y_qblk_size);
// Load src0 rows (reused across both src1 columns)
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8_full(r1_x_q + i * x_qblk_size);
// Compute 4 dot products: r0×c0, r0×c1, r1×c0, r1×c1
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy0_q));
@ -448,8 +547,8 @@ static void vec_dot_q4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy1_q));
// Load scales
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
@ -473,18 +572,18 @@ static void vec_dot_q4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8(y1_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_partial(y0_q + i * y_qblk_size, nloe);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_partial(y1_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_q4x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector_x8 r1_q = hvx_vec_load_q4x4x8_partial(r1_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy0_q, nloe));
HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy1_q, nloe));
HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy0_q, nloe));
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy1_q, nloe));
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy0_q, nloe));
HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy1_q, nloe));
HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy0_q, nloe));
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy1_q, nloe));
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
@ -545,7 +644,7 @@ static void vec_dot_q8x4x2_q8x4x2_1x1(const int n, float * restrict s0, const vo
const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size); // then scales
// Row sum (sf)
HVX_Vector r0_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_sum = Q6_V_vzero();
// Multiply and accumulate into int32.
// Compute combined scale (fp32).
@ -556,12 +655,12 @@ static void vec_dot_q8x4x2_q8x4x2_1x1(const int n, float * restrict s0, const vo
uint32_t i = 0;
for (; i < nb; i++) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
@ -571,19 +670,19 @@ static void vec_dot_q8x4x2_q8x4x2_1x1(const int n, float * restrict s0, const vo
r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
}
// Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial(y_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy_q, nloe));
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
// Zero out unused scales
// Zero out unused elements
HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
r0_dd = Q6_V_vand_QV(bmask, r0_dd);
r0_ia = Q6_V_vand_QV(bmask, r0_ia);
@ -625,8 +724,8 @@ static void vec_dot_q8x4x2_q8x4x2_2x1(const int n, float * restrict s0,
const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size); // then scales
// Row sum (qf32)
HVX_Vector r0_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_sum = Q6_V_vzero();
HVX_Vector r1_sum = Q6_V_vzero();
// Multiply and accumulate into int32.
// Compute combined scale (fp32).
@ -637,14 +736,14 @@ static void vec_dot_q8x4x2_q8x4x2_2x1(const int n, float * restrict s0,
uint32_t i = 0;
for (; i < nb; i++) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8_full(r1_x_q + i * x_qblk_size);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy_q));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
@ -658,14 +757,14 @@ static void vec_dot_q8x4x2_q8x4x2_2x1(const int n, float * restrict s0,
r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
}
// Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial(y_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8_partial(r1_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy_q, nloe));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy_q, nloe));
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy_q, nloe));
HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
@ -674,7 +773,7 @@ static void vec_dot_q8x4x2_q8x4x2_2x1(const int n, float * restrict s0,
HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
HVX_Vector r1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy_d)));
// Zero out unused scales
// Zero out unused elements
HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
r0_dd = Q6_V_vand_QV(bmask, r0_dd);
r1_dd = Q6_V_vand_QV(bmask, r1_dd);
@ -722,10 +821,10 @@ static void vec_dot_q8x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
const uint8_t * restrict y1_d = ((const uint8_t *) vy1) + y_qrow_size; // then scales
// Row sums (sf) - 4 accumulators for 2×2 tile
HVX_Vector r0_c0_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_c1_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_c0_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_c1_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_c0_sum = Q6_V_vzero();
HVX_Vector r0_c1_sum = Q6_V_vzero();
HVX_Vector r1_c0_sum = Q6_V_vzero();
HVX_Vector r1_c1_sum = Q6_V_vzero();
const uint32_t nb = n / qk; // num full blocks
const uint32_t nloe = n % qk; // num leftover elements
@ -733,12 +832,12 @@ static void vec_dot_q8x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
uint32_t i = 0;
for (; i < nb; i++) {
// Load src1 columns (reused across both src0 rows)
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8(y1_q + i * y_qblk_size);
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_full(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_full(y1_q + i * y_qblk_size);
// Load src0 rows (reused across both src1 columns)
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8_full(r1_x_q + i * x_qblk_size);
// Compute 4 dot products: r0×c0, r0×c1, r1×c0, r1×c1
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy0_q));
@ -747,8 +846,8 @@ static void vec_dot_q8x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy1_q));
// Load scales
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
@ -772,18 +871,18 @@ static void vec_dot_q8x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8(y1_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_partial(y0_q + i * y_qblk_size, nloe);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_partial(y1_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_q8x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector_x8 r1_q = hvx_vec_load_q8x4x8_partial(r1_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy0_q, nloe));
HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy1_q, nloe));
HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy0_q, nloe));
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy1_q, nloe));
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy0_q, nloe));
HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy1_q, nloe));
HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy0_q, nloe));
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy1_q, nloe));
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
@ -792,7 +891,7 @@ static void vec_dot_q8x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
HVX_Vector r1_c0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy0_d)));
HVX_Vector r1_c1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy1_d)));
// Zero out unused scales
// Zero out unused elements
HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
r0_c0_dd = Q6_V_vand_QV(bmask, r0_c0_dd);
r0_c1_dd = Q6_V_vand_QV(bmask, r0_c1_dd);
@ -844,7 +943,7 @@ static void vec_dot_mxfp4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const
const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size); // then scales
// Row sum (sf)
HVX_Vector r0_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_sum = Q6_V_vzero();
// Multiply and accumulate into int32.
// Compute combined scale (fp32).
@ -855,8 +954,8 @@ static void vec_dot_mxfp4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const
uint32_t i = 0;
for (; i < nb; i++) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full( y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
@ -887,12 +986,12 @@ static void vec_dot_mxfp4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial( y_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
HVX_Vector vy_d = *(const HVX_UVector *) (y_d + i * y_dblk_size);
HVX_Vector vy_d = *(const HVX_UVector *) (y_d + i * y_dblk_size);
HVX_Vector r0_d = *(const HVX_UVector *) (r0_x_d + i * x_dblk_size);
// Convert vy_d from fp16 to fp32 while applying 0.5 scaling which is used for e8m0 halving
@ -954,8 +1053,8 @@ static void vec_dot_mxfp4x4x2_q8x4x2_2x1(const int n, float * restrict s0,
const uint8_t * restrict y_d = ((const uint8_t *) vy0) + y_qrow_size; // then scales
// Row sum (sf)
HVX_Vector r0_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_sum = Q6_V_vzero();
HVX_Vector r1_sum = Q6_V_vzero();
// Multiply and accumulate into int32.
// Compute combined scale (fp32).
@ -966,9 +1065,9 @@ static void vec_dot_mxfp4x4x2_q8x4x2_2x1(const int n, float * restrict s0,
uint32_t i = 0;
for (; i < nb; i++) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full( y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8_full(r1_x_q + i * x_qblk_size);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy_q));
@ -1007,14 +1106,14 @@ static void vec_dot_mxfp4x4x2_q8x4x2_2x1(const int n, float * restrict s0,
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8(y_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial( y_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8_partial(r1_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy_q));
HVX_Vector vy_d = *(const HVX_UVector *) (y_d + i * y_dblk_size);
HVX_Vector vy_d = *(const HVX_UVector *) (y_d + i * y_dblk_size);
HVX_Vector r0_d = *(const HVX_UVector *) (r0_x_d + i * x_dblk_size);
HVX_Vector r1_d = *(const HVX_UVector *) (r1_x_d + i * x_dblk_size);
@ -1087,10 +1186,10 @@ static void vec_dot_mxfp4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float
const uint8_t * restrict y1_d = ((const uint8_t *) vy1) + y_qrow_size; // then scales
// Row sums (sf) - 4 accumulators for 2×2 tile
HVX_Vector r0_c0_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_c1_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_c0_sum = Q6_V_vsplat_R(0);
HVX_Vector r1_c1_sum = Q6_V_vsplat_R(0);
HVX_Vector r0_c0_sum = Q6_V_vzero();
HVX_Vector r0_c1_sum = Q6_V_vzero();
HVX_Vector r1_c0_sum = Q6_V_vzero();
HVX_Vector r1_c1_sum = Q6_V_vzero();
const uint32_t nb = n / qk; // num full blocks
const uint32_t nloe = n % qk; // num leftover elements
@ -1098,12 +1197,12 @@ static void vec_dot_mxfp4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float
uint32_t i = 0;
for (; i < nb; i++) {
// Load src1 columns (reused across both src0 rows)
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8(y1_q + i * y_qblk_size);
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_full(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_full(y1_q + i * y_qblk_size);
// Load src0 rows (reused across both src1 columns)
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8_full(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8_full(r1_x_q + i * x_qblk_size);
// Compute 4 dot products: r0×c0, r0×c1, r1×c0, r1×c1
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy0_q));
@ -1157,15 +1256,15 @@ static void vec_dot_mxfp4x4x2_q8x4x2_2x2(const int n, float * restrict s0, float
// Process leftovers
if (nloe) {
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8(y0_q + i * y_qblk_size);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8(y1_q + i * y_qblk_size);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8(r0_x_q + i * x_qblk_size);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8(r1_x_q + i * x_qblk_size);
HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_partial( y0_q + i * y_qblk_size, nloe);
HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_partial( y1_q + i * y_qblk_size, nloe);
HVX_Vector_x8 r0_q = hvx_vec_load_mxfp4x4x8_partial(r0_x_q + i * x_qblk_size, nloe);
HVX_Vector_x8 r1_q = hvx_vec_load_mxfp4x4x8_partial(r1_x_q + i * x_qblk_size, nloe);
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy0_q, nloe));
HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r0_q, vy1_q, nloe));
HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy0_q, nloe));
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_nloe(r1_q, vy1_q, nloe));
HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy0_q, nloe));
HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy1_q, nloe));
HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy0_q, nloe));
HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy1_q, nloe));
HVX_Vector vy0_d = *(const HVX_UVector *) (y0_d + i * y_dblk_size);
HVX_Vector vy1_d = *(const HVX_UVector *) (y1_d + i * y_dblk_size);
@ -1234,7 +1333,7 @@ static void vec_dot_f16_f16_aa_1x1(const int n, float * restrict s, const void *
uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
uint32_t nloe = n % VLEN_FP16; // leftover elements
HVX_VectorPair rsum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair rsum_p = Q6_W_vzero();
uint32_t i = 0;
@ -1264,8 +1363,8 @@ static void vec_dot_f16_f16_aa_2x1(const int n, float * restrict s0,
uint32_t nvec = n / VLEN_FP16;
uint32_t nloe = n % VLEN_FP16;
HVX_VectorPair rsum0_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair rsum1_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair rsum0_p = Q6_W_vzero();
HVX_VectorPair rsum1_p = Q6_W_vzero();
uint32_t i = 0;
@ -1303,10 +1402,10 @@ static void vec_dot_f16_f16_aa_2x2(const int n, float * restrict s0, float * res
uint32_t nloe = n % VLEN_FP16;
// Row sums (sf) - 4 accumulators for 2×2 tile
HVX_VectorPair r0_c0_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair r0_c1_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair r1_c0_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair r1_c1_sum_p = Q6_W_vcombine_VV(Q6_V_vsplat_R(0), Q6_V_vsplat_R(0));
HVX_VectorPair r0_c0_sum_p = Q6_W_vzero();
HVX_VectorPair r0_c1_sum_p = Q6_W_vzero();
HVX_VectorPair r1_c0_sum_p = Q6_W_vzero();
HVX_VectorPair r1_c1_sum_p = Q6_W_vzero();
uint32_t i = 0;
@ -1358,7 +1457,7 @@ static void vec_dot_f16_f16_uu_1x1(const int n, float * restrict s, const void *
uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
uint32_t nloe = n % VLEN_FP16; // leftover elements
HVX_Vector rsum = Q6_V_vsplat_R(0);
HVX_Vector rsum = Q6_V_vzero();
uint32_t i = 0;
@ -1388,9 +1487,9 @@ static void vec_dot_f16_f32_uu_1x1(const int n, float * restrict s, const void *
uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
uint32_t nloe = n % VLEN_FP16; // leftover elements
const HVX_Vector zero = Q6_V_vsplat_R(0);
const HVX_Vector zero = Q6_V_vzero();
HVX_Vector rsum = Q6_V_vsplat_R(0);
HVX_Vector rsum = Q6_V_vzero();
uint32_t i = 0;
@ -1973,7 +2072,7 @@ static inline void quantize_block_f32_q8x1(float * restrict x, uint8_t * restric
assert((unsigned long) y_q % 128 == 0);
HVX_Vector * vx = (HVX_Vector *) x;
HVX_Vector zero = Q6_V_vsplat_R(0);
HVX_Vector zero = Q6_V_vzero();
// Use reduce max fp32 to find max(abs(e)) first
HVX_Vector vmax0_sf = hvx_vec_reduce_max_f32(hvx_vec_abs_f32(vx[0]));
@ -2034,7 +2133,7 @@ static inline void quantize_block_f32_q8x2(float * restrict x, uint8_t * restric
HVX_Vector * vx = (HVX_Vector *) x;
// Load and convert into QF32
HVX_Vector zero = Q6_V_vsplat_R(0);
HVX_Vector zero = Q6_V_vzero();
HVX_Vector vx0_qf = Q6_Vqf32_vsub_VsfVsf(vx[0], zero); // 32 elements
HVX_Vector vx1_qf = Q6_Vqf32_vsub_VsfVsf(vx[1], zero); // 32 elements
HVX_Vector vx2_qf = Q6_Vqf32_vsub_VsfVsf(vx[2], zero); // 32 elements
@ -2077,7 +2176,7 @@ static inline void quantize_block_f32_q8x4(float * restrict x, uint8_t * restric
HVX_Vector * vx = (HVX_Vector *) x;
// Load and convert into QF32
HVX_Vector zero = Q6_V_vsplat_R(0);
HVX_Vector zero = Q6_V_vzero();
HVX_Vector vx0_qf = Q6_Vqf32_vsub_VsfVsf(vx[0], zero); // 32 elements
HVX_Vector vx1_qf = Q6_Vqf32_vsub_VsfVsf(vx[1], zero); // 32 elements
HVX_Vector vx2_qf = Q6_Vqf32_vsub_VsfVsf(vx[2], zero); // 32 elements

148
ggml/src/ggml-hexagon/htp/repeat-ops.c Normal file
View File

@ -0,0 +1,148 @@
#pragma clang diagnostic ignored "-Wunused-variable"
#pragma clang diagnostic ignored "-Wunused-function"
#pragma clang diagnostic ignored "-Wunused-but-set-variable"
#include <HAP_farf.h>
#include <HAP_perf.h>
#include <string.h>
#include "hvx-utils.h"
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "htp-ctx.h"
#include "htp-msg.h"
#include "htp-ops.h"
struct htp_repeat_context {
struct htp_ops_context * octx;
uint32_t nr0;
uint32_t nr1;
uint32_t nr2;
uint32_t nr3;
uint32_t nrows_per_thread;
uint32_t total_dst_rows; // ne1 * ne2 * ne3
size_t type_size;
};
static void repeat_job_per_thread(unsigned int nth, unsigned int ith, void * data) {
const struct htp_repeat_context * rctx = (const struct htp_repeat_context *) data;
struct htp_ops_context * octx = rctx->octx;
const struct htp_tensor * src = &octx->src0;
const struct htp_tensor * dst = &octx->dst;
const uint32_t ne00 = src->ne[0];
const uint32_t ne01 = src->ne[1];
const uint32_t ne02 = src->ne[2];
const uint32_t ne03 = src->ne[3];
const uint32_t nb00 = src->nb[0];
const uint32_t nb01 = src->nb[1];
const uint32_t nb02 = src->nb[2];
const uint32_t nb03 = src->nb[3];
const uint32_t ne0 = dst->ne[0];
const uint32_t ne1 = dst->ne[1];
const uint32_t ne2 = dst->ne[2];
const uint32_t ne3 = dst->ne[3];
const uint32_t nb0 = dst->nb[0];
const uint32_t nb1 = dst->nb[1];
const uint32_t nb2 = dst->nb[2];
const uint32_t nb3 = dst->nb[3];
const uint32_t nr0 = rctx->nr0;
const uint32_t nr1 = rctx->nr1;
const uint32_t nr2 = rctx->nr2;
const uint32_t nr3 = rctx->nr3;
const size_t row_bytes = ne00 * rctx->type_size;
const uint32_t row_start = rctx->nrows_per_thread * ith;
const uint32_t row_end = MIN(row_start + rctx->nrows_per_thread, rctx->total_dst_rows);
uint64_t t1, t2;
t1 = HAP_perf_get_qtimer_count();
for (uint32_t dst_row = row_start; dst_row < row_end; dst_row++) {
// Decompose flat dst row index into (i1, i2, i3)
const uint32_t i1 = dst_row % ne1;
const uint32_t i2 = (dst_row / ne1) % ne2;
const uint32_t i3 = dst_row / (ne1 * ne2);
// Map to source indices (tiling)
const uint32_t k1 = i1 % ne01;
const uint32_t k2 = i2 % ne02;
const uint32_t k3 = i3 % ne03;
const uint8_t * src_row = (const uint8_t *) src->data + k1 * nb01 + k2 * nb02 + k3 * nb03;
uint8_t * dst_base = (uint8_t *) dst->data + i1 * nb1 + i2 * nb2 + i3 * nb3;
// Tile along dimension 0
for (uint32_t i0 = 0; i0 < nr0; i0++) {
uint8_t * dst_ptr = dst_base + i0 * ne00 * nb0;
memcpy(dst_ptr, src_row, row_bytes);
}
}
t2 = HAP_perf_get_qtimer_count();
FARF(HIGH, "repeat %d/%d: (%ux%ux%ux%u) -> (%ux%ux%ux%u) rows %u:%u usec %u\n",
ith, nth, src->ne[0], src->ne[1], src->ne[2], src->ne[3],
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
row_start, row_end, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
}
int op_repeat(struct htp_ops_context * octx) {
const struct htp_tensor * src0 = &octx->src0;
struct htp_tensor * dst = &octx->dst;
// Validate that dst dims are multiples of src dims
if (dst->ne[0] % src0->ne[0] != 0 ||
dst->ne[1] % src0->ne[1] != 0 ||
dst->ne[2] % src0->ne[2] != 0 ||
dst->ne[3] % src0->ne[3] != 0) {
FARF(ERROR, "repeat: dst dims must be multiples of src dims\n");
return HTP_STATUS_INVAL_PARAMS;
}
size_t type_size;
switch (src0->type) {
case HTP_TYPE_F32: type_size = 4; break;
case HTP_TYPE_F16: type_size = 2; break;
default:
FARF(ERROR, "repeat: unsupported type %u\n", src0->type);
return HTP_STATUS_NO_SUPPORT;
}
const uint32_t total_dst_rows = dst->ne[1] * dst->ne[2] * dst->ne[3];
const uint32_t n_threads = MIN(octx->n_threads, total_dst_rows);
if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
return HTP_STATUS_OK;
}
struct htp_repeat_context rctx = {
.octx = octx,
.nr0 = dst->ne[0] / src0->ne[0],
.nr1 = dst->ne[1] / src0->ne[1],
.nr2 = dst->ne[2] / src0->ne[2],
.nr3 = dst->ne[3] / src0->ne[3],
.nrows_per_thread = (total_dst_rows + n_threads - 1) / n_threads,
.total_dst_rows = total_dst_rows,
.type_size = type_size,
};
FARF(HIGH, "repeat: (%ux%ux%ux%u) -> (%ux%ux%ux%u) nr=(%u,%u,%u,%u)\n",
src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
rctx.nr0, rctx.nr1, rctx.nr2, rctx.nr3);
worker_pool_run_func(octx->ctx->worker_pool, repeat_job_per_thread, &rctx, n_threads);
return HTP_STATUS_OK;
}

2
ggml/src/ggml-hexagon/htp/softmax-ops.c
View File

@ -195,7 +195,7 @@ static float hvx_softmax_f32(const uint8_t * restrict src,
const float max) {
hvx_sub_scalar_f32(spad, src, max, num_elems);
hvx_exp_f32(spad, dst, num_elems, false);
hvx_exp_f32(dst, spad, num_elems, false);
float sum = hvx_reduce_sum_f32(dst, num_elems);

95
ggml/src/ggml-hexagon/htp/unary-ops.c
View File

@ -9,6 +9,8 @@
#include <string.h>
#include "hex-dma.h"
#include "hvx-exp.h"
#include "hvx-sigmoid.h"
#include "hvx-utils.h"
#define GGML_COMMON_DECL_C
@ -166,6 +168,75 @@ static void sqrt_f32(const float * restrict src,
}
}
static void neg_f32(const float * restrict src,
float * restrict dst,
uint8_t * restrict spad,
const uint32_t num_rows,
const uint32_t row_elems,
const size_t row_size,
int32_t * op_params) {
for (uint32_t ir = 0; ir < num_rows; ir++) {
const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
uint8_t * restrict dst_local = (uint8_t *)dst + (ir * row_size);
hvx_scale_f32_aa(dst_local, src_local, row_elems, -1.0f);
}
}
static void exp_f32(const float * restrict src,
float * restrict dst,
uint8_t * restrict spad,
const uint32_t num_rows,
const uint32_t row_elems,
const size_t row_size,
int32_t * op_params) {
for (uint32_t ir = 0; ir < num_rows; ir++) {
const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
uint8_t * restrict dst_local = (uint8_t *)dst + (ir * row_size);
hvx_exp_f32(dst_local, src_local, row_elems, false);
}
}
static void sigmoid_f32(const float * restrict src,
float * restrict dst,
uint8_t * restrict spad,
const uint32_t num_rows,
const uint32_t row_elems,
const size_t row_size,
int32_t * op_params) {
for (uint32_t ir = 0; ir < num_rows; ir++) {
const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
uint8_t * restrict dst_local = (uint8_t *)dst + (ir * row_size);
hvx_sigmoid_f32_aa(dst_local, src_local, row_elems);
}
}
static void softplus_f32(const float * restrict src,
float * restrict dst,
uint8_t * restrict spad,
const uint32_t num_rows,
const uint32_t row_elems,
const size_t row_size,
int32_t * op_params) {
// softplus(x) = log(1 + exp(x))
// Match CPU reference: ggml_compute_softplus_f32() in ggml-impl.h
for (uint32_t ir = 0; ir < num_rows; ir++) {
const float * restrict src_f = (const float *)((const uint8_t *)src + (ir * row_size));
float * restrict dst_f = (float *)((uint8_t *)dst + (ir * row_size));
for (uint32_t i = 0; i < row_elems; i++) {
float x = src_f[i];
// For x > 20: softplus(x) ≈ x (avoids exp overflow)
dst_f[i] = (x > 20.0f) ? x : logf(1.0f + expf(x));
}
}
}
static void unary_job_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
const struct htp_unary_context * uctx = (const struct htp_unary_context *) data;
struct htp_ops_context * octx = uctx->octx;
@ -247,6 +318,18 @@ static void unary_job_f32_per_thread(unsigned int nth, unsigned int ith, void *
case HTP_OP_SQRT:
sqrt_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
break;
case HTP_OP_UNARY_NEG:
neg_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
break;
case HTP_OP_UNARY_EXP:
exp_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
break;
case HTP_OP_UNARY_SIGMOID:
sigmoid_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
break;
case HTP_OP_UNARY_SOFTPLUS:
softplus_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
break;
default:
break;
}
@ -295,6 +378,18 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
case HTP_OP_SQRT:
op_type = "sqrt-f32";
break;
case HTP_OP_UNARY_NEG:
op_type = "neg-f32";
break;
case HTP_OP_UNARY_EXP:
op_type = "exp-f32";
break;
case HTP_OP_UNARY_SIGMOID:
op_type = "sigmoid-f32";
break;
case HTP_OP_UNARY_SOFTPLUS:
op_type = "softplus-f32";
break;
default:
FARF(ERROR, "Unsupported unary Op %u\n", octx->op);

1
ggml/src/ggml-metal/ggml-metal-device.m
View File

@ -1142,6 +1142,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
op->src[0]->ne[0] != 128 &&
op->src[0]->ne[0] != 192 &&
op->src[0]->ne[0] != 256 &&
op->src[0]->ne[0] != 320 &&
op->src[0]->ne[0] != 576) {
return false;
}

28
ggml/src/ggml-metal/ggml-metal.metal
View File

@ -2469,13 +2469,14 @@ kernel void kernel_gated_delta_net_impl(
const float scale = 1.0f / sqrt((float)S_v);
device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
// state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
float ls[NSG];
FOR_UNROLL (short j = 0; j < NSG; j++) {
const short is = tx*NSG + j;
ls[j] = s_ptr[is*S_v];
ls[j] = s_ptr[is];
}
device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
@ -2536,11 +2537,11 @@ kernel void kernel_gated_delta_net_impl(
g_ptr += args.ne21*G;
}
device float * dst_state = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
device float * dst_state = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
FOR_UNROLL (short j = 0; j < NSG; j++) {
const short is = tx*NSG + j;
dst_state[is*S_v] = ls[j];
dst_state[is] = ls[j];
}
#undef S_v
@ -6175,6 +6176,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_f32_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4, 1, dequantize_f32, float4x4, 1, dequantize_f32, 192, 192>;
template [[host_name("kernel_flash_attn_ext_f32_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4, 1, dequantize_f32, float4x4, 1, dequantize_f32, 192, 128>;
template [[host_name("kernel_flash_attn_ext_f32_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4, 1, dequantize_f32, float4x4, 1, dequantize_f32, 256, 256>;
template [[host_name("kernel_flash_attn_ext_f32_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4, 1, dequantize_f32, float4x4, 1, dequantize_f32, 320, 256>;
template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4, 1, dequantize_f32, float4x4, 1, dequantize_f32, 576, 512>;
template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 32, 32>;
@ -6189,6 +6191,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_f16_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 192, 192>;
template [[host_name("kernel_flash_attn_ext_f16_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 192, 128>;
template [[host_name("kernel_flash_attn_ext_f16_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 256, 256>;
template [[host_name("kernel_flash_attn_ext_f16_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 320, 256>;
template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 576, 512>;
#if defined(GGML_METAL_HAS_BF16)
@ -6204,6 +6207,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_bf16_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 192, 192>;
template [[host_name("kernel_flash_attn_ext_bf16_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 192, 128>;
template [[host_name("kernel_flash_attn_ext_bf16_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 256, 256>;
template [[host_name("kernel_flash_attn_ext_bf16_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 320, 256>;
template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 576, 512>;
#endif
@ -6219,6 +6223,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q4_0_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q4_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q4_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q4_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 320, 256>;
template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;
template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32, 32>;
@ -6233,6 +6238,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q4_1_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q4_1_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q4_1_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q4_1_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 320, 256>;
template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;
template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32, 32>;
@ -6247,6 +6253,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q5_0_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q5_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q5_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q5_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 320, 256>;
template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;
template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32, 32>;
@ -6261,6 +6268,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q5_1_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q5_1_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q5_1_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q5_1_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 320, 256>;
template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;
template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32, 32>;
@ -6275,6 +6283,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_dk128_dv128")]] kernel flash_at
template [[host_name("kernel_flash_attn_ext_q8_0_dk192_dv192")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
template [[host_name("kernel_flash_attn_ext_q8_0_dk192_dv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
template [[host_name("kernel_flash_attn_ext_q8_0_dk256_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
template [[host_name("kernel_flash_attn_ext_q8_0_dk320_dv256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 320, 256>;
template [[host_name("kernel_flash_attn_ext_q8_0_dk576_dv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
#undef FA_TYPES
@ -6845,6 +6854,17 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_dk256_dv256")]] kernel flas
template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk256_dv256")]] 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, 256, 256, 1>;
template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk256_dv256")]] 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, 256, 256, 1>;
template [[host_name("kernel_flash_attn_ext_vec_f32_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES_F32, float4, 1, dequantize_f32_t4, float4, 1, dequantize_f32_t4, 320, 256, 2>;
template [[host_name("kernel_flash_attn_ext_vec_f16_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 320, 256, 2>;
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_dk320_dv256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4, 1, dequantize_bf16_t4, bfloat4, 1, dequantize_bf16_t4, 320, 256, 2>;
#endif
template [[host_name("kernel_flash_attn_ext_vec_q4_0_dk320_dv256")]] 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, 320, 256, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q4_1_dk320_dv256")]] 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, 320, 256, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q5_0_dk320_dv256")]] 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, 320, 256, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk320_dv256")]] 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, 320, 256, 2>;
template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk320_dv256")]] 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, 320, 256, 2>;
template [[host_name("kernel_flash_attn_ext_vec_f32_dk576_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES_F32, float4, 1, dequantize_f32_t4, float4, 1, dequantize_f32_t4, 576, 512, 2>;
template [[host_name("kernel_flash_attn_ext_vec_f16_dk576_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 576, 512, 2>;
#if defined(GGML_METAL_HAS_BF16)

2
ggml/src/ggml-opencl/kernels/l2_norm.cl
View File

@ -63,7 +63,7 @@ kernel void kernel_l2_norm_f32(
barrier(CLK_LOCAL_MEM_FENCE);
const float scale = 1.0f/sqrt(max(sum[0], eps));
const float scale = 1.0f/max(sqrt(sum[0]), eps);
for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
y[i00] = x[i00] * scale;

154
ggml/src/ggml-openvino/.clang-format Normal file
View File

@ -0,0 +1,154 @@
---
# Override root .clang-format
AlignConsecutiveAssignments: false
AlignConsecutiveDeclarations: false
Cpp11BracedListStyle: true
SpacesInContainerLiterals: false
BreakBeforeBraces: Attach
AccessModifierOffset: -4
IndentCaseBlocks: false
IndentCaseLabels: false
Language: Cpp
AlignAfterOpenBracket: Align
AlignArrayOfStructures: Left
AlignConsecutiveBitFields: AcrossComments
AlignConsecutiveMacros: AcrossComments
# AlignConsecutiveShortCaseStatements: AcrossComments
AlignEscapedNewlines: Left # LeftWithLastLine
AlignOperands: Align
AlignTrailingComments:
Kind: Always
OverEmptyLines: 1
AllowAllArgumentsOnNextLine: true
AllowAllParametersOfDeclarationOnNextLine: false
# AllowBreakBeforeNoexceptSpecifier: OnlyWithParen
AllowShortBlocksOnASingleLine: Never
AllowShortCaseLabelsOnASingleLine: false
AllowShortFunctionsOnASingleLine: Inline
AllowShortIfStatementsOnASingleLine: Never
AllowShortLambdasOnASingleLine: Inline
AllowShortLoopsOnASingleLine: false
AlwaysBreakBeforeMultilineStrings: true
# Treat CUDA keywords/attributes as "attribute macros" and avoid breaking lines inside them
AttributeMacros:
- __host__
- __device__
- __global__
- __forceinline__
- __launch_bounds__
BinPackArguments: true
BinPackParameters: false # OnePerLine
BitFieldColonSpacing: Both
# BreakAdjacentStringLiterals: true
BreakAfterAttributes: Never
BreakBeforeBinaryOperators: None
BreakBeforeInlineASMColon: OnlyMultiline
BreakBeforeTernaryOperators: false
# BreakBinaryOperations: Never
BreakConstructorInitializers: AfterColon
# BreakFunctionDefinitionParameters: false
BreakInheritanceList: AfterComma
BreakStringLiterals: true
# BreakTemplateDeclarations: Yes
ColumnLimit: 120
CommentPragmas: '^ IWYU pragma:'
CompactNamespaces: false
ConstructorInitializerIndentWidth: 4
ContinuationIndentWidth: 4
DerivePointerAlignment: false
DisableFormat: false
EmptyLineBeforeAccessModifier: Leave
EmptyLineAfterAccessModifier: Never
ExperimentalAutoDetectBinPacking: false
FixNamespaceComments: true
IncludeBlocks: Regroup
IncludeCategories:
- Regex: '".*"'
Priority: 1
SortPriority: 0
- Regex: '^<.*\.h>'
Priority: 2
SortPriority: 0
- Regex: '^<.*'
Priority: 3
SortPriority: 0
- Regex: '.*'
Priority: 4
SortPriority: 0
IncludeIsMainRegex: '([-_](test|unittest))?$'
IncludeIsMainSourceRegex: ''
IndentAccessModifiers: false
IndentExternBlock: NoIndent
IndentGotoLabels: false
IndentPPDirectives: AfterHash
IndentWidth: 4
IndentWrappedFunctionNames: false
InsertBraces: true # NOTE: may lead to incorrect formatting
InsertNewlineAtEOF: true
JavaScriptQuotes: Leave
JavaScriptWrapImports: true
KeepEmptyLinesAtTheStartOfBlocks: false
LambdaBodyIndentation: Signature
LineEnding: LF
MacroBlockBegin: ''
MacroBlockEnd: ''
MaxEmptyLinesToKeep: 1
NamespaceIndentation: None
ObjCBinPackProtocolList: Auto
ObjCBlockIndentWidth: 4
ObjCSpaceAfterProperty: true
ObjCSpaceBeforeProtocolList: true
PPIndentWidth: -1
PackConstructorInitializers: CurrentLine
PenaltyBreakAssignment: 2
PenaltyBreakBeforeFirstCallParameter: 1
PenaltyBreakComment: 300
PenaltyBreakFirstLessLess: 120
PenaltyBreakString: 1000
PenaltyBreakTemplateDeclaration: 10
PenaltyExcessCharacter: 1000000
PenaltyReturnTypeOnItsOwnLine: 200
PointerAlignment: Middle
QualifierAlignment: Left
#QualifierOrder: ['static', 'inline', 'friend', 'constexpr', 'const', 'volatile', 'type', 'restrict']
RawStringFormats:
- Language: Cpp
Delimiters:
- cc
- CC
- cpp
- Cpp
- CPP
- 'c++'
- 'C++'
CanonicalDelimiter: ''
ReferenceAlignment: Middle
ReflowComments: false # IndentOnly
SeparateDefinitionBlocks: Always
SortIncludes: CaseInsensitive
SortUsingDeclarations: LexicographicNumeric
SpaceAfterCStyleCast: true
SpaceAfterLogicalNot: false
SpaceAfterTemplateKeyword: true
SpaceBeforeAssignmentOperators: true
SpaceBeforeCpp11BracedList: false
SpaceBeforeCtorInitializerColon: true
SpaceBeforeInheritanceColon: true
SpaceBeforeParens: ControlStatements
SpaceBeforeRangeBasedForLoopColon: true
SpaceInEmptyBlock: false
SpaceInEmptyParentheses: false
SpacesBeforeTrailingComments: 2
SpacesInAngles: Never
SpacesInLineCommentPrefix:
Minimum: 1
Maximum: -1
SpacesInParentheses: false
SpacesInSquareBrackets: false
SpaceBeforeSquareBrackets: false
Standard: c++17
TabWidth: 4
UseTab: Never
WhitespaceSensitiveMacros: ['STRINGIZE']
...

22
ggml/src/ggml-openvino/CMakeLists.txt Normal file
View File

@ -0,0 +1,22 @@
find_package(OpenVINO REQUIRED)
find_package(OpenCL REQUIRED)
include("${OpenVINO_DIR}/../3rdparty/tbb/lib/cmake/TBB/TBBConfig.cmake")
file(GLOB_RECURSE GGML_HEADERS_OPENVINO "*.h" "*.hpp")
file(GLOB_RECURSE GGML_SOURCES_OPENVINO "*.cpp")
ggml_add_backend_library(ggml-openvino
${GGML_SOURCES_OPENVINO}
${GGML_HEADERS_OPENVINO}
)
target_link_libraries(ggml-openvino PRIVATE openvino::runtime TBB::tbb OpenCL::OpenCL)
if (GGML_OPENVINO)
if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "AMD64")
else()
message(FATAL_ERROR "OpenVINO: OpenVINO toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
endif()
endif()

975
ggml/src/ggml-openvino/ggml-decoder.cpp Normal file
View File

@ -0,0 +1,975 @@
#include "ggml-decoder.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-openvino-extra.h"
#include "ggml-openvino.h"
#include "ggml-quants.h"
#include <ggml-impl.h>
#include <ggml.h>
#include <algorithm>
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <cstdlib>
#include <execution>
#include <fstream>
#include <iomanip>
#include <map>
#include <memory>
#include <mutex>
#include <openvino/core/dimension.hpp>
#include <openvino/core/except.hpp>
#include <openvino/core/node.hpp>
#include <openvino/core/partial_shape.hpp>
#include <openvino/core/type/bfloat16.hpp>
#include <openvino/core/type/element_type.hpp>
#include <openvino/core/type/float16.hpp>
#include <openvino/op/constant.hpp>
#include <openvino/op/convert.hpp>
#include <openvino/op/parameter.hpp>
#include <openvino/runtime/tensor.hpp>
#include <optional>
#include <ostream>
#include <set>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <vector>
GgmlOvDecoder::GgmlOvDecoder(ggml_cgraph * cgraph,
ModelParams & model_params,
ComputeParams & compute_params,
std::map<std::string, std::shared_ptr<ov::Node>> & model_weights,
bool is_static,
bool is_stateful,
bool is_prefill,
int prefill_chunk_size) :
m_is_static(is_static),
m_is_stateful(is_stateful),
m_is_prefill(is_prefill),
m_naive(false),
m_prefill_chunk_size(prefill_chunk_size),
m_cgraph(cgraph),
m_model_weights(model_weights),
m_model_params(model_params),
m_compute_params(compute_params) {
if (auto * env = getenv("GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS"); env && std::string(env) != "0") {
#ifdef _WIN32
_putenv_s("GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS", "");
#else
unsetenv("GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS");
#endif
print_tensor_address_map(cgraph);
}
validate_cgraph();
set_input_output();
compute_model_inputs();
compute_model_outputs();
for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
m_node_info_list[node_n].node_op_case = compute_op_case(m_node_info_list[node_n].node);
m_node_info_list[node_n].node_op_type = compute_op_type(m_node_info_list[node_n].node);
}
add_extra_inputs();
}
void GgmlOvDecoder::update_io(ggml_cgraph * cgraph) {
m_cgraph = cgraph;
m_model_inputs.clear();
m_model_outputs.clear();
m_node_info_list.clear();
set_input_output();
compute_model_inputs();
compute_model_outputs();
}
GgmlOvDecoder::GgmlOvDecoder(ggml_cgraph * cgraph, std::map<std::string, std::shared_ptr<ov::Node>> & model_weights) {
m_cgraph = cgraph;
m_model_weights = model_weights;
m_naive = true;
set_input_output();
compute_model_inputs();
compute_model_outputs();
for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
m_node_info_list[node_n].node_op_case = compute_op_case(m_node_info_list[node_n].node);
m_node_info_list[node_n].node_op_type = compute_op_type(m_node_info_list[node_n].node);
}
}
void GgmlOvDecoder::set_input_output() {
for (int node_n = 0; node_n < m_cgraph->n_nodes; node_n++) {
auto node = m_cgraph->nodes[node_n];
NodeInfo current_node_info;
auto node_name = std::string(node->name);
auto node_output_name = node_name;
auto * node_output = node;
if (node->op == GGML_OP_SET_ROWS) {
// SET_ROWS updates the tensor in place. For later ov op that uses the
// the view_src of SET_ROWS, we need to make sure they get the updated tensor
// by putting the view_src name in the tensor_map in
// <openvino>/src/frontends/ggml/src/translate_session.cpp
node_output_name = std::string(node->view_src->name);
node_output = node->view_src;
}
current_node_info.node = node;
current_node_info.node_name = node_name;
current_node_info.node_output = node_output;
current_node_info.node_output_name = node_output_name;
current_node_info.node_op_case = 0;
current_node_info.data_addr = node->data;
for (int i = 0; i < GGML_MAX_SRC; i++) {
auto * src = node->src[i];
if (src == nullptr) {
continue;
}
auto src_name = std::string(src->name);
if (src->flags & GGML_TENSOR_FLAG_INPUT) {
src_name = get_graph_input_ov_name(src, node);
}
current_node_info.node_inputs[src_name] = src;
current_node_info.node_inputs_names.push_back(src_name);
}
m_node_info_list.push_back(current_node_info);
}
}
int GgmlOvDecoder::compute_op_case(const ggml_tensor * node) const {
int op_case = 0;
switch (node->op) {
case GGML_OP_RESHAPE: {
auto * src = node->src[0];
if (src->op == GGML_OP_RESHAPE && src->src[0]->ne[0] == node->ne[0] && src->src[0]->ne[1] == node->ne[1]) {
op_case = 4;
} else if (node->ne[0] * node->ne[1] == src->ne[0]) {
op_case = 1;
} else if (src->ne[0] * src->ne[1] == node->ne[0]) {
op_case = 2;
if (src->ne[2] * src->ne[3] == node->ne[1]) {
op_case = 5;
}
} else if (src->ne[0] * src->ne[1] == node->ne[1]) {
op_case = 3;
} else if (src->ne[1] * src->ne[2] == node->ne[1]) {
op_case = 6;
}
break;
}
case GGML_OP_CONT: {
if (node->src[0]->op == GGML_OP_PERMUTE) {
op_case = 1;
} else if (node->src[0]->op == GGML_OP_TRANSPOSE) {
op_case = 2;
} else if (node->src[0]->op == GGML_OP_VIEW) {
op_case = 3;
}
break;
}
case GGML_OP_PERMUTE: {
if (node->src[0]->op != GGML_OP_VIEW) {
op_case = 1;
} else if (node->src[0]->src[0]->op == GGML_OP_NONE) {
// kv cache tensor
std::string src_name(node->view_src->name);
int layer = extract_layer_from_name(src_name);
if (!is_swa_layer(layer)) {
op_case = 2;
} else {
op_case = 3;
}
} else {
// rope'ed query tensor
op_case = 4;
}
break;
}
case GGML_OP_MUL_MAT: {
if (node->src[0]->op == GGML_OP_CONT && node->src[0]->src[0]->op == GGML_OP_TRANSPOSE) {
op_case = 2;
} else if (node->src[0]->op == GGML_OP_VIEW && node->src[1]->op == GGML_OP_VIEW) {
op_case = 3;
}
break;
}
case GGML_OP_GET_ROWS: {
if (node->src[1]->op == GGML_OP_VIEW) {
op_case = 2;
}
break;
}
case GGML_OP_ROPE: {
if (node->src[0]->op == GGML_OP_VIEW) {
op_case = 2;
}
break;
}
case GGML_OP_VIEW: {
if (node->src[0]->op == GGML_OP_VIEW) {
auto * src = node->src[0];
if (ggml_nelements(node) != ggml_nelements(src)) {
throw std::runtime_error("Unsupported VIEW case");
}
op_case = 2;
}
{
auto * src = node->src[0];
if ((ggml_nelements(node) != ggml_nelements(src)) && m_naive) {
// Compare each dimension of node and src, if only one dimension differs then op_case=3
int diff_count = 0;
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (node->ne[i] != src->ne[i]) {
diff_count++;
}
}
if (diff_count == 1) {
op_case = 3;
}
}
}
break;
}
default:
break;
}
return op_case;
}
int extract_layer_from_name(const std::string & name) {
size_t pos1 = name.find("_l");
assert(pos1 != std::string::npos);
pos1 += 2;
size_t pos2 = name.find(' ', pos1);
if (pos2 == std::string::npos) {
pos2 = name.length();
}
std::string layer_str = name.substr(pos1, pos2 - pos1);
int layer = std::stoi(layer_str);
return layer;
}
std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgraph * cgraph, bool is_static) {
ModelParams model_params;
ComputeParams compute_params;
for (int i = 0; i < cgraph->n_nodes; i++) {
auto * node = cgraph->nodes[i];
std::string name = std::string(node->name);
if (node->op == GGML_OP_FLASH_ATTN_EXT) {
model_params.n_heads = node->src[0]->ne[2];
model_params.n_heads_kv = node->src[1]->ne[2];
model_params.head_size = node->src[0]->ne[0];
compute_params.input_len = node->src[0]->ne[1];
auto * cache_k_perm = node->src[1];
if (cache_k_perm->op == GGML_OP_CPY) {
cache_k_perm = cache_k_perm->src[0];
}
assert(cache_k_perm->op == GGML_OP_PERMUTE);
auto * cache_k_view = cache_k_perm->src[0];
assert(cache_k_view->op == GGML_OP_VIEW);
auto * cache_k = cache_k_view->src[0];
int layer = extract_layer_from_name(cache_k->name);
auto * mask = node->src[3];
std::string mask_name(mask->name);
model_params.kv_buffer_ctx_id = ggml_backend_openvino_buffer_get_ctx_id(cache_k->buffer);
if (mask_name.find("swa") != std::string::npos) {
model_params.swa_layers.push_back(layer);
model_params.ctx_per_seq_swa = cache_k->ne[1];
} else {
model_params.ctx_per_seq = cache_k->ne[1];
model_params.n_seq = cache_k->ne[2];
}
compute_params.n_seq_active = mask->ne[3];
auto seq_size = cache_k->ne[0] * cache_k->ne[1] * ggml_type_size(cache_k->type);
size_t offset;
memcpy(&offset, cache_k_view->op_params, sizeof(size_t));
compute_params.seq_active_start = offset / seq_size;
compute_params.token_len_per_seq = node->ne[2];
if (mask_name.find("swa") != std::string::npos) {
compute_params.attention_size_swa = mask->ne[0];
} else {
compute_params.attention_size = mask->ne[0];
}
if (is_static) {
compute_params.attention_size = model_params.ctx_per_seq;
compute_params.attention_size_swa = model_params.ctx_per_seq_swa;
compute_params.token_len_per_seq = 1;
}
break;
}
if (node->op == GGML_OP_ROPE) {
memcpy(model_params.rope_params, node->op_params, sizeof(int32_t) * 15);
}
}
auto * output_tensor = cgraph->nodes[cgraph->n_nodes - 1];
compute_params.output_len = output_tensor->ne[1];
// for NPU, output_len is always 1 except for llama-perplexity
if (is_static && compute_params.output_len == 0) {
compute_params.output_len = 1;
}
model_params.ctx = model_params.ctx_per_seq * model_params.n_seq;
model_params.ctx_swa = model_params.ctx_per_seq_swa * model_params.n_seq;
return {model_params, compute_params};
}
void GgmlOvDecoder::validate_cgraph() const {
if (m_model_params.n_seq > 1 && m_is_static == true) {
throw std::runtime_error("n_seq > 1 is not supported on NPU. Try setting -np 1.");
}
}
ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op, const ggml_tensor * input) const {
if (m_naive) {
return input!= nullptr ? ov::PartialShape{get_shape(input)} : ov::PartialShape{get_shape(op)};
}
auto name = std::string(input->name);
ov::PartialShape input_shape;
if (is_inp_tok(input, op) || is_inp_pos(input, op)) {
// tokens or positions
int len = m_is_static ? (m_is_prefill ? m_prefill_chunk_size : 1) : -1;
input_shape = ov::PartialShape{1, 1, 1, len};
} else if (is_output_idx(input, op)) {
// output index
input_shape = ov::PartialShape{1, 1, 1, m_is_static ? m_compute_params.output_len : -1};
} else if (is_inp_mask(input, op)) {
// mask
if (m_is_static) {
input_shape = ov::PartialShape{1, 1, m_is_prefill ? m_prefill_chunk_size : 1, m_model_params.ctx};
} else if (m_is_stateful) {
input_shape = ov::PartialShape{1, 1, -1, -1};
} else {
input_shape = ov::PartialShape{-1, 1, -1, -1};
}
} else if (is_kvcache(input, op)) {
// kvcache
input_shape = ov::PartialShape{get_shape(input)};
if (!m_is_static) {
// do not fix ctx size to make llama-bench work across test params
input_shape[2] = -1;
}
if (is_stateful()) {
// Convert stateless KV cache layout [1, 1, seq, n_heads_kv * head_size]
// to stateful layout [1, seq, n_heads_kv, head_size].
assert(input_shape.size() == 4 && input_shape[0] == 1 && input_shape[1] == 1 &&
input_shape[2].is_dynamic() &&
input_shape[3] == (m_model_params.n_heads_kv * m_model_params.head_size));
input_shape = {input_shape[0], ov::Dimension::dynamic(), m_model_params.n_heads_kv,
m_model_params.head_size};
}
} else if (is_kv_idx(input, op)) {
// kv update index
int len = m_is_static ? (m_is_prefill ? m_prefill_chunk_size : 1) : -1;
input_shape = ov::PartialShape{1, 1, 1, len};
} else {
input_shape = ov::PartialShape{get_shape(input)};
}
return input_shape;
}
void GgmlOvDecoder::add_extra_inputs() {
// Extra inputs:
// 1. `attention_size`, used in FLASH_ATTN where the shape of the matmul's are 256 aligned,
// see llama_kv_cache_unified::get_n_kv and llama_kv_cache_unified::get_padding.
// 2. `n_seq_active` and `seq_active_start`, used in FLASH_ATTN_EXT to indicate the active sequences in the batch
auto create_1d_input = [this](const std::string & name, int64_t value) {
if (m_is_static) {
auto constant =
std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{1}, std::vector<int64_t>{value});
constant->set_friendly_name(name);
m_model_extra_inputs[name] = constant;
} else {
auto param_node = std::make_shared<ov::op::v0::Parameter>(ov::element::i64, ov::Shape{1});
param_node->set_friendly_name(name);
param_node->output(0).get_tensor().set_names({name});
m_model_extra_inputs[name] = param_node;
auto tensor = std::make_shared<ov::Tensor>(ov::element::i64, ov::Shape{1});
*tensor->data<int64_t>() = value;
m_model_extra_input_values[name] = tensor;
}
};
create_1d_input("attention_size", m_compute_params.attention_size);
if (m_compute_params.attention_size_swa != -1) {
create_1d_input("attention_size_swa", m_compute_params.attention_size_swa);
}
create_1d_input("n_seq_active", m_compute_params.n_seq_active);
create_1d_input("seq_active_start", m_compute_params.seq_active_start);
create_1d_input("seq_active_end", m_compute_params.seq_active_start + m_compute_params.n_seq_active);
create_1d_input("token_len_per_seq", m_compute_params.token_len_per_seq);
// create_1d_input("token_len", m_token_len_per_seq * m_n_seq_active);
}
bool GgmlOvDecoder::node_is_used_as_src(const int node_idx) {
ggml_tensor * node = m_cgraph->nodes[node_idx];
for (int i = node_idx; i < m_cgraph->n_nodes; i++) {
ggml_tensor * other_node = m_cgraph->nodes[i];
for (int j = 0; j < GGML_MAX_SRC; j++) {
if (other_node->src[j] == node) {
return true;
}
}
}
return false;
}
void GgmlOvDecoder::compute_model_inputs() {
m_model_inputs.clear();
m_inputs.clear();
for (int i = 0; i < m_cgraph->n_nodes; i++) {
ggml_tensor * node = m_cgraph->nodes[i];
// the node op is NONE means this node maybe as input of later nodes, we should add it to model inputs for this node.
if (node->op == GGML_OP_NONE && node_is_used_as_src(i)) {
std::string node_name(node->name);
if (m_model_weights.find(node_name) == m_model_weights.end()) {
m_inputs[node_name] = node;
auto param_node =
std::make_shared<ov::op::v0::Parameter>(get_ov_type(node), get_graph_input_shape(node, nullptr));
param_node->set_friendly_name(node_name);
param_node->output(0).get_tensor().set_names({node_name});
m_model_inputs[node_name] = param_node;
}
continue;
}
for (int i = 0; i < GGML_MAX_SRC; i++) {
auto * src = node->src[i];
if (src == nullptr) {
continue;
}
std::string src_name = std::string(src->name);
if (src->flags & GGML_TENSOR_FLAG_INPUT) {
src_name = get_graph_input_ov_name(src, node);
}
if (m_model_weights.find(src_name) != m_model_weights.end()) {
continue;
}
bool is_intermediate_node = false;
for (const auto & node_info : m_node_info_list) {
if (node_info.node == src) {
is_intermediate_node = true;
break;
}
}
if (is_intermediate_node) {
continue;
}
if (m_model_inputs.find(src_name) != m_model_inputs.end()) {
continue;
}
m_inputs[src_name] = src;
ggml_backend_buffer * buffer = src->buffer;
// GGML_BACKEND_BUFFER_USAGE_ANY are kv caches
if (buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY) {
if (auto it = std::find(m_model_params.kv_names.begin(), m_model_params.kv_names.end(), src_name);
it == m_model_params.kv_names.end()) {
m_model_params.kv_names.push_back(src_name);
}
}
ov::PartialShape param_shape = get_graph_input_shape(node, src);
auto param_node = std::make_shared<ov::op::v0::Parameter>(get_ov_type(src), param_shape);
param_node->set_friendly_name(src_name);
param_node->output(0).get_tensor().set_names({src_name});
m_model_inputs[src_name] = param_node;
}
}
}
void GgmlOvDecoder::compute_model_outputs() {
m_model_outputs.clear();
m_model_output_names.clear();
for (int node_n = 0; node_n < m_cgraph->n_nodes; node_n++) {
auto * cur_node = m_cgraph->nodes[node_n];
// if the node op is NONE means this node is not used at all, we can skip it directly without adding to model outputs.
if (cur_node->op == GGML_OP_NONE) {
continue;
}
auto cur_node_use_count = m_cgraph->use_counts[ggml_hash_find(&m_cgraph->visited_hash_set, cur_node)];
if (cur_node_use_count == 0) {
// The output of SET_ROWS is the view_src tensor, which is updated in place. We should use the view_src name as the output name to make sure it can be correctly matched with the later ops that use the view_src.
if (cur_node != nullptr && cur_node->op == GGML_OP_SET_ROWS) {
cur_node = cur_node->view_src;
}
} else {
int input_use_count = 0;
for (int i = 0; i < m_cgraph->n_nodes; i++) {
ggml_tensor * node = m_cgraph->nodes[i];
for (int j = 0; j < GGML_MAX_SRC; j++) {
if (node->src[j] != NULL && node->src[j] == cur_node) {
input_use_count++;
}
}
}
if (input_use_count == cur_node_use_count) {
cur_node = nullptr;
}
}
if (cur_node != nullptr) {
std::string node_output_name(cur_node->name);
m_model_outputs[node_output_name] = cur_node;
m_model_output_names.push_back(node_output_name);
}
}
}
const ggml_tensor * GgmlOvDecoder::get_tensor_used_op(const ggml_tensor * tensor) const {
if (tensor == nullptr) {
return nullptr;
}
for (int i = 0; i < m_cgraph->n_nodes; i++) {
const auto * node = m_cgraph->nodes[i];
for (int j = 0; j < GGML_MAX_SRC; j++) {
if (node->src[j] == tensor) {
return node;
}
}
}
return nullptr;
}
const ggml_tensor * GgmlOvDecoder::get_tensor_from_name(const std::string & name) const {
for (int i = 0; i < m_cgraph->n_nodes; i++) {
const auto * node = m_cgraph->nodes[i];
for (int j = 0; j < GGML_MAX_SRC; j++) {
const auto * src = node->src[j];
if (src == nullptr) {
break;
}
if (std::string(src->name) == name) {
return src;
}
}
}
return nullptr;
}
std::map<std::string, std::string> GgmlOvDecoder::get_kv_param_res_names() const {
std::map<std::string, std::string> kv_param_res_names;
for (const auto & name : m_model_params.kv_names) {
kv_param_res_names[name] = name;
}
return kv_param_res_names;
}
std::map<std::string, std::shared_ptr<ov::Node>> GgmlOvDecoder::create_weight_nodes(ggml_cgraph * cgraph, bool naive) {
static std::mutex weights_mutex;
std::lock_guard<std::mutex> lock(weights_mutex);
std::map<std::string, std::shared_ptr<ov::Node>> model_weights;
auto * nodes = cgraph->nodes;
auto n_nodes = cgraph->n_nodes;
for (int node_i = 0; node_i < n_nodes; node_i++) {
auto * node = nodes[node_i];
for (int i = 0; i < GGML_MAX_SRC; i++) {
auto * src = node->src[i];
if (src == nullptr) {
continue;
}
std::string src_name(src->name);
if (is_rope_freqs_weight(src, node)) {
src_name = "rope_freqs.weight";
}
if (!src->view_src) {
ggml_backend_buffer * buffer = src->buffer;
if (buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS || ggml_is_quantized(src->type)) {
if (model_weights.find(src_name) == model_weights.end()) {
auto weight_node = create_weight_node(src, naive);
weight_node->set_friendly_name(src_name);
model_weights[src_name] = weight_node;
}
}
}
}
}
return model_weights;
}
std::shared_ptr<ov::Node> GgmlOvDecoder::create_weight_node(ggml_tensor * tensor, bool naive) {
const bool is_ov_buffer = ggml_backend_buffer_is_openvino(tensor->buffer);
// Check if we have a pre-built constant from the OpenVINO backend buffer
// This is set during ggml_backend_openvino_buffer_set_tensor
if (tensor->extra) {
OPENVINO_ASSERT(is_ov_buffer, "Unsupported weight tensor: " + std::string(tensor->name) +
" Possibly this is a cpu backend repacked quantized weights");
// Cast to our extra base type and check the type
auto * extra_base = static_cast<ggml_openvino_extra_base *>(tensor->extra);
if (extra_base->type == ggml_openvino_extra_base::Type::WEIGHT) {
// F16/F32/BF16 weight with shared-memory constant
auto * weight_extra = static_cast<ggml_openvino_weight_extra *>(tensor->extra);
if (weight_extra->weight_node) {
// GGML_LOG_DEBUG("%s: using pre-built weight node for %s\n", __func__, tensor->name);
return weight_extra->weight_node;
}
} else if (extra_base->type == ggml_openvino_extra_base::Type::QUANTIZED_WEIGHT) {
// Quantized weight with pre-extracted data
auto * quant_extra = static_cast<ggml_openvino_quantized_weight_extra *>(tensor->extra);
if (quant_extra->weight_node) {
// GGML_LOG_DEBUG("%s: using pre-extracted quantized weight node for %s\n", __func__, tensor->name);
return quant_extra->weight_node;
}
}
}
// There are three cases where we need to create a new weight node:
// 1. weights are in openvino_host_buffer. Weight loading to host buffer will not trigger backend_buffer_set_tensor
// 2. weights are in cpu/cpu_mapped buffer. On token_embd.weight goes to case 1 or 2, depending on whether mmap or direct_io is used
// 3. test-backend-ops. buffers in test-backend-ops does not set USAGE_WEIGHT so backend_buffer_set_tensor will not create weight node
// GGML_LOG_DEBUG("%s: creating new weight node for %s\n", __func__, tensor->name);
static const std::set<ggml_type> weight_types = {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16,
GGML_TYPE_Q8_0, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
GGML_TYPE_Q4_K, GGML_TYPE_Q5_K, GGML_TYPE_Q6_K};
if (weight_types.find(tensor->type) == weight_types.end()) {
throw std::runtime_error("Unexpected weight tensor type: " + std::string(tensor->name) + " with type " +
ggml_type_name(tensor->type));
}
OvWeight ov_weight;
if (ggml_is_quantized(tensor->type)) {
auto use_bias = naive;
if (is_ov_buffer) {
// For quantized weights, copy raw data to a temp buffer first because
// process_weight_tensor reads from data and writes extracted results
// (weights/scales/zp) to output_base_ptr — they would overlap if both
// point to tensor->data.
size_t raw_size = ggml_nbytes(tensor);
std::vector<uint8_t> tmp(raw_size);
memcpy(tmp.data(), tensor->data, raw_size);
ov_weight = process_weight_tensor(tensor, tmp.data(), tensor->data, use_bias);
} else {
ov_weight = process_weight_tensor(tensor, tensor->data, nullptr, use_bias);
}
} else {
// For non-quantized weights (F16/F32/BF16), data is already in tensor->data.
// process_weight_tensor will create an ov::Tensor wrapping tensor->data directly.
ov_weight = process_weight_tensor(tensor, tensor->data, tensor->data);
}
ov_weight.weight_node->set_friendly_name(tensor->name);
if (!is_ov_buffer) {
return ov_weight.weight_node;
}
ggml_openvino_extra_base * extra;
if (ov_weight.is_quantized()) {
extra = new ggml_openvino_quantized_weight_extra(std::move(ov_weight.weights), std::move(ov_weight.scales),
std::move(ov_weight.zp), ov_weight.weight_node);
} else {
extra = new ggml_openvino_weight_extra(std::move(ov_weight.weights), ov_weight.weight_node);
}
ggml_openvino_buffer_register_extra(tensor, extra);
return ov_weight.weight_node;
}
void GgmlOvDecoder::dump_cgraph(const ggml_cgraph * cgraph, std::string & filename) {
std::ofstream file(filename);
if (!file.is_open()) {
std::cerr << "Failed to open file" << std::endl;
return;
}
file << "=== GRAPH ===\n";
// clang-format off
file << "n_nodes = " << cgraph->n_nodes << "\n";
file << " " << std::setw(3) << "nodes"
<< std::setw(15) << "shape"
<< std::setw(20) << "op"
<< std::setw(20) << "name"
<< std::setw(3) << " "
<< std::setw(62) << "stride"
<< std::setw(20) << "buffer_type"
<< "\n";
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
// Get buffer type name
const char * buf_name = "none";
ggml_backend_buffer_t buf = node->view_src ? node->view_src->buffer : node->buffer;
if (buf) {
buf_name = ggml_backend_buffer_name(buf);
}
file << " - " << std::setw(3) << i << ": [ "
<< std::setw(5) << node->ne[0] << ", "
<< std::setw(5) << node->ne[1] << ", "
<< std::setw(5) << node->ne[2] << ", "
<< std::setw(5) << node->ne[3] << "] "
<< std::left << std::setw(20) << ggml_op_name(node->op) << std::right << " "
<< std::left << std::setw(45) << node->name << std::right
<< std::setw(2) << "[ "
<< std::setw(0) << node->nb[0] << ", "
<< std::setw(5) << node->nb[1] << ", "
<< std::setw(5) << node->nb[2] << ", "
<< std::setw(5) << node->nb[3] << "] "
<< std::right << std::setw(15) << buf_name << std::right
<< "\n";
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (auto* src = node->src[i]) {
// Get buffer type name for source
const char * src_buf_name = "none";
ggml_backend_buffer_t src_buf = src->view_src ? src->view_src->buffer : src->buffer;
if (src_buf) {
src_buf_name = ggml_backend_buffer_name(src_buf);
}
file << std::setw(10) << " [ "
<< std::setw(5) << src->ne[0] << ", "
<< std::setw(5) << src->ne[1] << ", "
<< std::setw(5) << src->ne[2] << ", "
<< std::setw(5) << src->ne[3] << "] "
<< std::setw(12)
<< i << ": " << std::left << std::setw(12) << ggml_op_name(src->op) << std::right;
file << std::left << std::setw(30) << src->name << std::right
<< std::setw(16) << "[ "
<< std::setw(0) << src->nb[0] << ", "
<< std::setw(5) << src->nb[1] << ", "
<< std::setw(5) << src->nb[2] << ", "
<< std::setw(5) << src->nb[3] << "] "
<< std::right << std::setw(15) << src_buf_name << std::right
<< "\n";
}
}
}
file << "n_leafs = " << cgraph->n_leafs << "\n";
for (int i = 0; i < cgraph->n_leafs; i++) {
ggml_tensor * node = cgraph->leafs[i];
// Get buffer type name for leaf
const char * leaf_buf_name = "none";
ggml_backend_buffer_t leaf_buf = node->view_src ? node->view_src->buffer : node->buffer;
if (leaf_buf) {
leaf_buf_name = ggml_backend_buffer_name(leaf_buf);
}
file << " - " << std::setw(3) << i << ": [ "
<< std::setw(5) << node->ne[0] << ", "
<< std::setw(5) << node->ne[1] << "] "
<< std::setw(8) << ggml_op_name(node->op) << " "
<< std::setw(16) << ggml_get_name(node)
<< std::setw(20) << leaf_buf_name << "\n";
}
// clang-format on
file << "========================================\n";
file.close();
}
void print_tensor_address_map(const ggml_cgraph * cgraph) {
std::map<void *, std::vector<std::string>> address_map;
for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
auto * node = cgraph->nodes[node_n];
if (node->data) {
auto it = address_map.find(node->data);
if (it == address_map.end()) {
address_map[node->data] = std::vector<std::string>();
}
address_map[node->data].push_back(node->name);
}
}
for (const auto & pair : address_map) {
std::cout << "Address: " << pair.first << std::endl;
for (const auto & name : pair.second) {
std::cout << name << " ; ";
}
std::cout << std::endl << std::endl;
}
}
ov::Shape GgmlOvDecoder::get_shape(const ggml_tensor * tensor) {
std::vector<size_t> shape;
for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
shape.push_back(static_cast<size_t>(tensor->ne[i]));
}
return shape;
}
std::vector<size_t> GgmlOvDecoder::get_stride(const ggml_tensor * tensor) {
std::vector<size_t> stride;
for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
stride.push_back(static_cast<size_t>(tensor->nb[i]));
}
return stride;
}
ov::element::Type GgmlOvDecoder::get_ov_type(const ggml_tensor * tensor) {
switch (tensor->type) {
case GGML_TYPE_F64:
return ov::element::f64;
case GGML_TYPE_F32:
return ov::element::f32;
case GGML_TYPE_F16:
return ov::element::f16;
case GGML_TYPE_BF16:
return ov::element::bf16;
case GGML_TYPE_I8:
return ov::element::i8;
case GGML_TYPE_I16:
return ov::element::i16;
case GGML_TYPE_I32:
return ov::element::i32;
case GGML_TYPE_I64:
return ov::element::i64;
default:
return ov::element::dynamic;
}
}
ov::PartialShape GgmlOvDecoder::get_input_shape(int node_idx, const std::string & name) const {
return ov::PartialShape(get_shape(m_node_info_list[node_idx].node_inputs.at(name)));
}
std::vector<size_t> GgmlOvDecoder::get_input_stride(int node_idx, const std::string & name) const {
return get_stride(m_node_info_list[node_idx].node_inputs.at(name));
}
ov::element::Type GgmlOvDecoder::get_input_type(int node_idx, const std::string & name) const {
return get_ov_type(m_node_info_list[node_idx].node_inputs.at(name));
}
size_t GgmlOvDecoder::get_input_size() const {
return m_model_inputs.size();
}
size_t GgmlOvDecoder::get_input_size(int node_idx) const {
return m_node_info_list[node_idx].node_inputs_names.size();
}
std::vector<std::string> GgmlOvDecoder::get_input_names(int node_idx) const {
return m_node_info_list[node_idx].node_inputs_names;
}
ov::PartialShape GgmlOvDecoder::get_output_shape(int node_idx) const {
auto * ggml_tensor = m_node_info_list[node_idx].node_output;
return ov::PartialShape(get_shape(ggml_tensor));
}
ov::element::Type GgmlOvDecoder::get_output_type(const int node_idx) const {
return get_ov_type(m_node_info_list[node_idx].node);
}
std::vector<std::string> GgmlOvDecoder::get_output_names(int node_idx) const {
return {m_node_info_list[node_idx].node_output_name};
}
const std::string & GgmlOvDecoder::get_op_name() const {
static const std::string unknown_name = "UNKNOWN_OP_NAME";
return unknown_name;
}
const std::string & GgmlOvDecoder::get_op_name(int node_idx) const {
return m_node_info_list[node_idx].node_name;
}
int32_t * GgmlOvDecoder::get_input_op_params(int node_idx, const std::string & name) const {
return m_node_info_list[node_idx].node_inputs.at(name)->op_params;
}
int32_t * GgmlOvDecoder::get_output_op_params(int node_idx) const {
return m_node_info_list[node_idx].node->op_params;
}
void GgmlOvDecoder::visit_subgraph(std::function<void(std::shared_ptr<GgmlDecoder>, int node_idx)> node_visitor) const {
for (int node_idx = 0; node_idx < m_cgraph->n_nodes; node_idx++) {
if (m_cgraph->nodes[node_idx]->op == GGML_OP_NONE) {
continue;
}
node_visitor(std::make_shared<GgmlOvDecoder>(*this), node_idx);
}
}
std::string GgmlOvDecoder::compute_op_type(const ggml_tensor * node) {
static const std::map<ggml_op, std::string> ops = {
{GGML_OP_NONE, "GGML_OP_NONE" },
{GGML_OP_ACC, "GGML_OP_ACC" },
{GGML_OP_ADD, "GGML_OP_ADD" },
{GGML_OP_ADD1, "GGML_OP_ADD1" },
{GGML_OP_CONT, "GGML_OP_CONT" },
{GGML_OP_DIV, "GGML_OP_DIV" },
{GGML_OP_DUP, "GGML_OP_DUP" },
{GGML_OP_GET_ROWS, "GGML_OP_GET_ROWS" },
{GGML_OP_MUL, "GGML_OP_MUL" },
{GGML_OP_MUL_MAT, "GGML_OP_MUL_MAT" },
{GGML_OP_PERMUTE, "GGML_OP_PERMUTE" },
{GGML_OP_RESHAPE, "GGML_OP_RESHAPE" },
{GGML_OP_RMS_NORM, "GGML_OP_RMS_NORM" },
{GGML_OP_ROPE, "GGML_OP_ROPE" },
{GGML_OP_SCALE, "GGML_OP_SCALE" },
{GGML_OP_SOFT_MAX, "GGML_OP_SOFT_MAX" },
{GGML_OP_SUB, "GGML_OP_SUB" },
{GGML_OP_TRANSPOSE, "GGML_OP_TRANSPOSE" },
{GGML_OP_VIEW, "GGML_OP_VIEW" },
{GGML_OP_SET_ROWS, "GGML_OP_SET_ROWS" },
{GGML_OP_CPY, "GGML_OP_CPY" },
{GGML_OP_FLASH_ATTN_EXT, "GGML_OP_FLASH_ATTN_EXT"},
};
static const std::map<ggml_unary_op, std::string> unary_ops = {
{GGML_UNARY_OP_ABS, "GGML_UNARY_OP_ABS" },
{GGML_UNARY_OP_SGN, "GGML_UNARY_OP_SGN" },
{GGML_UNARY_OP_NEG, "GGML_UNARY_OP_NEG" },
{GGML_UNARY_OP_STEP, "GGML_UNARY_OP_STEP" },
{GGML_UNARY_OP_TANH, "GGML_UNARY_OP_TANH" },
{GGML_UNARY_OP_ELU, "GGML_UNARY_OP_ELU" },
{GGML_UNARY_OP_RELU, "GGML_UNARY_OP_RELU" },
{GGML_UNARY_OP_SIGMOID, "GGML_UNARY_OP_SIGMOID" },
{GGML_UNARY_OP_GELU, "GGML_UNARY_OP_GELU" },
{GGML_UNARY_OP_GELU_QUICK, "GGML_UNARY_OP_GELU_QUICK" },
{GGML_UNARY_OP_SILU, "GGML_UNARY_OP_SILU" },
{GGML_UNARY_OP_HARDSWISH, "GGML_UNARY_OP_HARDSWISH" },
{GGML_UNARY_OP_HARDSIGMOID, "GGML_UNARY_OP_HARDSIGMOID"},
{GGML_UNARY_OP_EXP, "GGML_UNARY_OP_EXP" },
{GGML_UNARY_OP_COUNT, "GGML_UNARY_OP_COUNT" }
};
static const std::map<ggml_glu_op, std::string> glu_ops = {
{GGML_GLU_OP_SWIGLU, "GGML_GLU_OP_SWIGLU"},
{GGML_GLU_OP_GEGLU, "GGML_GLU_OP_GEGLU" },
{GGML_GLU_OP_REGLU, "GGML_GLU_OP_REGLU" }
};
switch (node->op) {
case GGML_OP_UNARY:
return unary_ops.at(ggml_get_unary_op(node));
case GGML_OP_GLU:
return glu_ops.at(ggml_get_glu_op(node));
default:
return ops.at(node->op);
}
static const std::string unknown_op = "UNKNOWN_GGML_OP";
return unknown_op;
}
const std::string & GgmlOvDecoder::get_op_type(int node_idx) const {
return m_node_info_list[node_idx].node_op_type;
}
const std::string & GgmlOvDecoder::get_op_type() const {
static const std::string unknown_op = "UNKNOWN_GGML_OP";
return unknown_op;
}

294
ggml/src/ggml-openvino/ggml-decoder.h Normal file
View File

@ -0,0 +1,294 @@
#pragma once
#include "ggml-quants.h"
#include "ggml.h"
#include "openvino/decoder.h"
#include <cstdint>
#include <cstring>
#include <map>
#include <memory>
#include <openvino/core/partial_shape.hpp>
#include <optional>
#include <vector>
struct ModelParams {
int ctx = -1;
int ctx_swa = -1;
int ctx_per_seq = -1;
int ctx_per_seq_swa = -1;
int n_seq = 1;
int n_heads = -1;
int n_heads_kv = -1;
int head_size = -1;
int32_t rope_params[15];
std::vector<int> swa_layers;
std::vector<std::string> kv_names;
size_t kv_buffer_ctx_id = 0;
bool same_rope_params(const ModelParams & other) const {
return memcmp(rope_params, other.rope_params, sizeof(int32_t) * 15) == 0;
}
bool can_reuse_dynamically(const ModelParams & other) const { return same_rope_params(other); }
bool can_reuse_statically(const ModelParams & other) const { return same_rope_params(other) && ctx == other.ctx; }
bool kv_buffer_changed(const ModelParams & other) const { return kv_buffer_ctx_id != other.kv_buffer_ctx_id; }
};
struct ComputeParams {
int n_seq_active = 1;
int seq_active_start = 0;
int attention_size = -1;
int attention_size_swa = -1;
int input_len = -1;
int token_len_per_seq = -1;
int past_kv_len = -1;
int output_len = 1;
};
class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
public:
struct NodeInfo {
ggml_tensor * node;
std::string node_name;
std::string node_op_type;
std::map<std::string, ggml_tensor *> node_inputs;
std::vector<std::string> node_inputs_names;
ggml_tensor * node_output;
std::string node_output_name;
int node_op_case = 0;
void * data_addr;
};
// Graph decoder
GgmlOvDecoder(ggml_cgraph * cgraph,
ModelParams & model_params,
ComputeParams & compute_params,
std::map<std::string, std::shared_ptr<ov::Node>> & model_weights,
bool is_static,
bool is_stateful = false,
bool is_prefill = false,
int prefill_chunk_size = 256);
// Naive graph decoder
GgmlOvDecoder(ggml_cgraph * cgraph, std::map<std::string, std::shared_ptr<ov::Node>> & model_weights);
virtual ov::Any get_attribute(const std::string & name) const override {
return nullptr;
GGML_UNUSED(name);
}
virtual ov::PartialShape get_input_shape(int node_idx, const std::string & name) const override;
virtual std::vector<size_t> get_input_stride(int node_idx, const std::string & name) const override;
virtual ov::element::Type get_input_type(int node_idx, const std::string & name) const override;
virtual size_t get_input_size() const override;
virtual size_t get_input_size(int node_idx) const override;
virtual void get_input_node(size_t input_port_idx,
std::string & producer_name,
std::string & producer_output_port_name,
size_t & producer_output_port_index) const override {
GGML_UNUSED(input_port_idx);
GGML_UNUSED(producer_name);
GGML_UNUSED(producer_output_port_name);
GGML_UNUSED(producer_output_port_index);
}
virtual std::vector<std::string> get_input_names(int node_idx) const override;
virtual ov::PartialShape get_output_shape(int node_idx) const override;
virtual ov::element::Type get_output_type(int node_idx) const override;
virtual int32_t * get_input_op_params(int node_idx, const std::string & name) const override;
virtual int32_t * get_output_op_params(int node_idx) const override;
virtual std::vector<std::string> get_output_names(int node_idx) const override;
virtual const std::string & get_op_type() const override;
virtual const std::string & get_op_type(int node_idx) const override;
virtual const std::string & get_op_name() const override;
virtual const std::string & get_op_name(int node_idx) const override;
virtual void visit_subgraph(std::function<void(std::shared_ptr<GgmlDecoder>, int node_idx)> node_visitor) const override;
ggml_tensor * get_input_ggml_tensor(const std::string & name) const { return m_inputs.at(name); }
virtual int get_op_case(int node_idx) const override { return m_node_info_list[node_idx].node_op_case; }
virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_inputs() const override {
return m_model_inputs;
}
virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_extra_inputs() const override {
return m_model_extra_inputs;
}
virtual const std::map<std::string, std::shared_ptr<ov::Tensor>> & get_model_extra_input_values() const {
return m_model_extra_input_values;
}
virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_weights() const override {
return m_model_weights;
}
virtual std::vector<std::string> get_model_output_names() const override {
return m_model_output_names;
}
const std::map<std::string, ggml_tensor *> & get_model_outputs() const { return m_model_outputs; }
virtual int get_ctx_size() const { return m_model_params.ctx; }
virtual int get_ctx_swa_size() const { return m_model_params.ctx_swa; }
virtual int get_ctx_per_seq() const { return m_model_params.ctx_per_seq; }
virtual int get_ctx_per_seq_swa() const { return m_model_params.ctx_per_seq_swa; }
virtual int get_n_seq() const { return m_model_params.n_seq; }
virtual int is_swa_layer(int layer) const override {
return std::find(m_model_params.swa_layers.begin(), m_model_params.swa_layers.end(), layer) !=
m_model_params.swa_layers.end();
}
int get_past_kv_len() const { return m_compute_params.past_kv_len; }
int get_input_len() const { return m_compute_params.input_len; }
virtual int32_t * get_rope_params() const override { return const_cast<int32_t *>(m_model_params.rope_params); }
virtual std::map<std::string, std::string> get_kv_param_res_names() const override;
virtual bool is_static() const override { return m_is_static; }
virtual bool is_stateful() const override { return m_is_stateful; }
ov::PartialShape get_graph_input_shape(const ggml_tensor * op, const ggml_tensor * input) const;
static void dump_cgraph(const ggml_cgraph * cgraph, std::string & filename);
static std::shared_ptr<ov::Node> create_weight_node(ggml_tensor * tensor, bool naive = false);
static std::map<std::string, std::shared_ptr<ov::Node>> create_weight_nodes(ggml_cgraph * cgraph,
bool naive = false);
const ggml_tensor * get_tensor_used_op(const ggml_tensor * tensor) const;
const ggml_tensor * get_tensor_from_name(const std::string & name) const;
void clear_model_weights() { m_model_weights.clear(); }
static std::pair<ModelParams, ComputeParams> compute_llm_params(ggml_cgraph * cgraph, bool is_static);
ModelParams get_model_params() const { return m_model_params; }
ComputeParams get_compute_params() const { return m_compute_params; }
void set_model_params(const ModelParams & model_params) { m_model_params = model_params; }
void set_compute_params(const ComputeParams & compute_params) { m_compute_params = compute_params; }
bool m_is_static = false;
bool m_is_stateful = false;
bool m_is_prefill = false;
bool m_naive = false;
int m_prefill_chunk_size = 0;
static ov::Shape get_shape(const ggml_tensor * tensor);
static std::vector<size_t> get_stride(const ggml_tensor * tensor);
static ov::element::Type get_ov_type(const ggml_tensor * tensor);
static std::string compute_op_type(const ggml_tensor * node);
void add_extra_inputs();
void update_io(ggml_cgraph * cgraph);
inline static bool is_inp_tok(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] && op->src[0]->op == GGML_OP_NONE;
}
inline static bool is_inp_pos(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_ROPE && tensor == op->src[1];
}
inline static bool is_inp_emb(const ggml_tensor * tensor, const ggml_tensor * op) {
return tensor->op == GGML_OP_GET_ROWS && op->op == GGML_OP_RMS_NORM;
}
inline static bool is_inp_mask(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_CPY || (op->op == GGML_OP_FLASH_ATTN_EXT && tensor == op->src[3]);
}
inline static bool is_rope_freqs_weight(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_ROPE && tensor == op->src[2];
}
inline static bool is_kvcache(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_SET_ROWS && op->src[2] == tensor;
}
inline static bool is_kv_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_SET_ROWS && op->src[1] == tensor;
}
inline static bool is_output_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] && op->src[0]->op != GGML_OP_NONE;
}
static std::string get_graph_input_ov_name(const ggml_tensor * tensor, const ggml_tensor * op) {
if (is_inp_tok(tensor, op)) {
return "inp_tokens";
}
if (is_inp_pos(tensor, op)) {
return "inp_pos";
}
if (is_inp_emb(tensor, op)) {
return "embd";
}
if (is_output_idx(tensor, op)) {
return "inp_out_ids";
}
if (is_inp_mask(tensor, op)) {
return std::string(tensor->name).find("swa") == std::string::npos ? "self_kq_mask" : "self_kq_mask_swa";
}
return tensor->name;
}
private:
void set_input_output();
int compute_op_case(const ggml_tensor * node) const;
bool node_is_used_as_src(const int node_idx);
void compute_model_inputs();
void compute_model_outputs();
void validate_cgraph() const;
ggml_cgraph * m_cgraph = nullptr;
std::map<std::string, ggml_tensor *> m_inputs;
std::map<std::string, std::shared_ptr<ov::Node>> m_model_inputs;
std::map<std::string, std::shared_ptr<ov::Node>> m_model_extra_inputs;
std::map<std::string, std::shared_ptr<ov::Tensor>> m_model_extra_input_values;
std::map<std::string, std::shared_ptr<ov::Node>> m_model_weights;
std::map<std::string, ggml_tensor *> m_model_outputs;
std::vector<std::string> m_model_output_names;
std::vector<NodeInfo> m_node_info_list;
ModelParams m_model_params;
ComputeParams m_compute_params;
};
void print_tensor_address_map(const ggml_cgraph * cgraph);
int extract_layer_from_name(const std::string & name);

373
ggml/src/ggml-openvino/ggml-openvino-extra.cpp Normal file
View File

@ -0,0 +1,373 @@
#include "ggml-openvino-extra.h"
#include "ggml-impl.h"
#include "ggml.h"
#include <cstring>
#include <openvino/runtime/intel_gpu/ocl/ocl.hpp>
#include <openvino/runtime/intel_npu/level_zero/level_zero.hpp>
#include <optional>
ov::Core & ov_singleton_core() {
static ov::Core core;
return core;
}
// =====================================================
// Device Configuration Implementations
// =====================================================
void ggml_openvino_device_config::init() {
if (initialized) {
return;
}
device_name = getenv("GGML_OPENVINO_DEVICE") ? getenv("GGML_OPENVINO_DEVICE") : "CPU";
auto available_devices = ov_singleton_core().get_available_devices();
if (std::find(available_devices.begin(), available_devices.end(), device_name) == available_devices.end()) {
GGML_LOG_WARN("GGML OpenVINO Backend: device %s is not available, fallback to CPU\n", device_name.c_str());
device_name = "CPU";
}
is_npu = (device_name == "NPU");
auto * cache_dir = getenv("GGML_OPENVINO_CACHE_DIR");
if (device_name == "NPU") {
compile_config = {
{"NPU_COMPILER_DYNAMIC_QUANTIZATION", "YES" },
{"NPU_USE_NPUW", "YES" },
{"NPUW_DEVICES", "NPU" },
{"NPUW_FOLD", "YES" },
{"NPUW_WEIGHTS_BANK", "shared"},
{"NPUW_FUNCALL_FOR_ALL", "YES" },
{"NPUW_FUNCALL_ASYNC", "YES" },
{"NPUW_DQ", "YES" },
{"NPUW_DQ_FULL", "NO" },
};
if (cache_dir) {
compile_config["NPUW_CACHE_DIR"] = cache_dir;
}
} else if (cache_dir) {
ov_singleton_core().set_property(ov::cache_dir(cache_dir));
}
// Initialize remote context with queue sharing for GPU
if (device_name == "GPU") {
// Create OpenCL context and queue
cl_int err;
cl_platform_id platform;
err = clGetPlatformIDs(1, &platform, nullptr);
if (err != CL_SUCCESS) {
GGML_LOG_ERROR("Failed to get OpenCL platform: %d\n", err);
return;
}
cl_device_id cl_device;
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &cl_device, nullptr);
if (err != CL_SUCCESS) {
GGML_LOG_ERROR("Failed to get OpenCL device: %d\n", err);
return;
}
cl_context cl_ctx = clCreateContext(nullptr, 1, &cl_device, nullptr, nullptr, &err);
if (err != CL_SUCCESS) {
GGML_LOG_ERROR("Failed to create OpenCL context: %d\n", err);
return;
}
cl_queue = clCreateCommandQueueWithProperties(cl_ctx, cl_device, nullptr, &err);
if (err != CL_SUCCESS) {
GGML_LOG_ERROR("Failed to create OpenCL command queue: %d\n", err);
clReleaseContext(cl_ctx);
return;
}
// Create OpenVINO remote context with queue sharing
remote_context = ov::intel_gpu::ocl::ClContext(ov_singleton_core(), cl_queue);
// Release the context (queue keeps a reference)
clReleaseContext(cl_ctx);
} else if (device_name == "NPU") {
// remote tensor is not used for NPU yet
// remote_context = ov_singleton_core().get_default_context(device_name);
}
initialized = true;
}
ggml_openvino_device_config::~ggml_openvino_device_config() {
if (cl_queue != nullptr) {
clReleaseCommandQueue(cl_queue);
cl_queue = nullptr;
}
}
// Get the global device config singleton
ggml_openvino_device_config & ggml_openvino_get_device_config() {
static ggml_openvino_device_config config;
return config;
}
// Initialize device config (call during backend init)
void ggml_openvino_init_device_config() {
ggml_openvino_get_device_config().init();
}
// Get the device name
const std::string & ggml_openvino_get_device_name() {
return ggml_openvino_get_device_config().device_name;
}
// Check if running on NPU
bool ggml_openvino_is_npu() {
return ggml_openvino_get_device_config().is_npu;
}
// Get the remote context for the current device (returns empty optional for CPU)
std::optional<ov::RemoteContext> ggml_openvino_get_remote_context() {
return ggml_openvino_get_device_config().remote_context;
}
// Get the compile config for the current device
const ov::AnyMap & ggml_openvino_get_compile_config() {
return ggml_openvino_get_device_config().compile_config;
}
// Get the OpenCL command queue for GPU operations
cl_command_queue ggml_openvino_get_cl_queue() {
return ggml_openvino_get_device_config().cl_queue;
}
// Get the clEnqueueMemFillINTEL function pointer (lazy load)
clEnqueueMemFillINTEL_fn ggml_openvino_get_clEnqueueMemFillINTEL() {
static clEnqueueMemFillINTEL_fn fn = nullptr;
static bool loaded = false;
if (!loaded) {
loaded = true;
cl_platform_id platform;
if (clGetPlatformIDs(1, &platform, nullptr) == CL_SUCCESS) {
fn = (clEnqueueMemFillINTEL_fn) clGetExtensionFunctionAddressForPlatform(platform, "clEnqueueMemFillINTEL");
}
}
return fn;
}
// Get the clEnqueueMemcpyINTEL function pointer (lazy load)
clEnqueueMemcpyINTEL_fn ggml_openvino_get_clEnqueueMemcpyINTEL() {
static clEnqueueMemcpyINTEL_fn fn = nullptr;
static bool loaded = false;
if (!loaded) {
loaded = true;
cl_platform_id platform;
if (clGetPlatformIDs(1, &platform, nullptr) == CL_SUCCESS) {
fn = (clEnqueueMemcpyINTEL_fn) clGetExtensionFunctionAddressForPlatform(platform, "clEnqueueMemcpyINTEL");
}
}
return fn;
}
// Get requantization type for a tensor type (returns nullopt if no requant needed)
std::optional<ExtraQuantType> ggml_openvino_get_requant_type(const ggml_tensor * tensor, bool no_requant) {
if (no_requant) {
return std::nullopt;
}
if (strncmp(tensor->name, "token_embd.weight", 17) == 0) {
return ((ggml_openvino_is_npu() && tensor->type == GGML_TYPE_Q6_K) ? ExtraQuantType::F16 : ExtraQuantType::Q8_0_C);
}
if (strncmp(tensor->name, "output.weight", 13) == 0) {
return ExtraQuantType::Q8_0_C;
}
if (ggml_openvino_is_npu()) {
return ExtraQuantType::Q4_0_128;
}
switch (tensor->type) {
case GGML_TYPE_Q6_K:
case GGML_TYPE_Q5_K:
return ExtraQuantType::Q8_0_C;
default:
return std::nullopt;
}
}
// =====================================================
// Extracted Layout Calculation
// =====================================================
ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_tensor * tensor, bool use_bias) {
ggml_openvino_extracted_layout layout = {};
layout.is_symmetric = false;
if (!ggml_is_quantized(tensor->type)) {
return layout;
}
// Only handle 2D weight tensors
if (tensor->ne[2] != 1 || tensor->ne[3] != 1) {
return layout;
}
int64_t n_elements = ggml_nelements(tensor);
const size_t alignment = 64; // Good for SIMD
// Check if requantization is needed (NPU-specific)
auto requant_type = ggml_openvino_get_requant_type(tensor, use_bias);
if (requant_type.has_value()) {
layout.is_requant = true;
layout.requant_type = requant_type;
// Special case: requant to F16 - just store F16 weights, no scales/zp
if (requant_type.value() == ExtraQuantType::F16) {
layout.weights_size = n_elements * sizeof(uint16_t); // F16 = 2 bytes
layout.total_size = layout.weights_size;
layout.weights_offset = 0;
// No scales/zp for F16
return layout;
}
// Requant to different quantized format (e.g., Q4_0_128)
switch (requant_type.value()) {
case ExtraQuantType::Q4_0_128:
layout.is_u4 = true;
layout.weights_per_block = 128;
layout.is_symmetric = true;
break;
case ExtraQuantType::Q4_0_C:
layout.is_u4 = true;
layout.weights_per_block = tensor->ne[0];
layout.is_symmetric = true;
break;
case ExtraQuantType::Q8_0_32:
layout.is_u4 = false;
layout.weights_per_block = 32;
layout.is_symmetric = true;
break;
case ExtraQuantType::Q8_0_C:
layout.is_u4 = false;
layout.weights_per_block = tensor->ne[0];
layout.is_symmetric = true;
break;
case ExtraQuantType::Q8_1_C:
layout.is_u4 = false;
layout.weights_per_block = tensor->ne[0];
break;
default:
layout.weights_per_block = -1;
GGML_ABORT("Code of re-quantizing to channel-wise is not updated");
break;
}
if (layout.is_requant) {
// Calculate sizes for requantized format
layout.weights_size = layout.is_u4 ? (n_elements / 2) : n_elements;
int64_t n_blocks = n_elements / layout.weights_per_block;
layout.scales_size = n_blocks * sizeof(uint16_t);
// For symmetric quantization, we only need one zp value (not one per block)
// Zero points are stored in U4 or U8 format matching the weight type
size_t n_zp_elements = layout.is_symmetric ? 1 : n_blocks;
layout.zp_size = layout.is_u4 ? ((n_zp_elements + 1) / 2) : n_zp_elements;
layout.weights_offset = 0;
layout.scales_offset = ((layout.weights_size + alignment - 1) / alignment) * alignment;
layout.zp_offset = layout.scales_offset + ((layout.scales_size + alignment - 1) / alignment) * alignment;
layout.total_size = layout.zp_offset + layout.zp_size;
layout.total_size = std::max(layout.total_size, ggml_nbytes(tensor));
return layout;
}
}
// Normal extraction (no requant) - determine format based on tensor type
layout.is_u4 = false;
layout.weights_per_block = 32;
layout.is_symmetric = false;
switch (tensor->type) {
case GGML_TYPE_Q4_0:
layout.is_u4 = true;
layout.is_symmetric = true;
break;
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q4_K:
layout.is_u4 = true;
break;
case GGML_TYPE_Q8_0:
layout.is_symmetric = true;
break;
case GGML_TYPE_Q6_K:
layout.weights_per_block = 16;
layout.is_symmetric = true;
break;
case GGML_TYPE_Q5_K:
break;
default:
// Unsupported quantization type
return layout;
}
// Calculate sizes
// Weights: U4 = n_elements/2 bytes, U8 = n_elements bytes
layout.weights_size = layout.is_u4 ? (n_elements / 2) : n_elements;
// Scales: F16 per block
int64_t n_blocks = n_elements / layout.weights_per_block;
layout.scales_size = n_blocks * sizeof(uint16_t); // F16 = 2 bytes
// Zero points: U4 or U8 matching weight type
// For symmetric quantization, we only need one zp value (not one per block)
size_t n_zp_elements = layout.is_symmetric ? 1 : n_blocks;
layout.zp_size = layout.is_u4 ? ((n_zp_elements + 1) / 2) : n_zp_elements;
// Layout in buffer: [weights | scales | zp] with alignment
layout.weights_offset = 0;
layout.scales_offset = ((layout.weights_size + alignment - 1) / alignment) * alignment;
layout.zp_offset = layout.scales_offset + ((layout.scales_size + alignment - 1) / alignment) * alignment;
layout.total_size = layout.zp_offset + layout.zp_size;
layout.total_size = std::max(layout.total_size, ggml_nbytes(tensor));
return layout;
}
ggml_openvino_tensor_extra * ggml_openvino_create_tensor_extra(const ggml_tensor * tensor, bool is_remote) {
ov::Shape shape;
for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
shape.push_back(static_cast<size_t>(tensor->ne[i]));
}
ov::element::Type element_type;
switch (tensor->type) {
case GGML_TYPE_F32:
element_type = ov::element::f32;
break;
case GGML_TYPE_F16:
element_type = ov::element::f16;
break;
case GGML_TYPE_BF16:
element_type = ov::element::bf16;
break;
case GGML_TYPE_I32:
element_type = ov::element::i32;
break;
case GGML_TYPE_I64:
element_type = ov::element::i64;
break;
default:
// GGML_LOG_WARN("%s: unsupported tensor type for ov::Tensor: %s\n", __func__, ggml_type_name(tensor->type));
return nullptr;
}
const auto & device_name = ggml_openvino_get_device_name();
auto remote_context = ggml_openvino_get_remote_context();
std::shared_ptr<ov::Tensor> ov_tensor;
if (is_remote) {
GGML_ASSERT(device_name == "GPU");
auto gpu_context = remote_context->as<ov::intel_gpu::ocl::ClContext>();
auto usm_tensor = gpu_context.create_tensor(element_type, shape, tensor->data);
ov_tensor = std::make_shared<ov::intel_gpu::ocl::USMTensor>(std::move(usm_tensor));
} else {
ov_tensor = std::make_shared<ov::Tensor>(element_type, shape, tensor->data);
}
return new ggml_openvino_tensor_extra(ov_tensor);
}

182
ggml/src/ggml-openvino/ggml-openvino-extra.h Normal file
View File

@ -0,0 +1,182 @@
#pragma once
#include "ggml.h"
#include "openvino/runtime/core.hpp"
#define CL_TARGET_OPENCL_VERSION 300
#include <CL/cl.h>
#include <cstdlib>
#include <memory>
#include <openvino/core/node.hpp>
#include <openvino/runtime/remote_context.hpp>
#include <openvino/runtime/tensor.hpp>
#include <optional>
#include <string>
// ExtraQuantType enum - defines requantization target formats
enum class ExtraQuantType { F16, Q4_0_C, Q8_1_C, Q4_0_128, Q8_0_C, Q8_0_32 };
ov::Core & ov_singleton_core();
// Get the remote context for the current device (returns empty optional for CPU)
std::optional<ov::RemoteContext> ggml_openvino_get_remote_context();
// Get the compile config for the current device
const ov::AnyMap & ggml_openvino_get_compile_config();
// Get the OpenCL command queue for GPU operations (returns nullptr for CPU/NPU)
cl_command_queue ggml_openvino_get_cl_queue();
// Intel USM extension function type
typedef cl_int(CL_API_CALL * clEnqueueMemFillINTEL_fn)(cl_command_queue queue,
void * dst_ptr,
const void * pattern,
size_t pattern_size,
size_t size,
cl_uint num_events_in_wait_list,
const cl_event * event_wait_list,
cl_event * event);
typedef cl_int(CL_API_CALL * clEnqueueMemcpyINTEL_fn)(cl_command_queue queue,
cl_bool blocking,
void * dst_ptr,
const void * src_ptr,
size_t size,
cl_uint num_events_in_wait_list,
const cl_event * event_wait_list,
cl_event * event);
// Get the clEnqueueMemFillINTEL function pointer (returns nullptr if not available)
clEnqueueMemFillINTEL_fn ggml_openvino_get_clEnqueueMemFillINTEL();
// Get the clEnqueueMemcpyINTEL function pointer (returns nullptr if not available)
clEnqueueMemcpyINTEL_fn ggml_openvino_get_clEnqueueMemcpyINTEL();
// =====================================================
// Global Device Configuration (singleton)
// =====================================================
// Initialized once during backend init from GGML_OPENVINO_DEVICE env var
struct ggml_openvino_device_config {
std::string device_name = "CPU";
bool is_npu = false;
bool initialized = false;
std::optional<ov::RemoteContext> remote_context;
ov::AnyMap compile_config;
cl_command_queue cl_queue = nullptr;
void init();
~ggml_openvino_device_config();
};
// Get the global device config singleton
ggml_openvino_device_config & ggml_openvino_get_device_config();
// Initialize device config (call during backend init)
void ggml_openvino_init_device_config();
// Get the device name
const std::string & ggml_openvino_get_device_name();
// Check if running on NPU
bool ggml_openvino_is_npu();
// Get requantization type for a tensor type (returns nullopt if no requant needed)
std::optional<ExtraQuantType> ggml_openvino_get_requant_type(const ggml_tensor * tensor, bool no_requant = false);
// =====================================================
// OpenVINO Tensor Extra Types
// =====================================================
// These types are stored in tensor->extra by the OpenVINO backend buffer.
// They allow:
// 1. Pre-built ov::Constant nodes for weights (avoiding memcpy during graph construction)
// 2. ov::Tensor wrappers for KV cache / compute tensors (for direct use with infer_request)
// Base class for OpenVINO tensor extra data
struct ggml_openvino_extra_base {
enum class Type { WEIGHT, QUANTIZED_WEIGHT, TENSOR };
Type type;
virtual ~ggml_openvino_extra_base() = default;
protected:
explicit ggml_openvino_extra_base(Type t) : type(t) {}
};
// Extra data for F16/F32/BF16 weight tensors - stores the pre-built weight node
struct ggml_openvino_weight_extra : public ggml_openvino_extra_base {
ov::Tensor weights; // The underlying weight data tensor
std::shared_ptr<ov::Node> weight_node; // Pre-built OpenVINO weight node
ggml_openvino_weight_extra(ov::Tensor w, std::shared_ptr<ov::Node> n) :
ggml_openvino_extra_base(Type::WEIGHT),
weights(std::move(w)),
weight_node(std::move(n)) {}
};
// Extra data for quantized weight tensors - stores extracted weights/scales/zp and weight node
struct ggml_openvino_quantized_weight_extra : public ggml_openvino_extra_base {
ov::Tensor weights; // U4 or U8 extracted weights
ov::Tensor scales; // F16 scales
ov::Tensor zp; // U4 or U8 zero points (same type as weights)
std::shared_ptr<ov::Node> weight_node; // Pre-built OpenVINO weight subgraph
ggml_openvino_quantized_weight_extra(ov::Tensor w, ov::Tensor s, ov::Tensor z, std::shared_ptr<ov::Node> n) :
ggml_openvino_extra_base(Type::QUANTIZED_WEIGHT),
weights(std::move(w)),
scales(std::move(s)),
zp(std::move(z)),
weight_node(std::move(n)) {}
};
// Extra data for KV cache / compute tensors - stores ov::Tensor for infer_request
struct ggml_openvino_tensor_extra : public ggml_openvino_extra_base {
std::shared_ptr<ov::Tensor> tensor; // For direct use with infer_request
explicit ggml_openvino_tensor_extra(std::shared_ptr<ov::Tensor> t)
: ggml_openvino_extra_base(Type::TENSOR), tensor(std::move(t)) {}
};
// =====================================================
// Extracted Size Calculation for Quantized Tensors
// =====================================================
// For quantized tensors, we need extra space to store extracted weights, scales, and zero points.
// Returns the total size needed in the buffer for extracted data.
struct ggml_openvino_extracted_layout {
size_t total_size = 0; // Total bytes needed
size_t weights_offset = 0; // Offset to weights in buffer
size_t weights_size = 0; // Size of weights in bytes
size_t scales_offset = 0; // Offset to scales in buffer
size_t scales_size = 0; // Size of scales in bytes
size_t zp_offset = 0; // Offset to zero points in buffer
size_t zp_size = 0; // Size of zero points in bytes (U4 or U8)
bool is_u4; // true for U4 weights, false for U8
int64_t weights_per_block; // weights per scale/zp block
bool is_symmetric; // true for symmetric quantization
// Requantization info
bool is_requant = false; // true if this tensor needs requantization
std::optional<ExtraQuantType> requant_type; // target requant type if is_requant
};
// Calculate the buffer layout for extracted quantized data
ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_tensor * tensor, bool use_bias = false);
ggml_openvino_tensor_extra * ggml_openvino_create_tensor_extra(const ggml_tensor * tensor, bool is_remote);
// Register an extra with the tensor's OpenVINO buffer context for proper lifetime management.
// This sets tensor->extra and tracks the extra in the buffer context for cleanup.
void ggml_openvino_buffer_register_extra(ggml_tensor * tensor, ggml_openvino_extra_base * extra);
// =====================================================
// OpenVINO Backend Context and Interface
// =====================================================
struct ggml_backend_openvino_context {
int device = 0;
std::string name = "OpenVINO";
std::string description = "OpenVINO Backend Context";
std::shared_ptr<void> runtime_context = nullptr;
ggml_backend_openvino_context() = default;
};

1110
ggml/src/ggml-openvino/ggml-openvino.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

884
ggml/src/ggml-openvino/ggml-quants.cpp Normal file
View File

@ -0,0 +1,884 @@
#include "ggml-quants.h"
#include "ggml-common.h"
#include "ggml-impl.h"
#include "ggml.h"
#include <algorithm>
#include <cassert>
#include <cmath>
#include <cstddef>
#include <cstdint>
#include <limits>
#include <memory>
#include <openvino/core/except.hpp>
#include <openvino/core/node.hpp>
#include <openvino/core/node_output.hpp>
#include <openvino/core/parallel.hpp>
#include <openvino/core/shape.hpp>
#include <openvino/core/type/element_type.hpp>
#include <openvino/core/type/element_type_traits.hpp>
#include <openvino/core/type/float16.hpp>
#include <openvino/op/add.hpp>
#include <openvino/op/constant.hpp>
#include <openvino/op/convert.hpp>
#include <openvino/op/multiply.hpp>
#include <openvino/op/reshape.hpp>
#include <openvino/op/subtract.hpp>
#include <openvino/op/util/attr_types.hpp>
#include <openvino/runtime/tensor.hpp>
#include <string>
#include <vector>
void unpack_32_4(const uint8_t * data, uint8_t * dst) {
std::fill_n(dst, 16, 0);
for (int j = 0; j < 16; ++j) {
uint8_t x = (data[j] & 0x0F);
uint8_t y = (data[j] >> 4);
if (j % 2 != 0) {
x <<= 4;
y <<= 4;
}
dst[j / 2] |= x;
dst[8 + j / 2] |= y; // Last 16 weights are in the higher bits
}
}
// Extracts (weight, scales, zp) from Q4_0 tensors.
// Data layout is: |16 bit scale|32 x 4bit weights|.
void extract_q4_0_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr) {
const uint64_t bytes_per_block = 18; // 2 bytes scale, 32x0.5 byte weights
auto * data = static_cast<uint8_t *>(tensor->data);
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
auto * zp = static_cast<uint8_t *>(zp_arr.data());
bool is_scalar_zp = (zp_arr.get_size() == 1); // Symmetric quantization
// For Q4_0, zero point is always 8
if (is_scalar_zp) {
zp[0] = 8 | (8 << 4); // Pack two 4-bit values
}
ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
scales[i] = ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block)));
// For asymmetric quantization, compute per-block zero points
if (!is_scalar_zp) {
// Pack two 4-bit zero points per byte
if (i % 2 == 0) {
zp[i / 2] = 8; // Lower nibble
} else {
zp[i / 2] |= (8 << 4); // Upper nibble
}
}
unpack_32_4(data + i * bytes_per_block + 2, weights + i * 16);
});
}
// Extracts (weight, scales, zp) from Q4_1 tensors.
// Data layout is: |16 bit scale|16 bit min|32 x 4bit weights|.
void extract_q4_1_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
bool use_bias) {
const uint64_t bytes_per_block = 20; // 2 bytes scale, 2 bytes min, 32x0.5 byte weights
auto * data = static_cast<uint8_t *>(tensor->data);
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
if (use_bias) {
// Store bias (min) directly as f16 instead of computing u4 zero points
auto * bias = zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
float scale = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block))));
float min = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block + 2))));
scales[i] = ov::float16(scale);
bias[i] = ov::float16(min); // bias = min, dequant: w*s + bias
unpack_32_4(data + i * bytes_per_block + 4, weights + i * 16);
});
} else {
auto * zp = static_cast<uint8_t *>(zp_arr.data());
ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
float scale = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block))));
float min = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block + 2))));
scales[i] = ov::float16(scale);
// zp = -min / scale (bias = min, so zp = -bias/scale)
uint8_t zp_val = (scale != 0.0f) ? (uint8_t) std::round(-min / scale) : 0;
// Pack two 4-bit zero points per byte
if (i % 2 == 0) {
zp[i / 2] = zp_val & 0x0F; // Lower nibble
} else {
zp[i / 2] |= (zp_val << 4); // Upper nibble
}
unpack_32_4(data + i * bytes_per_block + 4, weights + i * 16);
});
}
}
// Extracts (weight, scales, zp) from Q8_0 tensors.
// Data layout is: |16 bit scale|32 x 8bit weights|.
void extract_q8_0_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr) {
const uint64_t weights_per_block = 32;
const uint64_t bytes_per_block = 34; // 2 bytes scale, 32x1 byte weights
auto * data = static_cast<uint8_t *>(tensor->data);
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
auto * zp = static_cast<uint8_t *>(zp_arr.data());
bool is_scalar_zp = (zp_arr.get_size() == 1); // Symmetric quantization
// For Q8_0, zero point is always 128
if (is_scalar_zp) {
zp[0] = 128;
}
ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
uint8_t * block_data = data + i * bytes_per_block;
scales[i] = ov::float16::from_bits(*(uint16_t *) block_data);
// For asymmetric quantization, store per-block zero points
if (!is_scalar_zp) {
zp[i] = 128;
}
for (size_t j = 0; j < weights_per_block; ++j) {
uint8_t x = block_data[j + 2]; // j+2 to skip the scale bytes.
// Original data is in int8_t, so we add a bias of -128 and invert the first bit.
x ^= 1 << 7;
weights[i * weights_per_block + j] = x;
}
});
}
void unpack_256_4(const uint8_t * data, uint8_t * dst) {
// Initialize the output array with zeros
std::fill_n(dst, 128, 0);
for (size_t i = 0; i < 4; ++i) {
for (int j = 0; j < 32; ++j) {
uint8_t x = (data[i * 32 + j] & 0x0F);
uint8_t y = (data[i * 32 + j] >> 4);
if (j % 2 != 0) {
x <<= 4;
y <<= 4;
}
dst[i * 32 + j / 2] |= x;
dst[i * 32 + 16 + j / 2] |= y; // Last 16 weights are in the higher bits
}
}
}
void extract_q4_k_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
bool use_bias) {
const uint64_t bytes_per_block = 2 + 2 + 12 + 128;
const uint64_t n_super_block = tensor->nb[3] / bytes_per_block;
auto * data = static_cast<uint8_t *>(tensor->data);
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
// For bias path, zp_arr holds f16 bias values; for zp path, it holds packed u4 zero points
auto * zp_u4 = use_bias ? nullptr : static_cast<uint8_t *>(zp_arr.data());
auto * bias_f16 = use_bias ? zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>() : nullptr;
ov::parallel_for(n_super_block, [&](size_t i) {
uint8_t * block_data = data + i * bytes_per_block;
// Extract scale factors and offsets
float scale_scales = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data)));
float scale_mins = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 1)));
// Extract qs1 and qs2
uint8_t * qs1 = block_data + 4;
// Calculate scales
float scale_vals[8];
scale_vals[0] = scale_scales * static_cast<float>((*(qs1) & 0b111111));
scale_vals[1] = scale_scales * static_cast<float>((*(qs1 + 1) & 0b111111));
scale_vals[2] = scale_scales * static_cast<float>((*(qs1 + 2) & 0b111111));
scale_vals[3] = scale_scales * static_cast<float>((*(qs1 + 3) & 0b111111));
scale_vals[4] = scale_scales * static_cast<float>((*(qs1 + 8) & 0b00001111) | ((*(qs1) >> 6) << 4));
scale_vals[5] = scale_scales * static_cast<float>((*(qs1 + 9) & 0b00001111) | ((*(qs1 + 1) >> 6) << 4));
scale_vals[6] = scale_scales * static_cast<float>((*(qs1 + 10) & 0b00001111) | ((*(qs1 + 2) >> 6) << 4));
scale_vals[7] = scale_scales * static_cast<float>((*(qs1 + 11) & 0b00001111) | ((*(qs1 + 3) >> 6) << 4));
// Calculate min values (bias = -min)
float min_vals[8];
min_vals[0] = scale_mins * static_cast<float>((*(qs1 + 4) & 0b111111));
min_vals[1] = scale_mins * static_cast<float>((*(qs1 + 5) & 0b111111));
min_vals[2] = scale_mins * static_cast<float>((*(qs1 + 6) & 0b111111));
min_vals[3] = scale_mins * static_cast<float>((*(qs1 + 7) & 0b111111));
min_vals[4] = scale_mins * static_cast<float>((*(qs1 + 8) >> 4) | ((*(qs1 + 4) >> 6) << 4));
min_vals[5] = scale_mins * static_cast<float>((*(qs1 + 9) >> 4) | ((*(qs1 + 5) >> 6) << 4));
min_vals[6] = scale_mins * static_cast<float>((*(qs1 + 10) >> 4) | ((*(qs1 + 6) >> 6) << 4));
min_vals[7] = scale_mins * static_cast<float>((*(qs1 + 11) >> 4) | ((*(qs1 + 7) >> 6) << 4));
// Store scales and compute zero points or bias
for (int j = 0; j < 8; j++) {
scales[i * 8 + j] = ov::float16(scale_vals[j]);
if (use_bias) {
// Store bias = -min directly as f16, dequant: w*s + bias
bias_f16[i * 8 + j] = ov::float16(-min_vals[j]);
} else {
// zp = min / scale (since bias = -min and zp = -bias/scale)
uint8_t zp_val = (scale_vals[j] != 0.0f) ? (uint8_t) std::round(min_vals[j] / scale_vals[j]) : 0;
// Pack two 4-bit zero points per byte
size_t idx = i * 8 + j;
if (idx % 2 == 0) {
zp_u4[idx / 2] = zp_val & 0x0F;
} else {
zp_u4[idx / 2] |= (zp_val << 4);
}
}
}
unpack_256_4(block_data + 16, weights + i * 128);
});
}
void extract_q6_k_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr) {
const uint64_t bytes_per_block = 128 + 64 + 16 + 2;
const uint64_t n_super_block = tensor->nb[3] / bytes_per_block;
auto * data = static_cast<uint8_t *>(tensor->data);
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
auto * zp = static_cast<uint8_t *>(zp_arr.data());
bool is_scalar_zp = (zp_arr.get_size() == 1); // Symmetric quantization
// For Q6_K, zero point is always 32
if (is_scalar_zp) {
zp[0] = 32;
}
ov::parallel_for(n_super_block, [&](size_t i) {
uint8_t * block_data = data + i * bytes_per_block;
float scale_factor =
static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 104))); // (128+64+16)/2
for (size_t j = 0; j < 16; j++) {
scales[j + i * 16] =
ov::float16(scale_factor * static_cast<float>(*((int8_t *) (block_data + 128 + 64 + j))));
// For asymmetric quantization, store per-block zero points
if (!is_scalar_zp) {
zp[j + i * 16] = 32;
}
}
uint8_t * ql = block_data;
uint8_t * qh = block_data + 128;
for (int64_t j = 0; j < 32; ++j) {
weights[i * 256 + j] = (ql[j] & 0xF) | (((qh[j] >> 0) & 3) << 4);
weights[i * 256 + j + 32] = (ql[32 + j] & 0xF) | (((qh[j] >> 2) & 3) << 4);
weights[i * 256 + j + 64] = (ql[j] >> 4) | (((qh[j] >> 4) & 3) << 4);
weights[i * 256 + j + 96] = (ql[32 + j] >> 4) | (((qh[j] >> 6) & 3) << 4);
weights[i * 256 + j + 128] = (ql[64 + j] & 0xF) | (((qh[32 + j] >> 0) & 3) << 4);
weights[i * 256 + j + 160] = (ql[96 + j] & 0xF) | (((qh[32 + j] >> 2) & 3) << 4);
weights[i * 256 + j + 192] = (ql[64 + j] >> 4) | (((qh[32 + j] >> 4) & 3) << 4);
weights[i * 256 + j + 224] = (ql[96 + j] >> 4) | (((qh[32 + j] >> 6) & 3) << 4);
}
});
}
static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t * d, uint8_t * m) {
if (j < 4) {
*d = q[j] & 63;
*m = q[j + 4] & 63;
} else {
*d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
*m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
}
}
void extract_q5_k_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
bool use_bias) {
const uint64_t bytes_per_block = 4 + 12 + 32 + 128;
const uint64_t n_super_block = tensor->nb[3] / bytes_per_block;
auto * data = static_cast<uint8_t *>(tensor->data);
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
// For bias path, zp_arr holds f16 bias values; for zp path, it holds u8 zero points
auto * zp_u8 = use_bias ? nullptr : static_cast<uint8_t *>(zp_arr.data());
auto * bias_f16 = use_bias ? zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>() : nullptr;
ov::parallel_for(n_super_block, [&](size_t i) {
uint8_t * block_data = data + i * bytes_per_block;
const float d = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data)));
const float min_factor = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 1)));
const uint8_t * scales_data = block_data + 4; // 12 bytes of scales
const uint8_t * qh = block_data + 4 + 12; // 32 bytes of high bits
const uint8_t * ql = block_data + 4 + 12 + 32; // 128 bytes of low bits
int is = 0;
uint8_t u1 = 1;
uint8_t u2 = 2;
// Process 2 blocks in one iteration
for (int j = 0; j < 256; j += 64) { // 256 = QK_K, so 4 iterations of 64
uint8_t sc;
uint8_t m;
// Get scale and min for first 32 elements
get_scale_min_k4(is + 0, scales_data, &sc, &m);
const float d1 = d * sc;
const float m1 = min_factor * m;
// Get scale and min for second 32 elements
get_scale_min_k4(is + 1, scales_data, &sc, &m);
const float d2 = d * sc;
const float m2 = min_factor * m;
scales[i * 8 + is] = ov::float16(d1);
scales[i * 8 + is + 1] = ov::float16(d2);
if (use_bias) {
// Store bias = -min directly as f16, dequant: w*s + bias
bias_f16[i * 8 + is] = ov::float16(-m1);
bias_f16[i * 8 + is + 1] = ov::float16(-m2);
} else {
// zp = min / scale (since bias = -min and zp = -bias/scale)
zp_u8[i * 8 + is] = (d1 != 0.0f) ? (uint8_t) std::round(m1 / d1) : 0;
zp_u8[i * 8 + is + 1] = (d2 != 0.0f) ? (uint8_t) std::round(m2 / d2) : 0;
}
// Extract weights for first 32 elements (matching deq formula exactly)
for (int l = 0; l < 32; ++l) {
weights[i * 256 + j + l] = (ql[l] & 0xF) + ((qh[l] & u1) ? 16 : 0);
}
// Extract weights for second 32 elements
for (int l = 0; l < 32; ++l) {
weights[i * 256 + j + l + 32] = (ql[l] >> 4) + ((qh[l] & u2) ? 16 : 0);
}
ql += 32;
is += 2;
u1 <<= 2;
u2 <<= 2;
}
});
}
// TODO Reorder for make_intX_weights
ov::Output<ov::Node> make_int8_weights(ov::Tensor & weight,
ov::Tensor & scales,
ov::Tensor & zp,
size_t group_size,
bool use_bias) {
ov::Shape orig_shape = weight.get_shape();
// Expand dimensions for scales and zp/bias
auto scale_shape = scales.get_shape();
auto zp_shape = zp.get_shape();
bool is_scalar_zp = zp_shape.empty(); // Symmetric quantization
ov::Shape packed_shape = {orig_shape[0], orig_shape[1] / group_size, group_size};
if (packed_shape[1] == 1) {
// Requantized channel-wise case
packed_shape.erase(packed_shape.begin() + 1);
} else {
scale_shape.push_back(1);
scales.set_shape(scale_shape);
// For symmetric quantization, zp remains scalar (don't resize)
if (!is_scalar_zp) {
zp_shape.push_back(1);
zp.set_shape(zp_shape);
}
}
// Create graph nodes
auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::u8, packed_shape,
static_cast<uint8_t *>(weight.data()), nullptr);
weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
auto scales_f16 = std::make_shared<ov::op::v0::Constant>(scales);
auto weights_f16 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f16);
ov::Output<ov::Node> result;
if (use_bias && !is_scalar_zp) {
// Bias path: w * s + b (zp tensor holds f16 bias values)
auto bias_f16 = std::make_shared<ov::op::v0::Constant>(zp);
auto w_s = std::make_shared<ov::op::v1::Multiply>(weights_f16, scales_f16, ov::op::AutoBroadcastType::NUMPY);
result = std::make_shared<ov::op::v1::Add>(w_s, bias_f16, ov::op::AutoBroadcastType::NUMPY);
} else {
// Zero point path: (w - zp) * s
auto zero_point = std::make_shared<ov::op::v0::Constant>(zp);
float zp_value;
if (ov::op::util::get_single_value(zero_point, zp_value)) {
zero_point = ov::op::v0::Constant::create(zero_point->get_element_type(), {}, {zp_value});
}
auto zero_point_f16 = std::make_shared<ov::op::v0::Convert>(zero_point, ov::element::f16);
auto w_zp =
std::make_shared<ov::op::v1::Subtract>(weights_f16, zero_point_f16, ov::op::AutoBroadcastType::NUMPY);
result = std::make_shared<ov::op::v1::Multiply>(w_zp, scales_f16, ov::op::AutoBroadcastType::NUMPY);
}
if (packed_shape.size() != 2) {
// If not requantized channel-wise case, reshape back to original shape
auto final_shape =
std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{orig_shape.size()}, orig_shape);
result = std::make_shared<ov::op::v1::Reshape>(result, final_shape, false);
}
return std::make_shared<ov::op::v0::Convert>(result, ov::element::f32);
}
ov::Output<ov::Node> make_int4_weights(ov::Tensor & weight,
ov::Tensor & scales,
ov::Tensor & zp,
size_t group_size,
bool use_bias) {
ov::Shape orig_weight_shape = weight.get_shape();
// Expand dimensions for scales and zp/bias
ov::Shape scale_shape = scales.get_shape();
auto zp_shape = zp.get_shape();
bool is_scalar_zp = zp_shape.empty(); // Symmetric quantization
// Create INT4 weight tensor
ov::Shape packed_shape = {orig_weight_shape[0], orig_weight_shape[1] / group_size, group_size};
if (packed_shape[1] == 1) {
// Requantized channel-wise case
packed_shape.erase(packed_shape.begin() + 1);
} else {
scale_shape.push_back(1);
scales.set_shape(scale_shape);
// For symmetric quantization, zp remains scalar (don't resize)
if (!is_scalar_zp) {
zp_shape.push_back(1);
zp.set_shape(zp_shape);
}
}
auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::u4, packed_shape,
static_cast<uint8_t *>(weight.data()), nullptr);
weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
auto weights_f16 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f16);
auto scales_f16 = std::make_shared<ov::op::v0::Constant>(scales);
ov::Output<ov::Node> result;
if (use_bias && !is_scalar_zp) {
// Bias path: w * s + b (zp tensor holds f16 bias values)
auto bias_f16 = std::make_shared<ov::op::v0::Constant>(zp);
auto w_s = std::make_shared<ov::op::v1::Multiply>(weights_f16, scales_f16, ov::op::AutoBroadcastType::NUMPY);
result = std::make_shared<ov::op::v1::Add>(w_s, bias_f16, ov::op::AutoBroadcastType::NUMPY);
} else {
// Zero point path: (w - zp) * s
auto zero_points_node = std::make_shared<ov::op::v0::Constant>(zp);
float zp_value;
if (ov::op::util::get_single_value(zero_points_node, zp_value)) {
zero_points_node = ov::op::v0::Constant::create(zero_points_node->get_element_type(), {}, {zp_value});
}
auto zero_points_f16 = std::make_shared<ov::op::v0::Convert>(zero_points_node, ov::element::f16);
auto w_zp =
std::make_shared<ov::op::v1::Subtract>(weights_f16, zero_points_f16, ov::op::AutoBroadcastType::NUMPY);
result = std::make_shared<ov::op::v1::Multiply>(w_zp, scales_f16, ov::op::AutoBroadcastType::NUMPY);
}
if (packed_shape.size() != 2) {
// If not requantized channel-wise case, reshape back to original shape
auto final_shape = std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{orig_weight_shape.size()},
orig_weight_shape);
result = std::make_shared<ov::op::v1::Reshape>(result, final_shape, false);
}
return std::make_shared<ov::op::v0::Convert>(result, ov::element::f32);
}
// Extract quantized weights from tensor and create weight subgraph
std::shared_ptr<ov::Node> extract_quantized_weights(const ggml_tensor * tensor,
const void * data,
ov::Tensor & weights,
ov::Tensor & scales,
ov::Tensor & zp,
bool use_bias) {
// Create a temporary tensor for extraction functions that read from tensor->data
ggml_tensor temp_tensor = *tensor;
temp_tensor.data = const_cast<void *>(data);
// Determine block size based on tensor type
int64_t weights_per_block;
bool is_u4;
switch (tensor->type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q4_K:
is_u4 = true;
weights_per_block = 32;
break;
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q5_K:
is_u4 = false;
weights_per_block = 32;
break;
case GGML_TYPE_Q6_K:
is_u4 = false;
weights_per_block = 16;
break;
default:
throw std::runtime_error("Unsupported quantized type for extraction: " +
std::string(ggml_type_name(tensor->type)));
}
// Extract quantized data
switch (tensor->type) {
case GGML_TYPE_Q4_0:
extract_q4_0_data(&temp_tensor, weights, scales, zp);
break;
case GGML_TYPE_Q4_1:
extract_q4_1_data(&temp_tensor, weights, scales, zp, use_bias);
break;
case GGML_TYPE_Q4_K:
extract_q4_k_data(&temp_tensor, weights, scales, zp, use_bias);
break;
case GGML_TYPE_Q8_0:
extract_q8_0_data(&temp_tensor, weights, scales, zp);
break;
case GGML_TYPE_Q6_K:
extract_q6_k_data(&temp_tensor, weights, scales, zp);
break;
case GGML_TYPE_Q5_K:
extract_q5_k_data(&temp_tensor, weights, scales, zp, use_bias);
break;
default:
throw std::runtime_error("Unsupported quantized type: " + std::string(ggml_type_name(tensor->type)));
}
// Create the OpenVINO weight subgraph
ov::Output<ov::Node> weight_node;
if (is_u4) {
weight_node = make_int4_weights(weights, scales, zp, weights_per_block, use_bias);
} else {
weight_node = make_int8_weights(weights, scales, zp, weights_per_block, use_bias);
}
auto result = weight_node.get_node_shared_ptr();
result->set_friendly_name(tensor->name);
return result;
}
// Requantize weights to target format, writing to provided buffers
std::shared_ptr<ov::Node> requantize_to_buffers(const ggml_tensor * tensor,
const void * data,
ExtraQuantType requant_type,
int64_t block_size,
ov::Tensor & weights,
ov::Tensor & scales,
ov::Tensor & zp) {
int64_t n_elements = ggml_nelements(tensor);
// First dequantize to F32
std::vector<float> weights_f32(n_elements);
ggml_get_type_traits(tensor->type)->to_float(data, weights_f32.data(), n_elements);
// Handle F16 case - just convert and create constant
if (requant_type == ExtraQuantType::F16) {
ggml_get_type_traits(GGML_TYPE_F16)->from_float_ref(weights_f32.data(), weights.data(), n_elements);
auto result = std::make_shared<ov::op::v0::Constant>(weights);
result->set_friendly_name(tensor->name);
return result;
}
// Requantize to target quantized format
bool is_u4 = (requant_type == ExtraQuantType::Q4_0_C || requant_type == ExtraQuantType::Q4_0_128);
if (is_u4) {
quantize_q4_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
} else if (requant_type == ExtraQuantType::Q8_1_C) {
quantize_q8_1(weights_f32.data(), weights, scales, zp, n_elements, block_size);
} else {
quantize_q8_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
}
// Create the OpenVINO weight subgraph
ov::Output<ov::Node> weight_node;
if (is_u4) {
weight_node = make_int4_weights(weights, scales, zp, block_size);
} else {
weight_node = make_int8_weights(weights, scales, zp, block_size);
}
auto result = weight_node.get_node_shared_ptr();
result->set_friendly_name(tensor->name);
return result;
}
OvWeight process_weight_tensor(const ggml_tensor * tensor, const void * data, void * output_base_ptr, bool use_bias) {
GGML_ASSERT(tensor != nullptr);
GGML_ASSERT(data != nullptr);
OvWeight result;
// Get 2D shape for weights [rows, cols]
ov::Shape node_shape = {static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
// Handle F16/F32/BF16 weights
if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
ov::element::Type element_type;
switch (tensor->type) {
case GGML_TYPE_F32:
element_type = ov::element::f32;
break;
case GGML_TYPE_F16:
element_type = ov::element::f16;
break;
case GGML_TYPE_BF16:
element_type = ov::element::bf16;
break;
default:
OPENVINO_THROW("Unexpected tensor type in F16/F32/BF16 path");
}
if (output_base_ptr && output_base_ptr != data) {
// Using external buffer - copy data and create shared-memory constant
size_t tensor_bytes = ggml_nbytes(tensor);
memcpy(output_base_ptr, data, tensor_bytes);
result.weights = ov::Tensor(element_type, node_shape, output_base_ptr);
} else {
result.weights = ov::Tensor(element_type, node_shape, data);
}
result.weight_node = std::make_shared<ov::op::v0::Constant>(result.weights);
return result;
}
// Handle quantized weights
if (!ggml_is_quantized(tensor->type)) {
OPENVINO_THROW("Unsupported weight tensor type: ", ggml_type_name(tensor->type));
}
result.layout = ggml_openvino_get_extracted_layout(tensor, use_bias);
const auto & layout = result.layout;
if (layout.total_size == 0) {
OPENVINO_THROW("Unsupported quantized type: ", ggml_type_name(tensor->type));
}
if (use_bias) {
OPENVINO_ASSERT(!layout.is_requant,
"use_bias is only used for test-backend-ops, which should not have requantization");
// bias node will be created on the fly and not use backend buffer
output_base_ptr = nullptr;
}
// F16 requant path - no separate scales/zp needed in result
if (layout.is_requant && layout.requant_type.has_value() && layout.requant_type.value() == ExtraQuantType::F16) {
if (output_base_ptr) {
result.weights = ov::Tensor(ov::element::f16, node_shape,
static_cast<uint8_t *>(output_base_ptr) + layout.weights_offset);
} else {
result.weights = ov::Tensor(ov::element::f16, node_shape);
}
ov::Tensor dummy_scales, dummy_zp; // Not used for F16
result.weight_node =
requantize_to_buffers(tensor, data, ExtraQuantType::F16, 0, result.weights, dummy_scales, dummy_zp);
return result;
}
// Quantized path (normal extraction or quantized requant)
// Create weight/scale/zp tensors - shared between both paths
ov::element::Type weight_type = layout.is_u4 ? ov::element::u4 : ov::element::u8;
ov::Shape scale_shape = {node_shape[0], node_shape[1] / layout.weights_per_block};
ov::Shape zp_shape = layout.is_symmetric ? ov::Shape{} : scale_shape;
if (output_base_ptr) {
uint8_t * buf_base = static_cast<uint8_t *>(output_base_ptr);
result.weights = ov::Tensor(weight_type, node_shape, buf_base + layout.weights_offset);
result.scales = ov::Tensor(ov::element::f16, scale_shape, buf_base + layout.scales_offset);
result.zp = ov::Tensor(weight_type, zp_shape, buf_base + layout.zp_offset);
} else {
result.weights = ov::Tensor(weight_type, node_shape);
result.scales = ov::Tensor(ov::element::f16, scale_shape);
if (use_bias && !layout.is_symmetric) {
// bias only has effect for asymmetric quant
result.zp = ov::Tensor(ov::element::f16, zp_shape);
} else {
result.zp = ov::Tensor(weight_type, zp_shape);
}
}
if (layout.is_requant && layout.requant_type.has_value()) {
result.weight_node = requantize_to_buffers(tensor, data, layout.requant_type.value(), layout.weights_per_block,
result.weights, result.scales, result.zp);
} else {
result.weight_node =
extract_quantized_weights(tensor, data, result.weights, result.scales, result.zp, use_bias);
}
return result;
}
void quantize_q4_0(const float * x,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
int64_t k,
int64_t qk) {
assert(k % qk == 0);
const int nb = k / qk;
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
auto * zp = static_cast<uint8_t *>(zp_arr.data());
bool is_scalar_zp = (zp_arr.get_size() == 1); // Symmetric quantization
// For Q4_0, zero point is always 8
if (is_scalar_zp) {
zp[0] = 8 | (8 << 4); // Pack two 4-bit values
}
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int j = 0; j < qk; j++) {
const float v = x[i * qk + j];
if (amax < fabsf(v)) {
amax = fabsf(v);
max = v;
}
}
const float d = max / -8;
if (d == 0) {
scales[i] = ov::float16(1.0f);
// zp is already set to 8 for symmetric, or set per-block for asymmetric
if (!is_scalar_zp) {
if (i % 2 == 0) {
zp[i / 2] = 8;
} else {
zp[i / 2] |= (8 << 4);
}
}
memset(weights + i * qk / 2, 8 | (8 << 4), qk / 2);
continue;
}
const float id = 1.0f / d;
scales[i] = ov::float16(d);
// For asymmetric quantization, store per-block zero points
if (!is_scalar_zp) {
if (i % 2 == 0) {
zp[i / 2] = 8;
} else {
zp[i / 2] |= (8 << 4);
}
}
for (int j = 0; j < qk / 2; ++j) {
const float x0 = x[i * qk + 2 * j] * id;
const float x1 = x[i * qk + 2 * j + 1] * id;
const uint8_t xi0 = MIN(15, (int8_t) (x0 + 8.5f));
const uint8_t xi1 = MIN(15, (int8_t) (x1 + 8.5f));
weights[i * qk / 2 + j] = xi0 | (xi1 << 4);
}
}
}
void quantize_q8_0(const float * x,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
int64_t k,
int64_t qk) {
assert(k % qk == 0);
const int nb = k / qk;
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
auto * zp = static_cast<uint8_t *>(zp_arr.data());
bool is_scalar_zp = (zp_arr.get_size() == 1); // Symmetric quantization
// For Q8_0, zero point is always 128
if (is_scalar_zp) {
zp[0] = 128;
}
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < qk; j++) {
const float v = x[i * qk + j];
if (amax < fabsf(v)) {
amax = fabsf(v);
}
}
const float d = amax / 127.0f;
const float id = d ? 1.0f / d : 0.0f;
scales[i] = ov::float16(d);
// For asymmetric quantization, store per-block zero points
if (!is_scalar_zp) {
zp[i] = 128;
}
for (int j = 0; j < qk; ++j) {
const float x0 = x[i * qk + j] * id;
const int8_t xi0 = roundf(x0);
weights[i * qk + j] = (uint8_t) (xi0 + 128);
}
}
}
void quantize_q8_1(const float * x,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
int64_t k,
int64_t qk) {
assert(k % qk == 0);
const int nb = k / qk;
auto * weights = static_cast<uint8_t *>(weights_arr.data());
auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
auto * zp = static_cast<uint8_t *>(zp_arr.data());
for (int i = 0; i < nb; i++) {
float min = std::numeric_limits<float>::max();
float max = std::numeric_limits<float>::lowest();
for (int j = 0; j < qk; j++) {
const float v = x[i * qk + j];
if (v < min) {
min = v;
}
if (v > max) {
max = v;
}
}
const float d = (max - min) / ((1 << 8) - 1);
const float id = d ? 1.0f / d : 0.0f;
scales[i] = ov::float16(d);
// zp = -min / scale (Q8_1 is asymmetric)
zp[i] = (d != 0.0f) ? (uint8_t) std::round(-min / d) : 0;
for (int j = 0; j < qk; ++j) {
const float x0 = (x[i * qk + j] - min) * id;
const uint8_t xi0 = roundf(x0);
weights[i * qk + j] = xi0;
}
}
}

153
ggml/src/ggml-openvino/ggml-quants.h Normal file
View File

@ -0,0 +1,153 @@
#pragma once
#include "ggml-openvino-extra.h" // For ExtraQuantType
#include "ggml.h"
#include <cstdint>
#include <openvino/op/constant.hpp>
#include <openvino/runtime/tensor.hpp>
void unpack_32_4(const uint8_t* data, uint8_t* dst);
void extract_q4_0_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr);
void extract_q4_1_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
bool use_bias = false);
void extract_q8_0_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr);
void unpack_256_4(const uint8_t* data, uint8_t* dst);
void extract_q4_k_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
bool use_bias = false);
void extract_q5_k_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
bool use_bias = false);
void extract_q6_k_data(const ggml_tensor * tensor,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr);
static constexpr size_t GGML_QUANTIZATION_GROUP_SIZE = 32;
ov::Output<ov::Node> make_int8_weights(ov::Tensor & weight,
ov::Tensor & scales,
ov::Tensor & zp,
size_t group_size = GGML_QUANTIZATION_GROUP_SIZE,
bool use_bias = false);
ov::Output<ov::Node> make_int4_weights(ov::Tensor & weight,
ov::Tensor & scales,
ov::Tensor & zp,
size_t group_size = GGML_QUANTIZATION_GROUP_SIZE,
bool use_bias = false);
// Extract quantized weights from tensor and create weight subgraph
// If weights/scales/zp are provided (non-empty), uses them as output buffers
// Otherwise allocates new ov::Tensors internally
// Returns the weight node (make_int4_weights or make_int8_weights result)
std::shared_ptr<ov::Node> extract_quantized_weights(
const ggml_tensor * tensor,
const void * data, // Source data pointer (may differ from tensor->data)
ov::Tensor & weights,
ov::Tensor & scales,
ov::Tensor & zp,
bool use_bias = false); // Use fp bias instead of quantized zero_point (for test-backend-ops)
// Requantize weights from tensor to target format, writing to provided buffers
// For F16 target, only weights buffer is used (scales/zp ignored)
// Returns the weight node
std::shared_ptr<ov::Node> requantize_to_buffers(const ggml_tensor * tensor,
const void * data, // Source data pointer
ExtraQuantType requant_type,
int64_t block_size,
ov::Tensor & weights,
ov::Tensor & scales,
ov::Tensor & zp);
inline const char * extra_quant_type_name(ExtraQuantType t) {
switch (t) {
case ExtraQuantType::F16:
return "F16";
case ExtraQuantType::Q4_0_C:
return "Q4_0_C";
case ExtraQuantType::Q4_0_128:
return "Q4_0_128";
case ExtraQuantType::Q8_0_C:
return "Q8_0_C";
case ExtraQuantType::Q8_0_32:
return "Q8_0_32";
case ExtraQuantType::Q8_1_C:
return "Q8_1_C";
default:
return "unknown";
}
}
// Result from process_weight_tensor containing the weight node and tensors.
// For quantized weights, also contains the extracted layout and scale/zp tensors.
struct OvWeight {
std::shared_ptr<ov::Node> weight_node;
ggml_openvino_extracted_layout layout; // Only meaningful for quantized (layout.total_size > 0)
ov::Tensor weights;
ov::Tensor scales;
ov::Tensor zp;
bool is_quantized() const { return layout.scales_size > 0; }
};
// Process weight tensor and create an OpenVINO weight node
// Handles F16/F32/BF16 and quantized weights, with optional requantization
// If output_base_ptr is nullptr, allocates internal buffers (for decoder use)
// If output_base_ptr is provided, uses pre-allocated buffers at specified offsets (for backend buffer use)
// Returns OvWeight with the weight node and optional quantized tensors
OvWeight process_weight_tensor(
const ggml_tensor * tensor,
const void * data, // Source data pointer (may differ from tensor->data)
void * output_base_ptr = nullptr, // Base pointer for output buffers (or nullptr for internal allocation)
bool use_bias = false); // Use fp bias instead of quantized zero_point, only used in test-backend-ops
void quantize_q4_0(const float * x,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
int64_t k,
int64_t qk);
void quantize_q8_1(const float * x,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
int64_t k,
int64_t qk);
void quantize_q8_0(const float * x,
ov::Tensor & weights_arr,
ov::Tensor & scales_arr,
ov::Tensor & zp_arr,
int64_t k,
int64_t qk);
namespace ov {
namespace op {
namespace util {
// From <openvino>/src/common/transformations/include/transformations/utils/utils.hpp
bool get_single_value(const std::shared_ptr<ov::op::v0::Constant>& const_node,
float& value,
bool check_value_range = true);
} // namespace util
} // namespace op
} // namespace ov

74
ggml/src/ggml-openvino/openvino/decoder.h Normal file
View File

@ -0,0 +1,74 @@
#pragma once
#include <cstdint>
#include <map>
#include <openvino/core/node.hpp>
#include <openvino/frontend/decoder.hpp>
#include <string>
namespace ov {
namespace frontend {
namespace ggml {
class GgmlDecoder : public DecoderBase {
public:
virtual ov::Any get_attribute(const std::string& name) const = 0;
virtual PartialShape get_input_shape(int node_idx, const std::string& name) const = 0;
virtual std::vector<size_t> get_input_stride(int node_idx, const std::string& name) const = 0;
virtual element::Type get_input_type(int node_idx, const std::string& name) const = 0;
virtual size_t get_input_size() const = 0;
virtual size_t get_input_size(int node_idx) const = 0;
virtual void get_input_node(size_t input_port_idx,
std::string& producer_name,
std::string& producer_output_port_name,
size_t& producer_output_port_index) const = 0;
virtual std::vector<std::string> get_input_names(int node_idx) const = 0;
virtual PartialShape get_output_shape(int node_idx) const = 0;
virtual element::Type get_output_type(const int node_idx) const = 0;
virtual int32_t* get_input_op_params(int node_idx, const std::string& name) const = 0;
virtual int32_t * get_output_op_params(int node_idx) const = 0;
virtual std::vector<std::string> get_output_names(int node_idx) const = 0;
virtual const std::string& get_op_type() const = 0;
virtual const std::string& get_op_type(int node_idx) const = 0;
virtual const std::string& get_op_name() const = 0;
virtual const std::string& get_op_name(int node_idx) const = 0;
virtual void visit_subgraph(std::function<void(std::shared_ptr<GgmlDecoder>, int node_idx)> node_visitor) const = 0;
virtual int get_op_case(int node_idx) const = 0;
virtual const std::map<std::string, std::shared_ptr<ov::Node>>& get_model_inputs() const = 0;
virtual const std::map<std::string, std::shared_ptr<ov::Node>>& get_model_extra_inputs() const = 0;
virtual const std::map<std::string, std::shared_ptr<ov::Node>>& get_model_weights() const = 0;
virtual std::vector<std::string> get_model_output_names() const = 0;
virtual int32_t* get_rope_params() const = 0;
virtual std::map<std::string, std::string> get_kv_param_res_names() const = 0;
virtual bool is_static() const = 0;
virtual bool is_stateful() const = 0;
virtual int is_swa_layer(int layer) const = 0;
};
} // namespace ggml
} // namespace frontend
} // namespace ov

27
ggml/src/ggml-openvino/openvino/frontend.cpp Normal file
View File

@ -0,0 +1,27 @@
#include "frontend.h"
#include "input_model.h"
#include "op_table.h"
#include "translate_session.h"
namespace ov {
namespace frontend {
namespace ggml {
FrontEnd::FrontEnd() {}
std::shared_ptr<Model> FrontEnd::convert(const InputModel::Ptr & model, bool naive) {
auto ggml_model = std::dynamic_pointer_cast<ggml::InputModel>(model);
FRONT_END_GENERAL_CHECK(ggml_model, "Invalid input model");
std::shared_ptr<Model> converted_model;
const auto & supported_ops = get_supported_ops();
{
TranslateSession translate_session(model, supported_ops, naive);
converted_model = translate_session.get_converted_model();
}
return converted_model;
}
} // namespace ggml
} // namespace frontend
} // namespace ov

23
ggml/src/ggml-openvino/openvino/frontend.h Normal file
View File

@ -0,0 +1,23 @@
// Copyright (C) 2018-2024 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#pragma once
#include <openvino/frontend/frontend.hpp>
namespace ov {
namespace frontend {
namespace ggml {
class FrontEnd {
public:
using Ptr = std::shared_ptr<FrontEnd>;
FrontEnd();
static std::shared_ptr<Model> convert(const InputModel::Ptr& model, bool naive = false);
};
} // namespace ggml
} // namespace frontend
} // namespace ov

17
ggml/src/ggml-openvino/openvino/input_model.cpp Normal file
View File

@ -0,0 +1,17 @@
#include "input_model.h"
#include "decoder.h"
namespace ov {
namespace frontend {
namespace ggml {
InputModel::InputModel(const std::shared_ptr<GgmlDecoder> & gdecoder) : m_decoder(gdecoder) {}
const std::shared_ptr<GgmlDecoder> & InputModel::get_model_decoder() const {
return m_decoder;
}
} // namespace ggml
} // namespace frontend
} // namespace ov

29
ggml/src/ggml-openvino/openvino/input_model.h Normal file
View File

@ -0,0 +1,29 @@
#pragma once
#include <openvino/frontend/input_model.hpp>
#include "decoder.h"
namespace ov {
namespace frontend {
namespace ggml {
class FrontEnd;
class GgmlDecoder;
using ov::frontend::ggml::GgmlDecoder;
class InputModel : public ov::frontend::InputModel {
friend class ::ov::frontend::ggml::FrontEnd;
public:
explicit InputModel(const std::shared_ptr<GgmlDecoder>& gdecoder);
const std::shared_ptr<GgmlDecoder>& get_model_decoder() const;
private:
std::shared_ptr<GgmlDecoder> m_decoder;
};
} // namespace ggml
} // namespace frontend
} // namespace ov

112
ggml/src/ggml-openvino/openvino/node_context.h Normal file
View File

@ -0,0 +1,112 @@
#pragma once
#include <cstdint>
#include <openvino/frontend/node_context.hpp>
#include <string>
#include "decoder.h"
namespace ov {
namespace frontend {
namespace ggml {
class TranslateSession;
typedef std::map<std::string, Output<Node>> TensorMap;
class NodeContext : public frontend::NodeContext {
public:
NodeContext(const std::shared_ptr<GgmlDecoder>& decoder,
std::shared_ptr<TensorMap>& tensor_map,
int node_idx,
TranslateSession* translate_session = nullptr)
: ov::frontend::NodeContext(decoder->get_op_type(node_idx)),
m_decoder(decoder),
m_tensor_map(tensor_map),
m_node_idx(node_idx),
m_translate_session(translate_session) {
m_input_names = decoder->get_input_names(m_node_idx);
m_output_names = decoder->get_output_names(m_node_idx);
}
TranslateSession* get_translate_session() const {
return m_translate_session;
}
const std::vector<std::string>& get_input_names() const { return m_input_names; }
size_t get_input_size() const override {
return m_decoder->get_input_size(m_node_idx);
}
ov::element::Type get_input_type(size_t index) const {
return m_decoder->get_input_type(m_node_idx, m_input_names[index]);
}
PartialShape get_input_shape(size_t input_index) const {
return m_decoder->get_input_shape(m_node_idx, m_input_names[input_index]);
}
std::vector<size_t> get_input_stride(size_t index) const {
return m_decoder->get_input_stride(m_node_idx, m_input_names[index]);
}
std::string get_output_name() const { return m_output_names[0]; }
PartialShape get_output_shape() const { return m_decoder->get_output_shape(m_node_idx); }
int32_t* get_input_op_params(size_t index) const {
return m_decoder->get_input_op_params(m_node_idx, m_input_names[index]);
}
int32_t * get_output_op_params() const { return m_decoder->get_output_op_params(m_node_idx); }
ov::element::Type get_output_type() const {
return m_decoder->get_output_type(m_node_idx);
}
Output<Node> get_input(int idx) const override {
return m_tensor_map->at(m_input_names[idx]);
}
Output<Node> get_input(const std::string& name) const override {
if (m_tensor_map->find(name) == m_tensor_map->end()) {
throw std::runtime_error("'" + name + "' not found in tensor map.");
}
return m_tensor_map->at(name);
}
bool has_input(const std::string& name) const {
return m_tensor_map->find(name) != m_tensor_map->end();
}
const std::string& get_name() const override {
return m_decoder->get_op_name(m_node_idx);
}
ov::Any get_attribute_as_any(const std::string& name) const override {
return m_decoder->get_attribute(name);
}
int get_op_case() const {
return m_decoder->get_op_case(m_node_idx);
}
bool is_static() const { return m_decoder->is_static(); }
bool is_stateful() const { return m_decoder->is_stateful(); }
private:
std::shared_ptr<GgmlDecoder> m_decoder;
std::shared_ptr<TensorMap>& m_tensor_map;
int m_node_idx;
TranslateSession* m_translate_session;
std::vector<std::string> m_input_names;
std::vector<std::string> m_output_names;
};
using CreatorFunction = std::function<ov::OutputVector(const ov::frontend::ggml::NodeContext&)>;
} // namespace ggml
} // namespace frontend
} // namespace ov

48
ggml/src/ggml-openvino/openvino/op/cont.cpp Normal file
View File

@ -0,0 +1,48 @@
#include "../node_context.h"
#include "../op_table.h"
#include "../utils.h"
#include <climits>
#include <cstdint>
#include <memory>
#include <openvino/op/reshape.hpp>
#include <openvino/op/slice.hpp>
#include <vector>
namespace ov {
namespace frontend {
namespace ggml {
namespace op {
OutputVector translate_cont(const NodeContext & context) {
num_inputs_check(context, 1, 1);
int op_case = context.get_op_case();
FRONT_END_CHECK_IMPLEMENTED(op_case == 1 || op_case == 2 || op_case == 3, "Unsupported CONT case");
auto src_shape = context.get_input_shape(0).to_shape();
auto dst_shape = context.get_output_shape().to_shape();
ov::Output<Node> res;
if (op_case == 1) {
// The input comes from a PERMUTE
throw std::runtime_error("Code of this case might be outdated");
dst_shape[1] = -1;
res = std::make_shared<ov::op::v1::Reshape>(
context.get_input(0), ov::op::v0::Constant::create(ov::element::i64, {dst_shape.size()}, dst_shape), false);
} else if (op_case == 2) {
// The input comes from a TRANSPOSE
return {context.get_input(0)};
} else {
// The input comes from a VIEW
res = process_view_input(context, 0);
}
return rename_outputs_with_suffix({res}, context.get_name());
}
} // namespace op
} // namespace ggml
} // namespace frontend
} // namespace ov

21
ggml/src/ggml-openvino/openvino/op/cpy.cpp Normal file
View File

@ -0,0 +1,21 @@
#include "../node_context.h"
#include "../op_table.h"
#include "../utils.h"
#include <memory>
#include <openvino/op/convert.hpp>
namespace ov {
namespace frontend {
namespace ggml {
namespace op {
OutputVector translate_cpy(const NodeContext & context) {
auto res = std::make_shared<ov::op::v0::Convert>(context.get_input(0), context.get_output_type());
return rename_outputs_with_suffix({res}, context.get_name());
}
} // namespace op
} // namespace ggml
} // namespace frontend
} // namespace ov

90
ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp Normal file
View File

@ -0,0 +1,90 @@
#include "../node_context.h"
#include "../op_table.h"
#include "../utils.h"
#include <cstdint>
#include <memory>
#include <openvino/op/broadcast.hpp>
#include <openvino/op/concat.hpp>
#include <openvino/op/constant.hpp>
#include <openvino/op/convert.hpp>
#include <openvino/op/reshape.hpp>
#include <openvino/op/scaled_dot_product_attention.hpp>
#include <openvino/op/transpose.hpp>
#include <openvino/op/unsqueeze.hpp>
#include <string>
namespace ov {
namespace frontend {
namespace ggml {
namespace op {
OutputVector translate_flash_attn_ext(const NodeContext & context) {
num_inputs_check(context, 4, 4);
auto q_f32 = context.get_input(0);
auto k = context.get_input(1);
auto v = context.get_input(2);
auto mask = context.get_input(3);
float * params = reinterpret_cast<float *>(context.get_output_op_params());
float scale = params[0];
// float max_bias = params[1];
// float logit_softcap = params[2];
auto q = std::make_shared<ov::op::v0::Convert>(q_f32, ov::element::f16);
auto scale_node = std::make_shared<ov::op::v0::Constant>(ov::element::f16, ov::Shape{}, std::vector<float>{scale});
ov::Output<ov::Node> mask_sliced, res;
std::string mask_name = "KQ_mask_sliced";
if (context.get_input_names()[3].find("swa") != std::string::npos) {
mask_name = "KQ_mask_swa_sliced";
}
if (context.has_input(mask_name)) {
mask_sliced = context.get_input(mask_name);
} else {
auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
auto two = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
auto token_len = get_dimensions(q, {2});
mask_sliced = std::make_shared<ov::op::v8::Slice>(mask, zero, token_len, one, two);
}
if (mask_sliced.get_element_type() != ov::element::f16) {
mask_sliced = std::make_shared<ov::op::v0::Convert>(mask_sliced, ov::element::f16);
}
auto tile_kv = [&](int64_t num_heads, int64_t num_heads_kv, int64_t head_size, ov::Output<Node> kv) {
int64_t factor = num_heads / num_heads_kv;
if (factor > 1 && num_heads_kv > 1) {
ov::Output<ov::Node> kv_broadcast_shape, kv_unsqueezed, new_kv_shape;
auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, Shape{}, {2});
kv_unsqueezed = std::make_shared<ov::op::v0::Unsqueeze>(kv, unsqueeze_axes);
kv_broadcast_shape = ov::op::v0::Constant::create(
ov::element::i64, {5}, {(int64_t) 1, (int64_t) 1, factor, (int64_t) 1, (int64_t) 1});
new_kv_shape =
ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 0, num_heads, (int64_t) -1, head_size});
kv = std::make_shared<ov::op::v3::Broadcast>(kv_unsqueezed, kv_broadcast_shape,
ov::op::BroadcastType::BIDIRECTIONAL);
kv = std::make_shared<ov::op::v1::Reshape>(kv, new_kv_shape, true);
}
return kv;
};
auto q_shape = context.get_input_shape(0).to_shape();
auto k_shape = context.get_input_shape(1).to_shape();
k = tile_kv(q_shape[1], k_shape[1], q_shape[3], k);
v = tile_kv(q_shape[1], k_shape[1], q_shape[3], v);
auto sdpa = std::make_shared<ov::op::v13::ScaledDotProductAttention>(q, k, v, mask_sliced, scale_node, false);
res = std::make_shared<ov::op::v1::Transpose>(sdpa,
ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3}));
res = std::make_shared<ov::op::v0::Convert>(res, ov::element::f32);
return rename_outputs_with_suffix({res}, context.get_name());
}
} // namespace op
} // namespace ggml
} // namespace frontend
} // namespace ov

69
ggml/src/ggml-openvino/openvino/op/get_rows.cpp Normal file
View File

@ -0,0 +1,69 @@
#include "../node_context.h"
#include "../op_table.h"
#include "../utils.h"
#include <openvino/core/node.hpp>
#include <openvino/core/node_output.hpp>
#include <openvino/op/constant.hpp>
#include <openvino/op/convert.hpp>
#include <openvino/op/gather.hpp>
#include <openvino/op/squeeze.hpp>
#include <openvino/op/unsqueeze.hpp>
namespace ov {
namespace frontend {
namespace ggml {
namespace op {
OutputVector translate_get_rows(const NodeContext & context) {
num_inputs_check(context, 2, 2);
int op_case = context.get_op_case();
Output<Node> res;
auto data = context.get_input(0);
auto indices = context.get_input(1);
if (op_case == 2) {
// The input comes from a VIEW
indices = process_view_input(context, 1);
}
// data[1,b,x,y] ind[1,1,b,x'] test-backend-ops case
// data[x,y] ind[1,1,1,x'] normal case
indices =
std::make_shared<ov::op::v0::Squeeze>(indices, ov::op::v0::Constant::create(ov::element::i64, {2}, {0, 1}));
if (data.get_partial_shape().rank() == 4) {
if (!(data.get_partial_shape()[1].is_dynamic()) && data.get_partial_shape()[1].get_length() == 1) {
// Work-around for a bug in ov cpu plugin for test-backend-ops
data = std::make_shared<ov::op::v0::Squeeze>(data,
ov::op::v0::Constant::create(ov::element::i64, {2}, {0, 1}));
auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
res = std::make_shared<ov::op::v8::Gather>(data, indices, axis);
} else {
auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {1});
data =
std::make_shared<ov::op::v0::Squeeze>(data, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
res = std::make_shared<ov::op::v8::Gather>(data, indices, axis, 1);
}
} else if (context.is_stateful() && data.get_partial_shape().rank() == 3) {
auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {1});
res = std::make_shared<ov::op::v8::Gather>(data, indices, axis, 1);
} else {
auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
res = std::make_shared<ov::op::v8::Gather>(data, indices, axis);
}
if (res.get_element_type() != context.get_output_type()) {
res = std::make_shared<ov::op::v0::Convert>(res, context.get_output_type());
}
if (!(context.is_stateful())) {
res = std::make_shared<ov::op::v0::Unsqueeze>(res, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
}
return rename_outputs_with_suffix({res}, context.get_name());
}
} // namespace op
} // namespace ggml
} // namespace frontend
} // namespace ov

61
ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp Normal file
View File

@ -0,0 +1,61 @@
#include "../node_context.h"
#include "../op_table.h"
#include "../utils.h"
#include <memory>
#include <openvino/core/node_output.hpp>
#include <openvino/op/constant.hpp>
#include <openvino/op/gelu.hpp>
#include <openvino/op/multiply.hpp>
#include <openvino/op/sigmoid.hpp>
#include <openvino/op/slice.hpp>
namespace ov {
namespace frontend {
namespace ggml {
namespace op {
OutputVector translate_glu_geglu(const NodeContext & context) {
num_inputs_check(context, 1, 2);
ov::Output<ov::Node> src0;
ov::Output<ov::Node> src1;
if (context.get_input_size() == 2) {
src0 = context.get_input(0);
src1 = context.get_input(1);
} else {
// GGML splits along ne[0] (OV last axis) using floor division: nc = ne[0] / 2.
// Both halves are nc elements; if the dimension is odd, the last element is dropped.
// Use Slice instead of Split to handle odd dimensions correctly.
auto combined = context.get_input(0);
auto combined_shape = combined.get_partial_shape();
int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
int64_t nc = last_dim_val / 2;
auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
auto stop0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
auto stop1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
}
int32_t * params = context.get_output_op_params();
const int32_t swapped = params[1];
if (swapped) {
std::swap(src0, src1);
}
auto gelu = std::make_shared<ov::op::v7::Gelu>(src0);
auto res = std::make_shared<ov::op::v1::Multiply>(gelu, src1);
return rename_outputs_with_suffix({res}, context.get_name());
}
} // namespace op
} // namespace ggml
} // namespace frontend
} // namespace ov

62
ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp Normal file
View File

@ -0,0 +1,62 @@
#include "../node_context.h"
#include "../op_table.h"
#include "../utils.h"
#include <cstdint>
#include <memory>
#include <openvino/core/node_output.hpp>
#include <openvino/op/constant.hpp>
#include <openvino/op/multiply.hpp>
#include <openvino/op/sigmoid.hpp>
#include <openvino/op/slice.hpp>
namespace ov {
namespace frontend {
namespace ggml {
namespace op {
OutputVector translate_glu_swiglu(const NodeContext & context) {
num_inputs_check(context, 1, 2);
ov::Output<ov::Node> src0;
ov::Output<ov::Node> src1;
if (context.get_input_size() == 2) {
src0 = context.get_input(0);
src1 = context.get_input(1);
} else {
// GGML splits along ne[0] (OV last axis) using floor division: nc = ne[0] / 2.
// Both halves are nc elements; if the dimension is odd, the last element is dropped.
// Use Slice instead of Split to handle odd dimensions correctly.
auto combined = context.get_input(0);
auto combined_shape = combined.get_partial_shape();
int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
int64_t nc = last_dim_val / 2;
auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
auto stop0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
auto stop1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
}
int32_t * params = context.get_output_op_params();
const int32_t swapped = params[1];
if (swapped) {
std::swap(src0, src1);
}
auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(src0);
auto silu = std::make_shared<ov::op::v1::Multiply>(src0, sigmoid);
auto res = std::make_shared<ov::op::v1::Multiply>(silu, src1);
return rename_outputs_with_suffix({res}, context.get_name());
}
} // namespace op
} // namespace ggml
} // namespace frontend
} // namespace ov

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