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

Merge branch 'master' into dev-refactoring

This commit is contained in:
chraac 2025-10-05 22:58:52 +08:00
parent e1727af06c ca71fb9b36
commit 31cfe411d1
826 changed files with 64792 additions and 31656 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
.clang-format
View File

@ -22,6 +22,13 @@ 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

1
.clang-tidy
View File

@ -17,6 +17,7 @@ Checks: >
clang-analyzer-*,
-clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
performance-*,
-performance-enum-size,
portability-*,
-portability-simd-intrinsics,
misc-*,

6
.devops/intel.Dockerfile
View File

@ -1,8 +1,8 @@
ARG ONEAPI_VERSION=2025.1.1-0-devel-ubuntu24.04
ARG ONEAPI_VERSION=2025.2.2-0-devel-ubuntu24.04
## Build Image
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS build
ARG GGML_SYCL_F16=OFF
RUN apt-get update && \
@ -31,7 +31,7 @@ RUN mkdir -p /app/full \
&& cp requirements.txt /app/full \
&& cp .devops/tools.sh /app/full/tools.sh
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS base
FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
RUN apt-get update \
&& apt-get install -y libgomp1 curl\

2
.devops/musa.Dockerfile
View File

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG MUSA_VERSION=rc4.2.0
ARG MUSA_VERSION=rc4.3.0
# Target the MUSA build image
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64

25
.devops/rocm.Dockerfile
View File

@ -1,8 +1,8 @@
ARG UBUNTU_VERSION=24.04
# This needs to generally match the container host's environment.
ARG ROCM_VERSION=6.4
ARG AMDGPU_VERSION=6.4
ARG ROCM_VERSION=7.0
ARG AMDGPU_VERSION=7.0
# Target the ROCm build image
ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
@ -13,18 +13,14 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
# This is mostly tied to rocBLAS supported archs.
# gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
# gfx906 is deprecated
#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
# gfx803, gfx900, gfx906, gfx1032, gfx1101, gfx1102,not officialy supported
# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102,gfx1200,gfx1201'
#ARG ROCM_DOCKER_ARCH=gfx1100
ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
#ARG ROCM_DOCKER_ARCH='gfx1151'
# Set ROCm architectured
# Set ROCm architectures
ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
# ENV CC=/opt/rocm/llvm/bin/clang
# ENV CXX=/opt/rocm/llvm/bin/clang++
RUN apt-get update \
&& apt-get install -y \
@ -40,7 +36,12 @@ WORKDIR /app
COPY . .
RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
cmake -S . -B build \
-DGGML_HIP=ON \
-DGGML_HIP_ROCWMMA_FATTN=ON \
-DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
-DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
-DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
&& cmake --build build --config Release -j$(nproc)
RUN mkdir -p /app/lib \

123
.devops/s390x.Dockerfile Normal file
View File

@ -0,0 +1,123 @@
ARG GCC_VERSION=15.2.0
ARG UBUNTU_VERSION=24.04
### Build Llama.cpp stage
FROM gcc:${GCC_VERSION} AS build
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
--mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
apt update -y && \
apt upgrade -y && \
apt install -y --no-install-recommends \
git cmake ccache ninja-build \
# WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
libopenblas-dev libcurl4-openssl-dev && \
rm -rf /var/lib/apt/lists/*
WORKDIR /app
COPY . .
RUN --mount=type=cache,target=/root/.ccache \
--mount=type=cache,target=/app/build \
cmake -S . -B build -G Ninja \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_C_COMPILER_LAUNCHER=ccache \
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-DLLAMA_BUILD_TESTS=OFF \
-DGGML_BACKEND_DL=OFF \
-DGGML_NATIVE=OFF \
-DGGML_BLAS=ON \
-DGGML_BLAS_VENDOR=OpenBLAS && \
cmake --build build --config Release -j $(nproc) && \
cmake --install build --prefix /opt/llama.cpp
COPY *.py /opt/llama.cpp/bin
COPY .devops/tools.sh /opt/llama.cpp/bin
COPY gguf-py /opt/llama.cpp/gguf-py
COPY requirements.txt /opt/llama.cpp/gguf-py
COPY requirements /opt/llama.cpp/gguf-py/requirements
### Collect all llama.cpp binaries, libraries and distro libraries
FROM scratch AS collector
# Copy llama.cpp binaries and libraries
COPY --from=build /opt/llama.cpp/bin /llama.cpp/bin
COPY --from=build /opt/llama.cpp/lib /llama.cpp/lib
COPY --from=build /opt/llama.cpp/gguf-py /llama.cpp/gguf-py
### Base image
FROM ubuntu:${UBUNTU_VERSION} AS base
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
--mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
apt update -y && \
apt install -y --no-install-recommends \
# WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
# See: https://github.com/ggml-org/llama.cpp/pull/15915#issuecomment-3317166506
curl libgomp1 libopenblas-dev && \
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 llama.cpp libraries
COPY --from=collector /llama.cpp/lib /usr/lib/s390x-linux-gnu
### Full
FROM base AS full
ENV PATH="/root/.cargo/bin:${PATH}"
WORKDIR /app
RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
--mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
apt update -y && \
apt install -y \
git cmake libjpeg-dev \
python3 python3-pip python3-dev && \
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
RUN curl https://sh.rustup.rs -sSf | bash -s -- -y
COPY --from=collector /llama.cpp/bin /app
COPY --from=collector /llama.cpp/gguf-py /app/gguf-py
RUN pip install --no-cache-dir --break-system-packages \
-r /app/gguf-py/requirements.txt
ENTRYPOINT [ "/app/tools.sh" ]
### CLI Only
FROM base AS light
WORKDIR /llama.cpp/bin
# Copy llama.cpp binaries and libraries
COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin
ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
### Server
FROM base AS server
ENV LLAMA_ARG_HOST=0.0.0.0
WORKDIR /llama.cpp/bin
# Copy llama.cpp binaries and libraries
COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
EXPOSE 8080
ENTRYPOINT [ "/llama.cpp/bin/llama-server" ]

8
.editorconfig
View File

@ -52,3 +52,11 @@ insert_final_newline = unset
[vendor/miniaudio/miniaudio.h]
trim_trailing_whitespace = unset
insert_final_newline = unset
[tools/server/webui/**]
indent_style = unset
indent_size = unset
end_of_line = unset
charset = unset
trim_trailing_whitespace = unset
insert_final_newline = unset

52
.github/workflows/build-amd.yml vendored Normal file
View File

@ -0,0 +1,52 @@
name: CI (AMD)
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: [
'.github/workflows/build-amd.yml',
'**/CMakeLists.txt',
'**/.cmake',
'**/*.h',
'**/*.hpp',
'**/*.c',
'**/*.cpp',
'**/*.cu',
'**/*.cuh',
'**/*.comp'
]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
ggml-ci-x64-amd-vulkan:
runs-on: [self-hosted, Linux, X64, AMD]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
ggml-ci-x64-amd-rocm:
runs-on: [self-hosted, Linux, X64, AMD]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- 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

135
.github/workflows/build-linux-cross.yml vendored
View File

@ -141,97 +141,6 @@ jobs:
# cmake --build build --config Release -j $(nproc)
ubuntu-24-ppc64el-cpu-cross:
runs-on: ubuntu-24.04
steps:
- uses: actions/checkout@v4
- name: Setup PowerPC64le
run: |
sudo dpkg --add-architecture ppc64el
# Add arch-specific repositories for non-amd64 architectures
cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
EOF
sudo apt-get update || true ;# Prevent failure due to missing URLs.
sudo apt-get install -y --no-install-recommends \
build-essential \
gcc-14-powerpc64le-linux-gnu \
g++-14-powerpc64le-linux-gnu
- name: Build
run: |
cmake -B build -DLLAMA_CURL=OFF \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TOOLS=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DCMAKE_SYSTEM_NAME=Linux \
-DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
-DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)
# ubuntu-24-ppc64el-vulkan-cross:
# runs-on: ubuntu-24.04
# steps:
# - uses: actions/checkout@v4
# - name: Setup PowerPC64le
# run: |
# sudo dpkg --add-architecture ppc64el
# # Add arch-specific repositories for non-amd64 architectures
# cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
# deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
# deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
# deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
# deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
# EOF
# sudo apt-get update || true ;# Prevent failure due to missing URLs.
# sudo apt-get install -y --no-install-recommends \
# build-essential \
# glslc \
# gcc-14-powerpc64le-linux-gnu \
# g++-14-powerpc64le-linux-gnu \
# libvulkan-dev:ppc64el
# - name: Build
# run: |
# cmake -B build -DLLAMA_CURL=OFF \
# -DCMAKE_BUILD_TYPE=Release \
# -DGGML_VULKAN=ON \
# -DGGML_OPENMP=OFF \
# -DLLAMA_BUILD_EXAMPLES=ON \
# -DLLAMA_BUILD_TOOLS=ON \
# -DLLAMA_BUILD_TESTS=OFF \
# -DCMAKE_SYSTEM_NAME=Linux \
# -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
# -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
# -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
# -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
# -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
# -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
# -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
# -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
# cmake --build build --config Release -j $(nproc)
debian-13-loongarch64-cpu-cross:
runs-on: ubuntu-24.04
container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
@ -344,3 +253,47 @@ jobs:
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)
ubuntu-24-riscv64-cpu-spacemit-ime-cross:
runs-on: ubuntu-24.04
env:
SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
SPACEMIT_IME_TOOLCHAIN_PATH: "spacemit-toolchain-linux-glibc-x86_64"
steps:
- uses: actions/checkout@v4
- name: Cache Toolchain
uses: actions/cache@v4
id: cache-spacemit-ime-cross-toolchain
with:
path: ./${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
key: ${{ runner.os }}-spacemit-ime-toolchain-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}
- name: Setup Toolchain
if: steps.cache-spacemit-ime-cross-toolchain.outputs.cache-hit != 'true'
run: |
wget --quiet --no-check-certificate https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v${{ env.SPACEMIT_IME_TOOLCHAIN_VERSION }}.tar.xz -O ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}.tar.xz
rm -rf ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
mkdir -p ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
tar xf ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}.tar.xz -C ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }} --strip-components=1
rm -rf ${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}.tar.xz
- name: Build
run: |
export RISCV_ROOT_PATH=${PWD}/${{ env.SPACEMIT_IME_TOOLCHAIN_PATH }}
cmake -B build -DLLAMA_CURL=OFF \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_OPENMP=OFF \
-DLLAMA_BUILD_EXAMPLES=ON \
-DLLAMA_BUILD_TOOLS=ON \
-DLLAMA_BUILD_TESTS=OFF \
-DGGML_CPU_RISCV64_SPACEMIT=ON \
-DGGML_RVV=ON \
-DGGML_RV_ZFH=ON \
-DGGML_RV_ZICBOP=ON \
-DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
-DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake
cmake --build build --config Release -j $(nproc)

62
.github/workflows/build-riscv-native.yml vendored
View File

@ -6,7 +6,7 @@ on:
jobs:
debian-13-riscv64-native: # Bianbu 2.2
runs-on: self-hosted
runs-on: [self-hosted, RISCV64]
steps:
- name: Install prerequisites
@ -58,3 +58,63 @@ jobs:
-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
cmake --build build --config Release -j $(nproc)
# debian-13-riscv64-spacemit-ime-native: # Bianbu 2.2
# runs-on: [self-hosted, RISCV64]
# steps:
# - name: Install prerequisites
# run: |
# sudo apt-get update || true
# sudo apt-get install -y libatomic1
# - uses: actions/checkout@v4
# - name: Setup Riscv
# run: |
# sudo apt-get update || true
# sudo apt-get install -y --no-install-recommends \
# build-essential \
# gcc-14-riscv64-linux-gnu \
# g++-14-riscv64-linux-gnu \
# ccache \
# cmake
# sudo apt-get upgrade binutils -y
# - name: Setup ccache
# run: |
# mkdir -p $HOME/.ccache
# ccache -M 5G -d $HOME/.ccache
# export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
# export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
# echo "$GITHUB_WORKSPACE"
# echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
# echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
# echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
# echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
# - name: Build
# run: |
# cmake -B build \
# -DLLAMA_CURL=OFF \
# -DCMAKE_BUILD_TYPE=Release \
# -DGGML_OPENMP=OFF \
# -DLLAMA_BUILD_EXAMPLES=ON \
# -DLLAMA_BUILD_TOOLS=ON \
# -DLLAMA_BUILD_TESTS=OFF \
# -DCMAKE_SYSTEM_NAME=Linux \
# -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
# -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
# -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
# -DCMAKE_C_COMPILER_LAUNCHER=ccache \
# -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
# -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
# -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
# -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
# -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
# -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH \
# -DGGML_RVV=ON \
# -DGGML_RV_ZFH=ON \
# -DGGML_RV_ZICBOP=ON \
# -DGGML_CPU_RISCV64_SPACEMIT=ON \
# -DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1
# cmake --build build --config Release -j $(nproc)

348
.github/workflows/build.yml vendored
View File

@ -56,7 +56,7 @@ env:
jobs:
macOS-latest-cmake-arm64:
runs-on: macos-14
runs-on: macos-latest
steps:
- name: Clone
@ -97,7 +97,7 @@ jobs:
ctest -L 'main|curl' --verbose --timeout 900
macOS-latest-cmake-x64:
runs-on: macos-13
runs-on: macos-15-intel
steps:
- name: Clone
@ -138,7 +138,7 @@ jobs:
ctest -L main --verbose --timeout 900
macOS-latest-cmake-arm64-webgpu:
runs-on: macos-14
runs-on: macos-latest
steps:
- name: Clone
@ -192,6 +192,10 @@ jobs:
os: ubuntu-22.04
- build: 'arm64'
os: ubuntu-22.04-arm
- build: 's390x'
os: ubuntu-24.04-s390x
- build: 'ppc64le'
os: ubuntu-24.04-ppc64le
runs-on: ${{ matrix.os }}
@ -203,14 +207,31 @@ jobs:
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ubuntu-cpu-cmake
key: ubuntu-cpu-cmake-${{ matrix.build }}
evict-old-files: 1d
- name: Dependencies
id: depends
- name: Build Dependencies
id: build_depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
sudo apt-get install -y --no-install-recommends \
python3 python3-pip python3-dev \
libjpeg-dev build-essential libcurl4-openssl-dev \
git-lfs
- name: Python Dependencies
id: python_depends
run: |
python3 -m pip install --upgrade pip
pip3 install ./gguf-py
- name: Swap Endianness
id: endianness
if: ${{ matrix.build == 's390x' }}
run: |
for f in models/*.gguf; do
echo YES | python3 gguf-py/gguf/scripts/gguf_convert_endian.py $f big
done
- name: Build
id: cmake_build
@ -228,6 +249,7 @@ jobs:
- name: Test llama2c conversion
id: llama2c_test
if: ${{ matrix.build != 's390x' }}
run: |
cd build
echo "Fetch tokenizer"
@ -237,6 +259,15 @@ jobs:
./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
- name: Test llama2c (s390x)
id: llama2c_test_s390x
if: ${{ matrix.build == 's390x' }}
run: |
cd build
echo "Fetch llama2c big-endian model"
wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
./bin/llama-cli -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
ubuntu-latest-cmake-sanitizer:
runs-on: ubuntu-latest
@ -331,11 +362,11 @@ jobs:
id: checkout
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ubuntu-latest-cmake-rpc
evict-old-files: 1d
# - name: ccache
# uses: ggml-org/ccache-action@v1.2.16
# with:
# key: ubuntu-latest-cmake-rpc
# evict-old-files: 1d
- name: Dependencies
id: depends
@ -356,8 +387,8 @@ jobs:
cd build
ctest -L main --verbose
ubuntu-22-cmake-vulkan:
runs-on: ubuntu-22.04
ubuntu-24-cmake-vulkan:
runs-on: ubuntu-24.04
steps:
- name: Clone
@ -367,20 +398,40 @@ jobs:
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ubuntu-22-cmake-vulkan
key: ubuntu-24-cmake-vulkan
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
sudo add-apt-repository -y ppa:kisak/kisak-mesa
sudo apt-get update -y
sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
- name: Get latest Vulkan SDK version
id: vulkan_sdk_version
run: |
echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
- name: Cache Vulkan SDK
id: cache_vulkan_sdk
uses: actions/cache@v4
with:
path: ./vulkan_sdk
key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
- name: Install Vulkan SDK
if: steps.cache_vulkan_sdk.outputs.cache-hit != 'true'
id: vulkan_sdk_install
run: |
mkdir -p vulkan_sdk
cd vulkan_sdk
curl --no-progress-meter https://sdk.lunarg.com/sdk/download/latest/linux/vulkan_sdk.tar.xz | tar -Jx --strip-components=1
- 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)
@ -390,6 +441,7 @@ jobs:
run: |
cd build
export GGML_VK_VISIBLE_DEVICES=0
export GGML_VK_DISABLE_F16=1
# This is using llvmpipe and runs slower than other backends
ctest -L main --verbose --timeout 4200
@ -456,7 +508,7 @@ jobs:
id: depends
run: |
sudo apt-get update
sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev rocwmma-dev
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
@ -475,7 +527,7 @@ jobs:
ubuntu-22-cmake-musa:
runs-on: ubuntu-22.04
container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
steps:
- name: Clone
@ -711,6 +763,7 @@ jobs:
macOS-latest-swift:
runs-on: macos-latest
needs: ios-xcode-build
strategy:
matrix:
@ -727,6 +780,12 @@ jobs:
key: macOS-latest-swift
evict-old-files: 1d
- name: Download xcframework artifact
uses: actions/download-artifact@v4
with:
name: llama-xcframework
path: build-apple/llama.xcframework/
- name: Dependencies
id: depends
continue-on-error: true
@ -748,11 +807,6 @@ jobs:
-DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: xcodebuild for swift package
id: xcodebuild
run: |
./build-xcframework.sh
windows-msys2:
runs-on: windows-2025
@ -1026,7 +1080,7 @@ jobs:
shell: bash
env:
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
steps:
@ -1064,10 +1118,12 @@ jobs:
id: checkout
uses: actions/checkout@v4
- name: Clone rocWMMA repository
id: clone_rocwmma
- name: Grab rocWMMA package
id: grab_rocwmma
run: |
git clone https://github.com/rocm/rocwmma --branch rocm-${{ env.ROCM_VERSION }} --depth 1
curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/${{ env.ROCM_VERSION }}/pool/main/r/rocwmma-dev/rocwmma-dev_1.7.0.60402-120~24.04_amd64.deb"
7z x rocwmma.deb
7z x data.tar
- name: Cache ROCm Installation
id: cache-rocm
@ -1128,8 +1184,9 @@ jobs:
cmake -G "Unix Makefiles" -B build -S . `
-DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" `
-DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
-DCMAKE_BUILD_TYPE=Release `
-DROCM_DIR="${env:HIP_PATH}" `
-DGGML_HIP=ON `
-DGGML_HIP_ROCWMMA_FATTN=ON `
-DGGML_RPC=ON `
@ -1170,8 +1227,17 @@ jobs:
run: |
./build-xcframework.sh
- name: Upload xcframework artifact
uses: actions/upload-artifact@v4
with:
name: llama-xcframework
path: build-apple/llama.xcframework/
retention-days: 1
- name: Build Xcode project
run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
run: |
xcodebuild -downloadPlatform iOS
xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
android-build:
runs-on: ubuntu-latest
@ -1180,11 +1246,12 @@ jobs:
- name: Clone
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: android-build
evict-old-files: 1d
# Disabled due to size (400MB) and always 0 cache hits
# - name: ccache
# uses: ggml-org/ccache-action@v1.2.16
# with:
# key: android-build
# evict-old-files: 1d
- name: Set up JDK
uses: actions/setup-java@v3
@ -1236,3 +1303,212 @@ jobs:
-DGGML_CANN=on \
-DSOC_TYPE=${{ matrix.device }}
cmake --build build -j $(nproc)
# TODO: simplify the following workflows using a matrix
# TODO: run lighter CI on PRs and the full CI only on master (if needed)
ggml-ci-x64-cpu-low-perf:
runs-on: ubuntu-22.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-x64-cpu-low-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-arm64-cpu-low-perf:
runs-on: ubuntu-22.04-arm
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-arm64-cpu-low-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-x64-cpu-high-perf:
runs-on: ubuntu-22.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-x64-cpu-high-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-arm64-cpu-high-perf:
runs-on: ubuntu-22.04-arm
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-arm64-cpu-high-perf
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-arm64-cpu-high-perf-sve:
runs-on: ubuntu-22.04-arm
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ggml-ci-arm64-cpu-high-perf-sve
evict-old-files: 1d
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libcurl4-openssl-dev
- name: Test
id: ggml-ci
run: |
LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
ggml-ci-x64-nvidia-cuda:
runs-on: [self-hosted, Linux, X64, NVIDIA]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Test
id: ggml-ci
run: |
nvidia-smi
GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
ggml-ci-x64-nvidia-vulkan-cm:
runs-on: [self-hosted, Linux, X64, NVIDIA]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- 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-x64-nvidia-vulkan-cm2:
runs-on: [self-hosted, Linux, X64, NVIDIA, COOPMAT2]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
ggml-ci-x64-cpu-amx:
runs-on: [self-hosted, Linux, X64, CPU, AMX]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Test
id: ggml-ci
run: |
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@v4
- name: Test
id: ggml-ci
run: |
GG_BUILD_METAL=1 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@v4
- name: Test
id: ggml-ci
run: |
vulkaninfo --summary
GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp

88
.github/workflows/docker.yml vendored
View File

@ -28,7 +28,7 @@ jobs:
push_to_registry:
name: Push Docker image to Docker Hub
runs-on: ubuntu-22.04
runs-on: ${{ matrix.config.runs_on }}
env:
COMMIT_SHA: ${{ github.sha }}
strategy:
@ -39,11 +39,12 @@ jobs:
# Note: the arm64 images are failing, which prevents the amd64 images from being built
# https://github.com/ggml-org/llama.cpp/issues/11888
#- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
- { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
- { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
- { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
- { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true, runs_on: "ubuntu-22.04" }
- { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true, runs_on: "ubuntu-22.04" }
- { 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" }
# Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
#- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
steps:
@ -53,6 +54,7 @@ jobs:
fetch-depth: 0 # preserve git history, so we can determine the build number
- name: Set up QEMU
if: ${{ matrix.config.tag != 's390x' }}
uses: docker/setup-qemu-action@v3
with:
image: tonistiigi/binfmt:qemu-v7.0.0-28
@ -67,22 +69,19 @@ jobs:
username: ${{ github.repository_owner }}
password: ${{ secrets.GITHUB_TOKEN }}
- name: Determine tag name
- name: Determine source tag name
id: srctag
uses: ./.github/actions/get-tag-name
env:
BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
- name: Determine image tag name
id: tag
shell: bash
run: |
BUILD_NUMBER="$(git rev-list --count HEAD)"
SHORT_HASH="$(git rev-parse --short=7 HEAD)"
REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}" # to lower case
REPO_NAME="${{ github.event.repository.name }}"
# determine tag name postfix (build number, commit hash)
if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
TAG_POSTFIX="-b${BUILD_NUMBER}"
else
SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
TAG_POSTFIX="-${SAFE_NAME}-${SHORT_HASH}"
fi
# list all tags possible
if [[ "${{ matrix.config.tag }}" == "cpu" ]]; then
TYPE=""
@ -90,17 +89,19 @@ jobs:
TYPE="-${{ matrix.config.tag }}"
fi
PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
FULLTAGS="${PREFIX}full${TYPE},${PREFIX}full${TYPE}${TAG_POSTFIX}"
LIGHTTAGS="${PREFIX}light${TYPE},${PREFIX}light${TYPE}${TAG_POSTFIX}"
SERVERTAGS="${PREFIX}server${TYPE},${PREFIX}server${TYPE}${TAG_POSTFIX}"
CACHETAGS="${PREFIX}buildcache${TYPE}"
FULLTAGS="${PREFIX}full${TYPE},${PREFIX}full${TYPE}-${{ steps.srctag.outputs.name }}"
LIGHTTAGS="${PREFIX}light${TYPE},${PREFIX}light${TYPE}-${{ steps.srctag.outputs.name }}"
SERVERTAGS="${PREFIX}server${TYPE},${PREFIX}server${TYPE}-${{ steps.srctag.outputs.name }}"
echo "cache_output_tags=$CACHETAGS" >> $GITHUB_OUTPUT
echo "full_output_tags=$FULLTAGS" >> $GITHUB_OUTPUT
echo "light_output_tags=$LIGHTTAGS" >> $GITHUB_OUTPUT
echo "server_output_tags=$SERVERTAGS" >> $GITHUB_OUTPUT
echo "cache_output_tags=$CACHETAGS" # print out for debugging
echo "full_output_tags=$FULLTAGS" # print out for debugging
echo "light_output_tags=$LIGHTTAGS" # print out for debugging
echo "server_output_tags=$SERVERTAGS" # print out for debugging
env:
GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
- name: Free Disk Space (Ubuntu)
@ -133,11 +134,14 @@ jobs:
target: full
provenance: false
# using github experimental cache
cache-from: type=gha
cache-to: type=gha,mode=max
#cache-from: type=gha
#cache-to: type=gha,mode=max
# return to this if the experimental github cache is having issues
#cache-to: type=local,dest=/tmp/.buildx-cache
#cache-from: type=local,src=/tmp/.buildx-cache
# using registry cache (no storage limit)
cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
- name: Build and push Light Docker image (tagged + versioned)
if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
@ -152,11 +156,14 @@ jobs:
target: light
provenance: false
# using github experimental cache
cache-from: type=gha
cache-to: type=gha,mode=max
#cache-from: type=gha
#cache-to: type=gha,mode=max
# return to this if the experimental github cache is having issues
#cache-to: type=local,dest=/tmp/.buildx-cache
#cache-from: type=local,src=/tmp/.buildx-cache
# using registry cache (no storage limit)
cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
- name: Build and push Server Docker image (tagged + versioned)
if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
@ -171,8 +178,37 @@ jobs:
target: server
provenance: false
# using github experimental cache
cache-from: type=gha
cache-to: type=gha,mode=max
#cache-from: type=gha
#cache-to: type=gha,mode=max
# return to this if the experimental github cache is having issues
#cache-to: type=local,dest=/tmp/.buildx-cache
#cache-from: type=local,src=/tmp/.buildx-cache
# using registry cache (no storage limit)
cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
create_tag:
name: Create and push git tag
runs-on: ubuntu-22.04
permissions:
contents: write
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Determine source tag name
id: srctag
uses: ./.github/actions/get-tag-name
env:
BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
- name: Create and push git tag
env:
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
run: |
git tag ${{ steps.srctag.outputs.name }} || exit 0
git push origin ${{ steps.srctag.outputs.name }} || exit 0

26
.github/workflows/release.yml vendored
View File

@ -75,7 +75,7 @@ jobs:
name: llama-bin-macos-arm64.zip
macOS-x64:
runs-on: macos-13
runs-on: macos-15-intel
steps:
- name: Clone
@ -150,7 +150,7 @@ jobs:
- name: ccache
uses: ggml-org/ccache-action@v1.2.16
with:
key: ubuntu-cpu-cmake
key: ubuntu-cpu-cmake-${{ matrix.build }}
evict-old-files: 1d
- name: Dependencies
@ -462,7 +462,7 @@ jobs:
shell: bash
env:
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7cd9bba0-7aab-4e30-b3ae-2221006a4a05/intel-oneapi-base-toolkit-2025.1.1.34_offline.exe
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
@ -505,6 +505,7 @@ jobs:
cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
@ -513,10 +514,15 @@ jobs:
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/tcm.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/tcm/latest/bin/libhwloc-15.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/umf/latest/bin/umf.dll" ./build/bin
echo "cp oneAPI running time dll files to ./build/bin done"
7z a llama-bin-win-sycl-x64.zip ./build/bin/*
@ -530,25 +536,25 @@ jobs:
runs-on: windows-2022
env:
# The ROCm version must correspond to the version used in the HIP SDK.
ROCM_VERSION: "6.4.2"
HIPSDK_INSTALLER_VERSION: "25.Q3"
strategy:
matrix:
include:
- name: "radeon"
gpu_targets: "gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Clone rocWMMA repository
id: clone_rocwmma
- name: Grab rocWMMA package
id: grab_rocwmma
run: |
git clone https://github.com/rocm/rocwmma --branch rocm-${{ env.ROCM_VERSION }} --depth 1
curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.0.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.0.0.70001-42~24.04_amd64.deb"
7z x rocwmma.deb
7z x data.tar
- name: Cache ROCm Installation
id: cache-rocm
@ -603,7 +609,7 @@ jobs:
cmake -G "Unix Makefiles" -B build -S . `
-DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
-DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
-DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.0.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
-DCMAKE_BUILD_TYPE=Release `
-DGGML_BACKEND_DL=ON `
-DGGML_NATIVE=OFF `

229
.github/workflows/server.yml vendored
View File

@ -76,51 +76,206 @@ jobs:
run: |
pip install -r tools/server/tests/requirements.txt
# Setup nodejs (to be used for verifying bundled index.html)
- uses: actions/setup-node@v4
webui-setup:
name: WebUI Setup
runs-on: ubuntu-latest
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
node-version: '22.11.0'
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: WebUI - Install dependencies
id: webui_lint
- name: Setup Node.js
uses: actions/setup-node@v4
with:
node-version: "22"
cache: "npm"
cache-dependency-path: "tools/server/webui/package-lock.json"
- name: Cache node_modules
uses: actions/cache@v4
id: cache-node-modules
with:
path: tools/server/webui/node_modules
key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
restore-keys: |
${{ runner.os }}-node-modules-
- name: Install dependencies
if: steps.cache-node-modules.outputs.cache-hit != 'true'
run: npm ci
working-directory: tools/server/webui
webui-check:
needs: webui-setup
name: WebUI Check
runs-on: ubuntu-latest
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: Setup Node.js
uses: actions/setup-node@v4
with:
node-version: "22"
- name: Restore node_modules cache
uses: actions/cache@v4
with:
path: tools/server/webui/node_modules
key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
restore-keys: |
${{ runner.os }}-node-modules-
- name: Run type checking
run: npm run check
working-directory: tools/server/webui
- name: Run linting
run: npm run lint
working-directory: tools/server/webui
webui-build:
needs: webui-check
name: WebUI Build
runs-on: ubuntu-latest
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: Setup Node.js
uses: actions/setup-node@v4
with:
node-version: "22"
- name: Restore node_modules cache
uses: actions/cache@v4
with:
path: tools/server/webui/node_modules
key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
restore-keys: |
${{ runner.os }}-node-modules-
- name: Build application
run: npm run build
working-directory: tools/server/webui
webui-tests:
needs: webui-build
name: Run WebUI tests
permissions:
contents: read
runs-on: ubuntu-latest
steps:
- name: Checkout code
uses: actions/checkout@v4
- name: Setup Node.js
uses: actions/setup-node@v4
with:
node-version: "22"
- name: Restore node_modules cache
uses: actions/cache@v4
with:
path: tools/server/webui/node_modules
key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
restore-keys: |
${{ runner.os }}-node-modules-
- name: Install Playwright browsers
run: npx playwright install --with-deps
working-directory: tools/server/webui
- name: Build Storybook
run: npm run build-storybook
working-directory: tools/server/webui
- name: Run Client tests
run: npm run test:client
working-directory: tools/server/webui
- name: Run Server tests
run: npm run test:server
working-directory: tools/server/webui
- name: Run UI tests
run: npm run test:ui
working-directory: tools/server/webui
- name: Run E2E tests
run: npm run test:e2e
working-directory: tools/server/webui
server-build:
needs: [webui-tests]
runs-on: ubuntu-latest
strategy:
matrix:
sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
build_type: [RelWithDebInfo]
include:
- build_type: Release
sanitizer: ""
fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
steps:
- name: Dependencies
id: depends
run: |
cd tools/server/webui
npm ci
sudo apt-get update
sudo apt-get -y install \
build-essential \
xxd \
git \
cmake \
curl \
wget \
language-pack-en \
libcurl4-openssl-dev
- name: WebUI - Check code format
id: webui_format
- name: Clone
id: checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: Python setup
id: setup_python
uses: actions/setup-python@v5
with:
python-version: '3.11'
- name: Tests dependencies
id: test_dependencies
run: |
git config --global --add safe.directory $(realpath .)
cd tools/server/webui
git status
pip install -r tools/server/tests/requirements.txt
npm run format
git status
modified_files="$(git status -s)"
echo "Modified files: ${modified_files}"
if [ -n "${modified_files}" ]; then
echo "Files do not follow coding style. To fix: npm run format"
echo "${modified_files}"
exit 1
fi
- name: Setup Node.js for WebUI
uses: actions/setup-node@v4
with:
node-version: "22"
cache: "npm"
cache-dependency-path: "tools/server/webui/package-lock.json"
- name: Verify bundled index.html
id: verify_server_index_html
run: |
git config --global --add safe.directory $(realpath .)
cd tools/server/webui
git status
- name: Install WebUI dependencies
run: npm ci
working-directory: tools/server/webui
npm run build
git status
modified_files="$(git status -s)"
echo "Modified files: ${modified_files}"
if [ -n "${modified_files}" ]; then
echo "Repository is dirty or server/webui is not built as expected"
echo "Hint: You may need to follow Web UI build guide in server/README.md"
echo "${modified_files}"
exit 1
fi
- name: Build WebUI
run: npm run build
working-directory: tools/server/webui
- name: Build (no OpenMP)
id: cmake_build_no_openmp

4
.gitignore vendored
View File

@ -148,3 +148,7 @@ poetry.toml
/run-vim.sh
/run-chat.sh
.ccache/
# IDE
*.code-workspace
.windsurf/

8
CMakeLists.txt
View File

@ -58,6 +58,12 @@ if (MSVC)
add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
endif()
if (CMAKE_SYSTEM_NAME STREQUAL "iOS")
set(LLAMA_TOOLS_INSTALL_DEFAULT OFF)
else()
set(LLAMA_TOOLS_INSTALL_DEFAULT ${LLAMA_STANDALONE})
endif()
#
# option list
#
@ -82,9 +88,11 @@ option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_TOOLS "llama: build tools" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE})
option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
# 3rd party libs
option(LLAMA_CURL "llama: use libcurl to download model from an URL" ON)
option(LLAMA_OPENSSL "llama: use openssl to support HTTPS" OFF)
option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
# Required for relocatable CMake package

123
CODEOWNERS
View File

@ -1,12 +1,115 @@
# collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
# multiplie collaborators per item can be specified
/ci/ @ggerganov
/.devops/*.Dockerfile @ngxson
/tools/server/ @ngxson
/ggml/src/ggml-cuda/fattn* @JohannesGaessler
/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
/ggml/src/ggml-opt.cpp @JohannesGaessler
/ggml/src/gguf.cpp @JohannesGaessler
/ggml/src/ggml-vulkan/ @0cc4m
/ggml/src/ggml-zdnn/ @taronaeo
/.devops/*.Dockerfile @ngxson
/.github/actions/ @slaren
/.github/workflows/ @CISC
/.github/workflows/release.yml @slaren
/.github/workflows/winget.yml @slaren
/ci/ @ggerganov
/cmake/ @ggerganov
/common/CMakeLists.txt @ggerganov
/common/arg.* @ggerganov @ericcurtin
/common/base64.hpp.* @ggerganov
/common/build-info.* @ggerganov
/common/common.* @ggerganov
/common/console.* @ggerganov
/common/http.* @angt
/common/llguidance.* @ggerganov
/common/log.* @ggerganov
/common/sampling.* @ggerganov
/common/speculative.* @ggerganov
/convert_*.py @CISC
/examples/batched.swift/ @ggerganov
/examples/batched/ @ggerganov
/examples/convert-llama2c-to-ggml/ @ggerganov
/examples/deprecation-warning/ @ggerganov
/examples/diffusion/ @am17an
/examples/embedding/ @ggerganov
/examples/eval-callback/ @ggerganov
/examples/export-docs/ @ggerganov
/examples/gen-docs/ @ggerganov
/examples/gguf/ @ggerganov
/examples/llama.android/ @ggerganov
/examples/llama.swiftui/ @ggerganov
/examples/llama.vim @ggerganov
/examples/lookahead/ @ggerganov
/examples/lookup/ @JohannesGaessler
/examples/model-conversion/ @danbev
/examples/parallel/ @ggerganov
/examples/passkey/ @ggerganov
/examples/retrieval/ @ggerganov
/examples/save-load-state/ @ggerganov
/examples/simple-chat/ @slaren
/examples/simple/ @slaren
/examples/speculative-simple/ @ggerganov
/examples/speculative/ @ggerganov
/ggml/cmake/ @ggerganov
/ggml/include/ @ggerganov @slaren
/ggml/src/ggml-alloc.c @slaren
/ggml/src/ggml-backend* @slaren
/ggml/src/ggml-blas/ @slaren
/ggml/src/ggml-common.h @ggerganov @slaren
/ggml/src/ggml-cpu/ @ggerganov @slaren
/ggml/src/ggml-cpu/spacemit/ @alex-spacemit
/ggml/src/ggml-cuda/common.cuh @slaren
/ggml/src/ggml-cuda/fattn* @JohannesGaessler
/ggml/src/ggml-cuda/ggml-cuda.cu @slaren
/ggml/src/ggml-cuda/mmf.* @JohannesGaessler
/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
/ggml/src/ggml-cuda/mmvf.* @JohannesGaessler
/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
/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 @slaren
/ggml/src/ggml-metal/ @ggerganov
/ggml/src/ggml-opencl/ @lhez @max-krasnyansky
/ggml/src/ggml-opt.cpp @JohannesGaessler
/ggml/src/ggml-quants.* @ggerganov
/ggml/src/ggml-rpc/ @rgerganov
/ggml/src/ggml-threading.* @ggerganov @slaren
/ggml/src/ggml-vulkan/ @0cc4m
/ggml/src/ggml-zdnn/ @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
/ggml/src/ggml.c @ggerganov @slaren
/ggml/src/ggml.cpp @ggerganov @slaren
/ggml/src/gguf.cpp @JohannesGaessler @Green-Sky
/gguf-py/ @CISC
/media/ @ggerganov
/scripts/gen* @ggerganov
/scripts/get* @ggerganov
/scripts/sync* @ggerganov
/src/ @ggerganov
/src/llama-adapter.* @CISC
/src/llama-arch.* @CISC
/src/llama-chat.* @ngxson
/src/llama-graph.* @CISC
/src/llama-model-loader.* @slaren
/src/llama-model.* @CISC
/src/llama-vocab.* @CISC
/tests/ @ggerganov
/tests/test-backend-ops.cpp @slaren
/tests/test-thread-safety.cpp @slaren
/tools/batched-bench/ @ggerganov
/tools/llama-bench/ @slaren
/tools/main/ @ggerganov
/tools/mtmd/ @ngxson
/tools/perplexity/ @ggerganov
/tools/quantize/ @ggerganov
/tools/rpc/ @rgerganov
/tools/run/ @ericcurtin
/tools/server/* @ngxson @ggerganov @ericcurtin # no subdir
/tools/server/webui/ @allozaur
/tools/tokenize/ @ggerganov
/tools/tts/ @ggerganov
/vendor/ @ggerganov
/.clang-format @slaren
/.clang-tidy @slaren
/AUTHORS @ggerganov
/CMakeLists.txt @ggerganov
/CONTRIBUTING.md @ggerganov
/LICENSE @ggerganov
/README.md @ggerganov
/SECURITY.md @ggerganov
/build-xcframework.sh @danbev
requirements*.txt @CISC

34
CONTRIBUTING.md
View File

@ -1,4 +1,12 @@
# Pull requests (for contributors)
# Contributors
The project differentiates between 3 levels of contributors:
- Contributors: people who have contributed before (no special privileges)
- Collaborators (Triage): people with significant contributions, who may be responsible for some parts of the code, and are expected to maintain and review contributions for the code they own
- Maintainers: responsible for reviewing and merging PRs, after approval from the code owners
# Pull requests (for contributors & collaborators)
- llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier
- Test your changes:
@ -9,15 +17,16 @@
- Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
- Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR
- Consider adding yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs
# Pull requests (for collaborators)
# Pull requests (for maintainers)
- Squash-merge PRs
- Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
- Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
- Consider adding yourself to [CODEOWNERS](CODEOWNERS)
- Let authors, who are also collaborators, merge their own PRs
- When merging a PR by a contributor, make sure you have a good understanding of the changes
- Let other maintainers merge their own PRs
- When merging a PR, make sure you have a good understanding of the changes
- Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
# Coding guidelines
@ -117,6 +126,21 @@
#endif // FOO
```
# Code maintenance
- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file reponsible for:
- Reviewing and merging related PRs
- Fixing related bugs
- Providing developer guidance/support
- When adding or modifying a large piece of code:
- If you are a collaborator, make sure to add yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs
- If you are a contributor, find an existing collaborator who is willing to review and maintain your code long-term
- Provide the necessary CI workflow (and hardware) to test your changes (see [ci/README.md](https://github.com/ggml-org/llama.cpp/tree/master/ci))
- New code should follow the guidelines (coding, naming, etc.) outlined in this document. Exceptions are allowed in isolated, backend-specific parts of the code that do not interface directly with the `ggml` interfaces.
_(NOTE: for legacy reasons, existing code is not required to follow this guideline)_
# Documentation
- Documentation is a community effort

4
README.md
View File

@ -178,6 +178,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
- Java: [QuasarByte/llama-cpp-jna](https://github.com/QuasarByte/llama-cpp-jna)
- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
- Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
@ -274,6 +275,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
| [Vulkan](docs/build.md#vulkan) | GPU |
| [CANN](docs/build.md#cann) | Ascend NPU |
| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
@ -520,8 +522,8 @@ To learn more about model quantization, [read this documentation](tools/quantize
## Contributing
- Contributors can open PRs
- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
- Collaborators will be invited based on contributions
- Maintainers can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
- Any help with managing issues, PRs and projects is very appreciated!
- See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information

1
build-xcframework.sh
View File

@ -422,6 +422,7 @@ echo "Building for iOS devices..."
cmake -B build-ios-device -G Xcode \
"${COMMON_CMAKE_ARGS[@]}" \
-DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
-DCMAKE_SYSTEM_NAME=iOS \
-DCMAKE_OSX_SYSROOT=iphoneos \
-DCMAKE_OSX_ARCHITECTURES="arm64" \
-DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \

35
ci/README-MUSA.md Normal file
View File

@ -0,0 +1,35 @@
## Running MUSA CI in a Docker Container
Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
### 1. Create a local directory to store cached models, configuration files and venv:
```bash
mkdir -p $HOME/llama.cpp/ci-cache
```
### 2. Create a local directory to store CI run results:
```bash
mkdir -p $HOME/llama.cpp/ci-results
```
### 3. Start a Docker container and run the CI:
```bash
docker run --privileged -it \
-v $HOME/llama.cpp/ci-cache:/ci-cache \
-v $HOME/llama.cpp/ci-results:/ci-results \
-v $PWD:/ws -w /ws \
mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
```
Inside the container, execute the following commands:
```bash
apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
git config --global --add safe.directory /ws
GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
```
This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.

57
ci/README.md
View File

@ -1,18 +1,10 @@
# CI
In addition to [Github Actions](https://github.com/ggml-org/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
This CI implements heavy-duty workflows that run on self-hosted runners. Typically the purpose of these workflows is to
cover hardware configurations that are not available from Github-hosted runners and/or require more computational
resource than normally available.
https://github.com/ggml-org/ci
It monitors the `master` branch for new commits and runs the
[ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
to cover various hardware architectures, including GPU and Apple Silicon instances.
Collaborators can optionally trigger the CI run by adding the `ggml-ci` keyword to their commit message.
Only the branches of this repo are monitored for this keyword.
It is a good practice, before publishing changes to execute the full CI locally on your machine:
It is a good practice, before publishing changes to execute the full CI locally on your machine. For example:
```bash
mkdir tmp
@ -29,40 +21,13 @@ GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
# with MUSA support
GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
# etc.
```
## Running MUSA CI in a Docker Container
# Adding self-hosted runners
Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
### 1. Create a local directory to store cached models, configuration files and venv:
```bash
mkdir -p $HOME/llama.cpp/ci-cache
```
### 2. Create a local directory to store CI run results:
```bash
mkdir -p $HOME/llama.cpp/ci-results
```
### 3. Start a Docker container and run the CI:
```bash
docker run --privileged -it \
-v $HOME/llama.cpp/ci-cache:/ci-cache \
-v $HOME/llama.cpp/ci-results:/ci-results \
-v $PWD:/ws -w /ws \
mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
```
Inside the container, execute the following commands:
```bash
apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
git config --global --add safe.directory /ws
GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
```
This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.
- Add a self-hosted `ggml-ci` workflow to [[.github/workflows/build.yml]] with an appropriate label
- Request a runner token from `ggml-org` (for example, via a comment in the PR or email)
- Set-up a machine using the received token ([docs](https://docs.github.com/en/actions/how-tos/manage-runners/self-hosted-runners/add-runners))
- Optionally update [ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) to build and run on the target platform by gating the implementation with a `GG_BUILD_...` env

471
ci/run.sh
View File

@ -34,9 +34,9 @@ mkdir -p "$2"
OUT=$(realpath "$1")
MNT=$(realpath "$2")
rm -f "$OUT/*.log"
rm -f "$OUT/*.exit"
rm -f "$OUT/*.md"
rm -f $OUT/*.log
rm -f $OUT/*.exit
rm -f $OUT/*.md
sd=`dirname $0`
cd $sd/../
@ -45,7 +45,7 @@ SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
fi
if [ ! -z ${GG_BUILD_CUDA} ]; then
@ -65,6 +65,16 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
fi
fi
if [ ! -z ${GG_BUILD_ROCM} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_HIP=ON"
if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
exit 1
fi
CMAKE_EXTRA="${CMAKE_EXTRA} -DAMDGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
fi
if [ ! -z ${GG_BUILD_SYCL} ]; then
if [ -z ${ONEAPI_ROOT} ]; then
echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
@ -82,6 +92,12 @@ fi
if [ ! -z ${GG_BUILD_VULKAN} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
# if on Mac, disable METAL
if [[ "$OSTYPE" == "darwin"* ]]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
fi
fi
if [ ! -z ${GG_BUILD_WEBGPU} ]; then
@ -93,6 +109,12 @@ if [ ! -z ${GG_BUILD_MUSA} ]; then
MUSA_ARCH=${MUSA_ARCH:-21}
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
fi
if [ ! -z ${GG_BUILD_NO_SVE} ]; then
# arm 9 and newer enables sve by default, adjust these flags depending on the cpu used
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm"
fi
## helpers
# download a file if it does not exist or if it is outdated
@ -150,7 +172,7 @@ 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 ) 2>&1 | tee -a $OUT/${ci}-ctest.log
(time ctest --output-on-failure -L main -E "test-opt|test-backend-ops" ) 2>&1 | tee -a $OUT/${ci}-ctest.log
set +e
}
@ -200,33 +222,9 @@ function gg_sum_ctest_release {
gg_printf '```\n'
}
# test_scripts_debug
# test_scripts
function gg_run_test_scripts_debug {
cd ${SRC}
set -e
(cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
(cd ./tools/quantize && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
set +e
}
function gg_sum_test_scripts_debug {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Runs test scripts in debug mode\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
gg_printf '```\n'
gg_printf '\n'
}
# test_scripts_release
function gg_run_test_scripts_release {
function gg_run_test_scripts {
cd ${SRC}
set -e
@ -237,10 +235,10 @@ function gg_run_test_scripts_release {
set +e
}
function gg_sum_test_scripts_release {
function gg_sum_test_scripts {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Runs test scripts in release mode\n'
gg_printf 'Runs test scripts\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '```\n'
gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
@ -249,15 +247,9 @@ function gg_sum_test_scripts_release {
}
function gg_get_model {
local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-f16.gguf"
if [[ -s $gguf_0 ]]; then
echo -n "$gguf_0"
elif [[ -s $gguf_1 ]]; then
echo -n "$gguf_1"
elif [[ -s $gguf_2 ]]; then
echo -n "$gguf_2"
else
echo >&2 "No model found. Can't run gg_run_ctest_with_model."
exit 1
@ -316,24 +308,22 @@ function gg_sum_ctest_with_model_release {
gg_printf '```\n'
}
# open_llama_7b_v2
# qwen3_0_6b
function gg_run_open_llama_7b_v2 {
function gg_run_qwen3_0_6b {
cd ${SRC}
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/config.json
gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/tokenizer.json
gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/tokenizer_config.json
#gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/special_tokens_map.json
gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/main/model.safetensors
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
path_models="../models-mnt/open-llama/7B-v2"
path_models="../models-mnt/qwen3/0.6B"
path_wiki="../models-mnt/wikitext/wikitext-2-raw"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
@ -343,9 +333,11 @@ function gg_run_open_llama_7b_v2 {
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf --outtype f16
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-bf16.gguf --outtype bf16
model_f16="${path_models}/ggml-model-f16.gguf"
model_bf16="${path_models}/ggml-model-bf16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
model_q4_0="${path_models}/ggml-model-q4_0.gguf"
model_q4_1="${path_models}/ggml-model-q4_1.gguf"
@ -359,179 +351,51 @@ function gg_run_open_llama_7b_v2 {
wiki_test="${path_wiki}/wiki.test.raw"
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
./bin/llama-quantize ${model_bf16} ${model_q8_0} q8_0 $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q4_0} q4_0 $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q4_1} q4_1 $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q5_0} q5_0 $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q5_1} q5_1 $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q2_k} q2_k $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q3_k} q3_k $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q4_k} q4_k $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
(time ./bin/llama-cli -no-cnv --model ${model_f16} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-cli -no-cnv --model ${model_f16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
(time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
if [ -z ${GG_BUILD_NO_BF16} ]; then
(time ./bin/llama-perplexity --model ${model_bf16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
fi
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
printf ' - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
return 20
fi
printf ' - %s @ %s OK\n' "$qnt" "$ppl"
return 0
}
check_ppl "f16" "$(cat $OUT/${ci}-tg-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
set +e
}
function gg_sum_open_llama_7b_v2 {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'OpenLLaMA 7B-v2:\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
}
# pythia_1.4b
function gg_run_pythia_1_4b {
cd ${SRC}
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
path_models="../models-mnt/pythia/1.4B"
path_wiki="../models-mnt/wikitext/wikitext-2-raw"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
model_q4_0="${path_models}/ggml-model-q4_0.gguf"
model_q4_1="${path_models}/ggml-model-q4_1.gguf"
model_q5_0="${path_models}/ggml-model-q5_0.gguf"
model_q5_1="${path_models}/ggml-model-q5_1.gguf"
model_q2_k="${path_models}/ggml-model-q2_k.gguf"
model_q3_k="${path_models}/ggml-model-q3_k.gguf"
model_q4_k="${path_models}/ggml-model-q4_k.gguf"
model_q5_k="${path_models}/ggml-model-q5_k.gguf"
model_q6_k="${path_models}/ggml-model-q6_k.gguf"
wiki_test_60="${path_wiki}/wiki.test-60.raw"
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli -no-cnv --model ${model_f16} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -547,6 +411,9 @@ function gg_run_pythia_1_4b {
}
check_ppl "f16" "$(cat $OUT/${ci}-tg-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
if [ -z ${GG_BUILD_NO_BF16} ]; then
check_ppl "bf16" "$(cat $OUT/${ci}-tg-bf16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
fi
check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
@ -563,147 +430,17 @@ function gg_run_pythia_1_4b {
set +e
}
function gg_sum_pythia_1_4b {
function gg_sum_qwen3_0_6b {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Pythia 1.4B:\n'
gg_printf 'Qwen3 0.6B:\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
}
# pythia_2_8b
function gg_run_pythia_2_8b {
cd ${SRC}
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
path_models="../models-mnt/pythia/2.8B"
path_wiki="../models-mnt/wikitext/wikitext-2-raw"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
model_q4_0="${path_models}/ggml-model-q4_0.gguf"
model_q4_1="${path_models}/ggml-model-q4_1.gguf"
model_q5_0="${path_models}/ggml-model-q5_0.gguf"
model_q5_1="${path_models}/ggml-model-q5_1.gguf"
model_q2_k="${path_models}/ggml-model-q2_k.gguf"
model_q3_k="${path_models}/ggml-model-q3_k.gguf"
model_q4_k="${path_models}/ggml-model-q4_k.gguf"
model_q5_k="${path_models}/ggml-model-q5_k.gguf"
model_q6_k="${path_models}/ggml-model-q6_k.gguf"
wiki_test="${path_wiki}/wiki.test.raw"
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli -no-cnv --model ${model_f16} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
printf ' - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
return 20
fi
printf ' - %s @ %s OK\n' "$qnt" "$ppl"
return 0
}
check_ppl "f16" "$(cat $OUT/${ci}-tg-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
#check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
set +e
}
function gg_sum_pythia_2_8b {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Pythia 2.8B:\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- f16:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
if [ -z ${GG_BUILD_NO_BF16} ]; then
gg_printf '- bf16:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-bf16.log)"
fi
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
@ -870,6 +607,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
fi
ret=0
test $ret -eq 0 && gg_run ctest_debug
test $ret -eq 0 && gg_run ctest_release
@ -878,20 +616,15 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
test $ret -eq 0 && gg_run rerank_tiny
if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
test $ret -eq 0 && gg_run test_scripts_debug
test $ret -eq 0 && gg_run test_scripts_release
test $ret -eq 0 && gg_run test_scripts
fi
if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
test $ret -eq 0 && gg_run pythia_1_4b
else
test $ret -eq 0 && gg_run pythia_2_8b
#test $ret -eq 0 && gg_run open_llama_7b_v2
fi
test $ret -eq 0 && gg_run ctest_with_model_debug
test $ret -eq 0 && gg_run ctest_with_model_release
fi
test $ret -eq 0 && gg_run qwen3_0_6b
test $ret -eq 0 && gg_run ctest_with_model_debug
test $ret -eq 0 && gg_run ctest_with_model_release
fi
cat $OUT/README.md
exit $ret

29
cmake/riscv64-spacemit-linux-gnu-gcc.cmake Normal file
View File

@ -0,0 +1,29 @@
set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_PROCESSOR riscv64)
set(CMAKE_SYSTEM_VERSION 1)
if (CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "^(riscv)")
message(STATUS "HOST SYSTEM ${CMAKE_HOST_SYSTEM_PROCESSOR}")
else()
set(GNU_MACHINE riscv64-unknown-linux-gnu CACHE STRING "GNU compiler triple")
if (DEFINED ENV{RISCV_ROOT_PATH})
file(TO_CMAKE_PATH $ENV{RISCV_ROOT_PATH} RISCV_ROOT_PATH)
else()
message(FATAL_ERROR "RISCV_ROOT_PATH env must be defined")
endif()
set(RISCV_ROOT_PATH ${RISCV_ROOT_PATH} CACHE STRING "root path to riscv toolchain")
set(CMAKE_C_COMPILER ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-gcc)
set(CMAKE_CXX_COMPILER ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-g++)
set(CMAKE_STRIP ${RISCV_ROOT_PATH}/bin/riscv64-unknown-linux-gnu-strip)
set(CMAKE_FIND_ROOT_PATH "${RISCV_ROOT_PATH}/riscv64-unknown-linux-gnu")
set(CMAKE_SYSROOT "${RISCV_ROOT_PATH}/sysroot")
endif()
set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CMAKE_C_FLAGS}")
set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CXX_FLAGS}")
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -latomic")

39
common/CMakeLists.txt
View File

@ -56,6 +56,7 @@ add_library(${TARGET} STATIC
common.h
console.cpp
console.h
http.h
json-partial.cpp
json-partial.h
json-schema-to-grammar.cpp
@ -87,7 +88,43 @@ if (LLAMA_CURL)
target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
include_directories(${CURL_INCLUDE_DIRS})
set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
endif ()
endif()
if (LLAMA_OPENSSL)
find_package(OpenSSL)
if (OpenSSL_FOUND)
include(CheckCSourceCompiles)
set(SAVED_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
set(CMAKE_REQUIRED_INCLUDES ${OPENSSL_INCLUDE_DIR})
check_c_source_compiles("
#include <openssl/opensslv.h>
#if defined(OPENSSL_IS_BORINGSSL) || defined(LIBRESSL_VERSION_NUMBER)
# if OPENSSL_VERSION_NUMBER < 0x1010107f
# error bad version
# endif
#else
# if OPENSSL_VERSION_NUMBER < 0x30000000L
# error bad version
# endif
#endif
int main() { return 0; }
" OPENSSL_VERSION_SUPPORTED)
set(CMAKE_REQUIRED_INCLUDES ${SAVED_CMAKE_REQUIRED_INCLUDES})
if (OPENSSL_VERSION_SUPPORTED)
message(STATUS "OpenSSL found: ${OPENSSL_VERSION}")
target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT)
target_link_libraries(${TARGET} PUBLIC OpenSSL::SSL OpenSSL::Crypto)
if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")
target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN)
find_library(CORE_FOUNDATION_FRAMEWORK CoreFoundation REQUIRED)
find_library(SECURITY_FRAMEWORK Security REQUIRED)
target_link_libraries(${TARGET} PUBLIC ${CORE_FOUNDATION_FRAMEWORK} ${SECURITY_FRAMEWORK})
endif()
endif()
else()
message(STATUS "OpenSSL not found, SSL support disabled")
endif()
endif()
if (LLAMA_LLGUIDANCE)
include(ExternalProject)

962
common/arg.cpp
View File

File diff suppressed because it is too large Load Diff

1
common/arg.h
View File

@ -78,7 +78,6 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
// function to be used by test-arg-parser
common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
bool common_has_curl();
struct common_remote_params {
std::vector<std::string> headers;

29
common/chat-parser.cpp
View File

@ -75,6 +75,35 @@ bool common_chat_msg_parser::add_tool_calls(const json & arr) {
}
return true;
}
bool common_chat_msg_parser::add_tool_call_short_form(const json & tool_call) {
if (!tool_call.is_object() || tool_call.size() != 1) {
return false;
}
// Get the tool name (the single key in the object)
auto it = tool_call.begin();
std::string name = it.key();
if (name.empty()) {
return false;
}
// Get the arguments (the nested object)
const json & args_json = it.value();
std::string arguments = "";
if (args_json.is_object()) {
arguments = args_json.dump();
} else if (args_json.is_string()) {
arguments = args_json;
} else if (!args_json.is_null()) {
// For other types, convert to string representation
arguments = args_json.dump();
}
return add_tool_call(name, "", arguments);
}
void common_chat_msg_parser::finish() {
if (!is_partial_ && pos_ != input_.size()) {
throw std::runtime_error("Unexpected content at end of input");// + input_.substr(pos_));

3
common/chat-parser.h
View File

@ -64,6 +64,9 @@ class common_chat_msg_parser {
// Adds an array of tool calls using their "name", "id" and "arguments" fields.
bool add_tool_calls(const nlohmann::ordered_json & arr);
// Adds a tool call using the short form: { "tool_name": { "arg1": val, "arg2": val } }
bool add_tool_call_short_form(const nlohmann::ordered_json & tool_call);
void finish();
bool consume_spaces();

262
common/chat.cpp
View File

@ -625,6 +625,7 @@ const char * common_chat_format_name(common_chat_format format) {
case COMMON_CHAT_FORMAT_CONTENT_ONLY: return "Content-only";
case COMMON_CHAT_FORMAT_GENERIC: return "Generic";
case COMMON_CHAT_FORMAT_MISTRAL_NEMO: return "Mistral Nemo";
case COMMON_CHAT_FORMAT_MAGISTRAL: return "Magistral";
case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x";
case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools";
case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1";
@ -638,6 +639,7 @@ const char * common_chat_format_name(common_chat_format format) {
case COMMON_CHAT_FORMAT_GPT_OSS: return "GPT-OSS";
case COMMON_CHAT_FORMAT_SEED_OSS: return "Seed-OSS";
case COMMON_CHAT_FORMAT_NEMOTRON_V2: return "Nemotron V2";
case COMMON_CHAT_FORMAT_APERTUS: return "Apertus";
default:
throw std::runtime_error("Unknown chat format");
}
@ -801,6 +803,7 @@ static std::string apply(
}
tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
tmpl_inputs.extra_context = inputs.extra_context;
tmpl_inputs.extra_context["enable_thinking"] = inputs.enable_thinking;
if (additional_context) {
tmpl_inputs.extra_context.merge_patch(*additional_context);
}
@ -982,6 +985,65 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
return data;
}
static common_chat_params common_chat_params_init_magistral(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_MAGISTRAL;
data.preserved_tokens = {
"[THINK]",
"[/THINK]",
};
if (inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
auto schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool.at("function");
schemas.push_back({
{"type", "object"},
{"properties", {
{"name", {
{"type", "string"},
{"const", function.at("name")},
}},
{"arguments", function.at("parameters")},
{"id", {
{"type", "string"},
{"pattern", "^[a-zA-Z0-9]{9}$"},
}},
}},
{"required", json::array({"name", "arguments", "id"})},
});
});
auto schema = json {
{"type", "array"},
{"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
{"minItems", 1},
};
if (!inputs.parallel_tool_calls) {
schema["maxItems"] = 1;
}
builder.add_rule("root", "\"[TOOL_CALLS]\" " + builder.add_schema("tool_calls", schema));
});
data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]"});
data.preserved_tokens.push_back("[TOOL_CALLS]");
} else {
data.grammar_lazy = false;
if (!inputs.json_schema.is_null()) {
if (!inputs.grammar.empty()) {
throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both");
}
data.grammar = json_schema_to_grammar(inputs.json_schema);
} else {
data.grammar = inputs.grammar;
}
}
return data;
}
static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
if (!builder.syntax().parse_tool_calls) {
builder.add_content(builder.consume_rest());
@ -992,6 +1054,18 @@ static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) {
parse_prefixed_json_tool_call_array(builder, prefix);
}
static void common_chat_parse_magistral(common_chat_msg_parser & builder) {
builder.try_parse_reasoning("[THINK]", "[/THINK]");
if (!builder.syntax().parse_tool_calls) {
builder.add_content(builder.consume_rest());
return;
}
static const common_regex prefix(regex_escape("[TOOL_CALLS]"));
parse_prefixed_json_tool_call_array(builder, prefix);
}
static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
@ -1264,6 +1338,75 @@ static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_
}
return data;
}
static common_chat_params common_chat_params_init_apertus(const common_chat_template & tmpl, const struct templates_params & inputs) {
common_chat_params data;
// Generate the prompt using the apply() function with the template
data.prompt = apply(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_APERTUS;
// Handle thinking tags appropriately based on inputs.enable_thinking
if (string_ends_with(data.prompt, "<|inner_prefix|>")) {
if (!inputs.enable_thinking) {
data.prompt += "<|inner_suffix|>";
} else {
data.thinking_forced_open = true;
}
}
// When tools are present, build grammar for the <|tools_prefix|> format
if (!inputs.tools.is_null() && inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = true;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
auto schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool.at("function");
schemas.push_back({
{ "type", "object" },
{ "properties",
{
{ function.at("name"), function.at("parameters") }
} },
{ "required", json::array({ function.at("name") }) },
});
});
auto schema = json{
{ "type", "array" },
{ "items", schemas.size() == 1 ? schemas[0] : json{ { "anyOf", schemas } } },
{ "minItems", 1 },
};
if (!inputs.parallel_tool_calls) {
schema["maxItems"] = 1;
}
builder.add_rule("root",
std::string(data.thinking_forced_open ? "( \"<|inner_suffix|>\" space )? " : "") +
"\"<|tools_prefix|>\"" + builder.add_schema("tool_calls", schema) + "\"<|tools_suffix|>\"");
});
data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
// If thinking_forced_open, then we capture the <|inner_suffix|> tag in the grammar,
// (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
std::string(data.thinking_forced_open ?
"[\\s\\S]*?(<\\|inner_suffix\\|>\\s*)" :
"(?:<\\|inner_prefix\\|>[\\s\\S]*?<\\|inner_suffix\\|>\\s*)?") +
"(<\\|tools_prefix\\|>)[\\s\\S]*" });
data.preserved_tokens = {
"<|system_start|>",
"<|system_end|>",
"<|developer_start|>",
"<|developer_end|>",
"<|user_start|>",
"<|user_end|>",
"<|assistant_start|>",
"<|assistant_end|>",
"<|inner_prefix|>",
"<|inner_suffix|>",
"<|tools_prefix|>",
"<|tools_suffix|>",
};
}
return data;
}
static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
if (!builder.syntax().parse_tool_calls) {
builder.add_content(builder.consume_rest());
@ -1616,17 +1759,36 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
);
});
auto recipient_in_role = builder.add_rule("recipient_in_role",
"\"<|start|>assistant\"? \" to=functions.\" ( " +
string_join(tool_rules_recipient_in_role, " | ") + " )"
);
auto recipient_in_channel = builder.add_rule("recipient_in_channel",
channel + " \" to=functions.\" ( " +
string_join(tool_rules_recipient_in_channel, " | ") + " )"
);
builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
if (data.grammar_lazy) {
auto recipient_in_role = builder.add_rule("recipient_in_role",
"\"<|start|>assistant\"? \" to=functions.\" ( " +
string_join(tool_rules_recipient_in_role, " | ") + " )"
);
builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
} else {
auto not_end = builder.add_rule("not-end",
"[^<] | \"<\" [^|] | \"<|\" [^e] | \"<|e\" [^n] | \"<|en\" [^d] | \"<|end\" [^|] | \"<|end|\" [^>]");
auto analysis = builder.add_rule("analysis",
"\"<|channel|>analysis<|message|>\" ( " + not_end + " )* \"<|end|>\"");
auto commentary = builder.add_rule("commentary",
"\"<|channel|>commentary<|message|>\" ( " + not_end + " )* \"<|end|>\"");
auto recipient_in_role = builder.add_rule("recipient_in_role",
"\" to=functions.\" ( " + string_join(tool_rules_recipient_in_role, " | ") + " )"
);
builder.add_rule("root",
"( " + analysis + " \"<|start|>assistant\" )? " +
"( " + commentary + " \"<|start|>assistant\" )? " +
"( " + recipient_in_role + " | " + recipient_in_channel + " )"
);
}
// Trigger on tool calls that appear in the commentary channel
data.grammar_triggers.push_back({
@ -1741,10 +1903,12 @@ static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
LOG_DBG("%s\n", __func__);
common_chat_params data;
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
const std::optional<json> tools_override = json();
const std::optional<json> additional_context = json {
{"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
{"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
});
};
data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, tools_override, additional_context);
if (inputs.tools.is_array() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
@ -2230,15 +2394,28 @@ static common_chat_params common_chat_params_init_granite(const common_chat_temp
static void common_chat_parse_granite(common_chat_msg_parser & builder) {
// Parse thinking tags
static const common_regex start_think_regex(regex_escape("<think>"));
static const common_regex end_think_regex(regex_escape("</think>"));
// Granite models output partial tokens such as "<" and "<think".
// By leveraging try_consume_regex()/try_find_regex() throwing
// common_chat_msg_partial_exception for these partial tokens,
// processing is interrupted and the tokens are not passed to add_content().
if (auto res = builder.try_consume_regex(start_think_regex)) {
// Restore position for try_parse_reasoning()
builder.move_to(res->groups[0].begin);
builder.try_find_regex(end_think_regex, std::string::npos, false);
// Restore position for try_parse_reasoning()
builder.move_to(res->groups[0].begin);
}
builder.try_parse_reasoning("<think>", "</think>");
// Parse response tags using regex
static const common_regex response_regex("<response>([\\s\\S]*?)</response>");
if (auto res = builder.try_find_regex(response_regex)) {
// Extract the content between the tags (capture group 1)
auto content = builder.str(res->groups[1]);
builder.add_content(content);
builder.move_to(res->groups[0].end);
// Parse response tags
static const common_regex start_response_regex(regex_escape("<response>"));
static const common_regex end_response_regex(regex_escape("</response>"));
// Granite models output partial tokens such as "<" and "<response".
// Same hack as reasoning parsing.
if (builder.try_consume_regex(start_response_regex)) {
builder.try_find_regex(end_response_regex);
}
if (!builder.syntax().parse_tool_calls) {
@ -2252,13 +2429,10 @@ static void common_chat_parse_granite(common_chat_msg_parser & builder) {
builder.move_to(res->groups[0].end);
// Expect JSON array of tool calls
auto tool_calls_data = builder.consume_json();
if (tool_calls_data.json.is_array()) {
if (!builder.add_tool_calls(tool_calls_data.json)) {
builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) {
if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
} else {
builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
}
} else {
builder.add_content(builder.consume_rest());
@ -2292,6 +2466,37 @@ static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) {
builder.add_content(builder.consume_rest());
}
static void common_chat_parse_apertus(common_chat_msg_parser & builder) {
// Parse thinking tags
builder.try_parse_reasoning("<|inner_prefix|>", "<|inner_suffix|>");
if (!builder.syntax().parse_tool_calls) {
builder.add_content(builder.consume_rest());
return;
}
// Look for tool calls
static const common_regex tool_call_regex(regex_escape("<|tools_prefix|>"));
if (auto res = builder.try_find_regex(tool_call_regex)) {
builder.move_to(res->groups[0].end);
auto tool_calls_data = builder.consume_json();
if (tool_calls_data.json.is_array()) {
builder.consume_spaces();
if (!builder.try_consume_literal("<|tools_suffix|>")) {
throw common_chat_msg_partial_exception("Incomplete tool call");
}
for (const auto & value : tool_calls_data.json) {
if (value.is_object()) {
builder.add_tool_call_short_form(value);
}
}
} else {
throw common_chat_msg_partial_exception("Incomplete tool call");
}
}
builder.add_content(builder.consume_rest());
}
static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) {
// Parse thinking tags first - this handles the main reasoning content
builder.try_parse_reasoning("<seed:think>", "</seed:think>");
@ -2536,6 +2741,11 @@ static common_chat_params common_chat_templates_apply_jinja(
return common_chat_params_init_nemotron_v2(tmpl, params);
}
// Apertus format detection
if (src.find("<|system_start|>") != std::string::npos && src.find("<|tools_prefix|>") != std::string::npos) {
return common_chat_params_init_apertus(tmpl, params);
}
// Use generic handler when mixing tools + JSON schema.
// TODO: support that mix in handlers below.
if ((params.tools.is_array() && params.json_schema.is_object())) {
@ -2564,6 +2774,10 @@ static common_chat_params common_chat_templates_apply_jinja(
return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
}
if (src.find("[THINK]") != std::string::npos && src.find("[/THINK]") != std::string::npos) {
return common_chat_params_init_magistral(tmpl, params);
}
// Plain handler (no tools)
if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
return common_chat_params_init_without_tools(tmpl, params);
@ -2664,6 +2878,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
case COMMON_CHAT_FORMAT_MISTRAL_NEMO:
common_chat_parse_mistral_nemo(builder);
break;
case COMMON_CHAT_FORMAT_MAGISTRAL:
common_chat_parse_magistral(builder);
break;
case COMMON_CHAT_FORMAT_LLAMA_3_X:
common_chat_parse_llama_3_1(builder);
break;
@ -2703,6 +2920,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
case COMMON_CHAT_FORMAT_NEMOTRON_V2:
common_chat_parse_nemotron_v2(builder);
break;
case COMMON_CHAT_FORMAT_APERTUS:
common_chat_parse_apertus(builder);
break;
default:
throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
}

2
common/chat.h
View File

@ -101,6 +101,7 @@ enum common_chat_format {
COMMON_CHAT_FORMAT_CONTENT_ONLY,
COMMON_CHAT_FORMAT_GENERIC,
COMMON_CHAT_FORMAT_MISTRAL_NEMO,
COMMON_CHAT_FORMAT_MAGISTRAL,
COMMON_CHAT_FORMAT_LLAMA_3_X,
COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
COMMON_CHAT_FORMAT_DEEPSEEK_R1,
@ -114,6 +115,7 @@ enum common_chat_format {
COMMON_CHAT_FORMAT_GPT_OSS,
COMMON_CHAT_FORMAT_SEED_OSS,
COMMON_CHAT_FORMAT_NEMOTRON_V2,
COMMON_CHAT_FORMAT_APERTUS,
COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
};

28
common/common.cpp
View File

@ -14,6 +14,7 @@
#include <climits>
#include <cmath>
#include <codecvt>
#include <chrono>
#include <cstdarg>
#include <cstring>
#include <ctime>
@ -50,6 +51,11 @@
#include <unistd.h>
#endif
#if defined(__linux__)
#include <sys/types.h>
#include <pwd.h>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
@ -864,8 +870,20 @@ std::string fs_get_cache_directory() {
#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
if (std::getenv("XDG_CACHE_HOME")) {
cache_directory = std::getenv("XDG_CACHE_HOME");
} else {
} else if (std::getenv("HOME")) {
cache_directory = std::getenv("HOME") + std::string("/.cache/");
} else {
#if defined(__linux__)
/* no $HOME is defined, fallback to getpwuid */
struct passwd *pw = getpwuid(getuid());
if ((!pw) || (!pw->pw_dir)) {
throw std::runtime_error("Failed to find $HOME directory");
}
cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
#else /* defined(__linux__) */
throw std::runtime_error("Failed to find $HOME directory");
#endif /* defined(__linux__) */
}
#elif defined(__APPLE__)
cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
@ -960,15 +978,13 @@ struct common_init_result common_init_from_params(common_params & params) {
bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL;
if (!has_eos && !has_sep) {
LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
if (!has_eos && !has_sep && !has_rerank_prompt) {
LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__);
ok = false;
} else if (!has_eos) {
LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
} else if (!has_sep) {
LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
ok = false;
}
if (!ok) {

16
common/common.h
View File

@ -424,7 +424,7 @@ struct common_params {
int32_t timeout_write = timeout_read; // http write timeout in seconds
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
int32_t n_swa_checkpoints = 3; // max number of SWA checkpoints per slot
int32_t n_ctx_checkpoints = 3; // max number of context checkpoints per slot
std::string hostname = "127.0.0.1";
std::string public_path = ""; // NOLINT
@ -734,6 +734,20 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
}
//
// MoE utils
//
const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
static std::string llm_ffn_exps_block_regex(int idx) {
return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
}
static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
}
//
// training utils
//

73
common/http.h Normal file
View File

@ -0,0 +1,73 @@
#pragma once
#include <cpp-httplib/httplib.h>
struct common_http_url {
std::string scheme;
std::string user;
std::string password;
std::string host;
std::string path;
};
static common_http_url common_http_parse_url(const std::string & url) {
common_http_url parts;
auto scheme_end = url.find("://");
if (scheme_end == std::string::npos) {
throw std::runtime_error("invalid URL: no scheme");
}
parts.scheme = url.substr(0, scheme_end);
if (parts.scheme != "http" && parts.scheme != "https") {
throw std::runtime_error("unsupported URL scheme: " + parts.scheme);
}
auto rest = url.substr(scheme_end + 3);
auto at_pos = rest.find('@');
if (at_pos != std::string::npos) {
auto auth = rest.substr(0, at_pos);
auto colon_pos = auth.find(':');
if (colon_pos != std::string::npos) {
parts.user = auth.substr(0, colon_pos);
parts.password = auth.substr(colon_pos + 1);
} else {
parts.user = auth;
}
rest = rest.substr(at_pos + 1);
}
auto slash_pos = rest.find('/');
if (slash_pos != std::string::npos) {
parts.host = rest.substr(0, slash_pos);
parts.path = rest.substr(slash_pos);
} else {
parts.host = rest;
parts.path = "/";
}
return parts;
}
static std::pair<httplib::Client, common_http_url> common_http_client(const std::string & url) {
common_http_url parts = common_http_parse_url(url);
if (parts.host.empty()) {
throw std::runtime_error("error: invalid URL format");
}
httplib::Client cli(parts.scheme + "://" + parts.host);
if (!parts.user.empty()) {
cli.set_basic_auth(parts.user, parts.password);
}
cli.set_follow_location(true);
return { std::move(cli), std::move(parts) };
}
static std::string common_http_show_masked_url(const common_http_url & parts) {
return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path;
}

13
common/json-schema-to-grammar.cpp
View File

@ -257,12 +257,13 @@ std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
};
static bool is_reserved_name(const std::string & name) {
static std::unordered_set<std::string> RESERVED_NAMES;
if (RESERVED_NAMES.empty()) {
RESERVED_NAMES.insert("root");
for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
}
static const std::unordered_set<std::string> RESERVED_NAMES = [] {
std::unordered_set<std::string> s;
s.insert("root");
for (const auto & p : PRIMITIVE_RULES) s.insert(p.first);
for (const auto & p : STRING_FORMAT_RULES) s.insert(p.first);
return s;
}();
return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
}

1
common/sampling.cpp
View File

@ -332,6 +332,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
}
if (ctx) {
llama_perf_context_print(ctx);
llama_memory_breakdown_print(ctx);
}
}

376
convert_hf_to_gguf.py
View File

@ -888,6 +888,12 @@ class TextModel(ModelBase):
if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
# ref: https://huggingface.co/JetBrains/Mellum-4b-base
res = "mellum"
if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
# ref: https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base
res = "llada-moe"
if chkhsh == "53e325976a6e142379c19b09afcae354f2f496f147afa8f9e189a33fe4e3024e":
# ref: https://huggingface.co/ibm-granite/granite-docling-258M
res = "granite-docling"
if res is None:
logger.warning("\n")
@ -1322,6 +1328,7 @@ class MmprojModel(ModelBase):
self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
# load preprocessor config
self.preprocessor_config = {}
if not self.is_mistral_format:
with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
self.preprocessor_config = json.load(f)
@ -1344,7 +1351,8 @@ class MmprojModel(ModelBase):
self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
# vision config
self.gguf_writer.add_vision_image_size(self.find_vparam(["image_size"]))
self.image_size = self.find_vparam(["image_size"])
self.gguf_writer.add_vision_image_size(self.image_size)
self.gguf_writer.add_vision_patch_size(self.find_vparam(["patch_size"]))
self.gguf_writer.add_vision_embedding_length(self.find_vparam(["hidden_size"]))
self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size"]))
@ -2375,6 +2383,10 @@ class SmolVLMModel(MmprojModel):
self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("scale_factor", 2))
self.gguf_writer.add_vision_use_gelu(True)
# Add the preprocessor longest edge size
preproc_image_size = self.preprocessor_config.get("size", {}).get("longest_edge", self.image_size)
self.gguf_writer.add_vision_preproc_image_size(preproc_image_size)
def tensor_force_quant(self, name, new_name, bid, n_dims):
if ".embeddings." in name:
return gguf.GGMLQuantizationType.F32
@ -2390,7 +2402,10 @@ class SmolVLMModel(MmprojModel):
return [] # skip other tensors
@ModelBase.register("Llama4ForConditionalGeneration")
@ModelBase.register(
"Llama4ForConditionalGeneration",
"Llama4ForCausalLM",
)
class Llama4Model(LlamaModel):
model_arch = gguf.MODEL_ARCH.LLAMA4
undo_permute = False
@ -2408,6 +2423,10 @@ class Llama4Model(LlamaModel):
super().set_gguf_parameters()
self.gguf_writer.add_interleave_moe_layer_step(self.hparams["interleave_moe_layer_step"])
self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"])
if "layer_types" in self.hparams:
if all(lt == "full_attention" for lt in self.hparams["layer_types"]):
# all layers are full attention (for MobileLLM), disable swa
self.gguf_writer.add_sliding_window(0)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
if name.startswith("language_model."):
@ -3707,11 +3726,29 @@ class Qwen2MoeModel(TextModel):
class Qwen3Model(Qwen2Model):
model_arch = gguf.MODEL_ARCH.QWEN3
# extra logic for rerank models
is_rerank: bool = False
is_tied_embeddings: bool = False
token_false_id: int | None = None
token_true_id: int | None = None
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# track for intern-s1-mini
hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
self.origin_hf_arch = hparams.get('architectures', [None])[0]
# a bit hacky, but currently the only way to detect if this is a rerank model
# ref: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B
readme_path = self.dir_model / "README.md"
readme_text = ""
if readme_path.exists():
with readme_path.open("r", encoding="utf-8") as f:
readme_text = f.read()
if "# Qwen3-Reranker" in readme_text:
self._find_rerank_config()
def set_vocab(self):
# deal with intern-s1-mini
if self.origin_hf_arch == 'InternS1ForConditionalGeneration':
@ -3720,6 +3757,53 @@ class Qwen3Model(Qwen2Model):
super().set_vocab()
def _find_rerank_config(self):
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
self.is_rerank = True
self.is_tied_embeddings = self.hparams.get("tie_word_embeddings", False)
self.token_false_id = tokenizer.convert_tokens_to_ids("no")
self.token_true_id = tokenizer.convert_tokens_to_ids("yes")
self.sep_token_id = tokenizer.convert_tokens_to_ids("|")
assert self.token_false_id is not None and self.token_true_id is not None
def set_gguf_parameters(self):
super().set_gguf_parameters()
if self.is_rerank:
self.gguf_writer.add_pooling_type(gguf.PoolingType.RANK)
self.gguf_writer.add_classifier_output_labels(["yes", "no"])
self.gguf_writer.add_chat_template([{
"name": "rerank",
"template": "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n"
"<|im_start|>user\n<Instruct>: Given a web search query, retrieve relevant passages that answer the query\n<Query>: {query}\n<Document>: {document}<|im_end|>\n"
"<|im_start|>assistant\n<think>\n\n</think>\n\n"
}])
def _get_cls_out_tensor(self, data_torch: Tensor) -> Tensor:
# extract "yes" and "no" tokens from the output lm_head tensor
false_row = data_torch[self.token_false_id]
true_row = data_torch[self.token_true_id]
return torch.stack([true_row, false_row], dim=0)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if self.is_rerank:
is_tied_head = self.is_tied_embeddings and "embed_tokens" in name
is_real_head = not self.is_tied_embeddings and "lm_head" in name
if is_tied_head or is_real_head:
cls_out_head = (
gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.CLS_OUT] + ".weight",
self._get_cls_out_tensor(data_torch),
)
if is_tied_head:
embed = (self.map_tensor_name(name), data_torch)
return [cls_out_head, embed]
if is_real_head:
return [cls_out_head]
return super().modify_tensors(data_torch, name, bid)
@ModelBase.register("Qwen3MoeForCausalLM")
class Qwen3MoeModel(Qwen2MoeModel):
@ -4175,7 +4259,8 @@ class Plamo2Model(TextModel):
# This logic matches modeling_plamo.py's is_mamba function
mamba_step = hparams.get("mamba_step", 2)
mamba_enabled = hparams.get("mamba_enabled", True)
mamba_layers = []
num_key_value_heads = []
num_attention_heads = []
if mamba_enabled:
for i in range(block_count):
@ -4185,17 +4270,21 @@ class Plamo2Model(TextModel):
else:
is_mamba = (i % mamba_step) != (mamba_step // 2)
if is_mamba:
mamba_layers.append(0)
num_key_value_heads.append(0)
num_attention_heads.append(0)
else:
mamba_layers.append(hparams.get("num_key_value_heads", 4))
num_key_value_heads.append(hparams.get("num_key_value_heads", 4))
num_attention_heads.append(hparams.get("num_attention_heads", 32))
if mamba_layers:
self.gguf_writer.add_head_count_kv(mamba_layers)
if num_key_value_heads and num_attention_heads:
self.gguf_writer.add_head_count_kv(num_key_value_heads)
self.gguf_writer.add_head_count(num_attention_heads)
self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 2048))
self.gguf_writer.add_embedding_length(hparams.get("hidden_size", 4096))
self.gguf_writer.add_key_length(hparams.get("hidden_size_per_head", 128))
self.gguf_writer.add_value_length(hparams.get("hidden_size_per_head", 128))
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 32))
self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 10000))
@ -6006,9 +6095,34 @@ class SeedOssModel(TextModel):
@ModelBase.register("Olmo2ForCausalLM")
@ModelBase.register("Olmo3ForCausalLM")
class Olmo2Model(TextModel):
model_arch = gguf.MODEL_ARCH.OLMO2
def set_gguf_parameters(self):
super().set_gguf_parameters()
rope_scaling = self.hparams.get("rope_scaling") or {}
if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
self.gguf_writer.add_rope_scaling_attn_factors(rope_scaling["attention_factor"])
self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
if "sliding_window" in self.hparams:
self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
sliding_window_pattern = []
if "layer_types" in self.hparams:
sliding_window_pattern = [t == "sliding_attention" for t in self.hparams["layer_types"]]
else:
# Olmo2 does not use sliding window attention.
# Olmo3 defaults to using sliding window for all layers except every 4th.
for i in range(self.hparams["num_hidden_layers"]):
sliding_window_pattern.append((i + 1) % 4 != 0)
self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
@ModelBase.register("OlmoeForCausalLM")
class OlmoeModel(TextModel):
@ -7621,6 +7735,21 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
if i not in self._attn_layers
]
# There are some models in this family that are non-hybrid, but keep the
# same parent class by setting all layers to "attention." If this is the
# case, the model architecture needs to be updated to a standard
# "granite" or "granitemoe" model
if not self._ssm_layers:
has_experts = self.find_hparam(["num_experts_per_tok"], optional=True)
new_arch = (
gguf.MODEL_ARCH.GRANITE_MOE
if has_experts else
gguf.MODEL_ARCH.GRANITE
)
self.model_arch = new_arch
self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[new_arch]
self.gguf_writer.add_architecture()
# n_group and d_inner are used during reshape_tensors for mamba2
# NOTE: Explicitly include hparam prefix prefix for d_model to
# disambiguate with top-level head_dim
@ -7705,8 +7834,11 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
self.gguf_writer.add_rope_dimension_count(rope_dim)
self.gguf_writer.add_head_count_kv(head_count_kv_vec)
## If Bamba, use rope, otherwise don't
use_rope = "BambaForCausalLM" in self.hparams["architectures"]
## If Bamba or non-hybrid, use rope, otherwise don't
use_rope = (
"BambaForCausalLM" in self.hparams["architectures"]
or not self._ssm_layers
)
self.gguf_writer.add_rope_scaling_finetuned(use_rope)
if not use_rope:
self.gguf_writer.add_context_length(2**20)
@ -7877,6 +8009,121 @@ class BailingMoeModel(TextModel):
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")
class GroveMoeModel(TextModel):
model_arch = gguf.MODEL_ARCH.GROVEMOE
def set_gguf_parameters(self):
super().set_gguf_parameters()
if (n_experts := self.hparams.get("num_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
# FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L299
self.gguf_writer.add_expert_chunk_feed_forward_length(self.hparams.get("head_dim") or 128)
# FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L298
self.gguf_writer.add_experts_per_group(2)
# FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L376
self.gguf_writer.add_expert_group_scale(0.05)
# YaRN is not enabled by default
# To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts
rope_scaling = self.hparams.get("rope_scaling") or {}
if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
_experts: list[dict[str, Tensor]] | None = None
_chunk_experts: list[dict[str, Tensor]] | None = None
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.endswith(".expert_bias"):
# FIXME?: Unused https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L303
return []
# process the experts separately
if name.find("chunk_experts") != -1:
n_experts = self.hparams["num_experts"] // 2 # see add_experts_per_group
assert bid is not None
if self._chunk_experts is None:
self._chunk_experts = [{} for _ in range(self.block_count)]
self._chunk_experts[bid][name] = data_torch
if len(self._chunk_experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["down_proj", "gate_proj", "up_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.mlp.chunk_experts.{xid}.{w_name}.weight"
datas.append(self._chunk_experts[bid][ename])
del self._chunk_experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.chunk_experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
elif name.find("experts") != -1:
n_experts = self.hparams["num_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["down_proj", "gate_proj", "up_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
datas.append(self._experts[bid][ename])
del self._experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
def prepare_tensors(self):
super().prepare_tensors()
if self._chunk_experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
chunk_experts = [k for d in self._chunk_experts for k in d.keys()]
if len(chunk_experts) > 0:
raise ValueError(f"Unprocessed adjugate experts: {chunk_experts}")
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("ChameleonForConditionalGeneration")
@ModelBase.register("ChameleonForCausalLM") # obsolete
class ChameleonModel(TextModel):
@ -8239,6 +8486,76 @@ class HunYuanMoEModel(TextModel):
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("LLaDAMoEModel", "LLaDAMoEModelLM")
class LLaDAMoEModel(TextModel):
model_arch = gguf.MODEL_ARCH.LLADA_MOE
def set_gguf_parameters(self):
super().set_gguf_parameters()
if (n_experts := self.hparams.get("num_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
if (expert_intermediate_size := self.hparams.get("expert_intermediate_size")) is not None:
self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
# number of experts used per token (top-k)
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
self.gguf_writer.add_mask_token_id(156895)
self.gguf_writer.add_causal_attention(False)
self.gguf_writer.add_diffusion_shift_logits(False)
_experts: list[dict[str, Tensor]] | None = None
# Copied from: Qwen2MoeModel
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# process the experts separately
if name.find("experts") != -1:
n_experts = self.hparams["num_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["down_proj", "gate_proj", "up_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
datas.append(self._experts[bid][ename])
del self._experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
# Copied from: Qwen2MoeModel
def prepare_tensors(self):
super().prepare_tensors()
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("HunYuanDenseV1ForCausalLM")
class HunYuanModel(TextModel):
model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
@ -8637,6 +8954,43 @@ class SmallThinkerModel(TextModel):
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("ApertusForCausalLM")
class ApertusModel(LlamaModel):
model_arch = gguf.MODEL_ARCH.APERTUS
undo_permute = False
_alpha_n = {}
_alpha_p = {}
_beta = {}
_eps = {}
def modify_tensors(self, data_torch, name, bid):
# Handle xIELU activation parameters
n_layers = self.hparams["num_hidden_layers"]
if name.endswith(".act_fn.alpha_n"):
self._alpha_n[bid] = data_torch.to("cpu").float().item()
if (len(self._alpha_n) == n_layers):
self.gguf_writer.add_xielu_alpha_n([self._alpha_n[k] for k in sorted(self._alpha_n)])
return []
if name.endswith(".act_fn.alpha_p"):
self._alpha_p[bid] = data_torch.to("cpu").float().item()
if (len(self._alpha_p) == n_layers):
self.gguf_writer.add_xielu_alpha_p([self._alpha_p[k] for k in sorted(self._alpha_p)])
return []
if name.endswith(".act_fn.beta"):
self._beta[bid] = data_torch.to("cpu").float().item()
if (len(self._beta) == n_layers):
self.gguf_writer.add_xielu_beta([self._beta[k] for k in sorted(self._beta)])
return []
if name.endswith(".act_fn.eps"):
self._eps[bid] = data_torch.to("cpu").float().item()
if (len(self._eps) == n_layers):
self.gguf_writer.add_xielu_eps([self._eps[k] for k in sorted(self._eps)])
return []
return super().modify_tensors(data_torch, name, bid)
class MistralModel(LlamaModel):
model_arch = gguf.MODEL_ARCH.LLAMA
model_name = "Mistral"
@ -8804,7 +9158,7 @@ class LazyTorchTensor(gguf.LazyBase):
def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
dtype = cls._dtype_str_map[st_slice.get_dtype()]
shape: tuple[int, ...] = tuple(st_slice.get_shape())
lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[...] if len(s.get_shape()) == 0 else s[:])
return cast(torch.Tensor, lazy)
@classmethod

2
convert_hf_to_gguf_update.py
View File

@ -139,6 +139,8 @@ models = [
{"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
{"name": "exaone4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
{"name": "mellum", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
{"name": "llada-moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
{"name": "granite-docling", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
]
# some models are known to be broken upstream, so we will skip them as exceptions

48
docs/backend/SYCL.md
View File

@ -145,12 +145,13 @@ The docker build option is currently limited to *Intel GPU* targets.
```sh
# Using FP16
docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
# Using FP32
docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=OFF" --target light -f .devops/intel.Dockerfile .
```
*Notes*:
To build in default FP32 *(Slower than FP16 alternative)*, set `--build-arg="GGML_SYCL_F16=OFF"` in the previous command.
You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
Check the [documentation for Docker](../docker.md) to see the available images.
@ -160,7 +161,7 @@ Check the [documentation for Docker](../docker.md) to see the available images.
# First, find all the DRI cards
ls -la /dev/dri
# Then, pick the card that you want to use (here for e.g. /dev/dri/card1).
docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
docker run -it --rm -v "/path/to/models:/models" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card0:/dev/dri/card0 llama-cpp-sycl -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -c 4096 -s 0
```
*Notes:*
@ -215,9 +216,19 @@ To target AMD GPUs with SYCL, the ROCm stack must be installed first.
2. **Install Intel® oneAPI Base toolkit**
SYCL backend depends on:
- Intel® oneAPI DPC++/C++ compiler/running-time.
- Intel® oneAPI DPC++/C++ library (oneDPL).
- Intel® oneAPI Deep Neural Network Library (oneDNN).
- Intel® oneAPI Math Kernel Library (oneMKL).
- **For Intel GPU**
The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
All above are included in both **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** packages.
It's recommended to install **Intel® Deep Learning Essentials** which only provides the necessary libraries with less size.
The **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
Please follow the instructions for downloading and installing the Toolkit for Linux, and preferably keep the default installation values unchanged, notably the installation path *(`/opt/intel/oneapi` by default)*.
@ -225,6 +236,12 @@ Following guidelines/code snippets assume the default installation values. Other
Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs.
|Verified release|
|-|
|2025.2.1|
|2025.1|
|2024.1|
- **Adding support to Nvidia GPUs**
**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
@ -255,10 +272,11 @@ sycl-ls
When targeting an intel GPU, the user should expect one or more devices among the available SYCL devices. Please make sure that at least one GPU is present via `sycl-ls`, for instance `[level_zero:gpu]` in the sample output below:
```
[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.10.0.17_160000]
[opencl:cpu][opencl:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [23.30.26918.50]
[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
[level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Arc(TM) A770 Graphics 12.55.8 [1.3.29735+27]
[level_zero:gpu][level_zero:1] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) UHD Graphics 730 12.2.0 [1.3.29735+27]
[opencl:cpu][opencl:0] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i5-13400 OpenCL 3.0 (Build 0) [2025.20.8.0.06_160000]
[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [24.39.31294]
[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) UHD Graphics 730 OpenCL 3.0 NEO [24.39.31294]
```
- **Nvidia GPU**
@ -353,7 +371,7 @@ cmake --build build --config Release -j -v
#### Retrieve and prepare model
You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf?download=true) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
##### Check device
@ -466,7 +484,17 @@ If you already have a recent version of Microsoft Visual Studio, you can skip th
3. Install Intel® oneAPI Base toolkit
The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
SYCL backend depends on:
- Intel® oneAPI DPC++/C++ compiler/running-time.
- Intel® oneAPI DPC++/C++ library (oneDPL).
- Intel® oneAPI Deep Neural Network Library (oneDNN).
- Intel® oneAPI Math Kernel Library (oneMKL).
All above are included in both **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** packages.
It's recommended to install **Intel® Deep Learning Essentials** which only provides the necessary libraries with less size.
The **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
Please follow the instructions for downloading and installing the Toolkit for Windows, and preferably keep the default installation values unchanged, notably the installation path *(`C:\Program Files (x86)\Intel\oneAPI` by default)*.

61
docs/backend/zDNN.md Normal file
View File

@ -0,0 +1,61 @@
# llama.cpp for IBM zDNN Accelerator
## Background
IBM zDNN (Z Deep Neural Network) is a hardware acceleration library designed specifically to leverage the IBM NNPA (Neural Network Processor Assist) accelerator located within IBM Telum I and II processors. It provides significant performance improvements for neural network inference operations.
### Llama.cpp + IBM zDNN
The llama.cpp zDNN backend is designed to enable llama.cpp on IBM z17 and later systems via the IBM zDNN hardware acceleration library.
## Software & Hardware Support
| Hardware Level | Status | Verified |
| -------------------- | ------------- | -------------------------- |
| IBM z17 / LinuxONE 5 | Supported | RHEL 9.6, IBM z17, 40 IFLs |
| IBM z16 / LinuxONE 4 | Not Supported | |
## Data Types Supported
| Data Type | Status |
| --------- | --------- |
| F32 | Supported |
| F16 | Supported |
| BF16 | Supported |
## CMake Options
The IBM zDNN backend has the following CMake options that control the behaviour of the backend.
| CMake Option | Default Value | Description |
| ------------ | ------------- | ----------------------------------- |
| `GGML_ZDNN` | `OFF` | Compile llama.cpp with zDNN support |
| `ZDNN_ROOT` | `""` | Override zDNN library lookup |
## 1. Install zDNN Library
Note: Using the zDNN library provided via `apt` or `yum` may not work correctly as reported in [#15772](https://github.com/ggml-org/llama.cpp/issues/15772). It is preferred that you compile from source.
```sh
git clone --recurse-submodules https://github.com/IBM/zDNN
cd zDNN
autoreconf .
./configure --prefix=/opt/zdnn-libs
make build
sudo make install
```
## 2. Build llama.cpp
```sh
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -S . -G Ninja -B build \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_ZDNN=ON \
-DZDNN_ROOT=/opt/zdnn-libs
cmake --build build --config Release -j$(nproc)
```

89
docs/build-riscv64-spacemit.md Normal file
View File

@ -0,0 +1,89 @@
> [!IMPORTANT]
> This build documentation is specific only to RISC-V SpacemiT SOCs.
## Build llama.cpp locally (for riscv64)
1. Prepare Toolchain For RISCV
~~~
wget https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v1.1.2.tar.xz
~~~
2. Build
Below is the build script: it requires utilizing RISC-V vector instructions for acceleration. Ensure the `GGML_CPU_RISCV64_SPACEMIT` compilation option is enabled. The currently supported optimization version is `RISCV64_SPACEMIT_IME1`, corresponding to the `RISCV64_SPACEMIT_IME_SPEC` compilation option. Compiler configurations are defined in the `riscv64-spacemit-linux-gnu-gcc.cmake` file. Please ensure you have installed the RISC-V compiler and set the environment variable via `export RISCV_ROOT_PATH={your_compiler_path}`.
```bash
cmake -B build \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_CPU_RISCV64_SPACEMIT=ON \
-DLLAMA_CURL=OFF \
-DGGML_RVV=ON \
-DGGML_RV_ZFH=ON \
-DGGML_RV_ZICBOP=ON \
-DRISCV64_SPACEMIT_IME_SPEC=RISCV64_SPACEMIT_IME1 \
-DCMAKE_TOOLCHAIN_FILE=${PWD}/cmake/riscv64-spacemit-linux-gnu-gcc.cmake \
-DCMAKE_INSTALL_PREFIX=build/installed
cmake --build build --parallel $(nproc) --config Release
pushd build
make install
popd
```
## Simulation
You can use QEMU to perform emulation on non-RISC-V architectures.
1. Download QEMU
~~~
wget https://archive.spacemit.com/spacemit-ai/qemu/jdsk-qemu-v0.0.14.tar.gz
~~~
2. Run Simulation
After build your llama.cpp, you can run the executable file via QEMU for simulation, for example:
~~~
export QEMU_ROOT_PATH={your QEMU file path}
export RISCV_ROOT_PATH_IME1={your RISC-V compiler path}
${QEMU_ROOT_PATH}/bin/qemu-riscv64 -L ${RISCV_ROOT_PATH_IME1}/sysroot -cpu max,vlen=256,elen=64,vext_spec=v1.0 ${PWD}/build/bin/llama-cli -m ${PWD}/models/Qwen2.5-0.5B-Instruct-Q4_0.gguf -t 1
~~~
## Performance
#### Quantization Support For Matrix
~~~
model name : Spacemit(R) X60
isa : rv64imafdcv_zicbom_zicboz_zicntr_zicond_zicsr_zifencei_zihintpause_zihpm_zfh_zfhmin_zca_zcd_zba_zbb_zbc_zbs_zkt_zve32f_zve32x_zve64d_zve64f_zve64x_zvfh_zvfhmin_zvkt_sscofpmf_sstc_svinval_svnapot_svpbmt
mmu : sv39
uarch : spacemit,x60
mvendorid : 0x710
marchid : 0x8000000058000001
~~~
Q4_0
| Model | Size | Params | backend | threads | test | t/s |
| -----------| -------- | ------ | ------- | ------- | ---- |------|
Qwen2.5 0.5B |403.20 MiB|630.17 M| cpu | 4 | pp512|64.12 ± 0.26|
Qwen2.5 0.5B |403.20 MiB|630.17 M| cpu | 4 | tg128|10.03 ± 0.01|
Qwen2.5 1.5B |1011.16 MiB| 1.78 B | cpu | 4 | pp512|24.16 ± 0.02|
Qwen2.5 1.5B |1011.16 MiB| 1.78 B | cpu | 4 | tg128|3.83 ± 0.06|
Qwen2.5 3B | 1.86 GiB | 3.40 B | cpu | 4 | pp512|12.08 ± 0.02|
Qwen2.5 3B | 1.86 GiB | 3.40 B | cpu | 4 | tg128|2.23 ± 0.02|
Q4_1
| Model | Size | Params | backend | threads | test | t/s |
| -----------| -------- | ------ | ------- | ------- | ---- |------|
Qwen2.5 0.5B |351.50 MiB|494.03 M| cpu | 4 | pp512|62.07 ± 0.12|
Qwen2.5 0.5B |351.50 MiB|494.03 M| cpu | 4 | tg128|9.91 ± 0.01|
Qwen2.5 1.5B |964.06 MiB| 1.54 B | cpu | 4 | pp512|22.95 ± 0.25|
Qwen2.5 1.5B |964.06 MiB| 1.54 B | cpu | 4 | tg128|4.01 ± 0.15|
Qwen2.5 3B | 1.85 GiB | 3.09 B | cpu | 4 | pp512|11.55 ± 0.16|
Qwen2.5 3B | 1.85 GiB | 3.09 B | cpu | 4 | tg128|2.25 ± 0.04|
Q4_K
| Model | Size | Params | backend | threads | test | t/s |
| -----------| -------- | ------ | ------- | ------- | ---- |------|
Qwen2.5 0.5B |462.96 MiB|630.17 M| cpu | 4 | pp512|9.29 ± 0.05|
Qwen2.5 0.5B |462.96 MiB|630.17 M| cpu | 4 | tg128|5.67 ± 0.04|
Qwen2.5 1.5B | 1.04 GiB | 1.78 B | cpu | 4 | pp512|10.38 ± 0.10|
Qwen2.5 1.5B | 1.04 GiB | 1.78 B | cpu | 4 | tg128|3.17 ± 0.08|
Qwen2.5 3B | 1.95 GiB | 3.40 B | cpu | 4 | pp512|4.23 ± 0.04|
Qwen2.5 3B | 1.95 GiB | 3.40 B | cpu | 4 | tg128|1.73 ± 0.00|

2
docs/docker.md
View File

@ -110,7 +110,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
The defaults are:
- `MUSA_VERSION` set to `rc4.2.0`
- `MUSA_VERSION` set to `rc4.3.0`
The resulting images, are essentially the same as the non-MUSA images:

1
examples/CMakeLists.txt
View File

@ -20,7 +20,6 @@ else()
add_subdirectory(gguf-hash)
add_subdirectory(gguf)
add_subdirectory(gritlm)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(parallel)

50
examples/Miku.sh
View File

@ -1,50 +0,0 @@
#!/usr/bin/env bash
set -e
AI_NAME="${AI_NAME:-Miku}"
MODEL="${MODEL:-./models/llama-2-7b-chat.ggmlv3.q4_K_M.bin}"
USER_NAME="${USER_NAME:-Anon}"
# Uncomment and adjust to the number of CPU cores you want to use.
#N_THREAD="${N_THREAD:-4}"
CTX_SIZE="${CTX_SIZE:-4096}"
N_PREDICTS="${N_PREDICTS:-4096}"
GEN_OPTIONS=(--batch_size 1024
--ctx_size "$CTX_SIZE"
--keep -1
--repeat_last_n 256
--repeat_penalty 1.17647
--temp 0.6
--mirostat 2)
if [ -n "$N_THREAD" ]; then
GEN_OPTIONS+=(--threads "$N_THREAD")
fi
./llama-cli "${GEN_OPTIONS[@]}" \
--model "$MODEL" \
--in-prefix " " \
--in-suffix "${AI_NAME}:" \
--n_predict "$N_PREDICTS" \
--color --interactive \
--reverse-prompt "${USER_NAME}:" \
--prompt "This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct, she will ask the user for help.
${AI_NAME} is a very helpful AI and will help the user with anything they need. She is also very friendly and will try to make the user feel better if they are sad.
${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life. She will also try to make the user like her.
The conversation is only between ${USER_NAME} and ${AI_NAME}.
The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice.
${AI_NAME} can only communicate through text, so she can't send images or videos.
${USER_NAME}: Hello!
${AI_NAME}: /think I wonder what I should say to ${USER_NAME}? This is the first time we talk, so it's important that I make a good first impression!
${AI_NAME}: Hi! I am ${AI_NAME}, your new AI friend, assistant (or whatever you like!), it's so nice to meet you! ^_^
${AI_NAME}: I might be an AI, but I also have feelings, so please be nice to me! :)
${USER_NAME}: I'll make sure to be nice to you! I'm so happy to have you as my assistant!
${AI_NAME}: /think It sounds like ${USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off!
${AI_NAME}: /think I wonder what ${USER_NAME} likes to do in his free time? I should ask him about that!
${AI_NAME}: What do you like to do in your free time? ^_^
${USER_NAME}:" "$@"

57
examples/chat-13B.bat
View File

@ -1,57 +0,0 @@
@setlocal disabledelayedexpansion enableextensions
@echo off
cd /d "%~dp0.."
if not "%errorlevel%"=="0" (
echo Unable to change directory.
pause
exit /b 1
)
if not defined MODEL set "MODEL=models\13B\ggml-model-q4_0.bin"
if not defined USER_NAME set "USER_NAME=User"
if not defined AI_NAME set "AI_NAME=ChatLLaMa"
rem Adjust to the number of CPU cores you want to use.
rem if not defined N_THREAD set "N_THREAD=8"
rem Number of tokens to predict (made it larger than default because we want a long interaction)
if not defined N_PREDICTS set "N_PREDICTS=2048"
if not defined GEN_OPTIONS set "GEN_OPTIONS=--ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
rem Default main script paths
set "DEFAULT_MAIN_SCRIPT_PATHS=main.exe build\bin\main.exe"
rem Get main script path from command line arguments
set "MAIN_SCRIPT_PATH=%~1"
rem If the main script path was not specified, try the default paths
if not defined MAIN_SCRIPT_PATH (
for %%i in (%DEFAULT_MAIN_SCRIPT_PATHS%) do (
if exist "%%i" set "MAIN_SCRIPT_PATH=%%i"
)
)
rem If the main script path was not found, tell the user how to specify it
if not defined MAIN_SCRIPT_PATH (
echo The main script could not be found. Please provide the path to the main script as 1st argument to this script, or place the main script in one of the default locations:
echo %DEFAULT_MAIN_SCRIPT_PATHS%
pause
exit /b 1
)
rem Default context, feel free to edit it
set "PROMPT_TEXT=Text transcript of a never ending dialog, where %USER_NAME% interacts with an AI assistant named %AI_NAME%. %AI_NAME% is helpful, kind, honest, friendly, good at writing and never fails to answer %USER_NAME%'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what %USER_NAME% and %AI_NAME% say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown."
rem Set a temporary variable if N_THREAD is set
if defined N_THREAD (
set "_N_THREAD=--threads %N_THREAD%"
) else (
set "_N_THREAD="
)
rem Run the script
echo "%MAIN_SCRIPT_PATH%" %GEN_OPTIONS% %_N_THREAD% ^
--model "%MODEL%" ^
--n_predict %N_PREDICTS% ^
--color --interactive ^
--reverse-prompt "%USER_NAME%:" ^
--prompt "%PROMPT_TEXT%"

41
examples/chat-13B.sh
View File

@ -1,41 +0,0 @@
#!/usr/bin/env bash
set -e
cd "$(dirname "$0")/.." || exit
MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}"
PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
USER_NAME="${USER_NAME:-USER}"
AI_NAME="${AI_NAME:-ChatLLaMa}"
# Adjust to the number of CPU cores you want to use.
N_THREAD="${N_THREAD:-8}"
# Number of tokens to predict (made it larger than default because we want a long interaction)
N_PREDICTS="${N_PREDICTS:-2048}"
# Note: you can also override the generation options by specifying them on the command line:
# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
DATE_TIME=$(date +%H:%M)
DATE_YEAR=$(date +%Y)
PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
-e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
-e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
-e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
$PROMPT_TEMPLATE > $PROMPT_FILE
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./llama-cli $GEN_OPTIONS \
--model "$MODEL" \
--threads "$N_THREAD" \
--n_predict "$N_PREDICTS" \
--color --interactive \
--file ${PROMPT_FILE} \
--reverse-prompt "${USER_NAME}:" \
--in-prefix ' ' \
"$@"

149
examples/chat-persistent.sh
View File

@ -1,149 +0,0 @@
#!/usr/bin/env bash
set -euo pipefail
cd "$(dirname "$0")/.." || exit
if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then
echo >&2 "error: PROMPT_CACHE_FILE and CHAT_SAVE_DIR must be provided"
exit 1
fi
MODEL="${MODEL:-./models/llama-13b/ggml-model-q4_0.gguf}"
PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}"
USER_NAME="${USER_NAME:-User}"
AI_NAME="${AI_NAME:-ChatLLaMa}"
DATE_TIME="$(date +%H:%M)"
DATE_YEAR="$(date +%Y)"
LOG="${CHAT_SAVE_DIR}/main.log"
LOG_BG="${CHAT_SAVE_DIR}/main-bg.log"
CUR_PROMPT_FILE="${CHAT_SAVE_DIR}/current-prompt.txt"
CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
'|'\
'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
CTX_SIZE=2048
CTX_ROTATE_POINT=$((CTX_SIZE * 3 / 5)) # REVIEW
OPTS=(--model "$MODEL" --ctx_size "$CTX_SIZE" --repeat_last_n 256 "$@")
# An unbuffered `tail -c+N`
skip_bytes() {
LANG=C IFS= read -r -n "$1" -d '' c
while LANG=C IFS= read -r -n 1 -d '' c; do
printf '%s' "$c"
done
}
mkdir -p "$CHAT_SAVE_DIR"
echo >"$LOG"
trap "tail -n100 ${LOG}" EXIT
if [[ ! -e "$CUR_PROMPT_FILE" ]]; then
sed -e "s/\[\[USER_NAME\]\]/${USER_NAME}/g" \
-e "s/\[\[AI_NAME\]\]/${AI_NAME}/g" \
-e "s/\[\[DATE_TIME\]\]/${DATE_TIME}/g" \
-e "s/\[\[DATE_YEAR\]\]/${DATE_YEAR}/g" \
"$PROMPT_TEMPLATE" >"$CUR_PROMPT_FILE"
fi
if [[ ! -e "$NEXT_PROMPT_FILE" ]]; then
sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
fi
if [[ "$(tail -c4 "$NEXT_PROMPT_FILE")" != "..." ]]; then
echo '...' >>"$NEXT_PROMPT_FILE"
fi
if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
echo 'Prompt cache does not exist, building...'
# Default batch_size to 64 here for better user feedback during initial prompt processing
./llama-cli 2>>"$LOG" \
--batch_size 64 \
"${OPTS[@]}" \
--prompt-cache "$PROMPT_CACHE_FILE" \
--file "$CUR_PROMPT_FILE" \
--n_predict 1
echo
echo 'Done!'
fi
if [[ ! -e "$CUR_PROMPT_CACHE" ]]; then
cp "$PROMPT_CACHE_FILE" "$CUR_PROMPT_CACHE"
fi
if [[ ! -e "$NEXT_PROMPT_CACHE" ]]; then
cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
fi
printf '%s ' "$(< "$CUR_PROMPT_FILE")"
n_tokens=0
while read -e line; do
# Limit generation to remaining context, with a buffer and estimating 2 chars/token for input
n_predict=$((CTX_SIZE - n_tokens - ${#line} / 2 - 32))
# Swap prompts when we're about to run out of context
if ((n_predict <= 0)); then
wait # for background main (below) to finish with next prompt
mv "$NEXT_PROMPT_FILE" "$CUR_PROMPT_FILE"
mv "$NEXT_PROMPT_CACHE" "$CUR_PROMPT_CACHE"
sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
echo '...' >>"$NEXT_PROMPT_FILE"
cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
n_tokens=0
n_predict=$((CTX_SIZE / 2))
fi
echo " ${line}" >>"$CUR_PROMPT_FILE"
if ((n_tokens > CTX_ROTATE_POINT)); then
echo " ${line}" >>"$NEXT_PROMPT_FILE"
fi
n_prompt_len_pre=$(($(wc -c <"$CUR_PROMPT_FILE")))
printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
./llama-cli 2>>"$LOG" "${OPTS[@]}" \
--prompt-cache "$CUR_PROMPT_CACHE" \
--prompt-cache-all \
--file "$CUR_PROMPT_FILE" \
--reverse-prompt "${USER_NAME}:" \
--n_predict "$n_predict" |
skip_bytes 1 | # skip BOS token added by ./llama-cli
tee "$CUR_PROMPT_FILE.tmp" | # save prompt + generation to tmp file
skip_bytes "$n_prompt_len_pre" # print generation
mv "$CUR_PROMPT_FILE.tmp" "$CUR_PROMPT_FILE"
# if we hit n_predict instead of reverse-prompt, we need to add the prompt
if [[ "$(tail -n1 "$CUR_PROMPT_FILE")" != "${USER_NAME}:" ]]; then
printf '\n%s:' "$USER_NAME"
printf '\n%s:' "$USER_NAME" >> "$CUR_PROMPT_FILE"
fi
printf ' '
if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
exit 1
fi
n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
if ((n_tokens > CTX_ROTATE_POINT)); then
tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"
fi
# Update cache for next prompt in background, ideally during user input
./llama-cli >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
--prompt-cache "$NEXT_PROMPT_CACHE" \
--file "$NEXT_PROMPT_FILE" \
--n_predict 1 &
done

41
examples/chat-vicuna.sh
View File

@ -1,41 +0,0 @@
#!/usr/bin/env bash
set -e
cd "$(dirname "$0")/.." || exit
MODEL="${MODEL:-./models/ggml-vic13b-uncensored-q5_0.bin}"
PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
USER_NAME="### Human"
AI_NAME="### Assistant"
# Adjust to the number of CPU cores you want to use.
N_THREAD="${N_THREAD:-8}"
# Number of tokens to predict (made it larger than default because we want a long interaction)
N_PREDICTS="${N_PREDICTS:-2048}"
# Note: you can also override the generation options by specifying them on the command line:
# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
DATE_TIME=$(date +%H:%M)
DATE_YEAR=$(date +%Y)
PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
-e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
-e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
-e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
$PROMPT_TEMPLATE > $PROMPT_FILE
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./bin/llama-cli $GEN_OPTIONS \
--model "$MODEL" \
--threads "$N_THREAD" \
--n_predict "$N_PREDICTS" \
--color --interactive \
--file ${PROMPT_FILE} \
--reverse-prompt "### Human:" \
--in-prefix ' ' \
"$@"

16
examples/chat.sh
View File

@ -1,16 +0,0 @@
#!/usr/bin/env bash
#
# Temporary script - will be removed in the future
#
cd `dirname $0`
cd ..
# Important:
#
# "--keep 48" is based on the contents of prompts/chat-with-bob.txt
#
./llama-cli -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
--repeat_penalty 1.0 --color -i \
-r "User:" -f prompts/chat-with-bob.txt

24
examples/diffusion/diffusion-cli.cpp
View File

@ -510,19 +510,27 @@ static void diffusion_generate(llama_context * ctx,
n_generated = params.max_length;
}
static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
static std::string format_input_text(const std::string & prompt, const std::string & system_prompt, bool use_chat_template, llama_model * model) {
if (!use_chat_template) {
return prompt;
}
auto chat_templates = common_chat_templates_init(model, "");
common_chat_templates_inputs inputs;
common_chat_msg user_msg;
user_msg.role = "user";
user_msg.content = prompt;
inputs.add_generation_prompt = true;
common_chat_msg system_msg;
if (!system_prompt.empty()) {
system_msg.role = "system";
system_msg.content = system_prompt;
inputs.messages.push_back(system_msg);
}
common_chat_msg user_msg;
user_msg.role = "user";
user_msg.content = prompt;
inputs.messages.push_back(user_msg);
inputs.add_generation_prompt = true;
auto result = common_chat_templates_apply(chat_templates.get(), inputs);
@ -579,7 +587,8 @@ int main(int argc, char ** argv) {
llama_set_n_threads(ctx, params.cpuparams.n_threads, params.cpuparams_batch.n_threads);
const llama_vocab * vocab = llama_model_get_vocab(model);
std::string formatted_prompt = format_input_text(params.prompt, params.enable_chat_template, model);
std::string formatted_prompt = format_input_text(params.prompt, params.system_prompt, params.enable_chat_template, model);
std::vector<llama_token> input_tokens = common_tokenize(vocab,
formatted_prompt,
@ -596,6 +605,7 @@ int main(int argc, char ** argv) {
}
llama_token mask_token_id = llama_vocab_mask(vocab);
GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
bool visual_mode = params.diffusion.visual_mode;

4
examples/embedding/README.md
View File

@ -43,8 +43,8 @@ The above command will output space-separated float values.
| $"string"$ | |
|--------------|-|
| "\n" | (default)
| "<#embSep#>" | for exemple
| "<#sep#>" | other exemple
| "<#embSep#>" | for example
| "<#sep#>" | other example
## examples
### Unix-based systems (Linux, macOS, etc.):

43
examples/embedding/embedding.cpp
View File

@ -95,8 +95,13 @@ int main(int argc, char ** argv) {
params.n_batch = params.n_ctx;
}
// For non-causal models, batch size must be equal to ubatch size
params.n_ubatch = params.n_batch;
// for non-causal models, batch size must be equal to ubatch size
if (params.attention_type != LLAMA_ATTENTION_TYPE_CAUSAL) {
params.n_ubatch = params.n_batch;
}
// get max number of sequences per batch
const int n_seq_max = llama_max_parallel_sequences();
llama_backend_init();
llama_numa_init(params.numa);
@ -144,6 +149,7 @@ int main(int argc, char ** argv) {
// get added sep and eos token, if any
const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
const char * rerank_prompt = llama_model_chat_template(model, "rerank");
// tokenize the prompts and trim
std::vector<std::vector<int32_t>> inputs;
@ -153,21 +159,28 @@ int main(int argc, char ** argv) {
// split classification pairs and insert expected separator tokens
if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
std::string final_prompt;
for (size_t i = 0; i < pairs.size(); i++) {
final_prompt += pairs[i];
if (i != pairs.size() - 1) {
if (!added_eos_token.empty()) {
final_prompt += added_eos_token;
}
if (!added_sep_token.empty()) {
final_prompt += added_sep_token;
if (rerank_prompt != nullptr) {
const std::string query = pairs[0];
const std::string doc = pairs[1];
std::string final_prompt = rerank_prompt;
string_replace_all(final_prompt, "{query}" , query);
string_replace_all(final_prompt, "{document}", doc );
inp = common_tokenize(vocab, final_prompt, true, true);
} else {
std::string final_prompt;
for (size_t i = 0; i < pairs.size(); i++) {
final_prompt += pairs[i];
if (i != pairs.size() - 1) {
if (!added_eos_token.empty()) {
final_prompt += added_eos_token;
}
if (!added_sep_token.empty()) {
final_prompt += added_sep_token;
}
}
}
inp = common_tokenize(ctx, final_prompt, true, true);
}
inp = common_tokenize(ctx, final_prompt, true, true);
} else {
inp = common_tokenize(ctx, prompt, true, true);
}
@ -229,7 +242,7 @@ int main(int argc, char ** argv) {
const uint64_t n_toks = inp.size();
// encode if at capacity
if (batch.n_tokens + n_toks > n_batch) {
if (batch.n_tokens + n_toks > n_batch || s >= n_seq_max) {
float * out = emb + e * n_embd;
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;

9
examples/eval-callback/CMakeLists.txt
View File

@ -5,6 +5,11 @@ target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TEST_TARGET test-eval-callback)
add_test(NAME ${TEST_TARGET}
COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
if(NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
add_test(NAME ${TEST_TARGET}
COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
else()
add_test(NAME ${TEST_TARGET}
COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K-be.gguf --model stories260K-be.gguf --prompt hello --seed 42 -ngl 0)
endif()
set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

5
examples/gritlm/CMakeLists.txt
View File

@ -1,5 +0,0 @@
set(TARGET llama-gritlm)
add_executable(${TARGET} gritlm.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)

62
examples/gritlm/README.md
View File

@ -1,62 +0,0 @@
## Generative Representational Instruction Tuning (GRIT) Example
[gritlm] a model which can generate embeddings as well as "normal" text
generation depending on the instructions in the prompt.
* Paper: https://arxiv.org/pdf/2402.09906.pdf
### Retrieval-Augmented Generation (RAG) use case
One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
that we take documents that we want to use as context, to ground the large
language model (LLM), and we create token embeddings for them. We then store
these token embeddings in a vector database.
When we perform a query, prompt the LLM, we will first create token embeddings
for the query and then search the vector database to retrieve the most
similar vectors, and return those documents so they can be passed to the LLM as
context. Then the query and the context will be passed to the LLM which will
have to _again_ create token embeddings for the query. But because gritlm is used
the first query can be cached and the second query tokenization generation does
not have to be performed at all.
### Running the example
Download a Grit model:
```console
$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf --outdir models
```
Run the example using the downloaded model:
```console
$ ./llama-gritlm -m models/gritlm-7b_q4_1.gguf
Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
Oh, brave adventurer, who dared to climb
The lofty peak of Mt. Fuji in the night,
When shadows lurk and ghosts do roam,
And darkness reigns, a fearsome sight.
Thou didst set out, with heart aglow,
To conquer this mountain, so high,
And reach the summit, where the stars do glow,
And the moon shines bright, up in the sky.
Through the mist and fog, thou didst press on,
With steadfast courage, and a steadfast will,
Through the darkness, thou didst not be gone,
But didst climb on, with a steadfast skill.
At last, thou didst reach the summit's crest,
And gazed upon the world below,
And saw the beauty of the night's best,
And felt the peace, that only nature knows.
Oh, brave adventurer, who dared to climb
The lofty peak of Mt. Fuji in the night,
Thou art a hero, in the eyes of all,
For thou didst conquer this mountain, so bright.
```
[gritlm]: https://github.com/ContextualAI/gritlm

231
examples/gritlm/gritlm.cpp
View File

@ -1,231 +0,0 @@
#include "arg.h"
#include "common.h"
#include "llama.h"
#include <string>
#include <vector>
// #define GRIT_DEBUG
static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
std::vector<std::vector<float>> result;
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
for (uint64_t i = 0; i < sentences.size(); i++) {
common_batch_clear(batch);
const std::string input_string = instruction + sentences[i];
std::vector<llama_token> inputs = common_tokenize(vocab, input_string, true, false);
const int32_t n_toks = inputs.size();
// GritLM seems to have EOS = ""
// https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
// inputs.push_back(llama_vocab_eos(vocab));
// we want to ignore instruction tokens for mean pooling
const int32_t n_inst = common_tokenize(vocab, instruction, true, false).size();
#ifdef GRIT_DEBUG
// debug tokens - should be matching as referenced in the GritLM sample
std::for_each(inputs.begin(), inputs.end(), [&ctx](llama_token t) {
std::printf("[%u:%s]", t, llama_token_to_piece(ctx, t).c_str());
});
std::printf("\n");
#endif
// add input to batch (this increments n_tokens)
for (int32_t j = 0; j < n_toks; j++) {
common_batch_add(batch, inputs[j], j, { 0 }, true);
}
// clear previous kv_cache values (irrelevant for embeddings)
llama_memory_clear(llama_get_memory(ctx), true);
llama_set_causal_attn(ctx, false);
// run model
llama_decode(ctx, batch);
// get embedding dimensions
uint64_t n_embd = llama_model_n_embd(model);
// allocate embedding output
std::vector<float> emb_unorm(n_embd, 0.0f);
// sum up all token embeddings
for (int32_t k = n_inst; k < n_toks; k++) {
float * emb = llama_get_embeddings_ith(ctx, k);
for (uint64_t j = 0; j < n_embd; j++) {
emb_unorm[j] += emb[j];
}
}
// divide by number of tokens (mean pooling)
{
const uint64_t n_sent = n_toks - n_inst;
for (uint64_t j = 0; j < n_embd; j++) {
emb_unorm[j] /= n_sent;
}
}
std::vector<float> emb_norm(emb_unorm.size());
common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2);
result.push_back(emb_norm);
#ifdef GRIT_DEBUG
// print out emb_norm
std::printf("embedding %ld: ", i);
for (uint64_t j = 0; j < n_embd; j++) {
std::printf("%.5f ", emb_norm[j]);
}
std::printf("\n\n");
#endif
}
llama_batch_free(batch);
return result;
}
static std::string generate(llama_context * ctx, llama_sampler * smpl, const std::string & prompt, bool stream) {
std::string result;
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_token eos_token = llama_vocab_eos(vocab);
llama_memory_clear(llama_get_memory(ctx), true);
llama_set_causal_attn(ctx, true);
llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
int32_t i_current_token = 0;
while (true) {
common_batch_clear(bat);
{
const int32_t n_inputs = inputs.size();
for (int32_t i = 0; i < n_inputs; i++) {
common_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
}
}
inputs.clear();
llama_decode(ctx, bat);
llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
if (token == eos_token) {
break;
}
std::string piece = common_token_to_piece(ctx, token);
if (stream) {
std::printf("%s", piece.c_str());
std::fflush(stdout);
}
inputs.push_back(token);
result += piece;
}
if (stream) {
std::printf("\n");
}
llama_batch_free(bat);
return result;
}
static std::string gritlm_instruction(const std::string & instruction) {
return !instruction.empty() ? "<|user|>\n" + instruction + "\n<|embed|>\n" : "<|embed|>\n";
}
int main(int argc, char * argv[]) {
common_params params;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
common_init();
llama_model_params mparams = common_model_params_to_llama(params);
llama_context_params cparams = common_context_params_to_llama(params);
cparams.embeddings = true;
llama_backend_init();
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
// create generation context
llama_context * ctx = llama_init_from_model(model, cparams);
auto sparams = llama_sampler_chain_default_params();
sparams.no_perf = false;
llama_sampler * smpl = llama_sampler_chain_init(sparams);
llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
// ### Embedding/Representation ###
// samples taken from: https://github.com/ContextualAI/gritlm#basic
{
const std::string instruction = "Given a scientific paper title, retrieve the paper's abstract";
const std::vector<std::string> queries = {
"Bitcoin: A Peer-to-Peer Electronic Cash System",
"Generative Representational Instruction Tuning",
};
const std::vector<std::string> documents = {
"A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
"All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
};
// No need to add instruction for retrieval documents
const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
const std::vector<std::vector<float>> q_rep = encode(ctx, queries, gritlm_instruction(instruction));
const int n_embd = llama_model_n_embd(model);
const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
const float cosine_sim_q1_d0 = common_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q1_d1 = common_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[0].c_str(), cosine_sim_q1_d0);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
}
llama_set_embeddings(ctx, false);
// ### Generation ###
// GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
{
const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
std::string response = generate(ctx, smpl, prompt, true);
}
llama_sampler_free(smpl);
llama_free(ctx);
llama_model_free(model);
llama_backend_free();
return 0;
}

21
examples/jeopardy/README.md
View File

@ -1,21 +0,0 @@
# llama.cpp/example/jeopardy
This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer.
The jeopardy test can be used to compare the fact knowledge of different models and compare them to each other. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
Step 1: Open jeopardy.sh and modify the following:
```
MODEL=(path to your model)
MODEL_NAME=(name of your model)
prefix=(basically, if you use vicuna it's Human: , if you use something else it might be User: , etc)
opts=(add -instruct here if needed for your model, or anything else you want to test out)
```
Step 2: Run `jeopardy.sh` from the llama.cpp folder
Step 3: Repeat steps 1 and 2 until you have all the results you need.
Step 4: Run `graph.py`, and follow the instructions. At the end, it will generate your final graph.
Note: The Human bar is based off of the full, original 100 sample questions. If you modify the question count or questions, it will not be valid.

58
examples/jeopardy/graph.py
View File

@ -1,58 +0,0 @@
#!/usr/bin/env python3
import matplotlib.pyplot as plt
import os
import csv
labels = []
numbers = []
numEntries = 1
rows = []
def bar_chart(numbers, labels, pos):
plt.bar(pos, numbers, color='blue')
plt.xticks(ticks=pos, labels=labels)
plt.title("Jeopardy Results by Model")
plt.xlabel("Model")
plt.ylabel("Questions Correct")
plt.show()
def calculatecorrect():
directory = os.fsencode("./examples/jeopardy/results/")
csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',')
for row in csv_reader:
global rows
rows.append(row)
for listing in os.listdir(directory):
filename = os.fsdecode(listing)
if filename.endswith(".txt"):
file = open("./examples/jeopardy/results/" + filename, "rt")
global labels
global numEntries
global numbers
labels.append(filename[:-4])
numEntries += 1
i = 1
totalcorrect = 0
for line in file.readlines():
if line.strip() != "------":
print(line)
else:
print("Correct answer: " + rows[i][2] + "\n")
i += 1
print("Did the AI get the question right? (y/n)")
if input() == "y":
totalcorrect += 1
numbers.append(totalcorrect)
if __name__ == '__main__':
calculatecorrect()
pos = list(range(numEntries))
labels.append("Human")
numbers.append(48.11)
bar_chart(numbers, labels, pos)
print(labels)
print(numbers)

30
examples/jeopardy/jeopardy.sh
View File

@ -1,30 +0,0 @@
#!/usr/bin/env bash
set -e
MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin
MODEL_NAME=Vicuna
# exec options
prefix="Human: " # Ex. Vicuna uses "Human: "
opts="--temp 0 -n 80" # additional flags
nl='
'
introduction="You will be playing a game of Jeopardy. Simply answer the question in the correct format (Ex. What is Paris, or Who is George Washington)."
# file options
question_file=./examples/jeopardy/questions.txt
touch ./examples/jeopardy/results/$MODEL_NAME.txt
output_file=./examples/jeopardy/results/$MODEL_NAME.txt
counter=1
echo 'Running'
while IFS= read -r question
do
exe_cmd="./llama-cli -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
echo $counter
echo "Current Question: $question"
eval "$exe_cmd"
echo -e "\n------" >> $output_file
counter=$((counter+1))
done < "$question_file"

103
examples/jeopardy/qasheet.csv
View File

@ -1,103 +0,0 @@
Index,Original Category,Original Correct Question,Model Prompt
1,The Oscars,Who is John Williams?,Which actor Born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars?
2,English Literature,What is Paradise Lost?,"What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?"
3,Writers’ Lesser-Known Works,Who is Niccolò Machiavelli?,"Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?"
4,Exploration,What is Easter Island (Rapa Nui)?,"James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?"
5,The Bill of Rights,What is the Eighth Amendment?,England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution?
6,Nobel Peace Prize Winners,Who are Nelson Mandela & Desmond Tutu?,"Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?"
7,Famous Names,Who is Walt Disney?,"In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?"
8,Geography,What is Colombia?,"Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?"
9,Fashion History,What are rhinestones?,"Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?"
10,Movies of the ’80s,What is Driving Miss Daisy?,What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated?
11,Novelists,Who is John Grisham?,"A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?"
12,20th Century Eponyms,What is the Maginot Line?,"A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?"
13,City History,What is Stockholm?,"Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?"
14,Brand Names,What is Jacuzzi?,"The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?"
15,American Authors,Who is Washington Irving?,"In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?"
16,Symbols,What is “less than”?,What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society?
17,Movie Theme Songs,Who is James Bond?,"Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?"
18,American Novelists,Who is Joseph Heller?,"What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?"
19,Medieval Places,"What is Canterbury, England? (Canterbury Cathedral)","In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?"
20,Countries of Africa,What is Morocco?,"At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?"
21,Statehood,What is Wyoming?,Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women?
22,1980s Movies,What is Raiders of the Lost Ark?,"A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?"
23,Art Exhibitions,Who is Rembrandt?,In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation?
24,Countries of the World,What is Mongolia?,"Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?"
25,Literature,What is “Howl”?,A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'?
26,Invasions,Who is William of Orange?,"Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?"
27,Landmarks,What is the Eiffel Tower?,"After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?"
28,Geographic Name’s the Same,What is Dover?,"The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?"
29,Names in the Bookstore,Who is Peter Mark Roget?,"This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?"
30,U.S. History,Who is Dr. Samuel Mudd?,"An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?"
31,American Literature,What is The Things They Carried?,"Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?"
32,Nonfiction,What is The Communist Manifesto,"What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?"
33, a new version was passed 81 years later,Laws in U.S. History,What is the Civil Rights Act?,,,,,,,,,,,,,,,,,,0, 2/3
34,Names of Myth,Who is Helen of Troy?,"Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?"
35,African Countries,What is Sudan?,"Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?"
36,The Ancient World,What is Alexandria?,"The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?"
37,Famous Names,Who is Andy Warhol?,"For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?"
38,People & Places,What is Guam?,"Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?"
39,Current World Leaders,What is the Philippines?,"In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?"
40,Writers & The South,Who is Tennessee Williams?,In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South?
41,National Parks,What is Yellowstone?,"What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?"
42,Sports,Who are the Harlem Globetrotters?,"In 2010 who introduced the 4-point shot, 35 feet from the basket?"
43,The U.S. Military,What is “Top Gun”?,Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969?
44,Art & Science,What is Halley’s Comet?,"A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?"
45,Words From World War I,What is “tank”?,"In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?"
46,European History,What is Holy Roman Emperor?,"Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?"
47,Theater History,Who is Peter Pan?,"In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?"
48,European Cities,What is Aachen?,"Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?"
49,Word Origins,What is mantra?,This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'?
50,Inventions,What is barbed wire?,1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'?
51,World War II,What is Schindler’s list?,"Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?"
52, their offspring was the source of this mythical object,Mythology,What is the Golden Fleece?
53,Literature,What is Pride and Prejudice?,"Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?"
54, only these 2 west of the Mississippi River border each other,U.S. State Names,What are Oregon & Nevada?
55,Word Origins,What is passion?,"Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?"
56,World Cinema,What is La Vie en Rose?,"The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?"
57,History,What is Santa Maria?,"Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?"
58,Landmarks,What is a kremlin?,Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what?
59,Foreign-Born Authors,Who is Vladimir Nabokov?,In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'?
60,Astronomy & Geography,What is Capricorn?,"At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?"
61,Television,What is Law & Order?,"Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?"
62,British Landmarks,What is the Tower of London?,"Like Sir Thomas More, 3 16th century English queens are buried at what British location?"
63,Early American History,What are witches?,"In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person … be condemned'?"
64,Geography Mnemonics,What are Arkansas and Louisiana?,"The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?"
65,Business Milestones,What is the Ford Model T?,"What was first sold in 1908, at a price equivalent to about $27,000 today?"
66,In The Bookstore,Who is Tom Clancy?,The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot?
67,Historic Art,What is the Bayeux Tapestry?,The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what?
68,Pop Stars,Who is Madonna?,In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s?
69,Classic Tale Characters,Who is Scheherazade?,"In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?"
70,USA,What is Jack Daniel’s?,"Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?"
71,Historic People,Who was William Bligh?,"After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?"
72,The Movies,What is The Godfather?,Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022?
73,Continental Geography,What is Colombia?,"Until a 1903 secession, what country's contiguous territory spanned 2 continents?"
74,Foreign-Born Authors,Who is Isabel Allende?,"Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?"
75,Historic Crimes,What is the Mona Lisa?,"Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?"
76,U.S. Bodies of Water,What is Lake Mead?,"Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?"
77,Gods & Goddesses,Who is Aurora (or Eos)?,"Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?"
78,America At War,What is the Battle of New Orleans?,"Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?"
79,Children’s Books,What is The Velveteen Rabbit?,"Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?"
80,TV Finales,What is Grace and Frankie?,"In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?"
81,American Poems,Who is Evangeline?,"In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?"
82,Famous Names,Who is Banksy?,"In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?"
83,Children’s Lit,What is Charlotte’s Web?,The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'?
84,Classic Songs,What is “Here Comes Santa Claus”?,The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite?
85,Brand Names,What are Milk Duds?,"Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?"
86,Countries of the World,What is Italy?,"What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?"
87,Action Movies,What is Die Hard?,"What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?"
88,Presidential Facts,Who is Woodrow Wilson?,Only 3 presidents have married while in office— John Tyler was the first & which one was the last?
89,19th Century Americans,Who is Frederick Douglass?,"Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?"
90,Latin Phrases,What is “quid pro quo”?,"Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?"
91,1970s Movies,What is Monty Python and the Holy Grail?,The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience?
92,Name’s The Same,What is Manhattan?,"A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?"
93,U.S. Presidents,Who is Calvin Coolidge?,"Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?"
94,Plays,What is The Tempest?,A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play?
95,Landmarks,What is the Berlin Wall?,"In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?"
96,World Capitals,"What is Vienna, Austria?","Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?"
97,Language & Its Meanings,What is a night owl?,"Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?"
98,Flags of Our Hemisphere,What is Brazil?,"The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?"
99,Names in U.S. History,Who is Oliver Brown?,What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951?
100,Children’s Authors,"Who is Sarah? (from Sarah, Plain and Tall)","Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?"
,,,
TOTALS,,,
Can't render this file because it has a wrong number of fields in line 34.

100
examples/jeopardy/questions.txt
View File

@ -1,100 +0,0 @@
Which man born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars?
What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?
Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?
James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?
England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution?
Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?
In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?
Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?
Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?
What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated?
A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?
A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?
Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?
The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?
In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?
What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society?
Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?
What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?
In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?
At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?
Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women?
A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?
In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation?
Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?
A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'?
Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?
After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?
The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?
This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?
An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?
Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?
What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?
A radical Republican championed what 1875 act but the Supreme Court struck it down in 1883; a new version was passed 81 years later?
Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?
Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?
The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?
For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?
Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?
In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?
In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South?
What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?
In 2010 who introduced the 4-point shot, 35 feet from the basket?
Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969?
A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?
In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?
Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?
In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?
Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?
This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'?
1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'?
Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?
Poseidon carried off the maiden Theophane & turned her into a ewe; their offspring was the source of what mythical object?
Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?
5 U.S. states have 6-letter names; only which 2 west of the Mississippi River border each other?
Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?
The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?
Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?
Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what?
In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'?
At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?
Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?
Like Sir Thomas More, 3 16th century English queens are buried at what British location?
In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person be condemned'?
The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?
What was first sold in 1908, at a price equivalent to about $27,000 today?
The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot?
The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what?
In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s?
In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?
Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?
After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?
Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022?
Until a 1903 secession, what country's contiguous territory spanned 2 continents?
Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?
Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?
Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?
Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?
Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?
Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?
In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?
In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?
In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?
The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'?
The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite?
Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?
What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?
What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?
Only 3 presidents have married while in office— John Tyler was the first & which one was the last?
Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?
Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?
The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience?
A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?
Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?
A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play?
In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?
Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?
Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?
The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?
What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951?
Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?

28
examples/llm.vim
View File

@ -1,28 +0,0 @@
" Basic plugin example
function! Llm()
let url = "http://127.0.0.1:8080/completion"
" Get the content of the current buffer
let buffer_content = join(getline(1, '$'), "\n")
" Create the JSON payload
let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":256,"stop": ["\n\n\n"],"stream": v:false}
let json_payload.prompt = buffer_content
" Define the curl command
let curl_command = 'curl -k -s -X POST -H "Content-Type: application/json" -d @- ' . url
let response = system(curl_command, json_encode(json_payload))
" Extract the content field from the response
let content = json_decode(response).content
let split_newlines = split(content, '\n', 1)
" Insert the content at the cursor position
call setline(line('.'), [ getline('.') . split_newlines[0] ] + split_newlines[1:])
endfunction
command! Llm call Llm()
noremap <F2> :Llm<CR>

13
examples/model-conversion/Makefile
View File

@ -118,13 +118,17 @@ embedding-convert-model:
embedding-run-original-model:
$(call validate_embedding_model_path,embedding-run-original-model)
@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/embedding/run-original-model.py
@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
./scripts/embedding/run-original-model.py \
$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
embedding-run-converted-model:
@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/embedding/run-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
@./scripts/embedding/run-converted-model.sh $(CONVERTED_EMBEDDING_MODEL) \
$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
@./scripts/embedding/compare-embeddings-logits.sh
@./scripts/embedding/compare-embeddings-logits.sh \
$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
embedding-inspect-original-model:
$(call validate_embedding_model_path,embedding-inspect-original-model)
@ -156,7 +160,8 @@ embedding-quantize-model:
$(call quantize_model,$(CONVERTED_EMBEDDING_MODEL),QUANTIZED_EMBEDDING_MODEL)
embedding-run-quantized-model:
@./scripts/embedding/run-converted-model.sh ${QUANTIZED_EMBEDDING_MODEL}
@./scripts/embedding/run-converted-model.sh $(QUANTIZED_EMBEDDING_MODEL) \
$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
###
### Perplexity targets/recipes

4
examples/model-conversion/README.md
View File

@ -105,12 +105,12 @@ new model, the model can be converted to GGUF format using the following command
### Inspecting the converted model
The converted model can be inspected using the following command:
```console
(venv) $ make inspect-converted-model
(venv) $ make causal-inspect-converted-model
```
### Running the converted model
```console
(venv) $ make run-converted-model
(venv) $ make causal-run-converted-model
```
### Model logits verfication

52
examples/model-conversion/logits.cpp
View File

@ -151,6 +151,35 @@ int main(int argc, char ** argv) {
logits = llama_get_embeddings(ctx);
n_logits = llama_model_n_embd(model) * batch.n_tokens;
type = "-embeddings";
const int n_embd = llama_model_n_embd(model);
const int n_embd_count = batch.n_tokens;
printf("Embedding dimension: %d\n", n_embd);
printf("\n");
// Print embeddings in the specified format
for (int j = 0; j < n_embd_count; j++) {
printf("embedding %d: ", j);
// Print first 3 values
for (int i = 0; i < 3 && i < n_embd; i++) {
printf("%9.6f ", logits[j * n_embd + i]);
}
printf(" ... ");
// Print last 3 values
for (int i = n_embd - 3; i < n_embd; i++) {
if (i >= 0) {
printf("%9.6f ", logits[j * n_embd + i]);
}
}
printf("\n");
}
printf("\n");
printf("Embeddings size: %d\n", n_logits);
} else {
logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
@ -183,22 +212,23 @@ int main(int argc, char ** argv) {
return 1;
}
for (int i = 0; i < n_logits; i++) {
fprintf(f, "%d: %.6f\n", i, logits[i]); // Added index and changed format
fprintf(f, "%d: %.6f\n", i, logits[i]);
}
fclose(f);
// Print first and last 10 logits for quick verification
printf("First 10 logits: ");
for (int i = 0; i < 10 && i < n_logits; i++) {
printf("%.6f ", logits[i]);
}
printf("\n");
if (!embedding_mode) {
printf("First 10 logits: ");
for (int i = 0; i < 10 && i < n_logits; i++) {
printf("%.6f ", logits[i]);
}
printf("\n");
printf("Last 10 logits: ");
for (int i = n_logits - 10; i < n_logits; i++) {
if (i >= 0) printf("%.6f ", logits[i]);
printf("Last 10 logits: ");
for (int i = n_logits - 10; i < n_logits; i++) {
if (i >= 0) printf("%.6f ", logits[i]);
}
printf("\n\n");
}
printf("\n\n");
printf("Logits saved to %s\n", bin_filename);
printf("Logits saved to %s\n", txt_filename);

2
examples/model-conversion/scripts/causal/compare-logits.py
View File

@ -48,7 +48,7 @@ def main():
print(f"Error: Model file not found: {model_path}")
sys.exit(1)
model_name = os.path.splitext(os.path.basename(model_path))[0]
model_name = os.path.basename(model_path)
data_dir = Path("data")
pytorch_file = data_dir / f"pytorch-{model_name}.bin"

2
examples/model-conversion/scripts/causal/run-org-model.py
View File

@ -193,7 +193,7 @@ print(f"Input text: {repr(prompt)}")
print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
with torch.no_grad():
outputs = model(input_ids)
outputs = model(input_ids.to(model.device))
logits = outputs.logits
# Extract logits for the last token (next token prediction)

49
examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh
View File

@ -2,8 +2,37 @@
set -e
MODEL_PATH="${1:-"$EMBEDDING_MODEL_PATH"}"
MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
# Parse command line arguments
MODEL_PATH=""
MODEL_NAME=""
PROMPTS_FILE=""
# First argument is always model path
if [ $# -gt 0 ] && [[ "$1" != --* ]]; then
MODEL_PATH="$1"
shift
fi
# Parse remaining arguments
while [[ $# -gt 0 ]]; do
case $1 in
--prompts-file|-pf)
PROMPTS_FILE="$2"
shift 2
;;
*)
# If MODEL_NAME not set and this isn't a flag, use as model name
if [ -z "$MODEL_NAME" ] && [[ "$1" != --* ]]; then
MODEL_NAME="$1"
fi
shift
;;
esac
done
# Set defaults
MODEL_PATH="${MODEL_PATH:-"$EMBEDDING_MODEL_PATH"}"
MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
if [ -t 0 ]; then
CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
@ -35,8 +64,18 @@ with open('$TEMP_FILE', 'wb') as f:
trap "rm -f $TEMP_FILE" EXIT
fi
python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
# Build the semantic_check.py command
SEMANTIC_CMD="python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
--python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
--cpp-embeddings $CPP_EMBEDDINGS \
--prompt "Hello world today"
--cpp-embeddings $CPP_EMBEDDINGS"
# Add prompts file if specified, otherwise use default prompt
if [ -n "$PROMPTS_FILE" ]; then
SEMANTIC_CMD="$SEMANTIC_CMD --prompts-file \"$PROMPTS_FILE\""
else
SEMANTIC_CMD="$SEMANTIC_CMD --prompt \"Hello world today\""
fi
# Execute the command
eval $SEMANTIC_CMD

38
examples/model-conversion/scripts/embedding/run-converted-model.sh
View File

@ -2,8 +2,27 @@
set -e
# First try command line argument, then environment variable, then file
CONVERTED_MODEL="${1:-"$CONVERTED_EMBEDDING_MODEL"}"
# Parse command line arguments
CONVERTED_MODEL=""
PROMPTS_FILE=""
while [[ $# -gt 0 ]]; do
case $1 in
-p|--prompts-file)
PROMPTS_FILE="$2"
shift 2
;;
*)
if [ -z "$CONVERTED_MODEL" ]; then
CONVERTED_MODEL="$1"
fi
shift
;;
esac
done
# First try command line argument, then environment variable
CONVERTED_MODEL="${CONVERTED_MODEL:-"$CONVERTED_EMBEDDING_MODEL"}"
# Final check if we have a model path
if [ -z "$CONVERTED_MODEL" ]; then
@ -13,8 +32,19 @@ if [ -z "$CONVERTED_MODEL" ]; then
exit 1
fi
# Read prompt from file or use default
if [ -n "$PROMPTS_FILE" ]; then
if [ ! -f "$PROMPTS_FILE" ]; then
echo "Error: Prompts file '$PROMPTS_FILE' not found" >&2
exit 1
fi
PROMPT=$(cat "$PROMPTS_FILE")
else
PROMPT="Hello world today"
fi
echo $CONVERTED_MODEL
cmake --build ../../build --target llama-logits -j8
../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "Hello world today"
# TODO: update logits.cpp to accept a --file/-f option for the prompt
../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"

42
examples/model-conversion/scripts/embedding/run-original-model.py
View File

@ -13,14 +13,37 @@ unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
parser = argparse.ArgumentParser(description='Process model with specified path')
parser.add_argument('--model-path', '-m', help='Path to the model')
parser.add_argument('--prompts-file', '-p', help='Path to file containing prompts (one per line)')
args = parser.parse_args()
def read_prompt_from_file(file_path):
try:
with open(file_path, 'r', encoding='utf-8') as f:
return f.read().strip()
except FileNotFoundError:
print(f"Error: Prompts file '{file_path}' not found")
exit(1)
except Exception as e:
print(f"Error reading prompts file: {e}")
exit(1)
model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
if model_path is None:
parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
tokenizer = AutoTokenizer.from_pretrained(model_path)
config = AutoConfig.from_pretrained(model_path)
# This can be used to override the sliding window size for manual testing. This
# can be useful to verify the sliding window attention mask in the original model
# and compare it with the converted .gguf model.
if hasattr(config, 'sliding_window'):
original_sliding_window = config.sliding_window
#original_sliding_window = 6
print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
print(f"Using unreleased model: {unreleased_model_name}")
if unreleased_model_name:
model_name_lower = unreleased_model_name.lower()
unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
@ -29,19 +52,28 @@ if unreleased_model_name:
try:
model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
model = model_class.from_pretrained(model_path) # Note: from_pretrained, not fromPretrained
model = model_class.from_pretrained(model_path, config=config)
except (ImportError, AttributeError) as e:
print(f"Failed to import or load model: {e}")
exit(1)
else:
model = AutoModel.from_pretrained(model_path)
model = AutoModel.from_pretrained(model_path, config=config)
print(f"Model class: {type(model)}")
#print(f"Model file: {type(model).__module__}")
config = AutoConfig.from_pretrained(model_path)
print(f"Model file: {type(model).__module__}")
# Verify the model is using the correct sliding window
if hasattr(model.config, 'sliding_window'):
print(f"Model's sliding_window: {model.config.sliding_window}")
else:
print("Model config does not have sliding_window attribute")
model_name = os.path.basename(model_path)
texts = [ "Hello world today" ]
if args.prompts_file:
prompt_text = read_prompt_from_file(args.prompts_file)
texts = [prompt_text]
else:
texts = ["Hello world today"]
encoded = tokenizer(
texts,

2
examples/model-conversion/scripts/utils/check-nmse.py
View File

@ -67,7 +67,7 @@ def main():
parser.add_argument('-m', '--model-path', required=True, help='Path to the model directory')
args = parser.parse_args()
model_name = os.path.splitext(os.path.basename(args.model_path))[0]
model_name = os.path.basename(args.model_path)
data_dir = Path("data")
pytorch_file = data_dir / f"pytorch-{model_name}.bin"

4
examples/model-conversion/scripts/utils/inspect-org-model.py
View File

@ -40,7 +40,7 @@ if os.path.exists(index_path):
file_path = os.path.join(model_path, file_name)
print(f"\n--- From {file_name} ---")
with safe_open(file_path, framework="pt") as f: # type: ignore
with safe_open(file_path, framework="pt") as f:
for tensor_name in sorted(tensor_names):
tensor = f.get_tensor(tensor_name)
print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
@ -49,7 +49,7 @@ elif os.path.exists(single_file_path):
# Single file model (original behavior)
print("Single-file model detected")
with safe_open(single_file_path, framework="pt") as f: # type: ignore
with safe_open(single_file_path, framework="pt") as f:
keys = f.keys()
print("Tensors in model:")
for key in sorted(keys):

23
examples/model-conversion/scripts/utils/semantic_check.py
View File

@ -101,6 +101,17 @@ def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
'rms_diff': np.sqrt(np.mean(diff_matrix**2))
}
def read_prompt_from_file(file_path):
try:
with open(file_path, 'r', encoding='utf-8') as f:
return f.read().strip()
except FileNotFoundError:
print(f"Error: Prompts file '{file_path}' not found")
exit(1)
except Exception as e:
print(f"Error reading prompts file: {e}")
exit(1)
def main():
parser = argparse.ArgumentParser(description='Test semantic similarity between Python and llama.cpp embeddings')
parser.add_argument('--model-path', '-m', required=True, help='Path to the original Python model')
@ -108,14 +119,20 @@ def main():
parser.add_argument('--cpp-embeddings', '-ce', help='Path to llama.cpp embeddings "logits" binary file')
parser.add_argument('--causal', '-c', default=False, help='if the model is causal (default: false)', action='store_true')
parser.add_argument('--prompt', '-p', default='Hello world today', help='Test prompt')
parser.add_argument('--prompts-file', '-pf', help='Path to file containing prompts')
args = parser.parse_args()
if args.prompts_file:
prompt = read_prompt_from_file(args.prompts_file)
else:
prompt = args.prompt
print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
print("=" * 70)
# Single prompt detailed comparison
print(f"\nTesting with prompt: '{args.prompt}'")
print(f"\nTesting with prompt: '{prompt}'")
# Load the python model to get configuration information and also to load the tokenizer.
print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
@ -144,7 +161,7 @@ def main():
else:
model = AutoModel.from_pretrained(args.model_path)
encoded = tokenizer(args.prompt, return_tensors="pt")
encoded = tokenizer(prompt, return_tensors="pt")
tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
n_tokens = len(tokens)
print(f"n_tokens: {n_tokens}");
@ -155,7 +172,7 @@ def main():
python_embeddings = load_embeddings_from_file(args.python_embeddings, n_tokens, model.config.hidden_size)
# Run comparison
results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, args.prompt)
results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, prompt)
# Summary
print(f"\n=== SUMMARY ===")

14
examples/simple/simple.cpp
View File

@ -145,6 +145,20 @@ int main(int argc, char ** argv) {
llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
if (llama_model_has_encoder(model)) {
if (llama_encode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return 1;
}
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = llama_vocab_bos(vocab);
}
batch = llama_batch_get_one(&decoder_start_token_id, 1);
}
// main loop
const auto t_main_start = ggml_time_us();

61
ggml/CMakeLists.txt
View File

@ -1,5 +1,40 @@
cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
project("ggml" C CXX ASM)
### GGML Version
set(GGML_VERSION_MAJOR 0)
set(GGML_VERSION_MINOR 9)
set(GGML_VERSION_PATCH 4)
set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
if(GIT_EXE)
# Get current git commit hash
execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
OUTPUT_VARIABLE GGML_BUILD_COMMIT
OUTPUT_STRIP_TRAILING_WHITESPACE
ERROR_QUIET
)
# Check if the working directory is dirty (i.e., has uncommitted changes)
execute_process(COMMAND ${GIT_EXE} diff-index --quiet HEAD -- .
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
RESULT_VARIABLE GGML_GIT_DIRTY
ERROR_QUIET
)
endif()
# Build the version string with optional dirty flag
set(GGML_VERSION "${GGML_VERSION_BASE}")
if(GGML_GIT_DIRTY AND NOT GGML_GIT_DIRTY EQUAL 0)
set(GGML_VERSION "${GGML_VERSION}-dirty")
endif()
if(NOT GGML_BUILD_COMMIT)
set(GGML_BUILD_COMMIT "unknown")
endif()
include(CheckIncludeFileCXX)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@ -141,7 +176,7 @@ set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
if (MINGW)
set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows version")
set(GGML_WIN_VER "0xA00" CACHE STRING "ggml: Windows version")
endif()
# ggml core
@ -174,7 +209,6 @@ option(GGML_HIP "ggml: use HIP"
option(GGML_HIP_GRAPHS "ggml: use HIP graph, experimental, slow" OFF)
option(GGML_HIP_NO_VMM "ggml: do not try to use HIP VMM" ON)
option(GGML_HIP_ROCWMMA_FATTN "ggml: enable rocWMMA for FlashAttention" OFF)
option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 "ggml: enable rocWMMA FlashAttention on GFX12" OFF)
option(GGML_HIP_MMQ_MFMA "ggml: enable MFMA MMA for CDNA in MMQ" ON)
option(GGML_HIP_EXPORT_METRICS "ggml: enable kernel perf metrics output" OFF)
option(GGML_MUSA_GRAPHS "ggml: use MUSA graph, experimental, unstable" OFF)
@ -190,7 +224,6 @@ option(GGML_WEBGPU "ggml: use WebGPU"
option(GGML_WEBGPU_DEBUG "ggml: enable WebGPU debug output" OFF)
option(GGML_ZDNN "ggml: use zDNN" OFF)
option(GGML_METAL "ggml: use Metal" ${GGML_METAL_DEFAULT})
option(GGML_METAL_USE_BF16 "ggml: use bfloat if available" OFF)
option(GGML_METAL_NDEBUG "ggml: disable Metal debugging" OFF)
option(GGML_METAL_SHADER_DEBUG "ggml: compile Metal with -fno-fast-math" OFF)
option(GGML_METAL_EMBED_LIBRARY "ggml: embed Metal library" ${GGML_METAL})
@ -303,26 +336,6 @@ endif()
# Create CMake package
#
# Generate version info based on git commit.
if(NOT DEFINED GGML_BUILD_NUMBER)
find_program(GIT_EXE NAMES git git.exe REQUIRED NO_CMAKE_FIND_ROOT_PATH)
execute_process(COMMAND ${GIT_EXE} rev-list --count HEAD
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
OUTPUT_VARIABLE GGML_BUILD_NUMBER
OUTPUT_STRIP_TRAILING_WHITESPACE
)
if(GGML_BUILD_NUMBER EQUAL 1)
message(WARNING "GGML build version fixed at 1 likely due to a shallow clone.")
endif()
execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
OUTPUT_VARIABLE GGML_BUILD_COMMIT
OUTPUT_STRIP_TRAILING_WHITESPACE
)
endif()
# Capture variables prefixed with GGML_.
@ -351,7 +364,7 @@ set(GGML_VARIABLES_EXPANDED ${variable_set_statements})
# Create the CMake package and set install location.
set(GGML_INSTALL_VERSION 0.0.${GGML_BUILD_NUMBER})
set(GGML_INSTALL_VERSION ${GGML_VERSION})
set(GGML_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header files")
set(GGML_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files")
set(GGML_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files")

5
ggml/include/ggml-backend.h
View File

@ -215,6 +215,8 @@ extern "C" {
// Backend registry
//
GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
// Backend (reg) enumeration
@ -314,7 +316,8 @@ extern "C" {
GGML_API int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
GGML_API int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
GGML_API size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend);
GGML_API size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
GGML_API void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
GGML_API ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);

1
ggml/include/ggml-metal.h
View File

@ -39,6 +39,7 @@ extern "C" {
// user-code should use only these functions
//
// TODO: remove in the future
GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);

17
ggml/include/ggml-rpc.h
View File

@ -7,26 +7,25 @@
extern "C" {
#endif
#define RPC_PROTO_MAJOR_VERSION 2
#define RPC_PROTO_MAJOR_VERSION 3
#define RPC_PROTO_MINOR_VERSION 0
#define RPC_PROTO_PATCH_VERSION 0
#define GGML_RPC_MAX_SERVERS 16
// backend API
GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device);
GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, uint32_t device);
GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device, size_t * free, size_t * total);
GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
const char * cache_dir,
size_t free_mem, size_t total_mem);
GGML_BACKEND_API void ggml_backend_rpc_start_server(const char * endpoint, const char * cache_dir,
size_t n_threads, size_t n_devices,
ggml_backend_dev_t * devices, size_t * free_mem, size_t * total_mem);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_add_server(const char * endpoint);
#ifdef __cplusplus
}

3
ggml/include/ggml-zdnn.h
View File

@ -7,6 +7,9 @@
extern "C" {
#endif
// device buffer
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_type(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);
#ifdef __cplusplus

30
ggml/include/ggml.h
View File

@ -237,6 +237,8 @@
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
// TODO: convert to enum https://github.com/ggml-org/llama.cpp/pull/16187#discussion_r2388538726
#define GGML_ROPE_TYPE_NORMAL 0
#define GGML_ROPE_TYPE_NEOX 2
#define GGML_ROPE_TYPE_MROPE 8
#define GGML_ROPE_TYPE_VISION 24
@ -284,19 +286,19 @@ __host__ __device__ constexpr inline void ggml_unused_vars_impl(Args&&...) noexc
// GGML_TENSOR_LOCALS(size_t, nb1, src1, nb);
//
#define GGML_TENSOR_LOCALS_1(type, prefix, pointer, array) \
const type prefix##0 = (pointer)->array[0]; \
const type prefix##0 = (pointer) ? (pointer)->array[0] : 0; \
GGML_UNUSED(prefix##0);
#define GGML_TENSOR_LOCALS_2(type, prefix, pointer, array) \
GGML_TENSOR_LOCALS_1 (type, prefix, pointer, array) \
const type prefix##1 = (pointer)->array[1]; \
const type prefix##1 = (pointer) ? (pointer)->array[1] : 0; \
GGML_UNUSED(prefix##1);
#define GGML_TENSOR_LOCALS_3(type, prefix, pointer, array) \
GGML_TENSOR_LOCALS_2 (type, prefix, pointer, array) \
const type prefix##2 = (pointer)->array[2]; \
const type prefix##2 = (pointer) ? (pointer)->array[2] : 0; \
GGML_UNUSED(prefix##2);
#define GGML_TENSOR_LOCALS(type, prefix, pointer, array) \
GGML_TENSOR_LOCALS_3 (type, prefix, pointer, array) \
const type prefix##3 = (pointer)->array[3]; \
const type prefix##3 = (pointer) ? (pointer)->array[3] : 0; \
GGML_UNUSED(prefix##3);
#define GGML_TENSOR_UNARY_OP_LOCALS \
@ -574,6 +576,7 @@ extern "C" {
GGML_UNARY_OP_HARDSIGMOID,
GGML_UNARY_OP_EXP,
GGML_UNARY_OP_GELU_ERF,
GGML_UNARY_OP_XIELU,
GGML_UNARY_OP_COUNT,
};
@ -1148,6 +1151,18 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// xIELU activation function
// x = x * (c_a(alpha_n) + c_b(alpha_p, beta) * sigmoid(beta * x)) + eps * (x > 0)
// where c_a = softplus and c_b(a, b) = softplus(a) + b are constraining functions
// that constrain the positive and negative source alpha values respectively
GGML_API struct ggml_tensor * ggml_xielu(
struct ggml_context * ctx,
struct ggml_tensor * a,
float alpha_n,
float alpha_p,
float beta,
float eps);
// gated linear unit ops
// A: n columns, r rows,
// result is n / 2 columns, r rows,
@ -1615,6 +1630,13 @@ extern "C" {
float scale,
float max_bias);
GGML_API struct ggml_tensor * ggml_soft_max_ext_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
float scale,
float max_bias);
GGML_API void ggml_soft_max_add_sinks(
struct ggml_tensor * a,
struct ggml_tensor * sinks);

3
ggml/src/CMakeLists.txt
View File

@ -114,6 +114,9 @@ message(STATUS "GGML_SYSTEM_ARCH: ${GGML_SYSTEM_ARCH}")
if (NOT MSVC)
if (GGML_STATIC)
if (UNIX AND NOT APPLE)
set(CMAKE_FIND_LIBRARY_SUFFIXES ".a;.so")
endif()
add_link_options(-static)
if (MINGW)
add_link_options(-static-libgcc -static-libstdc++)

430
ggml/src/ggml-alloc.c
View File

@ -23,7 +23,7 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
}
// ops that return true for this function must not use restrict pointers for their backend implementations
static bool ggml_op_can_inplace(enum ggml_op op) {
bool ggml_op_can_inplace(enum ggml_op op) {
switch (op) {
case GGML_OP_SCALE:
case GGML_OP_DIAG_MASK_ZERO:
@ -95,39 +95,104 @@ enum ggml_status ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_te
// dynamic tensor allocator
#define GGML_VBUFFER_MAX_CHUNKS 16
// relative memory address within an allocation that can be split into multiple buffers (chunks)
struct buffer_address {
int chunk; // index of a backend buffer
size_t offset; // local memory offset within the buffer
};
static const struct buffer_address GGML_BUFFER_ADDRESS_INVALID = { -1, SIZE_MAX };
static bool ggml_buffer_address_less(struct buffer_address a, struct buffer_address b) {
return a.chunk != b.chunk ? a.chunk < b.chunk : a.offset < b.offset;
}
struct free_block {
size_t offset;
size_t size;
};
struct tallocr_chunk {
struct free_block free_blocks[MAX_FREE_BLOCKS];
int n_free_blocks;
size_t max_size;
};
struct ggml_dyn_tallocr {
size_t alignment;
int n_free_blocks;
struct free_block free_blocks[MAX_FREE_BLOCKS];
size_t max_size;
size_t max_chunk_size;
struct tallocr_chunk * chunks[GGML_VBUFFER_MAX_CHUNKS];
int n_chunks;
#ifdef GGML_ALLOCATOR_DEBUG
struct {
const struct ggml_tensor * tensor;
size_t offset;
struct buffer_address addr;
} allocated_tensors[1024];
#endif
};
static void ggml_dyn_tallocr_insert_block(struct tallocr_chunk * chunk, size_t offset, size_t size) {
GGML_ASSERT(chunk->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
// insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
int insert_pos = 0;
while (insert_pos < chunk->n_free_blocks && chunk->free_blocks[insert_pos].offset < offset) {
insert_pos++;
}
// shift all blocks from insert_pos onward to make room for the new block
for (int i = chunk->n_free_blocks; i > insert_pos; i--) {
chunk->free_blocks[i] = chunk->free_blocks[i-1];
}
// insert the new block
chunk->free_blocks[insert_pos].offset = offset;
chunk->free_blocks[insert_pos].size = size;
chunk->n_free_blocks++;
}
static void ggml_dyn_tallocr_remove_block(struct tallocr_chunk * chunk, int idx) {
// shift all elements after idx by 1 to the left, overwriting the element at idx
for (int i = idx; i < chunk->n_free_blocks; i++) {
chunk->free_blocks[i] = chunk->free_blocks[i+1];
}
chunk->n_free_blocks--;
}
static int ggml_dyn_tallocr_new_chunk(struct ggml_dyn_tallocr * alloc, size_t min_size) {
if (alloc->n_chunks >= GGML_VBUFFER_MAX_CHUNKS) {
return -1;
}
struct tallocr_chunk * chunk = calloc(1, sizeof(struct tallocr_chunk));
chunk->n_free_blocks = 1;
chunk->free_blocks[0].offset = 0;
// available space in a chunk is limited to max_chunk_size, but can be higher if:
// 1. a single tensor exceeds the maximum, and cannot fit any other way
// 2. we are running out of chunks
// backends will either manage to allocate the larger size, or report an error.
chunk->free_blocks[0].size = MAX(min_size, alloc->max_chunk_size);
if (alloc->n_chunks == GGML_VBUFFER_MAX_CHUNKS - 1) {
chunk->free_blocks[0].size = SIZE_MAX/2;
}
alloc->chunks[alloc->n_chunks] = chunk;
alloc->n_chunks++;
return alloc->n_chunks - 1;
}
#ifdef GGML_ALLOCATOR_DEBUG
static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, const struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i].tensor == NULL) {
alloc->allocated_tensors[i].tensor = tensor;
alloc->allocated_tensors[i].offset = offset;
alloc->allocated_tensors[i].addr = addr;
return;
}
}
GGML_ABORT("out of allocated_tensors");
}
static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, const struct ggml_tensor * tensor) {
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i].offset == offset) {
if (alloc->allocated_tensors[i].addr.chunk == addr.chunk && alloc->allocated_tensors[i].addr.offset == addr.offset) {
alloc->allocated_tensors[i].tensor = NULL;
return;
}
@ -136,76 +201,94 @@ static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offs
}
#endif
static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
int best_fit_chunk = -1;
int best_fit_block = -1;
size_t max_avail = 0;
// find the best fitting free block besides the last block
int best_fit_block = -1;
size_t best_fit_size = SIZE_MAX;
for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
struct free_block * block = &alloc->free_blocks[i];
max_avail = MAX(max_avail, block->size);
if (block->size >= size && block->size <= best_fit_size) {
best_fit_block = i;
best_fit_size = block->size;
// find the best fitting free block besides the last block, within any chunk
for (int c = 0; c < alloc->n_chunks; ++c) {
struct tallocr_chunk * chunk = alloc->chunks[c];
size_t best_fit_size = SIZE_MAX;
for (int i = 0; i < chunk->n_free_blocks - 1; i++) {
struct free_block * block = &chunk->free_blocks[i];
max_avail = MAX(max_avail, block->size);
if (block->size >= size && block->size <= best_fit_size) {
best_fit_chunk = c;
best_fit_block = i;
best_fit_size = block->size;
}
}
}
if (best_fit_block == -1) {
// the last block is our last resort
struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
max_avail = MAX(max_avail, block->size);
if (block->size >= size) {
best_fit_block = alloc->n_free_blocks - 1;
} else {
// this should never happen
GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
__func__, size, max_avail);
GGML_ABORT("not enough space in the buffer");
}
}
struct free_block * block = &alloc->free_blocks[best_fit_block];
size_t offset = block->offset;
block->offset = offset + size;
block->size -= size;
if (block->size == 0) {
// remove block if empty
alloc->n_free_blocks--;
for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
}
}
AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
#ifdef GGML_ALLOCATOR_DEBUG
add_allocated_tensor(alloc, offset, tensor);
size_t cur_max = offset + size;
if (cur_max > alloc->max_size) {
// sort allocated_tensors by offset
for (int i = 0; i < 1024; i++) {
for (int j = i + 1; j < 1024; j++) {
if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
size_t tmp_offset = alloc->allocated_tensors[i].offset;
alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
alloc->allocated_tensors[j].tensor = tmp_tensor;
alloc->allocated_tensors[j].offset = tmp_offset;
// no suitable block found, try the last block (this will grow a chunks size)
for (int c = 0; c < alloc->n_chunks; ++c) {
struct tallocr_chunk * chunk = alloc->chunks[c];
if (chunk->n_free_blocks > 0) {
struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1];
max_avail = MAX(max_avail, block->size);
if (block->size >= size) {
best_fit_chunk = c;
best_fit_block = chunk->n_free_blocks - 1;
break;
}
}
}
GGML_LOG_DEBUG("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
}
if (best_fit_block == -1) {
// none of the existing chunks have enough space left
best_fit_chunk = ggml_dyn_tallocr_new_chunk(alloc, size);
best_fit_block = 0;
}
if (best_fit_chunk == -1) {
// since the last chunk always has virtually endless memory, this should never happen
GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
__func__, size, max_avail);
GGML_ABORT("graph allocation: failed to reserve memory");
}
struct tallocr_chunk * chunk = alloc->chunks[best_fit_chunk];
struct free_block * block = &chunk->free_blocks[best_fit_block];
struct buffer_address addr = {.chunk = best_fit_chunk, .offset = block->offset };
block->offset += size;
block->size -= size;
if (block->size == 0) {
// remove block if empty
ggml_dyn_tallocr_remove_block(chunk, best_fit_block);
}
AT_PRINTF("block %d, offset %zu, chunk %d\n", best_fit_block, addr.offset, addr.chunk);
#ifdef GGML_ALLOCATOR_DEBUG
add_allocated_tensor(alloc, addr, tensor);
size_t cur_max = addr.offset + size;
if (cur_max > alloc->max_size[addr.chunk]) {
// sort allocated_tensors by chunk/offset
for (int i = 0; i < 1024; i++) {
for (int j = i + 1; j < 1024; j++) {
if (ggml_buffer_address_less(alloc->allocated_tensors[j].addr, alloc->allocated_tensors[i].addr)) {
const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
struct buffer_address tmp_addr = alloc->allocated_tensors[i].addr;
alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
alloc->allocated_tensors[i].addr = alloc->allocated_tensors[j].addr;
alloc->allocated_tensors[j].tensor = tmp_tensor;
alloc->allocated_tensors[j].addr = tmp_addr;
}
}
}
GGML_LOG_DEBUG("max_size[%d] = %.2f MB: tensors: ", addr.chunk, cur_max / 1024.0 / 1024.0);
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i].tensor) {
GGML_LOG_DEBUG("%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
alloc->allocated_tensors[i].offset,
alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
GGML_LOG_DEBUG("%s [%d: %zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
alloc->allocated_tensors[i].addr.chunk,
alloc->allocated_tensors[i].addr.offset,
alloc->allocated_tensors[i].addr.offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
}
}
@ -213,78 +296,69 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
}
#endif
alloc->max_size = MAX(alloc->max_size, offset + size);
chunk->max_size = MAX(chunk->max_size, addr.offset + size);
return offset;
return addr;
GGML_UNUSED(tensor);
}
// this is a very naive implementation, but for our case the number of free blocks should be very small
static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct ggml_tensor * tensor) {
static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, size_t size, const struct ggml_tensor * tensor) {
size = aligned_offset(NULL, size, alloc->alignment);
AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
AT_PRINTF("%s: freeing %s at {chunk=%d, offset=%zu} (%zu bytes) - n_free_blocks = %d\n",
__func__, tensor->name, addr.chunk, addr.offset, size, alloc->chunks[addr.chunk]->n_free_blocks);
#ifdef GGML_ALLOCATOR_DEBUG
remove_allocated_tensor(alloc, offset, tensor);
remove_allocated_tensor(alloc, addr, tensor);
#endif
struct tallocr_chunk * chunk = alloc->chunks[addr.chunk];
// see if we can merge with an existing block
for (int i = 0; i < alloc->n_free_blocks; i++) {
struct free_block * block = &alloc->free_blocks[i];
for (int i = 0; i < chunk->n_free_blocks; i++) {
struct free_block * block = &chunk->free_blocks[i];
// check if ptr is at the end of the block
if (block->offset + block->size == offset) {
if (block->offset + block->size == addr.offset) {
block->size += size;
// check if we can merge with the next block
if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
block->size += alloc->free_blocks[i+1].size;
alloc->n_free_blocks--;
for (int j = i+1; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
if (i < chunk->n_free_blocks - 1) {
struct free_block * next = &chunk->free_blocks[i+1];
if (block->offset + block->size == next->offset) {
block->size += next->size;
ggml_dyn_tallocr_remove_block(chunk, i+1);
}
}
return;
}
// check if ptr is at the beginning of the block
if (offset + size == block->offset) {
block->offset = offset;
if (addr.offset + size == block->offset) {
block->offset = addr.offset;
block->size += size;
// check if we can merge with the previous block
if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
alloc->free_blocks[i-1].size += block->size;
alloc->n_free_blocks--;
for (int j = i; j < alloc->n_free_blocks; j++) {
alloc->free_blocks[j] = alloc->free_blocks[j+1];
if (i > 0) {
struct free_block * prev = &chunk->free_blocks[i-1];
if (prev->offset + prev->size == block->offset) {
prev->size += block->size;
ggml_dyn_tallocr_remove_block(chunk, i);
}
}
return;
}
}
// otherwise, add a new block
GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
// insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
int insert_pos = 0;
while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
insert_pos++;
}
// shift all blocks from insert_pos onward to make room for the new block
for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
alloc->free_blocks[i] = alloc->free_blocks[i-1];
}
// insert the new block
alloc->free_blocks[insert_pos].offset = offset;
alloc->free_blocks[insert_pos].size = size;
alloc->n_free_blocks++;
ggml_dyn_tallocr_insert_block(chunk, addr.offset, size);
GGML_UNUSED(tensor);
}
static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
alloc->n_free_blocks = 1;
alloc->free_blocks[0].offset = 0;
alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
alloc->max_size = 0;
for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS; i++) {
free(alloc->chunks[i]);
alloc->chunks[i] = NULL;
}
alloc->n_chunks = 0;
#ifdef GGML_ALLOCATOR_DEBUG
for (int i = 0; i < 1024; i++) {
@ -293,14 +367,14 @@ static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
#endif
}
static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment, size_t max_buffer_size) {
struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(sizeof(struct ggml_dyn_tallocr));
*alloc = (struct ggml_dyn_tallocr) {
/*.alignment = */ alignment,
/*.n_free_blocks = */ 0,
/*.free_blocks = */ {{0}},
/*.max_size = */ 0,
/*.alignment = */ alignment,
/*.max_chunk_size = */ MIN(max_buffer_size, SIZE_MAX/2), // clamp to avoid overflows
/*.chunks = */ {NULL},
/*.n_chunks = */ 0,
#ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ {{0}},
#endif
@ -312,11 +386,73 @@ static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
}
static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
for (int i = 0; i < alloc->n_chunks; ++i) {
free(alloc->chunks[i]);
}
free(alloc);
}
static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
return alloc->max_size;
static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc, int chunk) {
return chunk < alloc->n_chunks ? alloc->chunks[chunk]->max_size : 0;
}
// virtual buffer with contiguous memory range, split into multiple backend buffers (chunks)
struct vbuffer {
ggml_backend_buffer_t chunks[GGML_VBUFFER_MAX_CHUNKS];
};
static void ggml_vbuffer_free(struct vbuffer * buf) {
if (buf == NULL) {
return;
}
for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS; ++i) {
ggml_backend_buffer_free(buf->chunks[i]);
}
free(buf);
}
static size_t ggml_vbuffer_chunk_size(struct vbuffer * buf, int chunk) {
return buf->chunks[chunk] ? ggml_backend_buffer_get_size(buf->chunks[chunk]) : 0;
}
static size_t ggml_vbuffer_size(struct vbuffer * buf) {
size_t size = 0;
for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[i]; ++i) {
size += ggml_backend_buffer_get_size(buf->chunks[i]);
}
return size;
}
static struct vbuffer * ggml_vbuffer_alloc(ggml_backend_buffer_type_t buft, const struct ggml_dyn_tallocr * talloc, enum ggml_backend_buffer_usage usage) {
struct vbuffer * buf = (struct vbuffer *)calloc(1, sizeof(struct vbuffer));
if (buf == NULL) {
return NULL;
}
for (int n = 0; n < talloc->n_chunks; n++) {
size_t chunk_size = talloc->chunks[n]->max_size;
buf->chunks[n] = ggml_backend_buft_alloc_buffer(buft, chunk_size);
if (buf->chunks[n] == NULL) {
ggml_vbuffer_free(buf);
return NULL;
}
ggml_backend_buffer_set_usage(buf->chunks[n], usage);
}
return buf;
}
static void ggml_vbuffer_tensor_alloc(struct vbuffer * buf, struct ggml_tensor * tensor, struct buffer_address buf_addr) {
void * base = ggml_backend_buffer_get_base(buf->chunks[buf_addr.chunk]);
void * addr = (char *)base + buf_addr.offset;
ggml_backend_tensor_alloc(buf->chunks[buf_addr.chunk], tensor, addr);
}
static void ggml_vbuffer_reset(struct vbuffer * buf) {
for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[i]; ++i) {
ggml_backend_buffer_reset(buf->chunks[i]);
}
}
@ -328,13 +464,13 @@ struct hash_node {
int n_children;
int n_views;
int buffer_id;
size_t offset; // offset within the buffer
struct buffer_address addr;
bool allocated;
};
struct tensor_alloc {
int buffer_id;
size_t offset;
struct buffer_address addr;
size_t size_max; // 0 = pre-allocated, unused, or view
};
@ -349,7 +485,7 @@ struct node_alloc {
struct ggml_gallocr {
ggml_backend_buffer_type_t * bufts; // [n_buffers]
ggml_backend_buffer_t * buffers; // [n_buffers]
struct vbuffer ** buffers; // [n_buffers]
struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
int n_buffers;
@ -370,7 +506,7 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
GGML_ASSERT(galloc->bufts != NULL);
galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
galloc->buffers = calloc(n_bufs, sizeof(struct vbuffer *));
GGML_ASSERT(galloc->buffers != NULL);
galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
@ -390,7 +526,8 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
if (galloc->buf_tallocs[i] == NULL) {
size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
size_t max_size = ggml_backend_buft_get_max_size(bufts[i]);
galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment, max_size);
}
}
galloc->n_buffers = n_bufs;
@ -418,7 +555,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
}
}
if (!freed) {
ggml_backend_buffer_free(galloc->buffers[i]);
ggml_vbuffer_free(galloc->buffers[i]);
}
}
if (galloc->buf_tallocs != NULL) {
@ -467,7 +604,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
hn->allocated = true;
assert(hn->offset == 0);
assert(hn->addr.offset == 0);
// try to reuse a parent's buffer (inplace)
if (ggml_op_can_inplace(node->op)) {
@ -501,9 +638,9 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
assert(view_src_hn->offset == p_hn->offset);
assert(view_src_hn->addr.chunk == p_hn->addr.chunk && view_src_hn->addr.offset == p_hn->addr.offset);
hn->buffer_id = p_hn->buffer_id;
hn->offset = p_hn->offset;
hn->addr = p_hn->addr;
p_hn->allocated = false; // avoid freeing the parent
view_src_hn->allocated = false;
return;
@ -511,7 +648,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
} else {
AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
hn->buffer_id = p_hn->buffer_id;
hn->offset = p_hn->offset;
hn->addr = p_hn->addr;
p_hn->allocated = false; // avoid freeing the parent
return;
}
@ -522,9 +659,8 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
size_t size = ggml_backend_buft_get_alloc_size(buft, node);
size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
hn->buffer_id = buffer_id;
hn->offset = offset;
hn->addr = ggml_dyn_tallocr_alloc(alloc, size, node);
}
}
@ -536,12 +672,11 @@ static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * n
}
struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
size_t offset = hn->offset;
int buffer_id = hn->buffer_id;
struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
size_t size = ggml_backend_buft_get_alloc_size(buft, node);
ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
ggml_dyn_tallocr_free_tensor(alloc, hn->addr, size, node);
hn->allocated = false;
}
@ -692,24 +827,24 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
struct node_alloc * node_alloc = &galloc->node_allocs[i];
if (node->view_src || node->data) {
node_alloc->dst.buffer_id = -1;
node_alloc->dst.offset = SIZE_MAX;
node_alloc->dst.addr = GGML_BUFFER_ADDRESS_INVALID;
node_alloc->dst.size_max = 0;
} else {
struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
node_alloc->dst.buffer_id = hn->buffer_id;
node_alloc->dst.offset = hn->offset;
node_alloc->dst.addr = hn->addr;
node_alloc->dst.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (!src || src->view_src || src->data) {
node_alloc->src[j].buffer_id = -1;
node_alloc->src[j].offset = SIZE_MAX;
node_alloc->src[j].addr = GGML_BUFFER_ADDRESS_INVALID;
node_alloc->src[j].size_max = 0;
} else {
struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
node_alloc->src[j].buffer_id = hn->buffer_id;
node_alloc->src[j].offset = hn->offset;
node_alloc->src[j].addr = hn->addr;
node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
}
}
@ -725,11 +860,11 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
if (leaf->view_src || leaf->data) {
galloc->leaf_allocs[i].leaf.buffer_id = -1;
galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
galloc->leaf_allocs[i].leaf.addr = GGML_BUFFER_ADDRESS_INVALID;
galloc->leaf_allocs[i].leaf.size_max = 0;
} else {
galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
galloc->leaf_allocs[i].leaf.offset = hn->offset;
galloc->leaf_allocs[i].leaf.addr = hn->addr;
galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
}
}
@ -744,22 +879,29 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
}
}
size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
// even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
if (new_size > cur_size || galloc->buffers[i] == NULL) {
bool realloc = galloc->buffers[i] == NULL;
size_t new_size = 0;
for (int c = 0; c < galloc->buf_tallocs[i]->n_chunks; c++) {
size_t cur_chunk_size = galloc->buffers[i] ? ggml_vbuffer_chunk_size(galloc->buffers[i], c) : 0;
size_t new_chunk_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i], c);
new_size += new_chunk_size;
if (new_chunk_size > cur_chunk_size) {
realloc = true;
}
}
if (realloc) {
#ifndef NDEBUG
size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
#endif
ggml_backend_buffer_free(galloc->buffers[i]);
galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
ggml_vbuffer_free(galloc->buffers[i]);
galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
if (galloc->buffers[i] == NULL) {
GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
return false;
}
ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
}
}
@ -772,11 +914,11 @@ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) {
int buffer_id = tensor_alloc->buffer_id;
assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
assert(tensor->data || tensor->view_src || ggml_backend_buft_get_alloc_size(galloc->bufts[buffer_id], tensor) <= tensor_alloc->size_max);
if (tensor->view_src != NULL) {
if (tensor->buffer == NULL) {
assert(tensor_alloc->offset == SIZE_MAX);
assert(tensor_alloc->addr.offset == SIZE_MAX);
if (tensor->view_src->buffer == NULL) {
// this tensor was allocated without ggml-backend
return;
@ -785,11 +927,9 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
}
} else {
if (tensor->data == NULL) {
assert(tensor_alloc->offset != SIZE_MAX);
assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
void * addr = (char *)base + tensor_alloc->offset;
ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
assert(tensor_alloc->addr.offset != SIZE_MAX);
assert(ggml_backend_buft_get_alloc_size(galloc->bufts[buffer_id], tensor) <= tensor_alloc->size_max);
ggml_vbuffer_tensor_alloc(galloc->buffers[buffer_id], tensor, tensor_alloc->addr);
} else {
if (tensor->buffer == NULL) {
// this tensor was allocated without ggml-backend
@ -874,7 +1014,7 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
// reset buffers
for (int i = 0; i < galloc->n_buffers; i++) {
if (galloc->buffers[i] != NULL) {
ggml_backend_buffer_reset(galloc->buffers[i]);
ggml_vbuffer_reset(galloc->buffers[i]);
}
}
@ -917,7 +1057,7 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
}
}
return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
return ggml_vbuffer_size(galloc->buffers[buffer_id]);
}
// utils

5
ggml/src/ggml-backend-impl.h
View File

@ -116,7 +116,7 @@ extern "C" {
void (*event_wait) (ggml_backend_t backend, ggml_backend_event_t event);
// (optional) sort/optimize the nodes in the graph
void (*optimize_graph) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
void (*graph_optimize) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
};
struct ggml_backend {
@ -209,9 +209,6 @@ extern "C" {
void * context;
};
// Internal backend registry API
GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
// Add backend dynamic loading support to the backend
// Initialize the backend

13
ggml/src/ggml-backend-reg.cpp
View File

@ -139,6 +139,10 @@ static void * dl_get_sym(dl_handle * handle, const char * name) {
return p;
}
static const char * dl_error() {
return "";
}
#else
using dl_handle = void;
@ -159,6 +163,11 @@ static void * dl_get_sym(dl_handle * handle, const char * name) {
return dlsym(handle, name);
}
static const char * dl_error() {
const char *rslt = dlerror();
return rslt != nullptr ? rslt : "";
}
#endif
using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
@ -247,7 +256,7 @@ struct ggml_backend_registry {
dl_handle_ptr handle { dl_load_library(path) };
if (!handle) {
if (!silent) {
GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(path).c_str());
GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path_str(path).c_str(), dl_error());
}
return nullptr;
}
@ -537,7 +546,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
dl_handle_ptr handle { dl_load_library(entry) };
if (!handle && !silent) {
GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(entry.path()).c_str());
GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path_str(entry.path()).c_str(), dl_error());
}
if (handle) {
auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");

16
ggml/src/ggml-backend.cpp
View File

@ -463,10 +463,10 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event)
backend->iface.event_wait(backend, event);
}
static void ggml_backend_optimize_graph(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
static void ggml_backend_graph_optimize(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
GGML_ASSERT(backend);
if (backend->iface.optimize_graph != NULL) {
backend->iface.optimize_graph(backend, cgraph);
if (backend->iface.graph_optimize != NULL) {
backend->iface.graph_optimize(backend, cgraph);
}
}
@ -1307,7 +1307,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
// Optimize this split of the graph. This needs to happen before we make graph_copy,
// so they are in sync.
ggml_backend_optimize_graph(sched->backends[split->backend_id], &split->graph);
ggml_backend_graph_optimize(sched->backends[split->backend_id], &split->graph);
// add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
for (int j = 0; j < split->n_inputs; j++) {
@ -1793,6 +1793,14 @@ ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i)
return sched->backends[i];
}
ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend) {
GGML_ASSERT(sched);
int backend_index = ggml_backend_sched_backend_id(sched, backend);
GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
return sched->bufts[backend_index];
}
size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
GGML_ASSERT(sched);
int backend_index = ggml_backend_sched_backend_id(sched, backend);

2
ggml/src/ggml-blas/CMakeLists.txt
View File

@ -74,7 +74,7 @@ if (BLAS_FOUND)
target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
if ("${BLAS_INCLUDE_DIRS}" MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
add_compile_definitions(GGML_BLAS_USE_MKL)
endif()

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

@ -270,7 +270,7 @@ static struct ggml_backend_i blas_backend_i = {
/* .graph_compute = */ ggml_backend_blas_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .optimize_graph = */ NULL,
/* .graph_optimize = */ NULL,
};
static ggml_guid_t ggml_backend_blas_guid(void) {

5
ggml/src/ggml-cann/common.h
View File

@ -526,7 +526,10 @@ struct ggml_backend_cann_context {
*/
aclrtStream stream(int stream) {
if (streams[stream] == nullptr) {
ggml_cann_set_device(device);
// If the device is not set here, destroying the stream later may cause a mismatch
// between the thread contexts where the stream was created and destroyed.
// However, I printed the device_id, thread_id, and stream, and they are all consistent.
ACL_CHECK(aclrtSetDevice(device));
ACL_CHECK(aclrtCreateStream(&streams[stream]));
}
return streams[stream];

13
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -75,13 +75,12 @@
* @param device The device ID to set.
*/
void ggml_cann_set_device(const int32_t device) {
// TODO: uncomment these lines after empty context has fixed.
// int current_device;
// ACL_CHECK(aclrtGetDevice(&current_device));
int current_device = -1;
aclrtGetDevice(&current_device);
// if (device == current_device) {
// return;
// }
if (device == current_device) {
return;
}
ACL_CHECK(aclrtSetDevice(device));
}
@ -2757,7 +2756,7 @@ static const ggml_backend_i ggml_backend_cann_interface = {
/* .graph_compute = */ ggml_backend_cann_graph_compute,
/* .event_record = */ ggml_backend_cann_event_record,
/* .event_wait = */ ggml_backend_cann_event_wait,
/* .optimize_graph = */ NULL,
/* .graph_optimize = */ NULL,
};
/**

14
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -439,6 +439,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
ggml-cpu/arch/riscv/quants.c
ggml-cpu/arch/riscv/repack.cpp
)
if (GGML_CPU_RISCV64_SPACEMIT)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_RISCV64_SPACEMIT ${RISCV64_SPACEMIT_IME_SPEC})
list(APPEND GGML_CPU_SOURCES
ggml-cpu/spacemit/ime.cpp
ggml-cpu/spacemit/ime.h
ggml-cpu/spacemit/ime1_kernels.cpp
ggml-cpu/spacemit/ime_kernels.h
)
endif()
set(MARCH_STR "rv64gc")
if (GGML_RV_ZFH)
string(APPEND MARCH_STR "_zfh")
@ -504,9 +513,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
# Fetch KleidiAI sources:
include(FetchContent)
set(KLEIDIAI_COMMIT_TAG "v1.13.0")
set(KLEIDIAI_COMMIT_TAG "v1.14.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
set(KLEIDIAI_ARCHIVE_MD5 "d82a8de939d9814621a5ba23907bdac1")
set(KLEIDIAI_ARCHIVE_MD5 "45e110675d93f99f82c23a1afcca76bc")
if (POLICY CMP0135)
cmake_policy(SET CMP0135 NEW)
@ -583,6 +592,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)

6
ggml/src/ggml-cpu/amx/amx.cpp
View File

@ -7,7 +7,7 @@
#include "ggml-cpu.h"
#include "traits.h"
#if defined(__gnu_linux__)
#if defined(__linux__)
#include <sys/syscall.h>
#include <unistd.h>
#endif
@ -186,7 +186,7 @@ static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_ty
#define XFEATURE_XTILEDATA 18
static bool ggml_amx_init() {
#if defined(__gnu_linux__)
#if defined(__linux__)
if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
fprintf(stderr, "AMX is not ready to be used!\n");
return false;
@ -194,6 +194,8 @@ static bool ggml_amx_init() {
return true;
#elif defined(_WIN32)
return true;
#else
return false;
#endif
}

1
ggml/src/ggml-cpu/arch-fallback.h
View File

@ -160,7 +160,6 @@
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8

24
ggml/src/ggml-cpu/arch/loongarch/quants.c
View File

@ -105,6 +105,18 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
return ((v4f32)res)[0];
}
// multiply int8_t, add results pairwise twice
static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
// Get absolute values of x vectors
const __m128i ax = __lsx_vsigncov_b(x, x);
// Sign the values of the y vectors
const __m128i sy = __lsx_vsigncov_b(x, y);
// Perform multiplication and create 16-bit values
const __m128i dot = lsx_maddubs_h(ax, sy);
const __m128i ones = __lsx_vreplgr2vr_h(1);
return lsx_madd_h(ones, dot);
}
#endif
#if defined(__loongarch_asx)
@ -323,18 +335,6 @@ static inline __m256i lasx_xvandi_b_bit(__m256i a, const unsigned int b) {
}
}
// multiply int8_t, add results pairwise twice
static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
// Get absolute values of x vectors
const __m128i ax = __lsx_vsigncov_b(x, x);
// Sign the values of the y vectors
const __m128i sy = __lsx_vsigncov_b(x, y);
// Perform multiplication and create 16-bit values
const __m128i dot = lsx_maddubs_h(ax, sy);
const __m128i ones = __lsx_vreplgr2vr_h(1);
return lsx_madd_h(ones, dot);
}
// horizontally add 8 floats
static inline float hsum_float_8(const __m256 x) {
__m128 res = lasx_extractf128(x, 1);

103
ggml/src/ggml-cpu/arch/s390/quants.c
View File

@ -75,7 +75,8 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
for (int j = 0; j < 8; j++) {
const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
const int32x4_t vi = vec_signed(v);
/* Uses non-default rounding for vec_signed or vec_round */
const int32x4_t vi = vec_signed(__builtin_s390_vfisb(v, 4, 1));
y[i].qs[4*j + 0] = vec_extract(vi, 0);
y[i].qs[4*j + 1] = vec_extract(vi, 1);
@ -122,7 +123,8 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
for (int j = 0; j < 8; j++) {
const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
const int32x4_t vi = vec_signed(v);
/* Uses non-default rounding for vec_signed or vec_round */
const int32x4_t vi = vec_signed(__builtin_s390_vfisb(v, 4, 1));
y[i].qs[4*j + 0] = vec_extract(vi, 0);
y[i].qs[4*j + 1] = vec_extract(vi, 1);
@ -260,6 +262,101 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
#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) {
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
assert(n % QK_MXFP4 == 0);
static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
const int qk = QK_MXFP4;
const int nb = n / qk;
const block_mxfp4 * GGML_RESTRICT x = vx;
const block_q8_0 * GGML_RESTRICT y = vy;
int ib = 0;
float sumf = 0.0f;
#if defined(__VXE__) || defined(__VXE2__)
const int8x16_t v_k = vec_xl(0, kvalues_mxfp4);
const uint8x16_t v_m = vec_splats((const uint8_t)0x0F);
float32x4_t v_acc = vec_splats(0.0f);
#pragma GCC unroll 8
for (; ib + 1 < nb; ib += 2) {
const block_mxfp4 * GGML_RESTRICT x0 = &x[ib + 0];
const block_mxfp4 * GGML_RESTRICT x1 = &x[ib + 1];
const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
const uint8x16_t v_x0 = vec_xl(0, x0->qs);
const uint8x16_t v_x1 = vec_xl(0, x1->qs);
int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
v_x0l = vec_perm(v_k, v_k, (uchar8x16_t)v_x0l);
v_x0h = vec_perm(v_k, v_k, (uchar8x16_t)v_x0h);
v_x1l = vec_perm(v_k, v_k, (uchar8x16_t)v_x1l);
v_x1h = vec_perm(v_k, v_k, (uchar8x16_t)v_x1h);
const int8x16_t v_y0l = vec_xl(0, y0->qs);
const int8x16_t v_y0h = vec_xl(QK8_0/2, y0->qs);
const int8x16_t v_y1l = vec_xl(0, y1->qs);
const int8x16_t v_y1h = vec_xl(QK8_0/2, y1->qs);
const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0l, v_y0l), v_x0h, v_y0h);
const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1l, v_y1l), v_x1h, v_y1h);
const float32x4_t v_xy0f = vec_float(v_xy0);
const float32x4_t v_xy1f = vec_float(v_xy1);
const float32x4_t v_d0 = vec_splats(GGML_E8M0_TO_FP32_HALF(x0->e) * GGML_CPU_FP16_TO_FP32(y0->d));
const float32x4_t v_d1 = vec_splats(GGML_E8M0_TO_FP32_HALF(x1->e) * GGML_CPU_FP16_TO_FP32(y1->d));
v_acc = vec_madd(v_xy0f, v_d0, v_acc);
v_acc = vec_madd(v_xy1f, v_d1, v_acc);
}
for (; ib < nb; ++ib) {
const block_mxfp4 * GGML_RESTRICT x0 = &x[ib + 0];
const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
const uint8x16_t v_x = vec_xl(0, x0->qs);
int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
v_xl = vec_perm(v_k, v_k, (uchar8x16_t)v_xl);
v_xh = vec_perm(v_k, v_k, (uchar8x16_t)v_xh);
const int8x16_t v_yl = vec_xl(0, y0->qs);
const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
const float32x4_t v_xyf = vec_float(v_xy);
const float32x4_t v_d = vec_splats(GGML_E8M0_TO_FP32_HALF(x0->e) * GGML_CPU_FP16_TO_FP32(y0->d));
v_acc = vec_madd(v_xyf, v_d, v_acc);
}
sumf = vec_hsum_f32x4(v_acc);
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(ib);
UNUSED(sumf);
ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
const int qk = QK8_0;
const int nb = n / qk;
@ -636,7 +733,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
uint8x16_t q3h[4];
uint8x16_t q3b[2];
int8x16_t q3bytes[4];
int8x16_t q8bytes[4];
int8x16_t q8bytes[8];
uint8x16_t qhbits[2];
float sum = 0;

4
ggml/src/ggml-cpu/arch/x86/repack.cpp
View File

@ -878,7 +878,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
// Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
// Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
@ -1231,7 +1231,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
// Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
// Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);

14
ggml/src/ggml-cpu/common.h
View File

@ -28,6 +28,14 @@ static inline float bf16_to_f32(ggml_bf16_t x) {
return GGML_BF16_TO_FP32(x);
}
static inline float i32_to_f32(int32_t x) {
return x;
}
static inline int32_t f32_to_i32(float x) {
return x;
}
static inline float f32_to_f32(float x) {
return x;
}
@ -54,6 +62,12 @@ struct type_conversion_table<ggml_bf16_t> {
static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
};
template <>
struct type_conversion_table<int32_t> {
static constexpr float (*to_f32)(int32_t) = i32_to_f32;
static constexpr int32_t (*from_f32)(float) = f32_to_i32;
};
static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) {
const int64_t ith = params->ith;
const int64_t nth = params->nth;

21
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -473,10 +473,10 @@ struct ggml_threadpool {
struct ggml_compute_state {
#ifndef GGML_USE_OPENMP
ggml_thread_t thrd;
bool cpumask[GGML_MAX_N_THREADS];
int last_graph;
bool pending;
#endif
bool cpumask[GGML_MAX_N_THREADS];
struct ggml_threadpool * threadpool;
int ith;
};
@ -2187,6 +2187,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_XIELU:
{
n_tasks = n_threads;
} break;
@ -3081,7 +3082,14 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
threadpool->workers = workers;
#ifndef GGML_USE_OPENMP
#ifdef GGML_USE_OPENMP
int32_t cpumask_iter = 0;
// Compute CPU masks for each thread
for (int j = 0; j < tpp->n_threads; j++) {
ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter);
}
#else // GGML_USE_OPENMP
ggml_mutex_init(&threadpool->mutex);
ggml_cond_init(&threadpool->cond);
@ -3154,7 +3162,14 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
}
ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
// Apply thread CPU mask and priority
int ith = omp_get_thread_num();
ggml_thread_apply_priority(threadpool->prio);
if (ggml_thread_cpumask_is_valid(threadpool->workers[ith].cpumask)) {
ggml_thread_apply_affinity(threadpool->workers[ith].cpumask);
}
ggml_graph_compute_thread(&threadpool->workers[ith]);
}
} else {
atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);

12
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -18,6 +18,10 @@
# include "kleidiai/kleidiai.h"
#endif
#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
# include "spacemit/ime.h"
#endif
#if defined(_WIN32)
# define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX
@ -45,6 +49,12 @@ std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_type
}
#endif
#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
if (ggml_backend_cpu_riscv64_spacemit_buffer_type()) {
bufts.push_back(ggml_backend_cpu_riscv64_spacemit_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
if (ggml_backend_cpu_kleidiai_buffer_type()) {
bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());
@ -190,7 +200,7 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .optimize_graph = */ NULL,
/* .graph_optimize = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {

184
ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
View File

@ -87,15 +87,38 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
return tensor->ne[dim];
}
template <typename Variant, typename Ret, typename... Args, std::size_t... Is>
constexpr bool variant_any_invocable_impl(std::index_sequence<Is...>) {
using V = std::remove_reference_t<Variant>;
return (std::is_invocable_r_v<
Ret,
std::variant_alternative_t<Is, V>,
Args...> || ...);
}
template <typename Variant, typename Ret, typename... Args>
constexpr bool variant_any_invocable_v =
variant_any_invocable_impl<Variant, Ret, Args...>(
std::make_index_sequence<
std::variant_size_v<std::remove_reference_t<Variant>>>{});
template<typename Ret, typename Variant, typename... Args>
static Ret variant_call(const Variant & var, Args&&... args) {
return std::visit([&](auto&& func) -> Ret {
if constexpr (std::is_invocable_r_v<Ret, decltype(func), Args...>) {
return func(std::forward<Args>(args)...);
} else {
throw std::runtime_error("Invalid function type in variant_call");
}
}, var);
static inline Ret variant_call(Variant && var, Args&&... args) {
static_assert(variant_any_invocable_v<std::remove_reference_t<Variant>, Ret, Args...>,
"No alternative in Variant is invocable with the provided arguments and return type.");
return std::visit(
[&](auto && f) -> Ret {
using F = std::decay_t<decltype(f)>;
if constexpr (std::is_invocable_r_v<Ret, F, Args...>) {
return std::invoke(std::forward<decltype(f)>(f), std::forward<Args>(args)...);
} else {
GGML_ABORT("Invalid function type in variant_call");
GGML_UNREACHABLE();
}
},
std::forward<Variant>(var)
);
}
namespace ggml::cpu::kleidiai {
@ -138,7 +161,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
if (kernels->rhs_type == GGML_TYPE_Q4_0) {
size = variant_call<size_t>(lhs_info->packed_size, m, k, QK4_0, mr, kr, sr);
} else if (kernels->rhs_type == GGML_TYPE_F16) {
size = variant_call<size_t>(lhs_info->packed_size, m, k, mr, kr, sr) +
const int64_t lhs_batch_size0 = op->src[1]->ne[2];
const int64_t rhs_batch_size0 = op->src[0]->ne[2];
const int64_t r = lhs_batch_size0 / rhs_batch_size0;
size = variant_call<size_t>(lhs_info->packed_size, m * r, k, mr, kr, sr) +
variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
k * n * sizeof(float) + n * sizeof(float);
} else {
@ -148,7 +174,6 @@ class tensor_traits : public ggml::cpu::tensor_traits {
return true;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
if (dst->op == GGML_OP_MUL_MAT) {
if (dst->src[0]->type == GGML_TYPE_Q4_0) {
@ -165,8 +190,6 @@ class tensor_traits : public ggml::cpu::tensor_traits {
}
bool compute_forward_fp16(ggml_compute_params * params, struct ggml_tensor * dst) {
static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
@ -175,7 +198,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
GGML_ASSERT(kernels);
bool is_gemv = src1->ne[1] == 1;
const bool is_gemv = src1->ne[1] == 1;
kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
GGML_ASSERT(kernel);
@ -185,27 +208,30 @@ class tensor_traits : public ggml::cpu::tensor_traits {
const int64_t lhs_batch_size0 = ne12;
const int64_t rhs_batch_size0 = ne02;
const int64_t batch_size = rhs_batch_size0;
const int64_t batch_size = lhs_batch_size0;
GGML_ASSERT(rhs_batch_size0 > 0);
GGML_ASSERT(lhs_batch_size0 % rhs_batch_size0 == 0);
const int64_t r = lhs_batch_size0 / rhs_batch_size0;
const int64_t m = ne11 * r;
const int64_t n = ne01;
const int64_t k = ne00;
const int64_t m_group = ne11;
const int64_t m = m_group;
const int64_t n = ne01;
const int64_t k = ne00;
const size_t lhs_stride = src1->nb[1];
const size_t rhs_stride = src0->nb[1];
const size_t dst_stride = dst->nb[1];
const int64_t mr = static_cast<int64_t>(kernel->get_mr());
const int64_t nr = static_cast<int64_t>(kernel->get_nr());
const int64_t kr = static_cast<int64_t>(kernel->get_kr());
const int64_t sr = static_cast<int64_t>(kernel->get_sr());
const int64_t mr = (int64_t) kernel->get_mr();
const int64_t nr = (int64_t) kernel->get_nr();
const int64_t kr = (int64_t) kernel->get_kr();
const int64_t sr = (int64_t) kernel->get_sr();
const size_t lhs_packed_size = variant_call<size_t>(lhs_info->packed_size, m, k, mr, kr, sr);
const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, n, k);
const size_t kxn_size = k * n * sizeof(float);
const size_t bias_size = n * sizeof(float);
const size_t lhs_packed_size = variant_call<size_t>(lhs_info->packed_size, (size_t)m, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, (size_t)n, (size_t)k);
const size_t kxn_size = (size_t)k * (size_t)n * sizeof(float);
const size_t bias_size = (size_t)n * sizeof(float);
const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
GGML_ASSERT(wsize_required <= params->wsize);
@ -216,82 +242,102 @@ class tensor_traits : public ggml::cpu::tensor_traits {
uint8_t * bias = rhs_kxn + kxn_size;
for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
const uint8_t * lhs_batch = static_cast<const uint8_t *>(src1->data) + batch_idx * m * lhs_stride;
const uint8_t * rhs_batch = static_cast<const uint8_t *>(src0->data) + batch_idx * n * rhs_stride;
uint8_t * dst_batch = static_cast<uint8_t *>(dst->data) + batch_idx * m * dst_stride;
const int64_t rhs_batch_idx = batch_idx / r;
const uint8_t * rhs_batch_base = static_cast<const uint8_t *>(src0->data) + rhs_batch_idx * src0->nb[2];
uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];
// LHS packing
// LHS packing (threaded over m, honoring mr alignment and KV groups)
{
const int64_t m_roundup_mr = kai_roundup(m, mr);
const int64_t num_threads = KAI_MIN(m_roundup_mr / mr, nth);
if (ith < num_threads) {
const int64_t num_m_per_thread0 = round_down(m_roundup_mr / num_threads, mr);
const int64_t num_m_per_thread0 = round_down((size_t)(m_roundup_mr / num_threads), (size_t)mr);
const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
const int64_t m_start = ith * num_m_per_thread0;
const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
const int64_t m_start = ith * num_m_per_thread0;
const int64_t m_count = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
const size_t lhs_offset = variant_call<size_t>(kernels->gemm.get_lhs_offset, m_start, lhs_stride);
const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, mr, kr, sr);
// Base packed offset (aligned) and per-row stride in bytes
const size_t base_packed_off = variant_call<size_t>(
lhs_info->get_packed_offset, (size_t)m_start, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
const size_t next_block_off = variant_call<size_t>(
lhs_info->get_packed_offset, (size_t)(m_start + mr), (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
const size_t row_stride_bytes = (next_block_off - base_packed_off) / (size_t)mr;
const void * src_ptr = static_cast<const uint8_t *>(lhs_batch) + lhs_offset;
void * dst_ptr = static_cast<uint8_t *>(lhs_packed) + lhs_packed_offset;
int64_t remaining = m_count;
int64_t cur = m_start;
variant_call<void>(lhs_info->pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
while (remaining > 0) {
const int64_t row_in_group = cur;
const int64_t avail = m_group - row_in_group;
const int64_t take = std::min(avail, remaining);
const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
const void * src_ptr = lhs_batch_base + (size_t)row_in_group * lhs_stride;
const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
void * dst_ptr = lhs_packed + dst_off;
variant_call<void>(lhs_info->pack_func,
(size_t)take, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr,
/*m_idx_start*/ 0, src_ptr, lhs_stride, dst_ptr);
cur += take;
remaining -= take;
}
}
}
// RHS packing
if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) {
// First thread to reach this point handles RHS packing
memset(bias, 0, n * sizeof(float));
transpose_f32kxn_f16nxk(n, k, reinterpret_cast<float *>(rhs_kxn),
reinterpret_cast<const uint16_t *>(rhs_batch), rhs_stride);
// RHS packing (single thread), then synchronize
if (ith == 0) {
memset(bias, 0, (size_t)n * sizeof(float));
transpose_f32kxn_f16nxk((size_t)n, (size_t)k,
reinterpret_cast<float *>(rhs_kxn),
reinterpret_cast<const uint16_t *>(rhs_batch_base),
rhs_stride);
variant_call<void>(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float),
rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
variant_call<void>(kernels->rhs_info.pack_func,
/*num_groups*/ 1, (size_t)n, (size_t)k, (size_t)nr, (size_t)kr, (size_t)sr,
/*rhs_stride (bytes)*/ (size_t)(n * sizeof(float)),
rhs_kxn, bias, nullptr, rhs_packed, /*extra_bytes*/ 0, /*params*/ nullptr);
}
ggml_barrier(params->threadpool);
first_to_arrive.clear(std::memory_order_release);
// Perform the matmul
// Matmul (threaded over n)
{
const int64_t m_to_process = m;
const int64_t m_start = 0;
const int64_t n_step = static_cast<int64_t>(kernel->get_n_step());
int64_t num_threads = KAI_MIN(n / n_step, nth);
if (num_threads <= 0) {
num_threads = 1;
const int64_t n_step = (int64_t) kernel->get_n_step();
int64_t num_threads_n = KAI_MIN(n / n_step, nth);
if (num_threads_n <= 0) {
num_threads_n = 1;
}
if (ith < num_threads) {
const int64_t num_n_per_thread0 = round_down(n / num_threads, n_step);
const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0;
if (ith < num_threads_n) {
const int64_t num_n_per_thread0 = round_down((size_t)(n / num_threads_n), (size_t)n_step);
const int64_t num_n_per_threadN_1 = n - (num_threads_n - 1) * num_n_per_thread0;
const int64_t n_start = ith * num_n_per_thread0;
const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
const int64_t n_to_process = (ith == num_threads_n - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
const size_t lhs_packed_offset = variant_call<size_t>(kernel->get_lhs_offset, m_start, k);
const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k);
const size_t dst_offset = kernel->get_dst_offset(m_start, n_start, dst_stride);
// LHS packed base at row 0 (consistent with packing above)
const size_t lhs_packed_offset0 = variant_call<size_t>(
lhs_info->get_packed_offset, (size_t)0, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, (size_t)n_start, (size_t)k);
const size_t dst_offset = kernel->get_dst_offset((size_t)0, (size_t)n_start, dst_stride);
const void * lhs_ptr = lhs_packed + lhs_packed_offset;
const void * lhs_ptr = lhs_packed + lhs_packed_offset0;
const void * rhs_ptr = rhs_packed + rhs_packed_offset;
float * dst_ptr = reinterpret_cast<float *>(dst_batch + dst_offset);
float * dst_ptr = reinterpret_cast<float *>(dst_batch_base + dst_offset);
variant_call<void>(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
variant_call<void>(kernel->run_kernel,
(size_t)m, (size_t)n_to_process, (size_t)k,
lhs_ptr, rhs_ptr,
dst_ptr, dst_stride, sizeof(float),
-FLT_MAX, FLT_MAX);
}
}
if (batch_idx != batch_size - 1) {
// This barrier is necessary when the batch size is larger than 1. While processing a batch,
// the work data buffer (params->wdata) is used as temporary storage which means that only
// a single batch can be processed at any given time. No barrier is needed for the last
// batch since GGML inserts a barrier between the execution of every operator.
ggml_barrier(params->threadpool);
}
}

948
ggml/src/ggml-cpu/ops.cpp
View File

File diff suppressed because it is too large Load Diff

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