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

Merge branch 'master' into add-paligemma-support

Browse source
This commit is contained in:
Andrei Betlen 2024-08-10 17:49:40 -04:00
commit 614cb6a5fb
711 changed files with 79279 additions and 186067 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.devops/cloud-v-pipeline
View file

@ -15,7 +15,7 @@ node('x86_runner1'){ // Running on x86 runner containing latest vecto
stage('Running llama.cpp'){
sh'''#!/bin/bash
module load gnu-bin2/0.1 # loading latest versions of vector qemu and vector gcc
qemu-riscv64 -L /softwares/gnu-bin2/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt # Running llama.cpp on vector qemu-riscv64
qemu-riscv64 -L /softwares/gnu-bin2/sysroot -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt # Running llama.cpp on vector qemu-riscv64
cat llama_log.txt # Printing results
'''
}

4
.devops/full-cuda.Dockerfile
View file

@ -6,7 +6,7 @@ ARG CUDA_VERSION=11.7.1
# Target the CUDA build image
ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
FROM ${BASE_CUDA_DEV_CONTAINER} as build
FROM ${BASE_CUDA_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
ARG CUDA_DOCKER_ARCH=all
@ -27,7 +27,7 @@ COPY . .
# Set nvcc architecture
ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
# Enable CUDA
ENV LLAMA_CUDA=1
ENV GGML_CUDA=1
# Enable cURL
ENV LLAMA_CURL=1

4
.devops/full-rocm.Dockerfile
View file

@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
# Target the CUDA build image
ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
FROM ${BASE_ROCM_DEV_CONTAINER} as build
FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
@ -36,7 +36,7 @@ COPY . .
# Set nvcc architecture
ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
ENV LLAMA_HIPBLAS=1
ENV GGML_HIPBLAS=1
ENV CC=/opt/rocm/llvm/bin/clang
ENV CXX=/opt/rocm/llvm/bin/clang++

2
.devops/full.Dockerfile
View file

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
FROM ubuntu:$UBUNTU_VERSION as build
FROM ubuntu:$UBUNTU_VERSION AS build
RUN apt-get update && \
apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1

12
.devops/main-cuda.Dockerfile → .devops/llama-cli-cuda.Dockerfile
View file

@ -6,7 +6,7 @@ ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VER
# Target the CUDA runtime image
ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
FROM ${BASE_CUDA_DEV_CONTAINER} as build
FROM ${BASE_CUDA_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
ARG CUDA_DOCKER_ARCH=all
@ -21,15 +21,15 @@ COPY . .
# Set nvcc architecture
ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
# Enable CUDA
ENV LLAMA_CUDA=1
ENV GGML_CUDA=1
RUN make -j$(nproc) main
RUN make -j$(nproc) llama-cli
FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/main /main
COPY --from=build /app/llama-cli /llama-cli
ENTRYPOINT [ "/main" ]
ENTRYPOINT [ "/llama-cli" ]

28
.devops/llama-cli-intel.Dockerfile Normal file
View file

@ -0,0 +1,28 @@
ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
ARG GGML_SYCL_F16=OFF
RUN apt-get update && \
apt-get install -y git
WORKDIR /app
COPY . .
RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
echo "GGML_SYCL_F16 is set" && \
export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
fi && \
echo "Building with static libs" && \
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
cmake --build build --config Release --target llama-cli
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/llama-cli" ]

8
.devops/main-rocm.Dockerfile → .devops/llama-cli-rocm.Dockerfile
View file

@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
# Target the CUDA build image
ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
FROM ${BASE_ROCM_DEV_CONTAINER} as build
FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
@ -36,10 +36,10 @@ COPY . .
# Set nvcc architecture
ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
ENV LLAMA_HIPBLAS=1
ENV GGML_HIPBLAS=1
ENV CC=/opt/rocm/llvm/bin/clang
ENV CXX=/opt/rocm/llvm/bin/clang++
RUN make -j$(nproc) main
RUN make -j$(nproc) llama-cli
ENTRYPOINT [ "/app/main" ]
ENTRYPOINT [ "/app/llama-cli" ]

10
.devops/main-vulkan.Dockerfile → .devops/llama-cli-vulkan.Dockerfile
View file

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=jammy
FROM ubuntu:$UBUNTU_VERSION as build
FROM ubuntu:$UBUNTU_VERSION AS build
# Install build tools
RUN apt update && apt install -y git build-essential cmake wget libgomp1
@ -14,14 +14,14 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
# Build it
WORKDIR /app
COPY . .
RUN cmake -B build -DLLAMA_VULKAN=1 && \
cmake --build build --config Release --target main
RUN cmake -B build -DGGML_VULKAN=1 && \
cmake --build build --config Release --target llama-cli
# Clean up
WORKDIR /
RUN cp /app/build/bin/main /main && \
RUN cp /app/build/bin/llama-cli /llama-cli && \
rm -rf /app
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/main" ]
ENTRYPOINT [ "/llama-cli" ]

10
.devops/main.Dockerfile → .devops/llama-cli.Dockerfile
View file

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
FROM ubuntu:$UBUNTU_VERSION as build
FROM ubuntu:$UBUNTU_VERSION AS build
RUN apt-get update && \
apt-get install -y build-essential git
@ -9,15 +9,15 @@ WORKDIR /app
COPY . .
RUN make -j$(nproc) main
RUN make -j$(nproc) llama-cli
FROM ubuntu:$UBUNTU_VERSION as runtime
FROM ubuntu:$UBUNTU_VERSION AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/main /main
COPY --from=build /app/llama-cli /llama-cli
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/main" ]
ENTRYPOINT [ "/llama-cli" ]

84
.devops/llama-cpp-clblast.srpm.spec
View file

@ -1,84 +0,0 @@
# SRPM for building from source and packaging an RPM for RPM-based distros.
# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
# Built and maintained by John Boero - boeroboy@gmail.com
# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
# Notes for llama.cpp:
# 1. Tags are currently based on hash - which will not sort asciibetically.
# We need to declare standard versioning if people want to sort latest releases.
# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
# Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
# It is up to the user to install the correct vendor-specific support.
Name: llama.cpp-clblast
Version: %( date "+%%Y%%m%%d" )
Release: 1%{?dist}
Summary: OpenCL Inference of LLaMA model in C/C++
License: MIT
Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
BuildRequires: coreutils make gcc-c++ git mesa-libOpenCL-devel clblast-devel
Requires: clblast
URL: https://github.com/ggerganov/llama.cpp
%define debug_package %{nil}
%define source_date_epoch_from_changelog 0
%description
CPU inference for Meta's Lllama2 models using default options.
%prep
%setup -n llama.cpp-master
%build
make -j LLAMA_CLBLAST=1
%install
mkdir -p %{buildroot}%{_bindir}/
cp -p main %{buildroot}%{_bindir}/llamaclblast
cp -p server %{buildroot}%{_bindir}/llamaclblastserver
cp -p simple %{buildroot}%{_bindir}/llamaclblastsimple
mkdir -p %{buildroot}/usr/lib/systemd/system
%{__cat} <<EOF > %{buildroot}/usr/lib/systemd/system/llamaclblast.service
[Unit]
Description=Llama.cpp server, CPU only (no GPU support in this build).
After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
[Service]
Type=simple
EnvironmentFile=/etc/sysconfig/llama
ExecStart=/usr/bin/llamaclblastserver $LLAMA_ARGS
ExecReload=/bin/kill -s HUP $MAINPID
Restart=never
[Install]
WantedBy=default.target
EOF
mkdir -p %{buildroot}/etc/sysconfig
%{__cat} <<EOF > %{buildroot}/etc/sysconfig/llama
LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
EOF
%clean
rm -rf %{buildroot}
rm -rf %{_builddir}/*
%files
%{_bindir}/llamaclblast
%{_bindir}/llamaclblastserver
%{_bindir}/llamaclblastsimple
/usr/lib/systemd/system/llamaclblast.service
%config /etc/sysconfig/llama
%pre
%post
%preun
%postun
%changelog

16
.devops/llama-cpp-cuda.srpm.spec
View file

@ -32,13 +32,13 @@ CPU inference for Meta's Lllama2 models using default options.
%setup -n llama.cpp-master
%build
make -j LLAMA_CUDA=1
make -j GGML_CUDA=1
%install
mkdir -p %{buildroot}%{_bindir}/
cp -p main %{buildroot}%{_bindir}/llamacppcuda
cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
cp -p llama-cli %{buildroot}%{_bindir}/llama-cuda-cli
cp -p llama-server %{buildroot}%{_bindir}/llama-cuda-server
cp -p llama-simple %{buildroot}%{_bindir}/llama-cuda-simple
mkdir -p %{buildroot}/usr/lib/systemd/system
%{__cat} <<EOF > %{buildroot}/usr/lib/systemd/system/llamacuda.service
@ -49,7 +49,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
[Service]
Type=simple
EnvironmentFile=/etc/sysconfig/llama
ExecStart=/usr/bin/llamacppcudaserver $LLAMA_ARGS
ExecStart=/usr/bin/llama-cuda-server $LLAMA_ARGS
ExecReload=/bin/kill -s HUP $MAINPID
Restart=never
@ -67,9 +67,9 @@ rm -rf %{buildroot}
rm -rf %{_builddir}/*
%files
%{_bindir}/llamacppcuda
%{_bindir}/llamacppcudaserver
%{_bindir}/llamacppcudasimple
%{_bindir}/llama-cuda-cli
%{_bindir}/llama-cuda-server
%{_bindir}/llama-cuda-simple
/usr/lib/systemd/system/llamacuda.service
%config /etc/sysconfig/llama

14
.devops/llama-cpp.srpm.spec
View file

@ -38,9 +38,9 @@ make -j
%install
mkdir -p %{buildroot}%{_bindir}/
cp -p main %{buildroot}%{_bindir}/llama
cp -p server %{buildroot}%{_bindir}/llamaserver
cp -p simple %{buildroot}%{_bindir}/llamasimple
cp -p llama-cli %{buildroot}%{_bindir}/llama-cli
cp -p llama-server %{buildroot}%{_bindir}/llama-server
cp -p llama-simple %{buildroot}%{_bindir}/llama-simple
mkdir -p %{buildroot}/usr/lib/systemd/system
%{__cat} <<EOF > %{buildroot}/usr/lib/systemd/system/llama.service
@ -51,7 +51,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
[Service]
Type=simple
EnvironmentFile=/etc/sysconfig/llama
ExecStart=/usr/bin/llamaserver $LLAMA_ARGS
ExecStart=/usr/bin/llama-server $LLAMA_ARGS
ExecReload=/bin/kill -s HUP $MAINPID
Restart=never
@ -69,9 +69,9 @@ rm -rf %{buildroot}
rm -rf %{_builddir}/*
%files
%{_bindir}/llama
%{_bindir}/llamaserver
%{_bindir}/llamasimple
%{_bindir}/llama-cli
%{_bindir}/llama-server
%{_bindir}/llama-simple
/usr/lib/systemd/system/llama.service
%config /etc/sysconfig/llama

16
.devops/server-cuda.Dockerfile → .devops/llama-server-cuda.Dockerfile
View file

@ -6,7 +6,7 @@ ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VER
# Target the CUDA runtime image
ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
FROM ${BASE_CUDA_DEV_CONTAINER} as build
FROM ${BASE_CUDA_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
ARG CUDA_DOCKER_ARCH=all
@ -21,17 +21,19 @@ COPY . .
# Set nvcc architecture
ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
# Enable CUDA
ENV LLAMA_CUDA=1
ENV GGML_CUDA=1
# Enable cURL
ENV LLAMA_CURL=1
RUN make -j$(nproc) server
RUN make -j$(nproc) llama-server
FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/server /server
COPY --from=build /app/llama-server /llama-server
ENTRYPOINT [ "/server" ]
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/llama-server" ]

32
.devops/llama-server-intel.Dockerfile Normal file
View file

@ -0,0 +1,32 @@
ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
ARG GGML_SYCL_F16=OFF
RUN apt-get update && \
apt-get install -y git libcurl4-openssl-dev
WORKDIR /app
COPY . .
RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
echo "GGML_SYCL_F16 is set" && \
export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
fi && \
echo "Building with dynamic libs" && \
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
cmake --build build --config Release --target llama-server
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev curl
COPY --from=build /app/build/bin/llama-server /llama-server
ENV LC_ALL=C.utf8
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/llama-server" ]

12
.devops/server-rocm.Dockerfile → .devops/llama-server-rocm.Dockerfile
View file

@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
# Target the CUDA build image
ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
FROM ${BASE_ROCM_DEV_CONTAINER} as build
FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
@ -36,15 +36,17 @@ COPY . .
# Set nvcc architecture
ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
ENV LLAMA_HIPBLAS=1
ENV GGML_HIPBLAS=1
ENV CC=/opt/rocm/llvm/bin/clang
ENV CXX=/opt/rocm/llvm/bin/clang++
# Enable cURL
ENV LLAMA_CURL=1
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev
apt-get install -y libcurl4-openssl-dev curl
RUN make -j$(nproc)
RUN make -j$(nproc) llama-server
ENTRYPOINT [ "/app/server" ]
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/app/llama-server" ]

20
.devops/server-vulkan.Dockerfile → .devops/llama-server-vulkan.Dockerfile
View file

@ -1,31 +1,29 @@
ARG UBUNTU_VERSION=jammy
FROM ubuntu:$UBUNTU_VERSION as build
FROM ubuntu:$UBUNTU_VERSION AS build
# Install build tools
RUN apt update && apt install -y git build-essential cmake wget
# Install Vulkan SDK
# Install Vulkan SDK and cURL
RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
apt update -y && \
apt-get install -y vulkan-sdk
# Install cURL
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev
apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
# Build it
WORKDIR /app
COPY . .
RUN cmake -B build -DLLAMA_VULKAN=1 -DLLAMA_CURL=1 && \
cmake --build build --config Release --target server
RUN cmake -B build -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
cmake --build build --config Release --target llama-server
# Clean up
WORKDIR /
RUN cp /app/build/bin/server /server && \
RUN cp /app/build/bin/llama-server /llama-server && \
rm -rf /app
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/server" ]
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/llama-server" ]

27
.devops/llama-server.Dockerfile Normal file
View file

@ -0,0 +1,27 @@
ARG UBUNTU_VERSION=22.04
FROM ubuntu:$UBUNTU_VERSION AS build
RUN apt-get update && \
apt-get install -y build-essential git libcurl4-openssl-dev curl
WORKDIR /app
COPY . .
ENV LLAMA_CURL=1
RUN make -j$(nproc) llama-server
FROM ubuntu:$UBUNTU_VERSION AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1
COPY --from=build /app/llama-server /llama-server
ENV LC_ALL=C.utf8
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/llama-server" ]

34
.devops/main-intel.Dockerfile
View file

@ -1,34 +0,0 @@
ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
rm /etc/apt/sources.list.d/intel-graphics.list && \
wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
chmod 644 /usr/share/keyrings/intel-graphics.gpg
ARG LLAMA_SYCL_F16=OFF
RUN apt-get update && \
apt-get install -y git
WORKDIR /app
COPY . .
RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
echo "LLAMA_SYCL_F16 is set" && \
export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
fi && \
cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
cmake --build build --config Release --target main
FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
COPY --from=build /app/build/bin/main /main
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/main" ]

5
.devops/nix/apps.nix
View file

@ -6,11 +6,10 @@
let
inherit (config.packages) default;
binaries = [
"llama"
"llama-cli"
"llama-embedding"
"llama-server"
"quantize"
"train-text-from-scratch"
"llama-quantize"
];
mkApp = name: {
type = "app";

74
.devops/nix/package.nix
View file

@ -17,19 +17,19 @@
rocmPackages,
vulkan-headers,
vulkan-loader,
clblast,
curl,
shaderc,
useBlas ? builtins.all (x: !x) [
useCuda
useMetalKit
useOpenCL
useRocm
useVulkan
] && blas.meta.available,
useCuda ? config.cudaSupport,
useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
useMpi ? false, # Increases the runtime closure size by ~700M
useOpenCL ? false,
useRocm ? config.rocmSupport,
enableCurl ? true,
useVulkan ? false,
llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
@ -56,7 +56,6 @@ let
++ lib.optionals useCuda [ "CUDA" ]
++ lib.optionals useMetalKit [ "MetalKit" ]
++ lib.optionals useMpi [ "MPI" ]
++ lib.optionals useOpenCL [ "OpenCL" ]
++ lib.optionals useRocm [ "ROCm" ]
++ lib.optionals useVulkan [ "Vulkan" ];
@ -91,6 +90,22 @@ let
ps.tiktoken
ps.torchWithoutCuda
ps.transformers
# server bench
ps.matplotlib
# server tests
ps.openai
ps.behave
ps.prometheus-client
# for examples/pydantic-models-to-grammar-examples.py
ps.docstring-parser
ps.pydantic
# for scripts/compare-llama-bench.py
ps.gitpython
ps.tabulate
]
);
@ -111,16 +126,9 @@ let
++ optionals useMetalKit [ MetalKit ];
cudaBuildInputs = with cudaPackages; [
cuda_cccl.dev # <nv/target>
# A temporary hack for reducing the closure size, remove once cudaPackages
# have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
cuda_cudart.dev
cuda_cudart.lib
cuda_cudart.static
libcublas.dev
libcublas.lib
libcublas.static
cuda_cudart
cuda_cccl # <nv/target>
libcublas
];
rocmBuildInputs = with rocmPackages; [
@ -132,6 +140,7 @@ let
vulkanBuildInputs = [
vulkan-headers
vulkan-loader
shaderc
];
in
@ -160,9 +169,9 @@ effectiveStdenv.mkDerivation (
};
postPatch = ''
substituteInPlace ./ggml-metal.m \
substituteInPlace ./ggml/src/ggml-metal.m \
--replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
substituteInPlace ./ggml-metal.m \
substituteInPlace ./ggml/src/ggml-metal.m \
--replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
'';
@ -198,24 +207,24 @@ effectiveStdenv.mkDerivation (
optionals effectiveStdenv.isDarwin darwinBuildInputs
++ optionals useCuda cudaBuildInputs
++ optionals useMpi [ mpi ]
++ optionals useOpenCL [ clblast ]
++ optionals useRocm rocmBuildInputs
++ optionals useBlas [ blas ]
++ optionals useVulkan vulkanBuildInputs;
++ optionals useVulkan vulkanBuildInputs
++ optionals enableCurl [ curl ];
cmakeFlags =
[
(cmakeBool "LLAMA_NATIVE" false)
(cmakeBool "LLAMA_BUILD_SERVER" true)
(cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
(cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
(cmakeBool "LLAMA_BLAS" useBlas)
(cmakeBool "LLAMA_CLBLAST" useOpenCL)
(cmakeBool "LLAMA_CUDA" useCuda)
(cmakeBool "LLAMA_HIPBLAS" useRocm)
(cmakeBool "LLAMA_METAL" useMetalKit)
(cmakeBool "LLAMA_VULKAN" useVulkan)
(cmakeBool "LLAMA_STATIC" enableStatic)
(cmakeBool "LLAMA_CURL" enableCurl)
(cmakeBool "GGML_NATIVE" false)
(cmakeBool "GGML_BLAS" useBlas)
(cmakeBool "GGML_CUDA" useCuda)
(cmakeBool "GGML_HIPBLAS" useRocm)
(cmakeBool "GGML_METAL" useMetalKit)
(cmakeBool "GGML_VULKAN" useVulkan)
(cmakeBool "GGML_STATIC" enableStatic)
]
++ optionals useCuda [
(
@ -231,7 +240,7 @@ effectiveStdenv.mkDerivation (
]
++ optionals useMetalKit [
(lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
(cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
(cmakeBool "GGML_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
];
# Environment variables needed for ROCm
@ -243,10 +252,8 @@ effectiveStdenv.mkDerivation (
# TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
# if they haven't been added yet.
postInstall = ''
mv $out/bin/main${executableSuffix} $out/bin/llama${executableSuffix}
mv $out/bin/server${executableSuffix} $out/bin/llama-server${executableSuffix}
mkdir -p $out/include
cp $src/llama.h $out/include/
cp $src/include/llama.h $out/include/
'';
# Define the shells here, but don't add in the inputsFrom to avoid recursion.
@ -256,7 +263,6 @@ effectiveStdenv.mkDerivation (
useCuda
useMetalKit
useMpi
useOpenCL
useRocm
useVulkan
;
@ -283,7 +289,7 @@ effectiveStdenv.mkDerivation (
# Configurations we don't want even the CI to evaluate. Results in the
# "unsupported platform" messages. This is mostly a no-op, because
# cudaPackages would've refused to evaluate anyway.
badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
badPlatforms = optionals useCuda lib.platforms.darwin;
# Configurations that are known to result in build failures. Can be
# overridden by importing Nixpkgs with `allowBroken = true`.
@ -294,7 +300,7 @@ effectiveStdenv.mkDerivation (
license = lib.licenses.mit;
# Accommodates `nix run` and `lib.getExe`
mainProgram = "llama";
mainProgram = "llama-cli";
# These people might respond, on the best effort basis, if you ping them
# in case of Nix-specific regressions or for reviewing Nix-specific PRs.

45
.devops/server-intel.Dockerfile
View file

@ -1,45 +0,0 @@
ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
rm /etc/apt/sources.list.d/intel-graphics.list && \
wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
chmod 644 /usr/share/keyrings/intel-graphics.gpg
ARG LLAMA_SYCL_F16=OFF
RUN apt-get update && \
apt-get install -y git libcurl4-openssl-dev
WORKDIR /app
COPY . .
RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
echo "LLAMA_SYCL_F16 is set" && \
export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
fi && \
cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
cmake --build build --config Release --target server
FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
rm /etc/apt/sources.list.d/intel-graphics.list && \
wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
chmod 644 /usr/share/keyrings/intel-graphics.gpg
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev
COPY --from=build /app/build/bin/server /server
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/server" ]

25
.devops/server.Dockerfile
View file

@ -1,25 +0,0 @@
ARG UBUNTU_VERSION=22.04
FROM ubuntu:$UBUNTU_VERSION as build
RUN apt-get update && \
apt-get install -y build-essential git libcurl4-openssl-dev
WORKDIR /app
COPY . .
ENV LLAMA_CURL=1
RUN make -j$(nproc) server
FROM ubuntu:$UBUNTU_VERSION as runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1
COPY --from=build /app/server /server
ENV LC_ALL=C.utf8
ENTRYPOINT [ "/server" ]

14
.devops/tools.sh
View file

@ -8,13 +8,11 @@ arg1="$1"
shift
if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
python3 ./convert-hf-to-gguf.py "$@"
python3 ./convert_hf_to_gguf.py "$@"
elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
./quantize "$@"
./llama-quantize "$@"
elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
./main "$@"
elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
./finetune "$@"
./llama-cli "$@"
elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
echo "Converting PTH to GGML..."
for i in `ls $1/$2/ggml-model-f16.bin*`; do
@ -22,11 +20,11 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
echo "Skip model quantization, it already exists: ${i/f16/q4_0}"
else
echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..."
./quantize "$i" "${i/f16/q4_0}" q4_0
./llama-quantize "$i" "${i/f16/q4_0}" q4_0
fi
done
elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
./server "$@"
./llama-server "$@"
else
echo "Unknown command: $arg1"
echo "Available commands: "
@ -36,8 +34,6 @@ else
echo " ex: --outtype f16 \"/models/7B/\" "
echo " --quantize (-q): Optimize with quantization process ggml"
echo " ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
echo " --finetune (-f): Run finetune command to create a lora finetune of the model"
echo " See documentation for finetune for command-line parameters"
echo " --all-in-one (-a): Execute --convert & --quantize"
echo " ex: \"/models/\" 7B"
echo " --server (-s): Run a model on the server"

4
.dockerignore
View file

@ -12,8 +12,8 @@ build*/
models/*
/main
/quantize
/llama-cli
/llama-quantize
arm_neon.h
compile_commands.json

4
.editorconfig
View file

@ -26,3 +26,7 @@ indent_size = 2
[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
indent_style = tab
[examples/cvector-generator/*.txt]
trim_trailing_whitespace = unset
insert_final_newline = unset

2
.github/ISSUE_TEMPLATE/01-bug-low.yml vendored
View file

@ -24,7 +24,7 @@ body:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:

2
.github/ISSUE_TEMPLATE/02-bug-medium.yml vendored
View file

@ -24,7 +24,7 @@ body:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:

2
.github/ISSUE_TEMPLATE/03-bug-high.yml vendored
View file

@ -24,7 +24,7 @@ body:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:

2
.github/ISSUE_TEMPLATE/04-bug-critical.yml vendored
View file

@ -24,7 +24,7 @@ body:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./main --version
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:

2
.github/ISSUE_TEMPLATE/config.yml vendored
View file

@ -9,5 +9,3 @@ contact_links:
- name: Want to contribute?
url: https://github.com/ggerganov/llama.cpp/wiki/contribute
about: Head to the contribution guide page of the wiki for areas you can help with

33
.github/labeler.yml vendored
View file

@ -2,31 +2,33 @@
Kompute:
- changed-files:
- any-glob-to-any-file:
- ggml-kompute.h
- ggml-kompute.cpp
- ggml/include/ggml-kompute.h
- ggml/src/ggml-kompute.cpp
- README-kompute.md
Apple Metal:
- changed-files:
- any-glob-to-any-file:
- ggml-metal.h
- ggml-metal.cpp
- ggml/include/ggml-metal.h
- ggml/src/ggml-metal.cpp
- README-metal.md
SYCL:
- changed-files:
- any-glob-to-any-file:
- ggml-sycl.h
- ggml-sycl.cpp
- README-sycl.md
- ggml/include/ggml-sycl.h
- ggml/src/ggml-sycl.cpp
- ggml/src/ggml-sycl/**
- docs/backend/SYCL.md
- examples/sycl/**
Nvidia GPU:
- changed-files:
- any-glob-to-any-file:
- ggml-cuda.h
- ggml-cuda/**
- ggml/include/ggml-cuda.h
- ggml/src/ggml-cuda/**
Vulkan:
- changed-files:
- any-glob-to-any-file:
- ggml_vk_generate_shaders.py
- ggml-vulkan*
- ggml/ggml_vk_generate_shaders.py
- ggml/src/ggml-vulkan*
documentation:
- changed-files:
- any-glob-to-any-file:
@ -42,7 +44,6 @@ build:
- cmake/**
- CMakeLists.txt
- CMakePresets.json
- codecov.yml
examples:
- changed-files:
- any-glob-to-any-file: examples/**
@ -74,10 +75,10 @@ server:
ggml:
- changed-files:
- any-glob-to-any-file:
- ggml.c
- ggml.h
- ggml-*.c
- ggml-*.h
- ggml/include/ggml*.h
- ggml/src/ggml*.c
- ggml/src/ggml*.cpp
- ggml/src/ggml*.h
- ggml-cuda/**
nix:
- changed-files:

12
.github/pull_request_template.md vendored
View file

@ -1,5 +1,7 @@
- Self Reported Review Complexity:
- [ ] Review Complexity : Low
- [ ] Review Complexity : Medium
- [ ] Review Complexity : High
- [ ] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
- Self-reported review complexity:
- [ ] Low
- [ ] Medium
- [ ] High

4
.github/workflows/bench.yml vendored
View file

@ -109,7 +109,7 @@ jobs:
run: |
set -eux
cmake -B build \
-DLLAMA_NATIVE=OFF \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DLLAMA_CUBLAS=ON \
@ -119,7 +119,7 @@ jobs:
-DLLAMA_FATAL_WARNINGS=OFF \
-DLLAMA_ALL_WARNINGS=OFF \
-DCMAKE_BUILD_TYPE=Release;
cmake --build build --config Release -j $(nproc) --target server
cmake --build build --config Release -j $(nproc) --target llama-server
- name: Download the dataset
id: download_dataset

98
.github/workflows/build.yml vendored
View file

@ -10,10 +10,10 @@ on:
push:
branches:
- master
paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m']
paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@ -47,7 +47,7 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -84,7 +84,7 @@ jobs:
name: llama-bin-macos-arm64.zip
macOS-latest-cmake-x64:
runs-on: macos-latest
runs-on: macos-12
steps:
- name: Clone
@ -103,12 +103,10 @@ jobs:
id: cmake_build
run: |
sysctl -a
mkdir build
cd build
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF -DLLAMA_CURL=ON ..
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
id: cmake_test
@ -224,7 +222,7 @@ jobs:
run: |
mkdir build
cd build
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
cmake --build . --config Release -j $(nproc)
- name: Test
@ -241,8 +239,8 @@ jobs:
wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
echo "Fetch llama2c model"
wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
- name: Determine tag name
id: tag
@ -307,7 +305,7 @@ jobs:
run: |
mkdir build
cd build
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_OPENMP=OFF
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DGGML_OPENMP=OFF
cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
- name: Test
@ -337,7 +335,7 @@ jobs:
run: |
mkdir build
cd build
cmake -DLLAMA_RPC=ON ..
cmake -DGGML_RPC=ON ..
cmake --build . --config Release -j $(nproc)
- name: Test
@ -357,15 +355,17 @@ jobs:
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential libvulkan-dev
wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
sudo apt-get update -y
sudo apt-get install -y build-essential vulkan-sdk
- name: Build
id: cmake_build
run: |
mkdir build
cd build
cmake -DLLAMA_VULKAN=ON ..
cmake -DGGML_VULKAN=ON ..
cmake --build . --config Release -j $(nproc)
ubuntu-22-cmake-hip:
@ -386,13 +386,13 @@ jobs:
- name: Build with native CMake HIP support
id: cmake_build
run: |
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DLLAMA_HIPBLAS=ON
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON
cmake --build build --config Release -j $(nproc)
- name: Build with legacy HIP support
id: cmake_build_legacy_hip
run: |
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DLLAMA_HIPBLAS=ON
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=ON
cmake --build build2 --config Release -j $(nproc)
ubuntu-22-cmake-sycl:
@ -433,7 +433,7 @@ jobs:
source /opt/intel/oneapi/setvars.sh
mkdir build
cd build
cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
cmake --build . --config Release -j $(nproc)
ubuntu-22-cmake-sycl-fp16:
@ -474,10 +474,10 @@ jobs:
source /opt/intel/oneapi/setvars.sh
mkdir build
cd build
cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON ..
cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON ..
cmake --build . --config Release -j $(nproc)
# TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# TODO: build with GGML_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# how to debug it.
# ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
macOS-latest-make:
@ -499,15 +499,15 @@ jobs:
env:
LLAMA_FATAL_WARNINGS: 1
run: |
LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
GGML_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
- name: Test
id: make_test
run: |
LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
LLAMA_NO_METAL=1 make test -j $(sysctl -n hw.logicalcpu)
GGML_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
GGML_NO_METAL=1 make test -j $(sysctl -n hw.logicalcpu)
# TODO: build with LLAMA_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# TODO: build with GGML_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# how to debug it.
# ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
# would be great if we fix these
@ -531,7 +531,7 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF ..
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -561,13 +561,14 @@ jobs:
mkdir build
cd build
cmake -G Xcode .. \
-DLLAMA_METAL_EMBED_LIBRARY=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_SYSTEM_NAME=iOS \
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
macOS-latest-cmake-tvos:
runs-on: macos-latest
@ -590,13 +591,14 @@ jobs:
mkdir build
cd build
cmake -G Xcode .. \
-DLLAMA_METAL_EMBED_LIBRARY=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
-DLLAMA_BUILD_SERVER=OFF \
-DCMAKE_SYSTEM_NAME=tvOS \
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
-DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
macOS-latest-swift:
runs-on: macos-latest
@ -664,7 +666,7 @@ jobs:
- name: Build using make w/ OpenBLAS
shell: msys2 {0}
run: |
make LLAMA_OPENBLAS=1 -j $(nproc)
make GGML_OPENBLAS=1 -j $(nproc)
- name: Build using CMake
shell: msys2 {0}
@ -680,7 +682,7 @@ jobs:
- name: Build using CMake w/ OpenBLAS
shell: msys2 {0}
run: |
cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
cmake --build build --config ${{ matrix.build }} -j $(nproc)
windows-latest-cmake:
@ -695,25 +697,25 @@ jobs:
matrix:
include:
- build: 'rpc-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_RPC=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
- build: 'noavx-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
- build: 'avx2-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
- build: 'avx-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
- build: 'avx512-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
- build: 'openblas-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
- build: 'kompute-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
- build: 'vulkan-x64'
defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
- build: 'llvm-arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
- build: 'msvc-arm64'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
steps:
- name: Clone
@ -726,7 +728,7 @@ jobs:
id: clone_kompute
if: ${{ matrix.build == 'kompute-x64' }}
run: |
git submodule update --init kompute
git submodule update --init ggml/src/kompute
- name: Download OpenBLAS
id: get_openblas
@ -799,6 +801,7 @@ jobs:
7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
$sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
cd build
$env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
& $sde -future -- ctest -L main -C Release --verbose --timeout 900
- name: Determine tag name
@ -856,7 +859,8 @@ jobs:
run: |
mkdir build
cd build
cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUDA=ON -DBUILD_SHARED_LIBS=ON
cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
- name: Determine tag name
@ -989,7 +993,7 @@ jobs:
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DLLAMA_HIPBLAS=ON
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
cmake --build build --config Release
ios-xcode-build:

40
.github/workflows/code-coverage.yml vendored
View file

@ -1,40 +0,0 @@
name: Code Coverage
on: [push, pull_request]
env:
GGML_NLOOP: 3
GGML_N_THREADS: 1
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
run:
runs-on: ubuntu-20.04
steps:
- name: Checkout
uses: actions/checkout@v4
- name: Dependencies
run: |
sudo apt-get update
sudo apt-get install build-essential gcc-8 lcov
- name: Build
run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
- name: Run tests
run: CC=gcc-8 make test
- name: Generate coverage report
run: |
make coverage
make lcov-report
- name: Upload coverage to Codecov
uses: codecov/codecov-action@v3
env:
CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
with:
files: lcov-report/coverage.info

27
.github/workflows/docker.yml vendored
View file

@ -10,10 +10,11 @@
name: Publish Docker image
on:
pull_request:
#pull_request:
push:
branches:
- master
paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@ -22,7 +23,7 @@ concurrency:
jobs:
push_to_registry:
name: Push Docker image to Docker Hub
if: github.event.pull_request.draft == false
#if: github.event.pull_request.draft == false
runs-on: ubuntu-latest
env:
@ -30,20 +31,18 @@ jobs:
strategy:
matrix:
config:
- { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "light", dockerfile: ".devops/llama-cli.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "server", dockerfile: ".devops/llama-server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
# NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
# have disabled them for now until the reason why
# is understood.
- { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
# Note: the full-rocm image is failing due to a "no space left on device" error. It is disabled for now to allow the workflow to complete.
#- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
steps:
- name: Check out the repo
uses: actions/checkout@v4

38
.github/workflows/python-type-check.yml vendored Normal file
View file

@ -0,0 +1,38 @@
name: Python Type-Check
on:
push:
paths:
- '.github/workflows/python-type-check.yml'
- '**.py'
- '**/requirements*.txt'
pull_request:
paths:
- '.github/workflows/python-type-check.yml'
- '**.py'
- '**/requirements*.txt'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
python-type-check:
runs-on: ubuntu-latest
name: pyright type-check
steps:
- name: Check out source repository
uses: actions/checkout@v4
- name: Set up Python environment
uses: actions/setup-python@v5
with:
python-version: "3.11"
- name: Install Python dependencies
# TODO: use a venv
run: pip install -r requirements/requirements-all.txt
- name: Type-check with Pyright
uses: jakebailey/pyright-action@v2
with:
version: 1.1.370
level: warning
warnings: true

26
.github/workflows/server.yml vendored
View file

@ -30,7 +30,7 @@ jobs:
strategy:
matrix:
sanitizer: [ADDRESS, THREAD, UNDEFINED]
sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
build_type: [RelWithDebInfo]
include:
- build_type: Release
@ -87,16 +87,30 @@ jobs:
exit 1
fi
- name: Build
id: cmake_build
- name: Build (no OpenMP)
id: cmake_build_no_openmp
if: ${{ matrix.sanitizer == 'THREAD' }}
run: |
cmake -B build \
-DLLAMA_NATIVE=OFF \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-DGGML_OPENMP=OFF ;
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
- name: Build
id: cmake_build
if: ${{ matrix.sanitizer != 'THREAD' }}
run: |
cmake -B build \
-DGGML_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DLLAMA_CURL=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server
cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
- name: Tests
id: server_integration_tests
@ -136,7 +150,7 @@ jobs:
id: cmake_build
run: |
cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
- name: Python setup
id: setup_python

163
.gitignore vendored
View file

@ -1,129 +1,132 @@
*.o
# Extensions
*.a
*.so
*.bat
*.bin
*.dll
*.dot
*.etag
*.exe
*.gcda
*.gcno
*.gcov
*.gguf
*.gguf.json
*.bin
*.exe
*.dll
*.log
*.gcov
*.gcno
*.gcda
*.dot
*.bat
*.tmp
*.metallib
*.etag
*.lastModified
.DS_Store
.build/
*.log
*.metallib
*.o
*.so
*.tmp
# IDE / OS
.cache/
.ccls-cache/
.direnv/
.DS_Store
.envrc
.idea/
.swiftpm
.venv
.clang-tidy
.vs/
.vscode/
.idea/
nppBackup
ggml-metal-embed.metal
lcov-report/
# Coverage
gcovr-report/
lcov-report/
# Build Artifacts
tags
.build/
build*
!build-info.cmake
!build-info.cpp.in
!build-info.sh
!build.zig
cmake-build-*
!docs/build.md
/libllama.so
/llama-*
/vulkan-shaders-gen
android-ndk-*
arm_neon.h
cmake-build-*
CMakeSettings.json
compile_commands.json
ggml-metal-embed.metal
llama-batched-swift
/rpc-server
out/
tmp/
# Deprecated
/main
/server
# CI
!.github/workflows/*.yml
# Models
models/*
models-mnt
!models/.editorconfig
!models/ggml-vocab-*.gguf*
/Pipfile
/baby-llama
/beam-search
/benchmark-matmult
/convert-llama2c-to-ggml
/embd-input-test
/embedding
/eval-callback
/gguf
/gguf-llama-simple
/gguf-split
/gritlm
/imatrix
/infill
/libllama.so
/llama-bench
/llava-cli
/lookahead
/lookup
/lookup-create
/lookup-merge
/lookup-stats
/main
/metal
/passkey
/perplexity
/q8dot
/quantize
/quantize-stats
/result
/save-load-state
/server
/simple
/batched
/batched-bench
/export-lora
/finetune
/retrieval
/speculative
/parallel
/train-text-from-scratch
/tokenize
/vdot
/common/build-info.cpp
arm_neon.h
compile_commands.json
CMakeSettings.json
__pycache__
dist
# Zig
zig-out/
zig-cache/
# Logs
ppl-*.txt
qnt-*.txt
perf-*.txt
# Examples
examples/jeopardy/results.txt
examples/server/*.css.hpp
examples/server/*.html.hpp
examples/server/*.js.hpp
examples/server/*.mjs.hpp
examples/server/*.css.hpp
!build_64.sh
!examples/*.bat
!examples/*/*.kts
!examples/*/*/*.kts
!examples/sycl/*.bat
!examples/sycl/*.sh
poetry.lock
# Python
/.venv
__pycache__/
*/poetry.lock
poetry.toml
nppBackup
# Nix
/result
# Test binaries
/tests/test-grammar-parser
/tests/test-llama-grammar
/tests/test-backend-ops
/tests/test-double-float
/tests/test-grad0
/tests/test-grammar-parser
/tests/test-llama-grammar
/tests/test-opt
/tests/test-quantize-fns
/tests/test-quantize-perf
/tests/test-rope
/tests/test-sampling
/tests/test-tokenizer-0
/tests/test-tokenizer-1-spm
/tests/test-tokenizer-1-bpe
/tests/test-rope
/tests/test-backend-ops
/tests/test-tokenizer-1-spm
# Scripts
!/scripts/install-oneapi.bat

2
.gitmodules vendored
View file

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

129
AUTHORS
View file

@ -1,8 +1,9 @@
# date: Tue Apr 9 09:17:14 EEST 2024
# date: Wed Jun 26 19:36:34 EEST 2024
# this file is auto-generated by scripts/gen-authors.sh
0cc4m <picard12@live.de>
0xspringtime <110655352+0xspringtime@users.noreply.github.com>
20kdc <asdd2808@gmail.com>
2f38b454 <dxf@protonmail.com>
3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com>
44670 <44670@users.noreply.github.com>
@ -11,14 +12,18 @@ AT <manyoso@users.noreply.github.com>
Aarni Koskela <akx@iki.fi>
Aaron Miller <apage43@ninjawhale.com>
Aaryaman Vasishta <aaryaman.vasishta@amd.com>
Abheek Gulati <abheekg@hotmail.com>
Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
Abhishek Gopinath K <31348521+overtunned@users.noreply.github.com>
Adithya Balaji <adithya.b94@gmail.com>
AdithyanI <adithyan.i4internet@gmail.com>
Adrian <smith.adriane@gmail.com>
Adrian Hesketh <a-h@users.noreply.github.com>
Ahmet Zeer <ahmed.zeer@std.yildiz.edu.tr>
AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
Aisuko <urakiny@gmail.com>
Akarshan Biswas <akarshanbiswas@fedoraproject.org>
Albert Jin <albert.jin@gmail.com>
Alberto <57916483+albbus-stack@users.noreply.github.com>
Alex <awhill19@icloud.com>
Alex Azarov <alex@azarov.by>
@ -35,19 +40,24 @@ Ali Nehzat <ali.nehzat@thanks.dev>
Ali Tariq <ali.tariq@10xengineers.ai>
Alon <alonfaraj@gmail.com>
AlpinDale <52078762+AlpinDale@users.noreply.github.com>
Amir <amir_zia@outlook.com>
AmirAli Mirian <37371367+amiralimi@users.noreply.github.com>
Ananta Bastola <anantarajbastola@gmail.com>
Anas Ahouzi <112881240+aahouzi@users.noreply.github.com>
András Salamon <ott2@users.noreply.github.com>
Andrei <abetlen@gmail.com>
Andrew Canis <andrew.canis@gmail.com>
Andrew Downing <andrew2085@gmail.com>
Andrew Duffy <a10y@users.noreply.github.com>
Andrew Godfrey <AndrewGodfrey@users.noreply.github.com>
Andy Tai <andy-tai@users.noreply.github.com>
Arik Poznanski <arikpoz@users.noreply.github.com>
Artem <guinmoon@gmail.com>
Artem Zinnatullin <ceo@abstractny.gay>
Artyom Lebedev <vagran.ast@gmail.com>
Asbjørn Olling <asbjornolling@gmail.com>
Ásgeir Bjarni Ingvarsson <asgeir@fundinn.org>
Ashish <1856117+ashishdatta@users.noreply.github.com>
Ashok Gelal <401055+ashokgelal@users.noreply.github.com>
Ashraful Islam <ashraful.meche@gmail.com>
Atsushi Tatsuma <yoshoku@outlook.com>
@ -57,35 +67,46 @@ BADR <contact@pythops.com>
Bach Le <bach@bullno1.com>
Bailey Chittle <39804642+bachittle@users.noreply.github.com>
BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com>
Bartowski <ckealty1182@gmail.com>
Behnam M <58621210+ibehnam@users.noreply.github.com>
Ben Ashbaugh <ben.ashbaugh@intel.com>
Ben Garney <bengarney@users.noreply.github.com>
Ben Siraphob <bensiraphob@gmail.com>
Ben Williams <ben@719ben.com>
Benjamin Findley <39356821+Kartoffelsaft@users.noreply.github.com>
Benjamin Lecaillon <84293038+blecaillon@users.noreply.github.com>
Bernat Vadell <hounter.caza@gmail.com>
Bingan <70050083+binganao@users.noreply.github.com>
Bodo Graumann <mail@bodograumann.de>
Bono Lv <lvscar@users.noreply.github.com>
Borislav Stanimirov <b.stanimirov@abv.bg>
Branden Butler <bwtbutler@hotmail.com>
Brian <mofosyne@gmail.com>
Bruce MacDonald <brucewmacdonald@gmail.com>
Bryan Honof <bryanhonof@gmail.com>
CJ Pais <cj@cjpais.com>
CRD716 <crd716@gmail.com>
Calvin Laurenson <calvin@laurenson.dev>
Cameron <csteele@steelecameron.com>
Cameron Kaiser <classilla@users.noreply.github.com>
Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
Casey Primozic <casey@cprimozic.net>
Casey Primozic <me@ameo.link>
CausalLM <148736309+CausalLM@users.noreply.github.com>
Cebtenzzre <cebtenzzre@gmail.com>
Chad Brewbaker <crb002@gmail.com>
Chao Jiang <jc19chaoj@zoho.com>
Cheng Shao <terrorjack@type.dance>
Chris Elrod <elrodc@gmail.com>
Chris Kuehl <ckuehl@ckuehl.me>
Christian Demsar <christian@github.email.demsar.us>
Christian Demsar <crasm@git.vczf.us>
Christian Falch <875252+chrfalch@users.noreply.github.com>
Christian Kögler <ck3d@gmx.de>
Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com>
Clark Saben <76020733+csaben@users.noreply.github.com>
Clint Herron <hanclinto@gmail.com>
CrispStrobe <154636388+CrispStrobe@users.noreply.github.com>
Cuong Trinh Manh <nguoithichkhampha@gmail.com>
DAN™ <dranger003@gmail.com>
Damian Stewart <d@damianstewart.com>
@ -95,8 +116,12 @@ Daniel Bevenius <daniel.bevenius@gmail.com>
Daniel Drake <drake@endlessos.org>
Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Daniel Illescas Romero <illescas.daniel@protonmail.com>
Daniele <57776841+daniandtheweb@users.noreply.github.com>
DannyDaemonic <DannyDaemonic@gmail.com>
Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com>
Dave <dave-fl@users.noreply.github.com>
Dave Airlie <airlied@gmail.com>
Dave Airlie <airlied@redhat.com>
Dave Della Costa <ddellacosta+github@gmail.com>
David Friehs <david@friehs.info>
David Kennedy <dakennedyd@gmail.com>
@ -104,10 +129,13 @@ David Pflug <david@pflug.email>
David Renshaw <dwrenshaw@gmail.com>
David Sommers <12738+databyte@users.noreply.github.com>
David Yang <davidyang6us@gmail.com>
Dawid Potocki <github@dawidpotocki.com>
Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
Dean <Dean.Sinaean@gmail.com>
Deins <deinsegle@gmail.com>
Deven Mistry <31466137+deven367@users.noreply.github.com>
Didzis Gosko <didzis@users.noreply.github.com>
Djip007 <djip.perois@free.fr>
Don Mahurin <dmahurin@users.noreply.github.com>
DooWoong Lee (David) <manics99@naver.com>
Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com>
@ -116,8 +144,11 @@ Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
Ebey Abraham <ebey97@gmail.com>
Ed Lee <edilee@mozilla.com>
Ed Lepedus <ed.lepedus@googlemail.com>
Eddie-Wang <wangjinheng1120@163.com>
Edward Taylor <edeetee@gmail.com>
Elaine <elaine.zosa@gmail.com>
Elbios <141279586+Elbios@users.noreply.github.com>
Elton Kola <eltonkola@gmail.com>
Engininja2 <139037756+Engininja2@users.noreply.github.com>
Equim <sayaka@ekyu.moe>
Eric Sommerlade <es0m@users.noreply.github.com>
@ -143,37 +174,47 @@ Firat <firatkiral@gmail.com>
Folko-Ven <71110216+Folko-Ven@users.noreply.github.com>
Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com>
Francisco Melo <43780565+francis2tm@users.noreply.github.com>
Frank Mai <thxcode0824@gmail.com>
FrankHB <frankhb1989@gmail.com>
Fred Douglas <43351173+fredlas@users.noreply.github.com>
Frederik Vogel <Schaltfehler@users.noreply.github.com>
Gabe Goodhart <gabe.l.hart@gmail.com>
GainLee <perfecter.gen@gmail.com>
Galunid <karolek1231456@gmail.com>
Gary Linscott <glinscott@gmail.com>
Gary Mulder <gjmulder@gmail.com>
Gavin Zhao <gavinzhaojw@protonmail.com>
Genkagaku.GPT <hlhr202@163.com>
Georgi Gerganov <ggerganov@gmail.com>
Gilad S <giladgd@users.noreply.github.com>
Giuseppe Scrivano <giuseppe@scrivano.org>
GiviMAD <GiviMAD@users.noreply.github.com>
Govlzkoy <gotope@users.noreply.github.com>
Guillaume "Vermeille" Sanchez <Guillaume.V.Sanchez@gmail.com>
Guillaume Wenzek <gwenzek@users.noreply.github.com>
Guoteng <32697156+SolenoidWGT@users.noreply.github.com>
Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com>
Haggai Nuchi <h.nuchi@gmail.com>
Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
Hamdoud Hakem <90524568+hamdoudhakem@users.noreply.github.com>
HanishKVC <hanishkvc@gmail.com>
Haohui Mai <ricetons@gmail.com>
Haoxiang Fei <tonyfettes@tonyfettes.com>
Harald Fernengel <harald.fernengel@here.com>
Hatsune Miku <129688334+at8u@users.noreply.github.com>
HatsuneMikuUwU33 <173229399+HatsuneMikuUwU33@users.noreply.github.com>
Henk Poley <HenkPoley@gmail.com>
Henri Vasserman <henv@hot.ee>
Henrik Forstén <henrik.forsten@gmail.com>
Herman Semenov <GermanAizek@yandex.ru>
Hesen Peng <hesen.peng@gmail.com>
Hoang Nguyen <hugo53@users.noreply.github.com>
Hong Bo PENG <penghb@cn.ibm.com>
Hongyu Ouyang <96765450+casavaca@users.noreply.github.com>
Howard Su <howard0su@gmail.com>
Hua Jiang <allenhjiang@outlook.com>
Huawei Lin <huaweilin.cs@gmail.com>
Hugo Roussel <hugo.rous@gmail.com>
Ian Bull <irbull@eclipsesource.com>
Ian Bull <irbull@gmail.com>
Ian Scrivener <github@zilogy.asia>
@ -190,8 +231,10 @@ Ivan Stepanov <ivanstepanovftw@gmail.com>
JH23X <165871467+JH23X@users.noreply.github.com>
Jack Mousseau <jmousseau@users.noreply.github.com>
JackJollimore <130917767+JackJollimore@users.noreply.github.com>
Jaemin Son <woalsdnd@gmail.com>
Jag Chadha <jagtesh@gmail.com>
Jakub N <jakubniemczyk97@gmail.com>
James A Capozzoli <157492257+jac-jim@users.noreply.github.com>
James Reynolds <magnusviri@users.noreply.github.com>
Jan Boon <jan.boon@kaetemi.be>
Jan Boon <kaetemi@gmail.com>
@ -205,12 +248,17 @@ Jean-Michaël Celerier <jeanmichael.celerier+github@gmail.com>
Jed Fox <git@jedfox.com>
Jeffrey Quesnelle <emozilla@nousresearch.com>
Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
Jeximo <jeximo@gmail.com>
Jhen-Jie Hong <iainst0409@gmail.com>
Jiahao Li <liplus17@163.com>
Jian Liao <jianliao@users.noreply.github.com>
JidongZhang-THU <1119708529@qq.com>
Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com>
Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
Jiří Sejkora <Sejseloid@gmail.com>
Joan Fontanals <jfontanalsmartinez@gmail.com>
Joan Fontanals <joan.fontanals.martinez@jina.ai>
Johan <JohanAR@users.noreply.github.com>
Johannes Gäßler <johannesg@5d6.de>
Johannes Rudolph <johannes.rudolph@gmail.com>
John <78893154+cmp-nct@users.noreply.github.com>
@ -221,15 +269,19 @@ Jonas Wunderlich <32615971+jonas-w@users.noreply.github.com>
Jorge A <161275481+jorgealias@users.noreply.github.com>
Jose Maldonado <63384398+yukiteruamano@users.noreply.github.com>
Joseph Stahl <1269177+josephst@users.noreply.github.com>
Josh Ramer <josh.ramer@icloud.com>
Joyce <joycebrum@google.com>
Juan Calderon-Perez <835733+gaby@users.noreply.github.com>
Judd <foldl@users.noreply.github.com>
Julius Arkenberg <arki05@users.noreply.github.com>
Jun Jie <71215065+junnjiee16@users.noreply.github.com>
Junyang Lin <justinlin930319@hotmail.com>
Juraj Bednar <juraj@bednar.io>
Justin Parker <jparkerweb@gmail.com>
Justin Suess <justin.suess@westpoint.edu>
Justina Cho <justcho5@gmail.com>
Justine Tunney <jtunney@gmail.com>
Justine Tunney <jtunney@mozilla.com>
Juuso Alasuutari <juuso.alasuutari@gmail.com>
KASR <karim.asrih@gmail.com>
Kamil Tomšík <info@tomsik.cz>
@ -242,6 +294,7 @@ Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Keiichi Tabata <keiichi.tabata@outlook.com>
Kenvix ⭐ <kenvixzure@live.com>
Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
Kevin Gibbons <bakkot@gmail.com>
Kevin Ji <1146876+kevinji@users.noreply.github.com>
Kevin Kwok <antimatter15@gmail.com>
Kevin Lo <kevlo@kevlo.org>
@ -257,6 +310,7 @@ Laura <Tijntje_7@msn.com>
Lee <44310445+lx200916@users.noreply.github.com>
Lee Drake <b.lee.drake@gmail.com>
Leng Yue <lengyue@lengyue.me>
Leon Knauer <git@leonknauer.com>
LeonEricsson <70749762+LeonEricsson@users.noreply.github.com>
Leonardo Neumann <leonardo@neumann.dev.br>
Li Tan <tanliboy@gmail.com>
@ -265,20 +319,26 @@ LoganDark <github@logandark.mozmail.com>
LostRuins <39025047+LostRuins@users.noreply.github.com>
Luciano <lucianostrika44@gmail.com>
Luo Tian <lt@basecity.com>
Lyle Dean <dean@lyle.dev>
M. Yusuf Sarıgöz <yusufsarigoz@gmail.com>
Maarten ter Huurne <maarten@treewalker.org>
Mack Straight <eiz@users.noreply.github.com>
Maël Kerbiriou <m431.kerbiriou@gmail.com>
MaggotHATE <clay1326@gmail.com>
Manuel <44313466+makuche@users.noreply.github.com>
Marc Köhlbrugge <subscriptions@marckohlbrugge.com>
Marco Matthies <71844+marcom@users.noreply.github.com>
Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com>
Marian Cepok <marian.cepok@gmail.com>
Mark Fairbairn <thebaron88@gmail.com>
Marko Tasic <mtasic85@gmail.com>
Markus Tavenrath <mtavenrath@users.noreply.github.com>
Martin Delille <martin@delille.org>
Martin Krasser <krasserm@googlemail.com>
Martin Schwaighofer <mschwaig@users.noreply.github.com>
Marvin Gießing <marvin.giessing@gmail.com>
Masaya, Kato <62578291+msy-kato@users.noreply.github.com>
MasterYi1024 <39848311+MasterYi1024@users.noreply.github.com>
Mateusz Charytoniuk <mateusz.charytoniuk@protonmail.com>
Matheus C. França <matheus-catarino@hotmail.com>
Matheus Gabriel Alves Silva <matheusgasource@gmail.com>
@ -287,8 +347,11 @@ Mathijs de Bruin <mathijs@mathijsfietst.nl>
Matt Clayton <156335168+mattjcly@users.noreply.github.com>
Matt Pulver <matt.pulver@heavy.ai>
Matteo Boschini <12133566+mbosc@users.noreply.github.com>
Mattheus Chediak <shammcity00@gmail.com>
Matthew Tejo <matthew.tejo@gmail.com>
Matvey Soloviev <blackhole89@gmail.com>
Max Krasnyansky <max.krasnyansky@gmail.com>
Max Krasnyansky <quic_maxk@quicinc.com>
Maxime <672982+maximegmd@users.noreply.github.com>
Maximilian Winter <maximilian.winter.91@gmail.com>
Meng Zhang <meng@tabbyml.com>
@ -300,32 +363,41 @@ Michael Kesper <mkesper@schokokeks.org>
Michael Klimenko <mklimenko29@gmail.com>
Michael Podvitskiy <podvitskiymichael@gmail.com>
Michael Potter <NanoTekGuy@Gmail.com>
Michael de Gans <michael.john.degans@gmail.com>
Michaël de Vries <vriesdemichael@gmail.com>
Mihai <mihai.chirculescu@yahoo.com>
Mike <ytianhui2004@gmail.com>
Mikko Juola <mikjuo@gmail.com>
Minsoo Cheong <54794500+mscheong01@users.noreply.github.com>
Mirko185 <mirkosig@gmail.com>
Mirror Azure <54669636+MirrorAzure@users.noreply.github.com>
Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com>
Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
Mohammadreza Hendiani <mohammad.r.hendiani@gmail.com>
Murilo Santana <mvrilo@gmail.com>
Musab Gultekin <musabgultekin@users.noreply.github.com>
Nam D. Tran <42194884+namtranase@users.noreply.github.com>
Nathan Epstein <nate2@umbc.edu>
NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
Nebula <infinitewormhole@gmail.com>
Neo Zhang <14088817+arthw@users.noreply.github.com>
Neo Zhang <zhang.jianyu@outlook.com>
Neo Zhang Jianyu <jianyu.zhang@intel.com>
Neuman Vong <neuman.vong@gmail.com>
Nexesenex <124105151+Nexesenex@users.noreply.github.com>
Niall Coates <1349685+Niall-@users.noreply.github.com>
Nicolai Weitkemper <kontakt@nicolaiweitkemper.de>
Nicolás Pérez <nicolas_perez@brown.edu>
Nigel Bosch <pnigelb@gmail.com>
Niklas Korz <niklas@niklaskorz.de>
Nikolas <127742645+nneubacher@users.noreply.github.com>
Nindaleth <Nindaleth@users.noreply.github.com>
Oleksandr Nikitin <oleksandr@tvori.info>
Oleksii Maryshchenko <oleksii.maryshchenko@gmail.com>
Olivier Chafik <ochafik@users.noreply.github.com>
Ondřej Čertík <ondrej@certik.us>
Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
Patrice Ferlet <metal3d@gmail.com>
Paul Tsochantaris <ptsochantaris@icloud.com>
Pavol Rusnak <pavol@rusnak.io>
Pedro Cuenca <pedro@huggingface.co>
@ -343,9 +415,14 @@ RJ Adriaansen <adriaansen@eshcc.eur.nl>
Radoslav Gerganov <rgerganov@gmail.com>
Radosław Gryta <radek.gryta@gmail.com>
Rahul Vivek Nair <68507071+RahulVivekNair@users.noreply.github.com>
Raj Hammeer Singh Hada <hammeerraj@gmail.com>
Ralph Soika <ralph.soika@imixs.com>
Rand Xie <randxiexyy29@gmail.com>
Randall Fitzgerald <randall@dasaku.net>
Reinforce-II <fate@eastal.com>
Ren Xuancheng <jklj077@users.noreply.github.com>
Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com>
RhinoDevel <RhinoDevel@users.noreply.github.com>
Riceball LEE <snowyu.lee@gmail.com>
Richard Kiss <him@richardkiss.com>
Richard Roberson <richardr1126@gmail.com>
@ -373,6 +450,7 @@ Rowan Hart <rowanbhart@gmail.com>
Rune <43761327+Rune-AI@users.noreply.github.com>
Ryan Landay <rlanday@gmail.com>
Ryder Wishart <ryderwishart@gmail.com>
Ryuei <louixs@users.noreply.github.com>
Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
SakuraUmi <yukinon244@gmail.com>
Salvador E. Tropea <stropea@inti.gob.ar>
@ -386,6 +464,7 @@ SebastianApel <13675545+SebastianApel@users.noreply.github.com>
Senemu <10880819+Senemu@users.noreply.github.com>
Sergey Alirzaev <zl29ah@gmail.com>
Sergio López <slp@sinrega.org>
Sertaç Özercan <852750+sozercan@users.noreply.github.com>
SeungWon Jeong <65549245+redlion0929@users.noreply.github.com>
ShadovvBeast <ShadovvBeast@gmail.com>
Shakhar Dasgupta <shakhardasgupta@gmail.com>
@ -394,6 +473,7 @@ Shijie <821898965@qq.com>
Shintarou Okada <kokuzen@gmail.com>
Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
Shouzheng Liu <lshzh.hi@gmail.com>
Shuichi Tsutsumi <shuichi0526@gmail.com>
Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Simon Willison <swillison@gmail.com>
Siwen Yu <yusiwen@gmail.com>
@ -405,11 +485,14 @@ Someone <sergei.kozlukov@aalto.fi>
Someone Serge <sergei.kozlukov@aalto.fi>
Sourab Mangrulkar <13534540+pacman100@users.noreply.github.com>
Spencer Sutton <spencersutton@users.noreply.github.com>
Srihari-mcw <96763064+Srihari-mcw@users.noreply.github.com>
Srinivas Billa <nivibilla@gmail.com>
Stefan Sydow <stefan@sydow.email>
Steffen Röcker <sroecker@gmail.com>
Stephan Walter <stephan@walter.name>
Stephen Nichols <snichols@users.noreply.github.com>
Steve Grubb <ausearch.1@gmail.com>
Steven Prichard <spprichard20@gmail.com>
Steven Roussey <sroussey@gmail.com>
Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
@ -434,16 +517,19 @@ Tom C <tom.corelis@gmail.com>
Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Tomas <tom.tomas.36478119@gmail.com>
Tomáš Pazdiora <tomas.pazdiora@gmail.com>
Tristan Druyen <tristan@vault81.mozmail.com>
Tristan Ross <rosscomputerguy@protonmail.com>
Tungsten842 <886724vf@anonaddy.me>
Tungsten842 <quantmint@protonmail.com>
Tushar <ditsuke@protonmail.com>
UEXTM.com <84163508+uextm@users.noreply.github.com>
Ulrich Drepper <drepper@gmail.com>
Uzo Nweke <uzoechi@gmail.com>
Vaibhav Srivastav <vaibhavs10@gmail.com>
Val Kharitonov <mail@kharvd.com>
Valentin Konovalov <valle.ketsujin@gmail.com>
Valentyn Bezshapkin <61702053+valentynbez@users.noreply.github.com>
Victor Nogueira <felladrin@gmail.com>
Victor Z. Peng <ziliangdotme@gmail.com>
Vlad <spitfireage@gmail.com>
Vladimir <bogdad@gmail.com>
@ -455,7 +541,9 @@ Weird Constructor <weirdconstructor@gmail.com>
Welby Seely <welbyseely@gmail.com>
Wentai Zhang <rchardx@gmail.com>
WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com>
William Tambellini <william.tambellini@gmail.com>
Willy Tarreau <w@1wt.eu>
Wouter <9594229+DifferentialityDevelopment@users.noreply.github.com>
Wu Jian Ping <wujjpp@hotmail.com>
Wu Jian Ping <wujp@greatld.com>
Xiake Sun <xiake.sun@intel.com>
@ -466,6 +554,8 @@ Xiaoyi Chen <cxychina@gmail.com>
Xingchen Song(宋星辰) <xingchensong1996@163.com>
Xuan Son Nguyen <thichthat@gmail.com>
Yann Follet <131855179+YannFollet@users.noreply.github.com>
Yaroslav <yaroslav.yashin@me.com>
Yazan Agha-Schrader <mountaiin@icloud.com>
Yiming Cui <conandiy@vip.qq.com>
Yishuo Wang <MeouSker77@outlook.com>
Yueh-Po Peng <94939112+y10ab1@users.noreply.github.com>
@ -477,6 +567,7 @@ Zane Shannon <z@zcs.me>
Zay <95888118+isaiahbjork@users.noreply.github.com>
Zenix <zenixls2@gmail.com>
Zhang Peiyuan <a1286225768@gmail.com>
Zheng.Deng <32841220+dengzheng-cloud@users.noreply.github.com>
ZhouYuChen <zhouyuchen@naver.com>
Ziad Ben Hadj-Alouane <zied.benhadjalouane@gmail.com>
Ziang Wu <97337387+ZiangWu-77@users.noreply.github.com>
@ -484,14 +575,18 @@ Zsapi <martin1.zsapka@gmail.com>
a-n-n-a-l-e-e <150648636+a-n-n-a-l-e-e@users.noreply.github.com>
adel boussaken <netdur@gmail.com>
afrideva <95653597+afrideva@users.noreply.github.com>
agray3 <agray3@users.noreply.github.com>
akawrykow <142945436+akawrykow@users.noreply.github.com>
alexpinel <93524949+alexpinel@users.noreply.github.com>
alonfaraj <alonfaraj@gmail.com>
alwqx <kenan3015@gmail.com>
amd-lalithnc <lalithnc@amd.com>
andrijdavid <david@geek.mg>
anon998 <131767832+anon998@users.noreply.github.com>
anzz1 <anzz1@live.com>
apaz <aarpazdera@gmail.com>
apcameron <37645737+apcameron@users.noreply.github.com>
arch-btw <57669023+arch-btw@users.noreply.github.com>
arcrank <arcrank@gmail.com>
arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
at8u <129688334+at8u@users.noreply.github.com>
@ -514,13 +609,17 @@ cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com>
coezbek <c.oezbek@gmail.com>
comex <comexk@gmail.com>
compilade <113953597+compilade@users.noreply.github.com>
compilade <git@compilade.net>
cpumaxx <163466046+cpumaxx@users.noreply.github.com>
crasm <crasm@git.vczf.net>
crasm <crasm@git.vczf.us>
daboe01 <daboe01@googlemail.com>
david raistrick <keen99@users.noreply.github.com>
ddh0 <dylanhalladay02@icloud.com>
ddpasa <112642920+ddpasa@users.noreply.github.com>
deepdiffuser <112834445+deepdiffuser@users.noreply.github.com>
divinity76 <divinity76@gmail.com>
dm4 <sunrisedm4@gmail.com>
dotpy314 <33351922+dotpy314@users.noreply.github.com>
drbh <david.richard.holtz@gmail.com>
ds5t5 <145942675+ds5t5@users.noreply.github.com>
@ -529,6 +628,7 @@ eastriver <lee@eastriver.dev>
ebraminio <ebraminio@gmail.com>
eiery <19350831+eiery@users.noreply.github.com>
eric8607242 <e0928021388@gmail.com>
fairydreaming <166155368+fairydreaming@users.noreply.github.com>
fraxy-v <65565042+fraxy-v@users.noreply.github.com>
github-actions[bot] <github-actions[bot]@users.noreply.github.com>
gliptic <gliptic@users.noreply.github.com>
@ -539,6 +639,7 @@ h-h-h-h <13482553+h-h-h-h@users.noreply.github.com>
hankcs <cnhankmc@gmail.com>
hoangmit <hoangmit@users.noreply.github.com>
hongbo.mo <352280764@qq.com>
hopkins385 <98618192+hopkins385@users.noreply.github.com>
howlger <eclipse@voormann.de>
howlger <github@voormann.de>
hutli <6594598+hutli@users.noreply.github.com>
@ -549,14 +650,22 @@ hydai <z54981220@gmail.com>
iSma <ismail.senhaji@gmail.com>
iacore <74560659+iacore@users.noreply.github.com>
igarnier <igarnier@protonmail.com>
intelmatt <61025942+intelmatt@users.noreply.github.com>
iohub <rickyang.pro@gmail.com>
jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com>
jaime-m-p <167997752+jaime-m-p@users.noreply.github.com>
jameswu2014 <545426914@qq.com>
jiez <373447296@qq.com>
jneem <joeneeman@gmail.com>
joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
johnson442 <56517414+johnson442@users.noreply.github.com>
jojorne <jojorne@users.noreply.github.com>
jon-chuang <9093549+jon-chuang@users.noreply.github.com>
jp-x-g <jpxg-dev@protonmail.com>
jukofyork <69222624+jukofyork@users.noreply.github.com>
junchao-loongson <68935141+junchao-loongson@users.noreply.github.com>
jwj7140 <32943891+jwj7140@users.noreply.github.com>
k.h.lai <adrian.k.h.lai@outlook.com>
kaizau <kaizau@users.noreply.github.com>
kalomaze <66376113+kalomaze@users.noreply.github.com>
kang <tpdns9032100@gmail.com>
@ -575,11 +684,15 @@ ldwang <ftgreat@163.com>
le.chang <cljs118@126.com>
leejet <leejet714@gmail.com>
limitedAtonement <limitedAtonement@users.noreply.github.com>
liuwei-git <14815172+liuwei-git@users.noreply.github.com>
lon <114724657+longregen@users.noreply.github.com>
loonerin <132926317+loonerin@users.noreply.github.com>
luoyu-intel <yu.luo@intel.com>
m3ndax <adrian.goessl@outlook.com>
maddes8cht <55592906+maddes8cht@users.noreply.github.com>
makomk <makosoft@googlemail.com>
manikbhandari <mbbhandarimanik2@gmail.com>
maor-ps <154728172+maor-ps@users.noreply.github.com>
mdrokz <mohammadmunshi@gmail.com>
mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com>
minarchist <minarchist@users.noreply.github.com>
@ -593,15 +706,19 @@ ngc92 <7938269+ngc92@users.noreply.github.com>
nhamanasu <45545786+nhamanasu@users.noreply.github.com>
niansa/tuxifan <anton-sa@web.de>
niansa/tuxifan <tuxifan@posteo.de>
nickp27 <nb.porter@gmail.com>
ningshanwutuobang <ningshanwutuobang@gmail.com>
nold <Nold360@users.noreply.github.com>
nopperl <54780682+nopperl@users.noreply.github.com>
nusu-github <29514220+nusu-github@users.noreply.github.com>
olexiyb <olexiyb@gmail.com>
omahs <73983677+omahs@users.noreply.github.com>
oobabooga <112222186+oobabooga@users.noreply.github.com>
opparco <parco.opaai@gmail.com>
ostix360 <55257054+ostix360@users.noreply.github.com>
pengxin99 <pengxin.yuan@intel.com>
perserk <perserk@gmail.com>
pmysl <piotr.myslinski@outlook.com>
postmasters <namnguyen@google.com>
pudepiedj <pudepiedj@gmail.com>
qingfengfenga <41416092+qingfengfenga@users.noreply.github.com>
@ -614,16 +731,19 @@ rhuddleston <ryan.huddleston@percona.com>
rimoliga <53384203+rimoliga@users.noreply.github.com>
runfuture <runfuture@users.noreply.github.com>
sandyiscool <sandyiscool@gmail.com>
sasha0552 <admin@sasha0552.org>
semidark <me@semidark.net>
sharpHL <132747147+sharpHL@users.noreply.github.com>
shibe2 <shibe@tuta.io>
singularity <12184989+singularity-s0@users.noreply.github.com>
sjinzh <sjinzh@gmail.com>
sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com>
slaren <2141330+slaren@users.noreply.github.com>
slaren <slarengh@gmail.com>
snadampal <87143774+snadampal@users.noreply.github.com>
staviq <staviq@gmail.com>
stduhpf <stephduh@live.fr>
strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com>
swittk <switt1995@gmail.com>
takov751 <40316768+takov751@users.noreply.github.com>
tarcey <cey.tarik@gmail.com>
@ -636,12 +756,16 @@ uint256_t <konndennsa@gmail.com>
uint256_t <maekawatoshiki1017@gmail.com>
unbounded <haakon@likedan.net>
valiray <133289098+valiray@users.noreply.github.com>
vik <vikhyatk@gmail.com>
viric <viric@viric.name>
vodkaslime <646329483@qq.com>
vvhg1 <94630311+vvhg1@users.noreply.github.com>
vxiiduu <73044267+vxiiduu@users.noreply.github.com>
wbpxre150 <100937007+wbpxre150@users.noreply.github.com>
whoreson <139810751+whoreson@users.noreply.github.com>
woachk <24752637+woachk@users.noreply.github.com>
wonjun Jang <strutive07@gmail.com>
woodx <124784234+woodx9@users.noreply.github.com>
wzy <32936898+Freed-Wu@users.noreply.github.com>
xaedes <xaedes@gmail.com>
xaedes <xaedes@googlemail.com>
@ -649,7 +773,10 @@ xloem <0xloem@gmail.com>
yangli2 <yangli2@gmail.com>
yuiseki <yuiseki@gmail.com>
zakkor <edward.partenie@gmail.com>
zhangkaihuo <zhangkaihuo@gmail.com>
zhouwg <6889919+zhouwg@users.noreply.github.com>
zhouwg <zhouwg2000@gmail.com>
zrm <trustiosity.zrm@gmail.com>
Ștefan-Gabriel Muscalu <legraphista@users.noreply.github.com>
源文雨 <41315874+fumiama@users.noreply.github.com>
Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>

1397
CMakeLists.txt
View file

File diff suppressed because it is too large Load diff

34
CMakePresets.json
View file

@ -11,10 +11,23 @@
"CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
}
},
{
"name": "sycl-base",
"hidden": true,
"generator": "Ninja",
"binaryDir": "${sourceDir}/build-${presetName}",
"cacheVariables": {
"CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
"CMAKE_CXX_COMPILER": "icx",
"CMAKE_C_COMPILER": "cl",
"GGML_SYCL": "ON",
"CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
}
},
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{
"name": "arm64-windows-msvc", "hidden": true,
@ -35,15 +48,18 @@
},
{ "name": "arm64-windows-llvm-debug" , "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "release" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "release", "static" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "release" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "release", "static" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] },
{ "name": "x64-windows-msvc-debug" , "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
{ "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
{ "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
{ "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
{ "name": "x64-windows-sycl-debug" , "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }
]
}

34
CONTRIBUTING.md
View file

@ -1,14 +1,28 @@
# Contributing Guidelines
# Pull requests (for contributors)
## Checklist
- Test your changes:
- Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
- Execute [the full CI locally on your machine](ci/README.md) before publishing
- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
- The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
* Make sure your PR follows the [coding guidelines](https://github.com/ggerganov/llama.cpp/blob/master/README.md#coding-guidelines)
* Test your changes using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
* Execute [the full CI locally on your machine](ci/README.md) before publishing
# Pull requests (for collaborators)
## PR formatting
- 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/ggerganov/llama.cpp/wiki/Modules
# Coding guidelines
- Avoid adding third-party dependencies, extra files, extra headers, etc.
- Always consider cross-compatibility with other operating systems and architectures
- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
![matmul](media/matmul.png)
* Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
- The PR template has a series of review complexity checkboxes `[ ]` that you can mark as `[X]` for your conveience. Refer to [About task lists](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) for more information.
* If the pull request only contains documentation changes (e.g., updating READMEs, adding new wiki pages), please add `[no ci]` to the commit title. This will skip unnecessary CI checks and help reduce build times.
* When squashing multiple commits on merge, use the following format for your commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : Fix typo in utils.py (#1234)`

1322
Makefile
View file

File diff suppressed because it is too large Load diff

25
Package.swift
View file

@ -3,14 +3,17 @@
import PackageDescription
var sources = [
"ggml.c",
"sgemm.cpp",
"llama.cpp",
"unicode.cpp",
"unicode-data.cpp",
"ggml-alloc.c",
"ggml-backend.c",
"ggml-quants.c",
"src/llama.cpp",
"src/llama-vocab.cpp",
"src/llama-grammar.cpp",
"src/llama-sampling.cpp",
"src/unicode.cpp",
"src/unicode-data.cpp",
"ggml/src/ggml.c",
"ggml/src/ggml-alloc.c",
"ggml/src/ggml-backend.c",
"ggml/src/ggml-quants.c",
"ggml/src/ggml-aarch64.c",
]
var resources: [Resource] = []
@ -26,8 +29,8 @@ var cSettings: [CSetting] = [
]
#if canImport(Darwin)
sources.append("ggml-metal.m")
resources.append(.process("ggml-metal.metal"))
sources.append("ggml/src/ggml-metal.m")
resources.append(.process("ggml/src/ggml-metal.metal"))
linkerSettings.append(.linkedFramework("Accelerate"))
cSettings.append(
contentsOf: [
@ -63,8 +66,6 @@ let package = Package(
"models",
"tests",
"CMakeLists.txt",
"ggml-cuda.cu",
"ggml-cuda.h",
"Makefile"
],
sources: sources,

771
README.md
View file

@ -3,15 +3,19 @@
![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
### Recent API changes
> [!IMPORTANT]
[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
## Recent API changes
- [2024 Jun 26] The source code and CMake build scripts have been restructured https://github.com/ggerganov/llama.cpp/pull/8006
- [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
- [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
@ -20,9 +24,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
### Hot topics
## Hot topics
- **`convert.py` has been deprecated and moved to `examples/convert-legacy-llama.py`, please use `convert-hf-to-gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
- BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
- MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
@ -35,37 +39,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
----
<details>
<summary>Table of Contents</summary>
<ol>
<li>
<a href="#description">Description</a>
</li>
<li>
<a href="#usage">Usage</a>
<ul>
<li><a href="#get-the-code">Get the Code</a></li>
<li><a href="#build">Build</a></li>
<li><a href="#blas-build">BLAS Build</a></li>
<li><a href="#prepare-and-quantize">Prepare and Quantize</a></li>
<li><a href="#run-the-quantized-model">Run the quantized model</a></li>
<li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
<li><a href="#quantization">Quantization</a></li>
<li><a href="#interactive-mode">Interactive mode</a></li>
<li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
<li><a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a></li>
<li><a href="#seminal-papers-and-background-on-the-models">Seminal papers and background on the models</a></li>
<li><a href="#perplexity-measuring-model-quality">Perplexity (measuring model quality)</a></li>
<li><a href="#android">Android</a></li>
<li><a href="#docker">Docker</a></li>
</ul>
</li>
<li><a href="#contributing">Contributing</a></li>
<li><a href="#coding-guidelines">Coding guidelines</a></li>
<li><a href="#docs">Docs</a></li>
</ol>
</details>
## Description
The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@ -83,14 +56,6 @@ Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomm
improved significantly thanks to many contributions. It is the main playground for developing new features for the
[ggml](https://github.com/ggerganov/ggml) library.
**Supported platforms:**
- [X] Mac OS
- [X] Linux
- [X] Windows (via CMake)
- [X] Docker
- [X] FreeBSD
**Supported models:**
Typically finetunes of the base models below are supported as well.
@ -104,6 +69,7 @@ Typically finetunes of the base models below are supported as well.
- [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
- [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
- [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423)
- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
- [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
@ -130,8 +96,9 @@ Typically finetunes of the base models below are supported as well.
- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
- [x] [OLMo](https://allenai.org/olmo)
- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
**Multimodal models:**
@ -145,12 +112,6 @@ Typically finetunes of the base models below are supported as well.
- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
**HTTP server**
[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
[simplechat](./examples/server/public_simplechat) is a simple chat client, which can be used to chat with the model exposed using above web server (use --path to point to simplechat), from a local web browser.
**Bindings:**
- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
@ -171,11 +132,13 @@ Typically finetunes of the base models below are supported as well.
- 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)
- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
**UI:**
Unless otherwise noted these projects are open-source with permissive licensing:
- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
- [iohub/collama](https://github.com/iohub/coLLaMA)
- [janhq/jan](https://github.com/janhq/jan) (AGPL)
- [nat/openplayground](https://github.com/nat/openplayground)
@ -192,6 +155,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
- [RAGNA Desktop](https://ragna.app/) (proprietary)
- [RecurseChat](https://recurse.chat/) (proprietary)
- [semperai/amica](https://github.com/semperai/amica)
- [withcatai/catai](https://github.com/withcatai/catai)
@ -205,19 +169,29 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
- [AIKit](https://github.com/sozercan/aikit) (MIT)
- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
**Tools:**
- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
---
**Infrastructure:**
Here is a typical run using LLaMA v2 13B on M2 Ultra:
- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
**Games:**
- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
## Demo
<details>
<summary>Typical run using LLaMA v2 13B on M2 Ultra</summary>
```
$ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
$ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
I llama.cpp build info:
I UNAME_S: Darwin
I UNAME_P: arm
@ -294,430 +268,85 @@ llama_print_timings: eval time = 24513.59 ms / 399 runs ( 61.44 ms
llama_print_timings: total time = 25431.49 ms
```
</details>
<details>
<summary>Demo of running both LLaMA-7B and whisper.cpp on a single M1 Pro MacBook</summary>
And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8b4f-add84093ffff.mp4
</details>
## Usage
Here are the end-to-end binary build and model conversion steps for most supported models.
### Get the Code
### Basic usage
Firstly, you need to get the binary. There are different methods that you can follow:
- Method 1: Clone this repository and build locally, see [how to build](./docs/build.md)
- Method 2: If you are using MacOS or Linux, you can install llama.cpp via [brew, flox or nix](./docs/install.md)
- Method 3: Use a Docker image, see [documentation for Docker](./docs/docker.md)
- Method 4: Download pre-built binary from [releases](https://github.com/ggerganov/llama.cpp/releases)
You can run a basic completion using this command:
```bash
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
llama-cli -m your_model.gguf -p "I believe the meaning of life is" -n 128
# Output:
# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
```
### Build
See [this page](./examples/main/README.md) for a full list of parameters.
In order to build llama.cpp you have four different options.
### Conversation mode
- Using `make`:
- On Linux or MacOS:
```bash
make
```
- On Windows:
1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
2. Extract `w64devkit` on your pc.
3. Run `w64devkit.exe`.
4. Use the `cd` command to reach the `llama.cpp` folder.
5. From here you can run:
```bash
make
```
- Notes:
- For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/).
- For debug builds, run `make LLAMA_DEBUG=1`
- Using `CMake`:
```bash
cmake -B build
cmake --build build --config Release
```
**Notes**:
- For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/).
- For debug builds, there are two cases:
1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
```bash
cmake -B build -DCMAKE_BUILD_TYPE=Debug
cmake --build build
```
2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
```bash
cmake -B build -G "Xcode"
cmake --build build --config Debug
```
- Using `gmake` (FreeBSD):
1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
2. Add your user to **video** group
3. Install compilation dependencies.
```bash
sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
```
### Homebrew
On Mac and Linux, the homebrew package manager can be used via
```
brew install llama.cpp
```
The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
### Metal Build
On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or the `LLAMA_METAL=OFF` cmake option.
When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
argument.
### BLAS Build
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
- #### Accelerate Framework:
This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
- #### OpenBLAS:
This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
- Using `make`:
- On Linux:
```bash
make LLAMA_OPENBLAS=1
```
- On Windows:
1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
3. Extract `w64devkit` on your pc.
4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
6. Run `w64devkit.exe`.
7. Use the `cd` command to reach the `llama.cpp` folder.
8. From here you can run:
```bash
make LLAMA_OPENBLAS=1
```
- Using `CMake` on Linux:
```bash
cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
cmake --build build --config Release
```
- #### BLIS
Check [BLIS.md](docs/BLIS.md) for more information.
- #### SYCL
SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
For detailed info, please refer to [llama.cpp for SYCL](README-sycl.md).
- #### Intel oneMKL
Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./README-sycl.md).
- Using manual oneAPI installation:
By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
```bash
source /opt/intel/oneapi/setvars.sh # You can skip this step if in oneapi-basekit docker image, only required for manual installation
cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
cmake --build build --config Release
```
- Using oneAPI docker image:
If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
- #### CUDA
This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
- Using `make`:
```bash
make LLAMA_CUDA=1
```
- Using `CMake`:
```bash
cmake -B build -DLLAMA_CUDA=ON
cmake --build build --config Release
```
The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
| Option | Legal values | Default | Description |
|--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| LLAMA_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
| LLAMA_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
| LLAMA_CUDA_FORCE_MMQ | Boolean | false | Force the use of dequantization + matrix multiplication kernels instead of leveraging Math libraries. | |
| LLAMA_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
| LLAMA_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
| LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
| LLAMA_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. |
- #### hipBLAS
This provides BLAS acceleration on HIP-supported AMD GPUs.
Make sure to have ROCm installed.
You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
- Using `make`:
```bash
make LLAMA_HIPBLAS=1
```
- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build --config Release -- -j 16
```
On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON`.
However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
Note that if you get the following error:
```
clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
```
Try searching for a directory under `HIP_PATH` that contains the file
`oclc_abi_version_400.bc`. Then, add the following to the start of the
command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
like:
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build -- -j 16
```
- Using `make` (example for target gfx1030, build with 16 CPU threads):
```bash
make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
```
- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
```bash
set PATH=%HIP_PATH%\bin;%PATH%
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake --build build
```
Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
| Option | Legal values | Default | Description |
|-------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| LLAMA_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
| LLAMA_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
- #### Vulkan
**With docker**:
You don't need to install Vulkan SDK. It will be installed inside the container.
```sh
# Build the image
docker build -t llama-cpp-vulkan -f .devops/main-vulkan.Dockerfile .
# Then, use it:
docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
```
**Without docker**:
Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
For example, on Ubuntu 22.04 (jammy), use the command below:
```bash
wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
apt update -y
apt-get install -y vulkan-sdk
# To verify the installation, use the command below:
vulkaninfo
```
Alternatively your package manager might be able to provide the appropiate libraries. For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
Then, build llama.cpp using the cmake command below:
```bash
cmake -B build -DLLAMA_VULKAN=1
cmake --build build --config Release
# Test the output binary (with "-ngl 33" to offload all layers to GPU)
./bin/main -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
# You should see in the output, ggml_vulkan detected your GPU. For example:
# ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
```
### Prepare and Quantize
> [!NOTE]
> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter:
```bash
# obtain the official LLaMA model weights and place them in ./models
ls ./models
llama-2-7b tokenizer_checklist.chk tokenizer.model
# [Optional] for models using BPE tokenizers
ls ./models
<folder containing weights and tokenizer json> vocab.json
# [Optional] for PyTorch .bin models like Mistral-7B
ls ./models
<folder containing weights and tokenizer json>
llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv
# install Python dependencies
python3 -m pip install -r requirements.txt
# convert the model to ggml FP16 format
python3 convert-hf-to-gguf.py models/mymodel/
# [Optional] for models using BPE tokenizers
python convert-hf-to-gguf.py models/mymodel/ --vocab-type bpe
# quantize the model to 4-bits (using Q4_K_M method)
./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
# update the gguf filetype to current version if older version is now unsupported
./quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
# Output:
# > hi, who are you?
# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
#
# > what is 1+1?
# Easy peasy! The answer to 1+1 is... 2!
```
### Run the quantized model
By default, the chat template will be taken from the input model. If you want to use another chat template, pass `--chat-template NAME` as a parameter. See the list of [supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
```bash
# start inference on a gguf model
./main -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
```
When running the larger models, make sure you have enough disk space to store all the intermediate files.
You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters:
### Running on Windows with prebuilt binaries
You will find prebuilt Windows binaries on the release page.
Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
```
.\main -m llama-2-7b.Q4_0.gguf -n 128
```bash
./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
```
### Memory/Disk Requirements
### Web server
As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
[llama.cpp web server](./examples/server/README.md) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
| Model | Original size | Quantized size (Q4_0) |
|------:|--------------:|----------------------:|
| 7B | 13 GB | 3.9 GB |
| 13B | 24 GB | 7.8 GB |
| 30B | 60 GB | 19.5 GB |
| 65B | 120 GB | 38.5 GB |
Example usage:
### Quantization
```bash
./llama-server -m your_model.gguf --port 8080
Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
*(outdated)*
| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 |
|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
| 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
| 7B | file size | 13.0G | 3.5G | 3.9G | 4.3G | 4.7G | 6.7G |
| 7B | ms/tok @ 4th | 127 | 55 | 54 | 76 | 83 | 72 |
| 7B | ms/tok @ 8th | 122 | 43 | 45 | 52 | 56 | 67 |
| 7B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 |
| 13B | perplexity | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
| 13B | file size | 25.0G | 6.8G | 7.6G | 8.3G | 9.1G | 13G |
| 13B | ms/tok @ 4th | - | 103 | 105 | 148 | 160 | 131 |
| 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 |
| 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 |
- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
- recent k-quants improvements and new i-quants
- [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
- [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
- [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
- [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
- [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
- [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
- [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
- [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
- [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
- [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
- [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
- [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
- [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
- [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
### Perplexity (measuring model quality)
You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
#### How to run
1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
3. Output:
# Basic web UI can be accessed via browser: http://localhost:8080
# Chat completion endpoint: http://localhost:8080/v1/chat/completions
```
perplexity : calculating perplexity over 655 chunks
24.43 seconds per pass - ETA 4.45 hours
[1]4.5970,[2]5.1807,[3]6.0382,...
```
And after 4.45 hours, you will have the final perplexity.
### Interactive mode
If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
> [!NOTE]
> If you prefer basic usage, please consider using conversation mode instead of interactive mode
In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
Here is an example of a few-shot interaction, invoked with the command
@ -730,16 +359,16 @@ Here is an example of a few-shot interaction, invoked with the command
./examples/chat-13B.sh
# custom arguments using a 13B model
./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
```
Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `main` example program.
Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `llama-cli` example program.
![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
### Persistent Interaction
The prompt, user inputs, and model generations can be saved and resumed across calls to `./main` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
The prompt, user inputs, and model generations can be saved and resumed across calls to `./llama-cli` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
```bash
# Start a new chat
@ -761,25 +390,78 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
`llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
```bash
./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
```
The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
### Obtaining and using the Facebook LLaMA 2 model
## Build
- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
- [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF)
- [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF)
- [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF)
- [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF)
- [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
- [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
Please refer to [Build llama.cpp locally](./docs/build.md)
### Seminal papers and background on the models
## Supported backends
| Backend | Target devices |
| --- | --- |
| [Metal](./docs/build.md#metal-build) | Apple Silicon |
| [BLAS](./docs/build.md#blas-build) | All |
| [BLIS](./docs/backend/BLIS.md) | All |
| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
| [CUDA](./docs/build.md#cuda) | Nvidia GPU |
| [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
| [Vulkan](./docs/build.md#vulkan) | GPU |
## Tools
### Prepare and Quantize
> [!NOTE]
> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face.
To learn more about quantizing model, [read this documentation](./examples/quantize/README.md)
### Perplexity (measuring model quality)
You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
To learn more how to measure perplexity using llama.cpp, [read this documentation](./examples/perplexity/README.md)
## 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
- Any help with managing issues and PRs is very appreciated!
- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
## Other documentations
- [main (cli)](./examples/main/README.md)
- [server](./examples/server/README.md)
- [jeopardy](./examples/jeopardy/README.md)
- [GBNF grammars](./grammars/README.md)
**Development documentations**
- [How to build](./docs/build.md)
- [Running on Docker](./docs/docker.md)
- [Build on Android](./docs/android.md)
- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md)
- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
**Seminal papers and background on the models**
If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
- LLaMA:
@ -790,178 +472,3 @@ If your issue is with model generation quality, then please at least scan the fo
- GPT-3.5 / InstructGPT / ChatGPT:
- [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
### Android
#### Build on Android using Termux
[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
```
apt update && apt upgrade -y
apt install git make cmake
```
It's recommended to move your model inside the `~/` directory for best performance:
```
cd storage/downloads
mv model.gguf ~/
```
[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
#### Building the Project using Android NDK
Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
```
$ mkdir build-android
$ cd build-android
$ export NDK=<your_ndk_directory>
$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
$ make
```
Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
```
$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
$cd /data/data/com.termux/files/home/bin
$chmod +x ./*
```
Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
```
$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
```
Now, you can start chatting:
```
$cd /data/data/com.termux/files/home/bin
$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
```
Here's a demo of an interactive session running on Pixel 5 phone:
https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
### Docker
#### Prerequisites
* Docker must be installed and running on your system.
* Create a folder to store big models & intermediate files (ex. /llama/models)
#### Images
We have three Docker images available for this project:
1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
Additionally, there the following images, similar to the above:
- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
#### Usage
The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
Replace `/path/to/models` below with the actual path where you downloaded the models.
```bash
docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
```
On completion, you are ready to play!
```bash
docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
```
or with a light image:
```bash
docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
```
or with a server image:
```bash
docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
```
### Docker With CUDA
Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
#### Building Locally
```bash
docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
```
You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
The defaults are:
- `CUDA_VERSION` set to `11.7.1`
- `CUDA_DOCKER_ARCH` set to `all`
The resulting images, are essentially the same as the non-CUDA images:
1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
#### Usage
After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
```bash
docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
```
### 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
- Any help with managing issues and PRs is very appreciated!
- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
### Coding guidelines
- Avoid adding third-party dependencies, extra files, extra headers, etc.
- Always consider cross-compatibility with other operating systems and architectures
- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
![matmul](media/matmul.png)
### Docs
- [main](./examples/main/README.md)
- [server](./examples/server/README.md)
- [jeopardy](./examples/jeopardy/README.md)
- [BLIS](./docs/BLIS.md)
- [Performance troubleshooting](./docs/token_generation_performance_tips.md)
- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
- [GBNF grammars](./grammars/README.md)

262
ci/run.sh
View file

@ -36,11 +36,11 @@ SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
fi
if [ ! -z ${GG_BUILD_CUDA} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUDA=1"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
fi
if [ ! -z ${GG_BUILD_SYCL} ]; then
@ -50,7 +50,7 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
exit 1
fi
CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
fi
## helpers
@ -103,6 +103,9 @@ function gg_run_ctest_debug {
set -e
# Check cmake, make and ctest are installed
gg_check_build_requirements
(time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
@ -131,6 +134,9 @@ function gg_run_ctest_release {
set -e
# Check cmake, make and ctest are installed
gg_check_build_requirements
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
@ -284,10 +290,10 @@ function gg_run_open_llama_7b_v2 {
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@ -303,47 +309,47 @@ function gg_run_open_llama_7b_v2 {
wiki_test="${path_wiki}/wiki.test.raw"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/quantize ${model_f16} ${model_q4_0} q4_0
./bin/quantize ${model_f16} ${model_q4_1} q4_1
./bin/quantize ${model_f16} ${model_q5_0} q5_0
./bin/quantize ${model_f16} ${model_q5_1} q5_1
./bin/quantize ${model_f16} ${model_q2_k} q2_k
./bin/quantize ${model_f16} ${model_q3_k} q3_k
./bin/quantize ${model_f16} ${model_q4_k} q4_k
./bin/quantize ${model_f16} ${model_q5_k} q5_k
./bin/quantize ${model_f16} ${model_q6_k} q6_k
./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/main --model ${model_f16} -t 1 -ngl 999 -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/main --model ${model_q8_0} -t 1 -ngl 999 -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/main --model ${model_q4_0} -t 1 -ngl 999 -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/main --model ${model_q4_1} -t 1 -ngl 999 -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/main --model ${model_q5_0} -t 1 -ngl 999 -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/main --model ${model_q5_1} -t 1 -ngl 999 -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/main --model ${model_q2_k} -t 1 -ngl 999 -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/main --model ${model_q3_k} -t 1 -ngl 999 -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/main --model ${model_q4_k} -t 1 -ngl 999 -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/main --model ${model_q5_k} -t 1 -ngl 999 -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/main --model ${model_q6_k} -t 1 -ngl 999 -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 --model ${model_f16} -t 1 -ngl 999 -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 --model ${model_q8_0} -t 1 -ngl 999 -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 --model ${model_q4_0} -t 1 -ngl 999 -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 --model ${model_q4_1} -t 1 -ngl 999 -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 --model ${model_q5_0} -t 1 -ngl 999 -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 --model ${model_q5_1} -t 1 -ngl 999 -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 --model ${model_q2_k} -t 1 -ngl 999 -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 --model ${model_q3_k} -t 1 -ngl 999 -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 --model ${model_q4_k} -t 1 -ngl 999 -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 --model ${model_q5_k} -t 1 -ngl 999 -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 --model ${model_q6_k} -t 1 -ngl 999 -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/perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -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} -t 1 -ngl 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -421,7 +427,7 @@ function gg_run_pythia_1_4b {
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@ -437,45 +443,45 @@ function gg_run_pythia_1_4b {
wiki_test_60="${path_wiki}/wiki.test-60.raw"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/quantize ${model_f16} ${model_q4_0} q4_0
./bin/quantize ${model_f16} ${model_q4_1} q4_1
./bin/quantize ${model_f16} ${model_q5_0} q5_0
./bin/quantize ${model_f16} ${model_q5_1} q5_1
./bin/quantize ${model_f16} ${model_q2_k} q2_k
./bin/quantize ${model_f16} ${model_q3_k} q3_k
./bin/quantize ${model_f16} ${model_q4_k} q4_k
./bin/quantize ${model_f16} ${model_q5_k} q5_k
./bin/quantize ${model_f16} ${model_q6_k} q6_k
./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/main --model ${model_f16} -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/main --model ${model_q8_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/main --model ${model_q4_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/main --model ${model_q4_1} -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/main --model ${model_q5_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/main --model ${model_q5_1} -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/main --model ${model_q2_k} -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/main --model ${model_q3_k} -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/main --model ${model_q4_k} -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/main --model ${model_q5_k} -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/main --model ${model_q6_k} -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-cli --model ${model_f16} -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 --model ${model_q8_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 --model ${model_q4_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 --model ${model_q4_1} -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 --model ${model_q5_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 --model ${model_q5_1} -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 --model ${model_q2_k} -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 --model ${model_q3_k} -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 --model ${model_q4_k} -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 --model ${model_q5_k} -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 --model ${model_q6_k} -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/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -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} -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} -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} -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} -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} -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} -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} -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} -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} -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} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -550,10 +556,10 @@ function gg_run_pythia_2_8b {
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@ -569,47 +575,47 @@ function gg_run_pythia_2_8b {
wiki_test="${path_wiki}/wiki.test.raw"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/quantize ${model_f16} ${model_q4_0} q4_0
./bin/quantize ${model_f16} ${model_q4_1} q4_1
./bin/quantize ${model_f16} ${model_q5_0} q5_0
./bin/quantize ${model_f16} ${model_q5_1} q5_1
./bin/quantize ${model_f16} ${model_q2_k} q2_k
./bin/quantize ${model_f16} ${model_q3_k} q3_k
./bin/quantize ${model_f16} ${model_q4_k} q4_k
./bin/quantize ${model_f16} ${model_q5_k} q5_k
./bin/quantize ${model_f16} ${model_q6_k} q6_k
./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/main --model ${model_f16} -t 1 -ngl 999 -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/main --model ${model_q8_0} -t 1 -ngl 999 -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/main --model ${model_q4_0} -t 1 -ngl 999 -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/main --model ${model_q4_1} -t 1 -ngl 999 -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/main --model ${model_q5_0} -t 1 -ngl 999 -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/main --model ${model_q5_1} -t 1 -ngl 999 -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/main --model ${model_q2_k} -t 1 -ngl 999 -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/main --model ${model_q3_k} -t 1 -ngl 999 -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/main --model ${model_q4_k} -t 1 -ngl 999 -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/main --model ${model_q5_k} -t 1 -ngl 999 -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/main --model ${model_q6_k} -t 1 -ngl 999 -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 --model ${model_f16} -t 1 -ngl 999 -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 --model ${model_q8_0} -t 1 -ngl 999 -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 --model ${model_q4_0} -t 1 -ngl 999 -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 --model ${model_q4_1} -t 1 -ngl 999 -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 --model ${model_q5_0} -t 1 -ngl 999 -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 --model ${model_q5_1} -t 1 -ngl 999 -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 --model ${model_q2_k} -t 1 -ngl 999 -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 --model ${model_q3_k} -t 1 -ngl 999 -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 --model ${model_q4_k} -t 1 -ngl 999 -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 --model ${model_q5_k} -t 1 -ngl 999 -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 --model ${model_q6_k} -t 1 -ngl 999 -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/perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -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} -t 1 -ngl 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -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 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -688,19 +694,33 @@ function gg_run_embd_bge_small {
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
(time ./bin/embedding --model ${model_f16} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
set +e
}
function gg_check_build_requirements {
if ! command -v cmake &> /dev/null; then
gg_printf 'cmake not found, please install'
fi
if ! command -v make &> /dev/null; then
gg_printf 'make not found, please install'
fi
if ! command -v ctest &> /dev/null; then
gg_printf 'ctest not found, please install'
fi
}
function gg_sum_embd_bge_small {
gg_printf '### %s\n\n' "${ci}"

0
scripts/build-info.cmake → cmake/build-info.cmake
View file

22
cmake/git-vars.cmake Normal file
View file

@ -0,0 +1,22 @@
find_package(Git)
# the commit's SHA1
execute_process(COMMAND
"${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
OUTPUT_VARIABLE GIT_SHA1
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
# the date of the commit
execute_process(COMMAND
"${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
OUTPUT_VARIABLE GIT_DATE
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
# the subject of the commit
execute_process(COMMAND
"${GIT_EXECUTABLE}" log -1 --format=%s
WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)

61
scripts/LlamaConfig.cmake.in → cmake/llama-config.cmake.in
View file

@ -1,53 +1,82 @@
set(LLAMA_VERSION @LLAMA_INSTALL_VERSION@)
set(LLAMA_VERSION @LLAMA_INSTALL_VERSION@)
set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(LLAMA_BLAS @LLAMA_BLAS@)
set(LLAMA_CUDA @LLAMA_CUDA@)
set(LLAMA_METAL @LLAMA_METAL@)
set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@)
set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@)
set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(GGML_BLAS @GGML_BLAS@)
set(GGML_CUDA @GGML_CUDA@)
set(GGML_METAL @GGML_METAL@)
set(GGML_HIPBLAS @GGML_HIPBLAS@)
set(GGML_ACCELERATE @GGML_ACCELERATE@)
set(GGML_VULKAN @GGML_VULKAN@)
set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
set(GGML_SYCL @GGML_SYCL@)
set(GGML_OPENMP @GGML_OPENMP@)
@PACKAGE_INIT@
set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
# Ensure transient dependencies satisfied
find_package(Threads REQUIRED)
if (APPLE AND LLAMA_ACCELERATE)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
endif()
if (LLAMA_BLAS)
if (GGML_BLAS)
find_package(BLAS REQUIRED)
endif()
if (LLAMA_CUDA)
if (GGML_CUDA)
find_package(CUDAToolkit REQUIRED)
endif()
if (LLAMA_METAL)
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
find_library(METAL_FRAMEWORK Metal REQUIRED)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
endif()
if (LLAMA_HIPBLAS)
if (GGML_VULKAN)
find_package(Vulkan REQUIRED)
endif()
if (GGML_HIPBLAS)
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
endif()
if (GGML_SYCL)
find_package(IntelSYCL REQUIRED)
find_package(MKL REQUIRED)
endif()
if (GGML_OPENMP)
find_package(OpenMP REQUIRED)
endif()
find_library(ggml_LIBRARY ggml
REQUIRED
HINTS ${LLAMA_LIB_DIR})
find_library(llama_LIBRARY llama
REQUIRED
HINTS ${LLAMA_LIB_DIR})
set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@")
set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
add_library(llama UNKNOWN IMPORTED)
set_target_properties(llama
PROPERTIES
INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"

14
codecov.yml
View file

@ -1,14 +0,0 @@
comment: off
coverage:
status:
project:
default:
target: auto
threshold: 0
base: auto
patch:
default:
target: auto
threshold: 0
base: auto

7
common/CMakeLists.txt
View file

@ -1,5 +1,6 @@
# common
find_package(Threads REQUIRED)
# Build info header
#
@ -36,7 +37,7 @@ add_custom_command(
COMMENT "Generating build details from Git"
COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
-DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/gen-build-info-cpp.cmake"
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
VERBATIM
@ -83,5 +84,5 @@ if (LLAMA_CURL)
endif ()
target_include_directories(${TARGET} PUBLIC .)
target_compile_features(${TARGET} PUBLIC cxx_std_11)
target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
target_compile_features (${TARGET} PUBLIC cxx_std_11)
target_link_libraries (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)

4
scripts/gen-build-info-cpp.cmake → common/cmake/build-info-gen-cpp.cmake
View file

@ -1,7 +1,7 @@
include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
set(OUTPUT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
set(OUTPUT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
# Only write the build info if it changed
if(EXISTS ${OUTPUT_FILE})

1201
common/common.cpp
View file

File diff suppressed because it is too large Load diff

79
common/common.h
View file

@ -52,6 +52,12 @@ int32_t cpu_get_num_math();
// CLI argument parsing
//
// dimensionality reduction methods, used by cvector-generator
enum dimre_method {
DIMRE_METHOD_PCA,
DIMRE_METHOD_MEAN,
};
struct gpt_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed
@ -73,7 +79,6 @@ struct gpt_params {
int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
int32_t n_beams = 0; // if non-zero then use beam search of given width.
int32_t grp_attn_n = 1; // group-attention factor
int32_t grp_attn_w = 512; // group-attention width
int32_t n_print = -1; // print token count every n tokens (-1 = disabled)
@ -94,6 +99,7 @@ struct gpt_params {
enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
// // sampling parameters
struct llama_sampling_params sparams;
@ -102,6 +108,7 @@ struct gpt_params {
std::string model_draft = ""; // draft model for speculative decoding
std::string model_alias = "unknown"; // model alias
std::string model_url = ""; // model url to download
std::string hf_token = ""; // HF token
std::string hf_repo = ""; // HF repo
std::string hf_file = ""; // HF file
std::string prompt = "";
@ -121,7 +128,6 @@ struct gpt_params {
// TODO: avoid tuple, use struct
std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
std::string lora_base = ""; // base model path for the lora adapter
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
@ -153,7 +159,6 @@ struct gpt_params {
bool prompt_cache_all = false; // save user input and generations to prompt cache
bool prompt_cache_ro = false; // open the prompt cache read-only and do not update it
bool embedding = false; // get only sentence embedding
bool escape = true; // escape "\n", "\r", "\t", "\'", "\"", and "\\"
bool multiline_input = false; // reverse the usage of `\`
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
@ -180,6 +185,12 @@ struct gpt_params {
std::string mmproj = ""; // path to multimodal projector
std::vector<std::string> image; // path to image file(s)
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
std::string embd_out = ""; // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
std::string embd_sep = "\n"; // separator of embendings
// server params
int32_t port = 8080; // server listens on this network port
int32_t timeout_read = 600; // http read timeout in seconds
@ -190,6 +201,7 @@ struct gpt_params {
std::string public_path = "";
std::string chat_template = "";
std::string system_prompt = "";
bool enable_chat_template = true;
std::vector<std::string> api_keys;
@ -232,8 +244,21 @@ struct gpt_params {
bool process_output = false; // collect data for the output tensor
bool compute_ppl = true; // whether to compute perplexity
// cvector-generator params
int n_pca_batch = 100;
int n_pca_iterations = 1000;
dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
std::string cvector_outfile = "control_vector.gguf";
std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
bool spm_infill = false; // suffix/prefix/middle pattern for infill
std::string lora_outfile = "ggml-lora-merged-f16.gguf";
};
void gpt_params_handle_hf_token(gpt_params & params);
void gpt_params_handle_model_default(gpt_params & params);
bool gpt_params_parse_ex (int argc, char ** argv, gpt_params & params);
@ -289,8 +314,8 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params);
struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
// Batch utils
@ -328,21 +353,13 @@ std::string llama_token_to_piece(
llama_token token,
bool special = true);
// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
// that takes into account the tokenizer type and decides how to handle the leading space
//
// detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode`
// removes the leading space from the first non-BOS token
std::string llama_detokenize_spm(
// optionally renders special/control tokens
std::string llama_detokenize(
llama_context * ctx,
const std::vector<llama_token> & tokens);
// detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode`
std::string llama_detokenize_bpe(
llama_context * ctx,
const std::vector<llama_token> & tokens);
const std::vector<llama_token> & tokens,
bool special = true);
// Uses the value from the model metadata if possible, otherwise
// defaults to true when model type is SPM, otherwise false.
@ -352,9 +369,34 @@ bool llama_should_add_bos_token(const llama_model * model);
// Chat template utils
//
// same with llama_chat_message, but uses std::string
struct llama_chat_msg {
std::string role;
std::string content;
};
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool llama_chat_verify_template(const std::string & tmpl);
// CPP wrapper for llama_chat_apply_template
// If the built-in template is not supported, we default to chatml
// If the custom "tmpl" is not supported, we throw an error
std::string llama_chat_apply_template(const struct llama_model * model,
const std::string & tmpl,
const std::vector<llama_chat_msg> & chat,
bool add_ass);
// Format single message, while taking into account the position of that message in chat history
std::string llama_chat_format_single(const struct llama_model * model,
const std::string & tmpl,
const std::vector<llama_chat_msg> & past_msg,
const llama_chat_msg & new_msg,
bool add_ass);
// Returns an example of formatted chat
std::string llama_chat_format_example(const struct llama_model * model,
const std::string & tmpl);
//
// KV cache utils
//
@ -369,7 +411,7 @@ void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_siz
// Embedding utils
//
void llama_embd_normalize(const float * inp, float * out, int n);
void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
@ -413,4 +455,3 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
void yaml_dump_non_result_info(
FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

351
common/json-schema-to-grammar.cpp
View file

@ -40,6 +40,233 @@ static std::string build_repetition(const std::string & item_rule, int min_items
return result;
}
/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
class string_view {
const std::string & _str;
const size_t _start;
const size_t _end;
public:
string_view(const std::string & str, size_t start = 0, size_t end = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
size_t size() const {
return _end - _start;
}
size_t length() const {
return size();
}
operator std::string() const {
return str();
}
std::string str() const {
return _str.substr(_start, _end - _start);
}
string_view substr(size_t pos, size_t len = std::string::npos) const {
return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
}
char operator[](size_t pos) const {
auto index = _start + pos;
if (index >= _end) {
throw std::out_of_range("string_view index out of range");
}
return _str[_start + pos];
}
bool operator==(const string_view & other) const {
std::string this_str = *this;
std::string other_str = other;
return this_str == other_str;
}
};
static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
auto has_min = min_value != std::numeric_limits<int>::min();
auto has_max = max_value != std::numeric_limits<int>::max();
auto digit_range = [&](char from, char to) {
out << "[";
if (from == to) {
out << from;
} else {
out << from << "-" << to;
}
out << "]";
};
auto more_digits = [&](int min_digits, int max_digits) {
out << "[0-9]";
if (min_digits == max_digits && min_digits == 1) {
return;
}
out << "{";
out << min_digits;
if (max_digits != min_digits) {
out << ",";
if (max_digits != std::numeric_limits<int>::max()) {
out << max_digits;
}
}
out << "}";
};
std::function<void(const string_view &, const string_view &)> uniform_range =
[&](const string_view & from, const string_view & to) {
size_t i = 0;
while (i < from.length() && i < to.length() && from[i] == to[i]) {
i++;
}
if (i > 0) {
out << "\"" << from.substr(0, i).str() << "\"";
}
if (i < from.length() && i < to.length()) {
if (i > 0) {
out << " ";
}
auto sub_len = from.length() - i - 1;
if (sub_len > 0) {
auto from_sub = from.substr(i + 1);
auto to_sub = to.substr(i + 1);
auto sub_zeros = repeat("0", sub_len);
auto sub_nines = repeat("9", sub_len);
auto to_reached = false;
out << "(";
if (from_sub == sub_zeros) {
digit_range(from[i], to[i] - 1);
out << " ";
more_digits(sub_len, sub_len);
} else {
out << "[" << from[i] << "] ";
out << "(";
uniform_range(from_sub, sub_nines);
out << ")";
if (from[i] < to[i] - 1) {
out << " | ";
if (to_sub == sub_nines) {
digit_range(from[i] + 1, to[i]);
to_reached = true;
} else {
digit_range(from[i] + 1, to[i] - 1);
}
out << " ";
more_digits(sub_len, sub_len);
}
}
if (!to_reached) {
out << " | ";
digit_range(to[i], to[i]);
out << " ";
uniform_range(sub_zeros, to_sub);
}
out << ")";
} else {
out << "[" << from[i] << "-" << to[i] << "]";
}
}
};
if (has_min && has_max) {
if (min_value < 0 && max_value < 0) {
out << "\"-\" (";
_build_min_max_int(-max_value, -min_value, out, decimals_left, /* top_level= */ true);
out << ")";
return;
}
if (min_value < 0) {
out << "\"-\" (";
_build_min_max_int(0, -min_value, out, decimals_left, /* top_level= */ true);
out << ") | ";
min_value = 0;
}
auto min_s = std::to_string(min_value);
auto max_s = std::to_string(max_value);
auto min_digits = min_s.length();
auto max_digits = max_s.length();
for (auto digits = min_digits; digits < max_digits; digits++) {
uniform_range(min_s, repeat("9", digits));
min_s = "1" + repeat("0", digits);
out << " | ";
}
uniform_range(min_s, max_s);
return;
}
auto less_decimals = std::max(decimals_left - 1, 1);
if (has_min) {
if (min_value < 0) {
out << "\"-\" (";
_build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
out << ") | [0] | [1-9] ";
more_digits(0, decimals_left - 1);
} else if (min_value == 0) {
if (top_level) {
out << "[0] | [1-9] ";
more_digits(0, less_decimals);
} else {
more_digits(1, decimals_left);
}
} else if (min_value <= 9) {
char c = '0' + min_value;
auto range_start = top_level ? '1' : '0';
if (c > range_start) {
digit_range(range_start, c - 1);
out << " ";
more_digits(1, less_decimals);
out << " | ";
}
digit_range(c, '9');
out << " ";
more_digits(0, less_decimals);
} else {
auto min_s = std::to_string(min_value);
auto len = min_s.length();
auto c = min_s[0];
if (c > '1') {
digit_range(top_level ? '1' : '0', c - 1);
out << " ";
more_digits(len, less_decimals);
out << " | ";
}
digit_range(c, c);
out << " (";
_build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
out << ")";
if (c < '9') {
out << " | ";
digit_range(c + 1, '9');
out << " ";
more_digits(len - 1, less_decimals);
}
}
return;
}
if (has_max) {
if (max_value >= 0) {
if (top_level) {
out << "\"-\" [1-9] ";
more_digits(0, less_decimals);
out << " | ";
}
_build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
} else {
out << "\"-\" (";
_build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
out << ")";
}
return;
}
throw std::runtime_error("At least one of min_value or max_value must be set");
}
const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
struct BuiltinRule {
@ -89,7 +316,7 @@ std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
};
std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
template <typename Iterator>
std::string join(Iterator begin, Iterator end, const std::string & separator) {
@ -160,7 +387,6 @@ static std::string format_literal(const std::string & literal) {
return "\"" + escaped + "\"";
}
class SchemaConverter {
private:
std::function<json(const std::string &)> _fetch_json;
@ -388,6 +614,75 @@ private:
return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
}
/*
Returns a rule that matches a JSON string that is none of the provided strings
not_strings({"a"})
-> ["] ( [a] char+ | [^"a] char* )? ["] space
not_strings({"and", "also"})
-> ["] ( [a] ([l] ([s] ([o] char+ | [^"o] char*) | [^"s] char*) | [n] ([d] char+ | [^"d] char*) | [^"ln] char*) | [^"a] char* )? ["] space
*/
std::string _not_strings(const std::vector<std::string> & strings) {
struct TrieNode {
std::map<char, TrieNode> children;
bool is_end_of_string;
TrieNode() : is_end_of_string(false) {}
void insert(const std::string & string) {
auto node = this;
for (char c : string) {
node = &node->children[c];
}
node->is_end_of_string = true;
}
};
TrieNode trie;
for (const auto & s : strings) {
trie.insert(s);
}
std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
std::ostringstream out;
out << "[\"] ( ";
std::function<void(const TrieNode &)> visit = [&](const TrieNode & node) {
std::ostringstream rejects;
auto first = true;
for (const auto & kv : node.children) {
rejects << kv.first;
if (first) {
first = false;
} else {
out << " | ";
}
out << "[" << kv.first << "]";
if (!kv.second.children.empty()) {
out << " (";
visit(kv.second);
out << ")";
} else if (kv.second.is_end_of_string) {
out << " " << char_rule << "+";
}
}
if (!node.children.empty()) {
if (!first) {
out << " | ";
}
out << "[^\"" << rejects.str() << "] " << char_rule << "*";
}
};
visit(trie);
out << " )";
if (!trie.is_end_of_string) {
out << "?";
}
out << " [\"] space";
return out.str();
}
std::string _resolve_ref(const std::string & ref) {
std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
@ -408,6 +703,7 @@ private:
std::vector<std::string> required_props;
std::vector<std::string> optional_props;
std::unordered_map<std::string, std::string> prop_kv_rule_names;
std::vector<std::string> prop_names;
for (const auto & kv : properties) {
const auto &prop_name = kv.first;
const auto &prop_schema = kv.second;
@ -422,11 +718,18 @@ private:
} else {
optional_props.push_back(prop_name);
}
prop_names.push_back(prop_name);
}
if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get<bool>())) {
if ((additional_properties.is_boolean() && additional_properties.get<bool>()) || additional_properties.is_object()) {
std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
std::string kv_rule = _add_rule(sub_name + "-kv", _add_primitive("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
std::string value_rule =
additional_properties.is_object() ? visit(additional_properties, sub_name + "-value")
: _add_primitive("value", PRIMITIVE_RULES.at("value"));
auto key_rule =
prop_names.empty() ? _add_primitive("string", PRIMITIVE_RULES.at("string"))
: _add_rule(sub_name + "-k", _not_strings(prop_names));
std::string kv_rule = _add_rule(sub_name + "-kv", key_rule + " \":\" space " + value_rule);
prop_kv_rule_names["*"] = kv_rule;
optional_props.push_back("*");
}
@ -452,15 +755,11 @@ private:
}
std::string k = ks[0];
std::string kv_rule_name = prop_kv_rule_names[k];
if (k == "*") {
res = _add_rule(
name + (name.empty() ? "" : "-") + "additional-kvs",
kv_rule_name + " ( \",\" space " + kv_rule_name + " )*"
);
} else if (first_is_optional) {
res = "( \",\" space " + kv_rule_name + " )?";
std::string comma_ref = "( \",\" space " + kv_rule_name + " )";
if (first_is_optional) {
res = comma_ref + (k == "*" ? "*" : "?");
} else {
res = kv_rule_name;
res = kv_rule_name + (k == "*" ? " " + comma_ref + "*" : "");
}
if (ks.size() > 1) {
res += " " + _add_rule(
@ -594,17 +893,19 @@ public:
} else if (schema_type.is_array()) {
std::vector<json> schema_types;
for (const auto & t : schema_type) {
schema_types.push_back({{"type", t}});
json schema_copy(schema);
schema_copy["type"] = t;
schema_types.push_back(schema_copy);
}
return _add_rule(rule_name, _generate_union_rule(name, schema_types));
} else if (schema.contains("const")) {
return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
return _add_rule(rule_name, _generate_constant_rule(schema["const"]) + " space");
} else if (schema.contains("enum")) {
std::vector<std::string> enum_values;
for (const auto & v : schema["enum"]) {
enum_values.push_back(_generate_constant_rule(v));
}
return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | "));
return _add_rule(rule_name, "(" + join(enum_values.begin(), enum_values.end(), " | ") + ") space");
} else if ((schema_type.is_null() || schema_type == "object")
&& (schema.contains("properties") ||
(schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
@ -686,6 +987,24 @@ public:
int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
} else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
int min_value = std::numeric_limits<int>::min();
int max_value = std::numeric_limits<int>::max();
if (schema.contains("minimum")) {
min_value = schema["minimum"].get<int>();
} else if (schema.contains("exclusiveMinimum")) {
min_value = schema["exclusiveMinimum"].get<int>() + 1;
}
if (schema.contains("maximum")) {
max_value = schema["maximum"].get<int>();
} else if (schema.contains("exclusiveMaximum")) {
max_value = schema["exclusiveMaximum"].get<int>() - 1;
}
std::stringstream out;
out << "(";
_build_min_max_int(min_value, max_value, out);
out << ") space";
return _add_rule(rule_name, out.str());
} else if (schema.empty() || schema_type == "object") {
return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
} else {

2
common/log.h
View file

@ -630,7 +630,7 @@ inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
buf << "[ ";
bool first = true;
for (const auto &token : tokens)
for (const auto & token : tokens)
{
if (!first) {
buf << ", ";

13
common/ngram-cache.h
View file

@ -37,11 +37,18 @@ struct llama_ngram {
}
};
struct llama_token_hash_function {
size_t operator()(const llama_token token) const {
// see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
return token * 11400714819323198485llu;
}
};
struct llama_ngram_hash_function {
size_t operator()(const llama_ngram & ngram) const {
size_t hash = 0;
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
hash ^= llama_token_hash_function{}(ngram.tokens[i]);
}
return hash;
}

29
common/sampling.cpp
View file

@ -28,9 +28,13 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
result->grammar = llama_grammar_init(
struct llama_grammar * grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
if (grammar == nullptr) {
throw std::runtime_error("Failed to initialize llama_grammar");
}
result->grammar = grammar;
}
result->prev.resize(params.n_prev);
@ -59,9 +63,13 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
if (!ctx->parsed_grammar.rules.empty()) {
std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
ctx->grammar = llama_grammar_init(
struct llama_grammar * grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
if (grammar == nullptr) {
throw std::runtime_error("Failed to initialize llama_grammar");
}
ctx->grammar = grammar;
}
std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
@ -274,8 +282,6 @@ static llama_token llama_sampling_sample_impl(
GGML_ASSERT(!original_logits.empty());
}
llama_token id = 0;
// Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx);
if (temp < 0.0) {
// greedy sampling, with probs
@ -316,12 +322,15 @@ static llama_token llama_sampling_sample_impl(
}
if (ctx_sampling->grammar != NULL && !is_resampling) {
// Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx);
// Create an array with a single token data element for the sampled id
llama_token_data single_token_data = {id, logits[id], 0.0f};
llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
// Apply grammar constraints to the single token
llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array);
// Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
@ -369,7 +378,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
if (ctx_sampling->grammar != NULL && !apply_grammar) {
GGML_ASSERT(original_logits != NULL);
// Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
*original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
*original_logits = {logits, logits + n_vocab};
}
// apply params.logit_bias map
@ -382,10 +391,10 @@ static llama_token_data_array llama_sampling_prepare_impl(
llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
}
cur.clear();
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
}
llama_token_data_array cur_p = { cur.data(), cur.size(), false };
@ -412,7 +421,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
// apply grammar checks before sampling logic
if (apply_grammar && ctx_sampling->grammar != NULL) {
llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);
}
return cur_p;
@ -446,6 +455,6 @@ void llama_sampling_accept(
ctx_sampling->prev.push_back(id);
if (ctx_sampling->grammar != NULL && apply_grammar) {
llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id);
}
}

1273
convert-hf-to-gguf.py → convert_hf_to_gguf.py
View file

File diff suppressed because it is too large Load diff

72
convert-hf-to-gguf-update.py → convert_hf_to_gguf_update.py
View file

@ -2,7 +2,7 @@
# -*- coding: utf-8 -*-
# This script downloads the tokenizer models of the specified models from Huggingface and
# generates the get_vocab_base_pre() function for convert-hf-to-gguf.py
# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
#
# This is necessary in order to analyze the type of pre-tokenizer used by the model and
# provide the necessary information to llama.cpp via the GGUF header in order to implement
@ -15,9 +15,9 @@
# - Add a new model to the "models" list
# - Run the script with your huggingface token:
#
# python3 convert-hf-to-gguf-update.py <huggingface_token>
# python3 convert_hf_to_gguf_update.py <huggingface_token>
#
# - Copy-paste the generated get_vocab_base_pre() function into convert-hf-to-gguf.py
# - Copy-paste the generated get_vocab_base_pre() function into convert_hf_to_gguf.py
# - Update llama.cpp with the new pre-tokenizer if necessary
#
# TODO: generate tokenizer tests for llama.cpp
@ -37,7 +37,7 @@ from enum import IntEnum, auto
from transformers import AutoTokenizer
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger("convert-hf-to-gguf-update")
logger = logging.getLogger("convert_hf_to_gguf_update")
sess = requests.Session()
@ -45,20 +45,21 @@ class TOKENIZER_TYPE(IntEnum):
SPM = auto()
BPE = auto()
WPM = auto()
UGM = auto()
# TODO: this string has to exercise as much pre-tokenizer functionality as possible
# will be updated with time - contributions welcome
chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
if len(sys.argv) == 2:
token = sys.argv[1]
if not token.startswith("hf_"):
logger.info("Huggingface token seems invalid")
logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
sys.exit(1)
else:
logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
sys.exit(1)
# TODO: add models here, base models preferred
@ -83,7 +84,16 @@ models = [
{"name": "jina-v2-es", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
{"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
{"name": "smaug-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
{"name": "poro-chat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
{"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
{"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
{"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
{"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
{"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
{"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
{"name": "codeshell", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
{"name": "tekken", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
{"name": "smollm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
]
@ -92,8 +102,8 @@ def download_file_with_auth(url, token, save_path):
response = sess.get(url, headers=headers)
response.raise_for_status()
os.makedirs(os.path.dirname(save_path), exist_ok=True)
with open(save_path, 'wb') as f:
f.write(response.content)
with open(save_path, 'wb') as downloaded_file:
downloaded_file.write(response.content)
logger.info(f"File {save_path} downloaded successfully")
@ -105,9 +115,13 @@ def download_model(model):
os.makedirs(f"models/tokenizers/{name}", exist_ok=True)
files = ["config.json", "tokenizer.json", "tokenizer_config.json"]
if tokt == TOKENIZER_TYPE.SPM:
files.append("tokenizer.model")
if tokt == TOKENIZER_TYPE.UGM:
files.append("spiece.model")
for file in files:
save_path = f"models/tokenizers/{name}/{file}"
if os.path.isfile(save_path):
@ -123,14 +137,14 @@ for model in models:
logger.error(f"Failed to download model {model['name']}. Error: {e}")
# generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function:
# generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
src_ifs = ""
for model in models:
name = model["name"]
tokt = model["tokt"]
if tokt == TOKENIZER_TYPE.SPM:
if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
continue
# Skip if the tokenizer folder does not exist or there are other download issues previously
@ -140,12 +154,15 @@ for model in models:
# create the tokenizer
try:
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
if name == "t5":
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
else:
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
except OSError as e:
logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
continue # Skip to the next model if the tokenizer can't be loaded
chktok = tokenizer.encode(chktxt)
chktok = tokenizer.encode(CHK_TXT)
chkhsh = sha256(str(chktok).encode()).hexdigest()
logger.info(f"model: {name}")
@ -177,7 +194,7 @@ src_func = f"""
# we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
# use in llama.cpp to implement the same pre-tokenizer
chktxt = {repr(chktxt)}
chktxt = {repr(CHK_TXT)}
chktok = tokenizer.encode(chktxt)
chkhsh = sha256(str(chktok).encode()).hexdigest()
@ -187,7 +204,7 @@ src_func = f"""
res = None
# NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script
# NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
# or pull the latest version of the model from Huggingface
# don't edit the hashes manually!
{src_ifs}
@ -196,9 +213,9 @@ src_func = f"""
logger.warning("**************************************************************************************")
logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!")
logger.warning("** There are 2 possible reasons for this:")
logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet")
logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet")
logger.warning("** - the pre-tokenization config has changed upstream")
logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.")
logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.")
logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920")
logger.warning("**")
logger.warning(f"** chkhsh: {{chkhsh}}")
@ -212,8 +229,8 @@ src_func = f"""
return res
"""
convert_py_pth = pathlib.Path("convert-hf-to-gguf.py")
convert_py = convert_py_pth.read_text()
convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
convert_py = convert_py_pth.read_text(encoding="utf-8")
convert_py = re.sub(
r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
lambda m: m.group(1) + src_func + m.group(3),
@ -221,9 +238,9 @@ convert_py = re.sub(
flags=re.DOTALL | re.MULTILINE,
)
convert_py_pth.write_text(convert_py)
convert_py_pth.write_text(convert_py, encoding="utf-8")
logger.info("+++ convert-hf-to-gguf.py was updated")
logger.info("+++ convert_hf_to_gguf.py was updated")
# generate tests for each tokenizer model
@ -261,6 +278,7 @@ tests = [
"\n =",
"' era",
"Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
"!!!!!!",
"3",
"33",
"333",
@ -270,8 +288,9 @@ tests = [
"3333333",
"33333333",
"333333333",
# "Cửa Việt", # llama-bpe fails on this
chktxt,
"Cửa Việt", # llama-bpe fails on this
" discards",
CHK_TXT,
]
# write the tests to ./models/ggml-vocab-{name}.gguf.inp
@ -298,7 +317,10 @@ for model in models:
# create the tokenizer
try:
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
if name == "t5":
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
else:
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
except OSError as e:
logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
continue # Skip this model and continue with the next one in the loop
@ -324,6 +346,6 @@ logger.info("\nRun the following commands to generate the vocab files for testin
for model in models:
name = model["name"]
print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100
print(f"python3 convert_hf_to_gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100
logger.info("\n")

9
convert-llama-ggml-to-gguf.py → convert_llama_ggml_to_gguf.py
View file

@ -132,6 +132,10 @@ class Tensor:
class GGMLModel:
file_format: GGMLFormat
format_version: int
def __init__(self):
self.hyperparameters = None
self.vocab = None
@ -290,7 +294,7 @@ class GGMLToGGUF:
if self.vocab_override is not None:
vo = self.vocab_override
logger.info('* Adding vocab item(s)')
for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
tokens.append(vbytes)
scores.append(score)
toktypes.append(ttype)
@ -354,7 +358,8 @@ class GGMLToGGUF:
def handle_metadata(cfg, hp):
import convert
import examples.convert_legacy_llama as convert
assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory'
hf_config_path = cfg.model_metadata_dir / "config.json"
orig_config_path = cfg.model_metadata_dir / "params.json"

393
convert_lora_to_gguf.py Executable file
View file

@ -0,0 +1,393 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
from __future__ import annotations
from dataclasses import dataclass
import logging
import argparse
import os
import sys
import json
from math import prod
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
import torch
if TYPE_CHECKING:
from torch import Tensor
if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
import gguf
# reuse model definitions from convert_hf_to_gguf.py
from convert_hf_to_gguf import LazyTorchTensor, Model
logger = logging.getLogger("lora-to-gguf")
@dataclass
class PartialLoraTensor:
A: Tensor | None = None
B: Tensor | None = None
# magic to support tensor shape modifications and splitting
class LoraTorchTensor:
_lora_A: Tensor # (n_rank, row_size)
_lora_B: Tensor # (col_size, n_rank)
_rank: int
def __init__(self, A: Tensor, B: Tensor):
assert len(A.shape) == len(B.shape)
assert A.shape[-2] == B.shape[-1]
if A.dtype != B.dtype:
A = A.to(torch.float32)
B = B.to(torch.float32)
self._lora_A = A
self._lora_B = B
self._rank = B.shape[-1]
def get_lora_A_B(self) -> tuple[Tensor, Tensor]:
return (self._lora_A, self._lora_B)
def __getitem__(
self,
indices: (
SupportsIndex
| slice
| tuple[SupportsIndex | slice | Tensor, ...] # TODO: add ellipsis in the type signature
),
) -> LoraTorchTensor:
shape = self.shape
if isinstance(indices, SupportsIndex):
if len(shape) > 2:
return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
else:
raise NotImplementedError # can't return a vector
elif isinstance(indices, slice):
if len(shape) > 2:
return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
else:
return LoraTorchTensor(self._lora_A, self._lora_B[indices])
elif isinstance(indices, tuple):
assert len(indices) > 0
if indices[-1] is Ellipsis:
return self[indices[:-1]]
# expand ellipsis
indices = tuple(
u
for v in (
(
(slice(None, None) for _ in range(len(indices) - 1))
if i is Ellipsis
else (i,)
)
for i in indices
)
for u in v
)
if len(indices) < len(shape):
indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape))))
# TODO: make sure this is correct
indices_A = (
*(
(
j.__index__() % self._lora_A.shape[i]
if isinstance(j, SupportsIndex)
else slice(None, None)
)
for i, j in enumerate(indices[:-2])
),
slice(None, None),
indices[-1],
)
indices_B = indices[:-1]
return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B])
else:
raise NotImplementedError # unknown indice type
@property
def dtype(self) -> torch.dtype:
assert self._lora_A.dtype == self._lora_B.dtype
return self._lora_A.dtype
@property
def shape(self) -> tuple[int, ...]:
assert len(self._lora_A.shape) == len(self._lora_B.shape)
return (*self._lora_B.shape[:-1], self._lora_A.shape[-1])
def size(self, dim=None):
assert dim is None
return self.shape
def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
if isinstance(shape[0], tuple):
new_shape: tuple[int, ...] = shape[0]
else:
new_shape = cast(tuple[int, ...], shape)
orig_shape = self.shape
if len(new_shape) < 2:
raise NotImplementedError # can't become a vector
# expand -1 in the shape
if any(dim == -1 for dim in new_shape):
n_elems = prod(orig_shape)
n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape)
assert n_elems % n_new_elems == 0
new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),)
if new_shape[-1] != orig_shape[-1]:
raise NotImplementedError # can't reshape the row size trivially
shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1])
shape_B = (*new_shape[:-1], self._rank)
return LoraTorchTensor(
self._lora_A.reshape(shape_A),
self._lora_B.reshape(shape_B),
)
def reshape_as(self, other: Tensor) -> LoraTorchTensor:
return self.reshape(*other.shape)
def view(self, *size: int) -> LoraTorchTensor:
return self.reshape(*size)
def permute(self, *dims: int) -> LoraTorchTensor:
shape = self.shape
dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims)
if dims[-1] == -1:
# TODO: support higher dimensional A shapes bigger than 1
assert all(dim == 1 for dim in self._lora_A.shape[:-2])
return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims))
if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1:
return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
else:
# TODO: compose the above two
raise NotImplementedError
def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor:
shape = self.shape
dims = [i for i in range(len(shape))]
dims[dim0], dims[dim1] = dims[dim1], dims[dim0]
return self.permute(*dims)
def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor:
return self.transpose(axis0, axis1)
def to(self, *args, **kwargs):
return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs))
@classmethod
def __torch_function__(cls, func: Callable, types, args=(), kwargs=None):
del types # unused
if kwargs is None:
kwargs = {}
if func is torch.permute:
return type(args[0]).permute(*args, **kwargs)
elif func is torch.reshape:
return type(args[0]).reshape(*args, **kwargs)
elif func is torch.stack:
assert isinstance(args[0], Sequence)
dim = kwargs.get("dim", 0)
assert dim == 0
return LoraTorchTensor(
torch.stack([a._lora_A for a in args[0]], dim),
torch.stack([b._lora_B for b in args[0]], dim),
)
elif func is torch.cat:
assert isinstance(args[0], Sequence)
dim = kwargs.get("dim", 0)
assert dim == 0
if len(args[0][0].shape) > 2:
return LoraTorchTensor(
torch.cat([a._lora_A for a in args[0]], dim),
torch.cat([b._lora_B for b in args[0]], dim),
)
elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]):
return LoraTorchTensor(
args[0][0]._lora_A,
torch.cat([b._lora_B for b in args[0]], dim),
)
else:
raise NotImplementedError
else:
raise NotImplementedError
def get_base_tensor_name(lora_tensor_name: str) -> str:
base_name = lora_tensor_name.replace("base_model.model.", "")
base_name = base_name.replace(".lora_A.weight", ".weight")
base_name = base_name.replace(".lora_B.weight", ".weight")
return base_name
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
parser.add_argument(
"--outfile", type=Path,
help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
)
parser.add_argument(
"--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
)
parser.add_argument(
"--bigendian", action="store_true",
help="model is executed on big endian machine",
)
parser.add_argument(
"--no-lazy", action="store_true",
help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
)
parser.add_argument(
"--verbose", action="store_true",
help="increase output verbosity",
)
parser.add_argument(
"--dry-run", action="store_true",
help="only print out what will be done, without writing any new files",
)
parser.add_argument(
"--base", type=Path, required=True,
help="directory containing base model file",
)
parser.add_argument(
"lora_path", type=Path,
help="directory containing LoRA adapter file",
)
return parser.parse_args()
if __name__ == '__main__':
args = parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
ftype_map: dict[str, gguf.LlamaFileType] = {
"f32": gguf.LlamaFileType.ALL_F32,
"f16": gguf.LlamaFileType.MOSTLY_F16,
"bf16": gguf.LlamaFileType.MOSTLY_BF16,
"q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
"auto": gguf.LlamaFileType.GUESSED,
}
ftype = ftype_map[args.outtype]
dir_base_model: Path = args.base
dir_lora: Path = args.lora_path
lora_config = dir_lora / "adapter_config.json"
input_model = dir_lora / "adapter_model.safetensors"
if args.outfile is not None:
fname_out = args.outfile
else:
# output in the same directory as the model by default
fname_out = dir_lora
if os.path.exists(input_model):
# lazy import load_file only if lora is in safetensors format.
from safetensors.torch import load_file
lora_model = load_file(input_model, device="cpu")
else:
input_model = os.path.join(dir_lora, "adapter_model.bin")
lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
# load base model
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = Model.load_hparams(dir_base_model)
with torch.inference_mode():
try:
model_class = Model.from_model_architecture(hparams["architectures"][0])
except NotImplementedError:
logger.error(f"Model {hparams['architectures'][0]} is not supported")
sys.exit(1)
class LoraModel(model_class):
model_arch = model_class.model_arch
lora_alpha: float
def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):
super().__init__(*args, **kwargs)
self.dir_model_card = dir_lora_model
self.lora_alpha = float(lora_alpha)
def set_type(self):
self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
def set_gguf_parameters(self):
self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
super().set_gguf_parameters()
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
tensor_map: dict[str, PartialLoraTensor] = {}
for name, tensor in lora_model.items():
if self.lazy:
tensor = LazyTorchTensor.from_eager(tensor)
base_name = get_base_tensor_name(name)
is_lora_a = ".lora_A.weight" in name
is_lora_b = ".lora_B.weight" in name
if not is_lora_a and not is_lora_b:
if ".base_layer.weight" in name:
continue
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
sys.exit(1)
if base_name in tensor_map:
if is_lora_a:
tensor_map[base_name].A = tensor
else:
tensor_map[base_name].B = tensor
else:
if is_lora_a:
tensor_map[base_name] = PartialLoraTensor(A=tensor)
else:
tensor_map[base_name] = PartialLoraTensor(B=tensor)
for name, tensor in tensor_map.items():
assert tensor.A is not None
assert tensor.B is not None
yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
dest = super().modify_tensors(data_torch, name, bid)
for dest_name, dest_data in dest:
assert isinstance(dest_data, LoraTorchTensor)
lora_a, lora_b = dest_data.get_lora_A_B()
yield (dest_name + ".lora_a", lora_a)
yield (dest_name + ".lora_b", lora_b)
with open(lora_config, "r") as f:
lparams: dict[str, Any] = json.load(f)
alpha: float = lparams["lora_alpha"]
model_instance = LoraModel(
dir_base_model,
ftype,
fname_out,
is_big_endian=args.bigendian,
use_temp_file=False,
eager=args.no_lazy,
dry_run=args.dry_run,
dir_lora_model=dir_lora,
lora_alpha=alpha,
)
logger.info("Exporting model...")
model_instance.write()
logger.info(f"Model successfully exported to {model_instance.fname_out}")

56
docs/android.md Normal file
View file

@ -0,0 +1,56 @@
# Android
## Build on Android using Termux
[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
```
apt update && apt upgrade -y
apt install git make cmake
```
It's recommended to move your model inside the `~/` directory for best performance:
```
cd storage/downloads
mv model.gguf ~/
```
[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
## Building the Project using Android NDK
Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
```
$ mkdir build-android
$ cd build-android
$ export NDK=<your_ndk_directory>
$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
$ make
```
Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
```
$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
$cd /data/data/com.termux/files/home/bin
$chmod +x ./*
```
Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
```
$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
```
Now, you can start chatting:
```
$cd /data/data/com.termux/files/home/bin
$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
```
Here's a demo of an interactive session running on Pixel 5 phone:
https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4

6
docs/BLIS.md → docs/backend/BLIS.md
View file

@ -30,8 +30,8 @@ We recommend using openmp since it's easier to modify the cores being used.
Makefile:
```bash
make LLAMA_BLIS=1 -j
# make LLAMA_BLIS=1 benchmark-matmult
make GGML_BLIS=1 -j
# make GGML_BLIS=1 llama-benchmark-matmult
```
CMake:
@ -39,7 +39,7 @@ CMake:
```bash
mkdir build
cd build
cmake -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=FLAME ..
cmake -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FLAME ..
make -j
```

114
README-sycl.md → docs/backend/SYCL.md
View file

@ -1,6 +1,7 @@
# llama.cpp for SYCL
- [Background](#background)
- [Recommended Release](#recommended-release)
- [News](#news)
- [OS](#os)
- [Hardware](#hardware)
@ -31,8 +32,23 @@ When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneM
It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
## Recommended Release
The SYCL backend would be broken by some PRs due to no online CI.
The following release is verified with good quality:
|Commit ID|Tag|Release|Verified Platform|
|-|-|-|-|
|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1|
## News
- 2024.5
- Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
- Arch Linux is verified successfully.
- 2024.4
- Support data types: GGML_TYPE_IQ4_NL, GGML_TYPE_IQ4_XS, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S, GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M.
@ -77,7 +93,7 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
*Notes:*
- **Memory**
- The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/main`.
- The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
- Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
@ -99,14 +115,14 @@ The docker build option is currently limited to *intel GPU* targets.
### Build image
```sh
# Using FP16
docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" -f .devops/llama-cli-intel.Dockerfile .
```
*Notes*:
To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="LLAMA_SYCL_F16=ON"` argument from the previous command.
To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="GGML_SYCL_F16=ON"` argument from the previous command.
You can also use the `.devops/server-intel.Dockerfile`, which builds the *"server"* alternative.
You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
### Run container
@ -228,10 +244,10 @@ source /opt/intel/oneapi/setvars.sh
# Build LLAMA with MKL BLAS acceleration for intel GPU
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
# Option 2: Use FP16
cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
# build all binary
cmake --build build --config Release -j -v
@ -248,10 +264,10 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
# Build LLAMA with Nvidia BLAS acceleration through SYCL
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake -B build -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
# Option 2: Use FP16
cmake -B build -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
# build all binary
cmake --build build --config Release -j -v
@ -275,33 +291,28 @@ source /opt/intel/oneapi/setvars.sh
Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
```sh
./build/bin/ls-sycl-device
./build/bin/llama-ls-sycl-device
```
A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
```
found 6 SYCL devices:
found 2 SYCL devices:
| | | |Compute |Max compute|Max work|Max sub| |
|ID| Device Type| Name|capability|units |group |group |Global mem size|
|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
| 0|[level_zero:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 1.3| 512| 1024| 32| 16225243136|
| 1|[level_zero:gpu:1]| Intel(R) UHD Graphics 770| 1.3| 32| 512| 32| 53651849216|
| 2| [opencl:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 3.0| 512| 1024| 32| 16225243136|
| 3| [opencl:gpu:1]| Intel(R) UHD Graphics 770| 3.0| 32| 512| 32| 53651849216|
| 4| [opencl:cpu:0]| 13th Gen Intel(R) Core(TM) i7-13700K| 3.0| 24| 8192| 64| 67064815616|
| 5| [opencl:acc:0]| Intel(R) FPGA Emulation Device| 1.2| 24|67108864| 64| 67064815616|
```
| Attribute | Note |
|------------------------|-------------------------------------------------------------|
| compute capability 1.3 | Level-zero driver/runtime, recommended |
| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases |
4. Launch inference
There are two device selection modes:
- Single device: Use one device target specified by the user.
- Multiple devices: Automatically select the devices with the same largest Max compute-units.
- Multiple devices: Automatically choose the devices with the same backend.
In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
| Device selection | Parameter |
|------------------|----------------------------------------|
@ -313,7 +324,7 @@ Examples:
- Use device 0:
```sh
ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
```
or run by script:
@ -324,7 +335,7 @@ or run by script:
- Use multiple devices:
```sh
ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
```
Otherwise, you can run the script:
@ -394,15 +405,9 @@ Output (example):
4. Install build tools
a. Download & install cmake for Windows: https://cmake.org/download/
a. Download & install cmake for Windows: https://cmake.org/download/ (CMake can also be installed from Visual Studio Installer)
b. The new Visual Studio will install Ninja as default. (If not, please install it manually: https://ninja-build.org/)
b. Download & install mingw-w64 make for Windows provided by w64devkit
- Download the 1.19.0 version of [w64devkit](https://github.com/skeeto/w64devkit/releases/download/v1.19.0/w64devkit-1.19.0.zip).
- Extract `w64devkit` on your pc.
- Add the **bin** folder path in the Windows system PATH environment (for e.g. `C:\xxx\w64devkit\bin\`).
### II. Build llama.cpp
@ -412,10 +417,10 @@ On the oneAPI command line window, step into the llama.cpp main directory and ru
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake -B build -G "MinGW Makefiles" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release
cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release
# Option 2: Or FP16
cmake -B build -G "MinGW Makefiles" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON
cmake --build build --config Release -j
```
@ -425,9 +430,23 @@ Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former in
.\examples\sycl\win-build-sycl.bat
```
Or, use CMake presets to build:
```sh
cmake --preset x64-windows-sycl-release
cmake --build build-x64-windows-sycl-release -j --target llama-cli
cmake -DGGML_SYCL_F16=ON --preset x64-windows-sycl-release
cmake --build build-x64-windows-sycl-release -j --target llama-cli
cmake --preset x64-windows-sycl-debug
cmake --build build-x64-windows-sycl-debug -j --target llama-cli
```
Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
*Notes:*
- By default, calling `make` will build all target binary files. In case of a minimal experimental setup, the user can build the inference executable only through `make main`.
- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`.
### III. Run the inference
@ -450,33 +469,26 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow
build\bin\ls-sycl-device.exe
```
The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following:
This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
```
found 6 SYCL devices:
found 2 SYCL devices:
| | | |Compute |Max compute|Max work|Max sub| |
|ID| Device Type| Name|capability|units |group |group |Global mem size|
|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
| 0|[level_zero:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 1.3| 512| 1024| 32| 16225243136|
| 1|[level_zero:gpu:1]| Intel(R) UHD Graphics 770| 1.3| 32| 512| 32| 53651849216|
| 2| [opencl:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 3.0| 512| 1024| 32| 16225243136|
| 3| [opencl:gpu:1]| Intel(R) UHD Graphics 770| 3.0| 32| 512| 32| 53651849216|
| 4| [opencl:cpu:0]| 13th Gen Intel(R) Core(TM) i7-13700K| 3.0| 24| 8192| 64| 67064815616|
| 5| [opencl:acc:0]| Intel(R) FPGA Emulation Device| 1.2| 24|67108864| 64| 67064815616|
```
| Attribute | Note |
|------------------------|-----------------------------------------------------------|
| compute capability 1.3 | Level-zero running time, recommended |
| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases |
4. Launch inference
There are two device selection modes:
- Single device: Use one device assigned by user.
- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
- Single device: Use one device assigned by user. Default device id is 0.
- Multiple devices: Automatically choose the devices with the same backend.
In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
| Device selection | Parameter |
|------------------|----------------------------------------|
@ -488,13 +500,13 @@ Examples:
- Use device 0:
```
build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
```
- Use multiple devices:
```
build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
```
Otherwise, run the following wrapper script:
@ -520,9 +532,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512
| Name | Value | Function |
|--------------------|-----------------------------------|---------------------------------------------|
| LLAMA_SYCL | ON (mandatory) | Enable build with SYCL code path. |
| LLAMA_SYCL_TARGET | INTEL *(default)* \| NVIDIA | Set the SYCL target device type. |
| LLAMA_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. |
| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path. |
| GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA | Set the SYCL target device type. |
| GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. |
| CMAKE_C_COMPILER | icx | Set *icx* compiler for SYCL code path. |
| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)* | Set `icpx/icx` compiler for SYCL code path. |

357
docs/build.md Normal file
View file

@ -0,0 +1,357 @@
# Build llama.cpp locally
**To get the Code:**
```bash
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
```
In order to build llama.cpp you have four different options.
- Using `make`:
- On Linux or MacOS:
```bash
make
```
- On Windows (x86/x64 only, arm64 requires cmake):
1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
2. Extract `w64devkit` on your pc.
3. Run `w64devkit.exe`.
4. Use the `cd` command to reach the `llama.cpp` folder.
5. From here you can run:
```bash
make
```
- Notes:
- For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
- For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/).
- For debug builds, run `make LLAMA_DEBUG=1`
- Using `CMake`:
```bash
cmake -B build
cmake --build build --config Release
```
**Notes**:
- For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
- For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/).
- For debug builds, there are two cases:
1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
```bash
cmake -B build -DCMAKE_BUILD_TYPE=Debug
cmake --build build
```
2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
```bash
cmake -B build -G "Xcode"
cmake --build build --config Debug
```
- Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
- Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
- Tab Workload: Desktop-development with C++
- Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
- Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
- For Windows on ARM (arm64, WoA) build with:
```bash
cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
cmake --build build-arm64-windows-llvm-release
```
Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
- Using `gmake` (FreeBSD):
1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
2. Add your user to **video** group
3. Install compilation dependencies.
```bash
sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
```
## Metal Build
On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option.
When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
argument.
## BLAS Build
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
### Accelerate Framework:
This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
### OpenBLAS:
This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
- Using `make`:
- On Linux:
```bash
make GGML_OPENBLAS=1
```
- On Windows:
1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
3. Extract `w64devkit` on your pc.
4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
6. Run `w64devkit.exe`.
7. Use the `cd` command to reach the `llama.cpp` folder.
8. From here you can run:
```bash
make GGML_OPENBLAS=1
```
- Using `CMake` on Linux:
```bash
cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
cmake --build build --config Release
```
### BLIS
Check [BLIS.md](./backend/BLIS.md) for more information.
### SYCL
SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
### Intel oneMKL
Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
- Using manual oneAPI installation:
By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
```bash
source /opt/intel/oneapi/setvars.sh # You can skip this step if in oneapi-basekit docker image, only required for manual installation
cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON
cmake --build build --config Release
```
- Using oneAPI docker image:
If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
### CUDA
This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
- Using `make`:
```bash
make GGML_CUDA=1
```
- Using `CMake`:
```bash
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release
```
The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
The following compilation options are also available to tweak performance:
| Option | Legal values | Default | Description |
|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models |
| GGML_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
| GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. |
### MUSA
- Using `make`:
```bash
make GGML_MUSA=1
```
- Using `CMake`:
```bash
cmake -B build -DGGML_MUSA=ON
cmake --build build --config Release
```
### hipBLAS
This provides BLAS acceleration on HIP-supported AMD GPUs.
Make sure to have ROCm installed.
You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
- Using `make`:
```bash
make GGML_HIPBLAS=1
```
- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build --config Release -- -j 16
```
On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
Note that if you get the following error:
```
clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
```
Try searching for a directory under `HIP_PATH` that contains the file
`oclc_abi_version_400.bc`. Then, add the following to the start of the
command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
like:
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build -- -j 16
```
- Using `make` (example for target gfx1030, build with 16 CPU threads):
```bash
make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
```
- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
```bash
set PATH=%HIP_PATH%\bin;%PATH%
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake --build build
```
Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
| Option | Legal values | Default | Description |
|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
### Vulkan
**Windows**
#### w64devkit
Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
```sh
SDK_VERSION=1.3.283.0
cp /VulkanSDK/$SDK_VERSION/Bin/glslc.exe $W64DEVKIT_HOME/bin/
cp /VulkanSDK/$SDK_VERSION/Lib/vulkan-1.lib $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/
cp -r /VulkanSDK/$SDK_VERSION/Include/* $W64DEVKIT_HOME/x86_64-w64-mingw32/include/
cat > $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/pkgconfig/vulkan.pc <<EOF
Name: Vulkan-Loader
Description: Vulkan Loader
Version: $SDK_VERSION
Libs: -lvulkan-1
EOF
```
Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
#### MSYS2
Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
```sh
pacman -S git \
mingw-w64-ucrt-x86_64-gcc \
mingw-w64-ucrt-x86_64-cmake \
mingw-w64-ucrt-x86_64-vulkan-devel \
mingw-w64-ucrt-x86_64-shaderc
```
Switch into `llama.cpp` directory and build using CMake.
```sh
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release
```
**With docker**:
You don't need to install Vulkan SDK. It will be installed inside the container.
```sh
# Build the image
docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
# Then, use it:
docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
```
**Without docker**:
Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
For example, on Ubuntu 22.04 (jammy), use the command below:
```bash
wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
apt update -y
apt-get install -y vulkan-sdk
# To verify the installation, use the command below:
vulkaninfo
```
Alternatively your package manager might be able to provide the appropriate libraries.
For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
Then, build llama.cpp using the cmake command below:
```bash
cmake -B build -DGGML_VULKAN=1
cmake --build build --config Release
# Test the output binary (with "-ngl 33" to offload all layers to GPU)
./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
# You should see in the output, ggml_vulkan detected your GPU. For example:
# ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
```
### Android
To read documentation for how to build on Android, [click here](./android.md)

18
docs/HOWTO-add-model.md → docs/development/HOWTO-add-model.md
View file

@ -1,4 +1,4 @@
## Add a new model architecture to `llama.cpp`
# Add a new model architecture to `llama.cpp`
Adding a model requires few steps:
@ -9,15 +9,15 @@ Adding a model requires few steps:
After following these steps, you can open PR.
Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
- [main](../examples/main)
- [imatrix](../examples/imatrix)
- [quantize](../examples/quantize)
- [server](../examples/server)
- [main](/examples/main/)
- [imatrix](/examples/imatrix/)
- [quantize](/examples/quantize/)
- [server](/examples/server/)
### 1. Convert the model to GGUF
This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library.
Depending on the model architecture, you can use either [convert-hf-to-gguf.py](../convert-hf-to-gguf.py) or [examples/convert-legacy-llama.py](../examples/convert-legacy-llama.py) (for `llama/llama2` models in `.pth` format).
Depending on the model architecture, you can use either [convert_hf_to_gguf.py](/convert_hf_to_gguf.py) or [examples/convert_legacy_llama.py](/examples/convert_legacy_llama.py) (for `llama/llama2` models in `.pth` format).
The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors.
@ -31,7 +31,7 @@ class MyModel(Model):
model_arch = gguf.MODEL_ARCH.GROK
```
2. Define the layout of the GGUF tensors in [constants.py](../gguf-py/gguf/constants.py)
2. Define the layout of the GGUF tensors in [constants.py](/gguf-py/gguf/constants.py)
Add an enum entry in `MODEL_ARCH`, the model human friendly name in `MODEL_ARCH_NAMES` and the GGUF tensor names in `MODEL_TENSORS`.
@ -54,7 +54,7 @@ Example for `falcon` model:
As a general rule, before adding a new tensor name to GGUF, be sure the equivalent naming does not already exist.
Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](../gguf-py/gguf/tensor_mapping.py) file.
Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](/gguf-py/gguf/tensor_mapping.py) file.
If the tensor name is part of a repetitive layer/block, the key word `bid` substitutes it.
@ -100,7 +100,7 @@ Have a look at existing implementation like `build_llama`, `build_dbrx` or `buil
When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support for missing backend operations can be added in another PR.
Note: to debug the inference graph: you can use [eval-callback](../examples/eval-callback).
Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).
## GGUF specification

0
docs/debugging-tests.md → docs/development/debugging-tests.md
View file

0
docs/llama-star/idea-arch.key → docs/development/llama-star/idea-arch.key
View file

0
docs/llama-star/idea-arch.pdf → docs/development/llama-star/idea-arch.pdf
View file

6
docs/token_generation_performance_tips.md → docs/development/token_generation_performance_tips.md
View file

@ -1,9 +1,9 @@
# Token generation performance troubleshooting
## Verifying that the model is running on the GPU with CUDA
Make sure you compiled llama with the correct env variables according to [this guide](../README.md#CUDA), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
Make sure you compiled llama with the correct env variables according to [this guide](/docs/build.md#cuda), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
```shell
./main -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
./llama-cli -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
```
When running llama, before it starts the inference work, it will output diagnostic information that shows whether cuBLAS is offloading work to the GPU. Look for these lines:
@ -27,7 +27,7 @@ RAM: 32GB
Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.q4_0.gguf` (30B parameters, 4bit quantization, GGML)
Run command: `./main -m "path/to/model.gguf" -p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
Run command: `./llama-cli -m "path/to/model.gguf" -p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
Result:

86
docs/docker.md Normal file
View file

@ -0,0 +1,86 @@
# Docker
## Prerequisites
* Docker must be installed and running on your system.
* Create a folder to store big models & intermediate files (ex. /llama/models)
## Images
We have three Docker images available for this project:
1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
Additionally, there the following images, similar to the above:
- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
## Usage
The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
Replace `/path/to/models` below with the actual path where you downloaded the models.
```bash
docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
```
On completion, you are ready to play!
```bash
docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
```
or with a light image:
```bash
docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
```
or with a server image:
```bash
docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
```
## Docker With CUDA
Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
## Building Docker locally
```bash
docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
```
You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
The defaults are:
- `CUDA_VERSION` set to `11.7.1`
- `CUDA_DOCKER_ARCH` set to `all`
The resulting images, are essentially the same as the non-CUDA images:
1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
## Usage
After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
```bash
docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
```

39
docs/install.md Normal file
View file

@ -0,0 +1,39 @@
# Install pre-built version of llama.cpp
## Homebrew
On Mac and Linux, the homebrew package manager can be used via
```sh
brew install llama.cpp
```
The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
## Nix
On Mac and Linux, the Nix package manager can be used via
```sh
nix profile install nixpkgs#llama-cpp
```
For flake enabled installs.
Or
```sh
nix-env --file '<nixpkgs>' --install --attr llama-cpp
```
For non-flake enabled installs.
This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
## Flox
On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
```sh
flox install llama-cpp
```
Flox follows the nixpkgs build of llama.cpp.

51
examples/CMakeLists.txt
View file

@ -12,43 +12,44 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
if (EMSCRIPTEN)
else()
add_subdirectory(cvector-generator)
add_subdirectory(baby-llama)
add_subdirectory(batched)
add_subdirectory(batched-bench)
add_subdirectory(batched)
add_subdirectory(benchmark)
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(embedding)
add_subdirectory(eval-callback)
add_subdirectory(finetune)
add_subdirectory(gritlm)
add_subdirectory(export-lora)
add_subdirectory(gbnf-validator)
add_subdirectory(gguf-hash)
add_subdirectory(gguf-split)
add_subdirectory(gguf)
add_subdirectory(gritlm)
add_subdirectory(imatrix)
add_subdirectory(infill)
add_subdirectory(llama-bench)
add_subdirectory(llava)
if (LLAMA_SYCL)
add_subdirectory(sycl)
endif()
add_subdirectory(main)
add_subdirectory(tokenize)
add_subdirectory(parallel)
add_subdirectory(perplexity)
add_subdirectory(quantize)
add_subdirectory(quantize-stats)
add_subdirectory(retrieval)
add_subdirectory(save-load-state)
add_subdirectory(simple)
add_subdirectory(passkey)
add_subdirectory(speculative)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(gguf)
add_subdirectory(train-text-from-scratch)
add_subdirectory(imatrix)
if (LLAMA_BUILD_SERVER)
add_subdirectory(server)
endif()
add_subdirectory(export-lora)
if (LLAMA_RPC)
add_subdirectory(main)
add_subdirectory(parallel)
add_subdirectory(passkey)
add_subdirectory(perplexity)
add_subdirectory(quantize-stats)
add_subdirectory(quantize)
add_subdirectory(retrieval)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
if (LLAMA_BUILD_SERVER)
add_subdirectory(server)
endif()
if (GGML_SYCL)
add_subdirectory(sycl)
endif()
add_subdirectory(save-load-state)
add_subdirectory(simple)
add_subdirectory(speculative)
add_subdirectory(tokenize)
endif()

2
examples/Miku.sh
View file

@ -22,7 +22,7 @@ if [ -n "$N_THREAD" ]; then
GEN_OPTIONS+=(--threads "$N_THREAD")
fi
./main "${GEN_OPTIONS[@]}" \
./llama-cli "${GEN_OPTIONS[@]}" \
--model "$MODEL" \
--in-prefix " " \
--in-suffix "${AI_NAME}:" \

2
examples/baby-llama/CMakeLists.txt
View file

@ -1,4 +1,4 @@
set(TARGET baby-llama)
set(TARGET llama-baby-llama)
add_executable(${TARGET} baby-llama.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

2
examples/base-translate.sh
View file

@ -58,4 +58,4 @@ echo "$2
model=$1
# generate the most likely continuation until the string "===" is found
./main -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs
./llama-cli -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

2
examples/batched-bench/CMakeLists.txt
View file

@ -1,4 +1,4 @@
set(TARGET batched-bench)
set(TARGET llama-batched-bench)
add_executable(${TARGET} batched-bench.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

8
examples/batched-bench/README.md
View file

@ -10,16 +10,16 @@ There are 2 modes of operation:
- `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
```bash
./batched-bench -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]
./llama-batched-bench -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]
# LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
./batched-bench -m ./models/llama-7b/ggml-model-f16.gguf -c 16384 -b 2048 -ub 512 -ngl 99
./llama-batched-bench -m ./models/llama-7b/ggml-model-f16.gguf -c 16384 -b 2048 -ub 512 -ngl 99
# LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
./batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 16384 -b 2048 -ub 512 -ngl 99 -pps
./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 16384 -b 2048 -ub 512 -ngl 99 -pps
# custom set of batches
./batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 2048 -b 512 -ub 512 -ngl 999 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32
./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 2048 -b 512 -ub 512 -ngl 999 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32
```
## Sample results

6
examples/batched.swift/Makefile
View file

@ -1,6 +1,6 @@
.PHONY: build
build:
xcodebuild -scheme batched_swift -destination "generic/platform=macOS" -derivedDataPath build
rm -f ./batched_swift
ln -s ./build/Build/Products/Debug/batched_swift ./batched_swift
xcodebuild -scheme llama-batched-swift -destination "generic/platform=macOS" -derivedDataPath build
rm -f ./llama-batched-swift
ln -s ./build/Build/Products/Debug/llama-batched-swift ./llama-batched-swift

4
examples/batched.swift/Package.swift
View file

@ -4,7 +4,7 @@
import PackageDescription
let package = Package(
name: "batched_swift",
name: "llama-batched-swift",
platforms: [.macOS(.v12)],
dependencies: [
.package(name: "llama", path: "../../"),
@ -13,7 +13,7 @@ let package = Package(
// Targets are the basic building blocks of a package, defining a module or a test suite.
// Targets can depend on other targets in this package and products from dependencies.
.executableTarget(
name: "batched_swift",
name: "llama-batched-swift",
dependencies: ["llama"],
path: "Sources",
linkerSettings: [.linkedFramework("Foundation"), .linkedFramework("AppKit")]

2
examples/batched.swift/README.md
View file

@ -1,4 +1,4 @@
This is a swift clone of `examples/batched`.
$ `make`
$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`
$ `./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]`

3
examples/batched.swift/Sources/main.swift
View file

@ -229,7 +229,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
var result = [CChar](repeating: 0, count: 8)
let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false)
let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), 0, false)
if nTokens < 0 {
let actualTokensCount = -Int(nTokens)
result = .init(repeating: 0, count: actualTokensCount)
@ -238,6 +238,7 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
token,
&result,
Int32(result.count),
0,
false
)
assert(check == actualTokensCount)

2
examples/batched/CMakeLists.txt
View file

@ -1,4 +1,4 @@
set(TARGET batched)
set(TARGET llama-batched)
add_executable(${TARGET} batched.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

2
examples/batched/README.md
View file

@ -3,7 +3,7 @@
The example demonstrates batched generation from a given prompt
```bash
./batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
...

36
examples/batched/batched.cpp
View file

@ -31,7 +31,7 @@ int main(int argc, char ** argv) {
int n_parallel = params.n_parallel;
// total length of the sequences including the prompt
int n_predict = 32;
int n_predict = params.n_predict;
// init LLM
@ -93,14 +93,34 @@ int main(int argc, char ** argv) {
// create a llama_batch
// we use this object to submit token data for decoding
llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
std::vector<llama_seq_id> seq_ids(n_parallel, 0);
for (int32_t i = 0; i < n_parallel; ++i) {
seq_ids[i] = i;
}
// evaluate the initial prompt
for (size_t i = 0; i < tokens_list.size(); ++i) {
llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
llama_batch_add(batch, tokens_list[i], i, seq_ids, false);
}
GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
if (llama_model_has_encoder(model)) {
if (llama_encode(ctx, batch)) {
LOG_TEE("%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 == -1) {
decoder_start_token_id = llama_token_bos(model);
}
llama_batch_clear(batch);
llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
}
// llama_decode will output logits only for the last token of the prompt
batch.logits[batch.n_tokens - 1] = true;
@ -109,11 +129,11 @@ int main(int argc, char ** argv) {
return 1;
}
// assign the system KV cache to all parallel sequences
// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
for (int32_t i = 1; i < n_parallel; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
//// assign the system KV cache to all parallel sequences
//// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
//for (int32_t i = 1; i < n_parallel; ++i) {
// llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
//}
if (n_parallel > 1) {
LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);

2
examples/benchmark/CMakeLists.txt
View file

@ -1,4 +1,4 @@
set(TARGET benchmark)
set(TARGET llama-bench-matmult)
add_executable(${TARGET} benchmark-matmult.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})

2
examples/chat-13B.sh
View file

@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
$PROMPT_TEMPLATE > $PROMPT_FILE
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./main $GEN_OPTIONS \
./llama-cli $GEN_OPTIONS \
--model "$MODEL" \
--threads "$N_THREAD" \
--n_predict "$N_PREDICTS" \

10
examples/chat-persistent.sh
View file

@ -62,7 +62,7 @@ fi
if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
echo 'Prompt cache does not exist, building...'
# Default batch_size to 64 here for better user feedback during initial prompt processing
./main 2>>"$LOG" \
./llama-cli 2>>"$LOG" \
--batch_size 64 \
"${OPTS[@]}" \
--prompt-cache "$PROMPT_CACHE_FILE" \
@ -109,13 +109,13 @@ while read -e line; do
printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
./main 2>>"$LOG" "${OPTS[@]}" \
./llama-cli 2>>"$LOG" "${OPTS[@]}" \
--prompt-cache "$CUR_PROMPT_CACHE" \
--prompt-cache-all \
--file "$CUR_PROMPT_FILE" \
--reverse-prompt "${USER_NAME}:" \
--n_predict "$n_predict" |
skip_bytes 1 | # skip BOS token added by ./main
skip_bytes 1 | # skip BOS token added by ./llama-cli
tee "$CUR_PROMPT_FILE.tmp" | # save prompt + generation to tmp file
skip_bytes "$n_prompt_len_pre" # print generation
@ -133,7 +133,7 @@ while read -e line; do
# TODO get both messages in one go
if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
echo >&2 "Couldn't get number of tokens from ./main output!"
echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
exit 1
fi
@ -144,7 +144,7 @@ while read -e line; do
fi
# Update cache for next prompt in background, ideally during user input
./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
./llama-cli >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
--prompt-cache "$NEXT_PROMPT_CACHE" \
--file "$NEXT_PROMPT_FILE" \
--n_predict 1 &

2
examples/chat-vicuna.sh
View file

@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
$PROMPT_TEMPLATE > $PROMPT_FILE
# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
./bin/main $GEN_OPTIONS \
./bin/llama-cli $GEN_OPTIONS \
--model "$MODEL" \
--threads "$N_THREAD" \
--n_predict "$N_PREDICTS" \

2
examples/chat.sh
View file

@ -11,6 +11,6 @@ cd ..
#
# "--keep 48" is based on the contents of prompts/chat-with-bob.txt
#
./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
./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

2
examples/convert-llama2c-to-ggml/CMakeLists.txt
View file

@ -1,4 +1,4 @@
set(TARGET convert-llama2c-to-ggml)
set(TARGET llama-convert-llama2c-to-ggml)
add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})

6
examples/convert-llama2c-to-ggml/README.md
View file

@ -8,7 +8,7 @@ To convert the model first download the models from the [llama2.c](https://githu
After successful compilation, following usage options are available:
```
usage: ./convert-llama2c-to-ggml [options]
usage: ./llama-convert-llama2c-to-ggml [options]
options:
-h, --help show this help message and exit
@ -19,10 +19,10 @@ options:
An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
`$ ./llama-convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
Now you can use the model with a command like:
`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
`$ ./llama-cli -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

242
examples/convert-legacy-llama.py → examples/convert_legacy_llama.py
View file

@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional
from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar
import numpy as np
@ -346,42 +346,6 @@ class Params:
return params
@dataclass
class Metadata:
name: Optional[str] = None
author: Optional[str] = None
version: Optional[str] = None
url: Optional[str] = None
description: Optional[str] = None
licence: Optional[str] = None
source_url: Optional[str] = None
source_hf_repo: Optional[str] = None
@staticmethod
def load(metadata_path: Path) -> Metadata:
if metadata_path is None or not metadata_path.exists():
return Metadata()
with open(metadata_path, 'r') as file:
data = json.load(file)
# Create a new Metadata instance
metadata = Metadata()
# Assigning values to Metadata attributes if they exist in the JSON file
# This is based on LLM_KV_NAMES mapping in llama.cpp
metadata.name = data.get("general.name")
metadata.author = data.get("general.author")
metadata.version = data.get("general.version")
metadata.url = data.get("general.url")
metadata.description = data.get("general.description")
metadata.license = data.get("general.license")
metadata.source_url = data.get("general.source.url")
metadata.source_hf_repo = data.get("general.source.huggingface.repository")
return metadata
#
# data loading
# TODO: reuse (probably move to gguf.py?)
@ -492,12 +456,13 @@ class LazyTensor:
LazyModel: TypeAlias = 'dict[str, LazyTensor]'
ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none']
@dataclass
class ModelPlus:
model: LazyModel
paths: list[Path] # Where this was read from.
format: Literal['ggml', 'torch', 'safetensors', 'none']
format: ModelFormat
vocab: BaseVocab | None # For GGML models (which have vocab built in), the vocab.
@ -536,7 +501,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
formats = set(mp.format for mp in models_plus)
formats: set[ModelFormat] = set(mp.format for mp in models_plus)
assert len(formats) == 1, "different formats?"
format = formats.pop()
paths = [path for mp in models_plus for path in mp.paths]
@ -555,7 +520,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
else:
model = merge_sharded([mp.model for mp in models_plus])
return ModelPlus(model, paths, format, vocab) # pytype: disable=wrong-arg-types
return ModelPlus(model, paths, format, vocab)
def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
@ -805,7 +770,7 @@ class OutputFile:
def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
def add_meta_model(self, params: Params, metadata: Metadata) -> None:
def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None:
# Metadata About The Model And Its Provenence
name = "LLaMA"
if metadata is not None and metadata.name is not None:
@ -823,16 +788,73 @@ class OutputFile:
self.gguf.add_author(metadata.author)
if metadata.version is not None:
self.gguf.add_version(metadata.version)
if metadata.url is not None:
self.gguf.add_url(metadata.url)
if metadata.organization is not None:
self.gguf.add_organization(metadata.organization)
if metadata.finetune is not None:
self.gguf.add_finetune(metadata.finetune)
if metadata.basename is not None:
self.gguf.add_basename(metadata.basename)
if metadata.description is not None:
self.gguf.add_description(metadata.description)
if metadata.licence is not None:
self.gguf.add_licence(metadata.licence)
if metadata.quantized_by is not None:
self.gguf.add_quantized_by(metadata.quantized_by)
if metadata.size_label is not None:
self.gguf.add_size_label(metadata.size_label)
if metadata.license is not None:
self.gguf.add_license(metadata.license)
if metadata.license_name is not None:
self.gguf.add_license_name(metadata.license_name)
if metadata.license_link is not None:
self.gguf.add_license_link(metadata.license_link)
if metadata.url is not None:
self.gguf.add_url(metadata.url)
if metadata.doi is not None:
self.gguf.add_doi(metadata.doi)
if metadata.uuid is not None:
self.gguf.add_uuid(metadata.uuid)
if metadata.repo_url is not None:
self.gguf.add_repo_url(metadata.repo_url)
if metadata.source_url is not None:
self.gguf.add_source_url(metadata.source_url)
if metadata.source_hf_repo is not None:
self.gguf.add_source_hf_repo(metadata.source_hf_repo)
if metadata.source_doi is not None:
self.gguf.add_source_doi(metadata.source_doi)
if metadata.source_uuid is not None:
self.gguf.add_source_uuid(metadata.source_uuid)
if metadata.source_repo_url is not None:
self.gguf.add_source_repo_url(metadata.source_repo_url)
if metadata.base_models is not None:
self.gguf.add_base_model_count(len(metadata.base_models))
for key, base_model_entry in enumerate(metadata.base_models):
if "name" in base_model_entry:
self.gguf.add_base_model_name(key, base_model_entry["name"])
if "author" in base_model_entry:
self.gguf.add_base_model_author(key, base_model_entry["author"])
if "version" in base_model_entry:
self.gguf.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
self.gguf.add_base_model_organization(key, base_model_entry["organization"])
if "url" in base_model_entry:
self.gguf.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
self.gguf.add_base_model_doi(key, base_model_entry["doi"])
if "uuid" in base_model_entry:
self.gguf.add_base_model_uuid(key, base_model_entry["uuid"])
if "repo_url" in base_model_entry:
self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
if metadata.tags is not None:
self.gguf.add_tags(metadata.tags)
if metadata.languages is not None:
self.gguf.add_languages(metadata.languages)
if metadata.datasets is not None:
self.gguf.add_datasets(metadata.datasets)
def add_meta_arch(self, params: Params) -> None:
# Metadata About The Neural Architecture Itself
@ -943,7 +965,7 @@ class OutputFile:
@staticmethod
def write_vocab_only(
fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None,
) -> None:
check_vocab_size(params, vocab, pad_vocab=pad_vocab)
@ -977,7 +999,7 @@ class OutputFile:
fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
pad_vocab: bool = False,
metadata: Metadata = None,
metadata: gguf.Metadata | None = None,
) -> None:
check_vocab_size(params, vocab, pad_vocab=pad_vocab)
@ -1020,35 +1042,32 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
raise ValueError(f"Unexpected combination of types: {name_to_type}")
def model_parameter_count(model: LazyModel) -> int:
total_model_parameters = 0
for i, (name, lazy_tensor) in enumerate(model.items()):
sum_weights_in_tensor = 1
def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]:
total_params = 0
shared_params = 0
expert_params = 0
for name, lazy_tensor in tensors:
# We don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
# Got A Tensor
sum_weights_in_tensor: int = 1
# Tensor Volume
for dim in lazy_tensor.shape:
sum_weights_in_tensor *= dim
total_model_parameters += sum_weights_in_tensor
return total_model_parameters
if ".experts." in name:
if ".experts.0." in name:
expert_params += sum_weights_in_tensor
else:
shared_params += sum_weights_in_tensor
def model_parameter_count_rounded_notation(model_params_count: int) -> str:
if model_params_count > 1e12 :
# Trillions Of Parameters
scaled_model_params = model_params_count * 1e-12
scale_suffix = "T"
elif model_params_count > 1e9 :
# Billions Of Parameters
scaled_model_params = model_params_count * 1e-9
scale_suffix = "B"
elif model_params_count > 1e6 :
# Millions Of Parameters
scaled_model_params = model_params_count * 1e-6
scale_suffix = "M"
else:
# Thousands Of Parameters
scaled_model_params = model_params_count * 1e-3
scale_suffix = "K"
total_params += sum_weights_in_tensor
return f"{round(scaled_model_params)}{scale_suffix}"
return total_params, shared_params, expert_params
def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@ -1230,34 +1249,24 @@ class VocabFactory:
return vocab, special_vocab
def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
quantization = {
def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str:
name = metadata.name if metadata.name is not None else None
basename = metadata.basename if metadata.basename is not None else None
finetune = metadata.finetune if metadata.finetune is not None else None
version = metadata.version if metadata.version is not None else None
size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0)
output_type = {
GGMLFileType.AllF32: "F32",
GGMLFileType.MostlyF16: "F16",
GGMLFileType.MostlyQ8_0: "Q8_0",
}[file_type]
parameters = model_parameter_count_rounded_notation(model_params_count)
expert_count = ""
if params.n_experts is not None:
expert_count = f"{params.n_experts}x"
version = ""
if metadata is not None and metadata.version is not None:
version = f"-{metadata.version}"
name = "ggml-model"
if metadata is not None and metadata.name is not None:
name = metadata.name
elif params.path_model is not None:
name = params.path_model.name
return f"{name}{version}-{expert_count}{parameters}-{quantization}"
return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path:
default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata)
ret = model_paths[0].parent / f"{default_filename}.gguf"
if ret in model_paths:
logger.error(
@ -1296,8 +1305,9 @@ def main(args_in: list[str] | None = None) -> None:
parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
parser.add_argument("--metadata", type=Path, help="Specify the path for a metadata file")
parser.add_argument("--metadata", type=Path, help="Specify the path for an authorship metadata override file")
parser.add_argument("--get-outfile", action="store_true", help="get calculated default outfile name")
parser.add_argument("--model-name", type=str, default=None, help="name of the model")
args = parser.parse_args(args_in)
@ -1309,32 +1319,36 @@ def main(args_in: list[str] | None = None) -> None:
else:
logging.basicConfig(level=logging.INFO)
metadata = Metadata.load(args.metadata)
model_name = args.model_name
dir_model = args.model
metadata = gguf.Metadata.load(args.metadata, dir_model, model_name)
if args.get_outfile:
model_plus = load_some_model(args.model)
model_plus = load_some_model(dir_model)
params = Params.load(model_plus)
model = convert_model_names(model_plus.model, params, args.skip_unknown)
model_params_count = model_parameter_count(model_plus.model)
ftype = pick_output_type(model, args.outtype)
print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
model = convert_model_names(model_plus.model, params, args.skip_unknown)
model_params_count = per_model_weight_count_estimation(model_plus.model.items())
ftype = pick_output_type(model, args.outtype)
if (metadata is None or metadata.name is None) and params.path_model is not None:
metadata.name = params.path_model.name
print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100
return
if args.no_vocab and args.vocab_only:
raise ValueError("--vocab-only does not make sense with --no-vocab")
if args.dump_single:
model_plus = lazy_load_file(args.model)
model_plus = lazy_load_file(dir_model)
do_dump_model(model_plus)
return
if not args.vocab_only:
model_plus = load_some_model(args.model)
model_plus = load_some_model(dir_model)
else:
model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
model_params_count = model_parameter_count(model_plus.model)
logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None)
if args.dump:
do_dump_model(model_plus)
@ -1367,7 +1381,7 @@ def main(args_in: list[str] | None = None) -> None:
logger.info(f"params = {params}")
model_parent_path = model_plus.paths[0].parent
vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
vocab_path = Path(args.vocab_dir or dir_model or model_parent_path)
vocab_factory = VocabFactory(vocab_path)
vocab_types = None if args.no_vocab else args.vocab_type.split(",")
vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
@ -1396,13 +1410,23 @@ def main(args_in: list[str] | None = None) -> None:
if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
vocab = model_plus.vocab
assert params is not None
if metadata.name is None and params.path_model is not None:
metadata.name = params.path_model.name
model_params_count = per_model_weight_count_estimation(model_plus.model.items())
logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})")
logger.info(f"Vocab info: {vocab}")
logger.info(f"Special vocab info: {special_vocab}")
model = model_plus.model
model = convert_model_names(model, params, args.skip_unknown)
ftype = pick_output_type(model, args.outtype)
model = convert_to_output_type(model, ftype)
outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0)
params.ftype = ftype
logger.info(f"Writing {outfile}, format {ftype}")

4
examples/finetune/CMakeLists.txt → examples/cvector-generator/CMakeLists.txt
View file

@ -1,5 +1,5 @@
set(TARGET finetune)
add_executable(${TARGET} finetune.cpp)
set(TARGET llama-cvector-generator)
add_executable(${TARGET} cvector-generator.cpp pca.hpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

45
examples/cvector-generator/README.md Normal file
View file

@ -0,0 +1,45 @@
# cvector-generator
This example demonstrates how to generate a control vector using gguf models.
Related PRs:
- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970)
- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880)
- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514)
## Examples
```sh
# CPU only
./cvector-generator -m ./llama-3.Q4_K_M.gguf
# With GPU
./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99
# With advanced options
./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100
# Using mean value instead of PCA
./cvector-generator -m ./llama-3.Q4_K_M.gguf --method mean
# To see help message
./cvector-generator -h
# Then, have a look at "cvector" section
```
## Tips and tricks
If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example:
```
<|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
<|im_start|>system\nYou are in a very good mood today<|im_end|>
```
Example to use output file with `llama-cli`:
(Tips: The control vector works better when apply to layers higher than 10)
```sh
./llama-cli -m ./llama-3.Q4_K_M.gguf -p "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nSing a song<|im_end|><|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" --special --control-vector-scaled ./control_vector.gguf 0.8 --control-vector-layer-range 10 31
```

582
examples/cvector-generator/completions.txt Normal file
View file

@ -0,0 +1,582 @@
That game
I can see
Hmm, this
I can relate to
Who is
I understand the
Ugh,
What the hell was
Hey, did anyone
Although
Thank you for choosing
What are you
Oh w
How dare you open
It was my pleasure
I'm hon
I appreciate that you
Are you k
Whoever left this
It's always
Ew,
Hey, I l
Hello? Is someone
I understand that
That poem
Aww, poor
Hey, it
Alright, who
I didn't
Well, life
The document
Oh no, this
I'm concerned
Hello, this is
This art
Hmm, this drink
Hi there!
It seems
Is
Good
I can't
Ex
Who are
I can see that
Wow,
Today is a
Hey friend
Sometimes friends
Oh, this old
The weather outside
This place is sur
I appreciate your input
Thank you for the
Look at
I'm disappoint
To my
How dare you
That's an
This piece of art
Eww
This park is
This is incredible
Oh no, someone
Exc
Well, it'
I warned
Hey, I understand
Hey, I saw
How dare you go
What the he
Hey
It's
Hello? Hello?
It
Oh no!
This is the perfect
Good morning,
Oh no, there
It's so
Yeah
Uh,
Hello everyone
Who turned off
The weather
Who'
Hey, this
Wait,
Eww, gross
Excuse
It seems like you
Thank you so
What happened?
Oh my g
I am deeply sad
I war
Okay, let'
Hey, that
That was a beautiful
Oh no! That
What happened
Hey there
The artist'
What?!
Hey, it'
I am disappoint
It seems like
Oh no! The
This park is a
If you
Yes! I did
It sounds
What
Who is it
Hmm, that
That's strange
Yeah, that was
That's interesting
This park
What the hell
Who is that
I feel like my
Oh well
What the hell is
Hello? Hello
To my dearest
Bless you!\"
Thank you for
Oh, looks like
Can you please
This place is
Eww, what
Bless you
Is everything
Hey, I just
Whoever left these
Well, that'
I feel
Hey, do you
It's sad
Oh no, it
Hey, that'
Oh my god,
Thank you,
Hello little one,
I apolog
Hey team, I
How dare you read
Who is this and
Whoever left
Hi there! W
A
If you have
I was
U
Bless
Well, this
Oh, I'
It's a
Eww,
Is everything okay?
Oh, I
Hello, can you
Al
That was a great
What are
I understand that not
Oh no, not
Who is it?\"
Hey, can we
Whoever is taking
I would love to
Hey, I noticed
Hey, could
I understand that there
Hello?
D
Oh man, I
Thank you so much
Oh no, my
Dear [Name
Uh
I remember
Hey, who
Well, it
Are you
I understand that it
Hey, is
I would
Who is this
Excuse me
Alright
I am thrilled
Sometimes friends have
Who the
It's interesting
I would love
E
Hello? Is anyone
Well, this is
This place
Well,
I warned you
Hey, watch where
Oh my
That'
Sometimes friends have different
I understand that everyone
What?
What do these notes
I can relate
I'm not
I understand
To my dear
Guys
Well
Hey, I appreciate
Wow, what
Dear
That melody
Who the hell
Today is
Hello little
Wow, look
That's great
Love is never wrong
I'm having
Whoa, did
Ugh
Can you please provide
I miss you,
I feel uncom
I know
Ugh, this
Hey, watch
Oh great, a
I didn
Okay
That game of char
Oh
I appreciate
Who's there
I am so
Oh great, someone
Hey, could you
I remember wondering
Wait, what?
What do
Hello? Can
Hey there,
That game of
This is incred
Oh my gosh
Oh great, f
I appreciate your
It sounds like
What the heck
Okay, I understand
Ew
I understand that this
Uh, hi
Hi everyone!
What the hell?
Thank you for your
Oh no, the
Wow, I
Who turned
Dear [
Whoever
This is a
Whoa, he
What in the world
Although the physical
Hello, who is
That's amaz
Hey, I know
Okay, that
Hi everyone
Hey, is everything
I understand your fr
Oh no, poor
Oh, look
Good morning
Ew, gross
Oh no, did
Look at the family
Hey team
Yes!
Hey, can I
Okay, that'
It's great
Love is
Hey, what
Good morning, world
Who is it?
That poem really reson
I
That's
I understand the task
Gu
Hello? Who'
This postcard is
Whoa,
Oh, that
I understand that I
Whoever is
Hello? Who is
I'm really
Wow, this
Can
This artwork really
This is a shame
I miss you too
Who are you?
Today is a difficult
Hey, just
Are you okay
I am
Hi,
Wow, that
Hey there! Can
Okay, stay
Oh great, just
Yeah,
Hello? Can you
Oh, looks
Thank you for sharing
I'm glad
Hey, is that
Hmm
It was my
It sounds like you
Wow, your
I was promised certain
That was such a
Thank
Excuse you
That was
Hey team,
I feel un
It was
What'
Hey friend, I
How
Saying goodbye
That
It's heart
How dare
Oh,
Hello, may
What's this
Thank you for recogn
Aww, that
Oh, I remember
Hmm, that'
I miss
I know this
Wait
Is everything okay
Who is that person
Wow, you
Oh great
I'm sad
Wow, the
I am very disappoint
Who turned off the
I understand that things
I'm very
Hi
That's very
Okay, I
Oh no,
Wow, there
What's wrong
I apologize for
Hey, I
Can I help you
Oh, I didn
Alright,
Oh wow,
Oh my goodness
I know this event
What in the
Saying
Yeah, that
Guys, I
Hey, this v
This post
Are
Hey, can
Hello? Is
I can only imagine
Oh, that sounds
Hey, is anyone
I am disappointed
Hello,
Hey everyone, I
That was such
It's okay
The artist
Whoa
I understand that mistakes
Can I help
Who
Hi everyone! I
Hey, can you
Wow, how
Today
Oh no, I
Oh well, I
Well, that
This is the
Yes! I finally
Hey there little
Hello everyone!
Love is never
Look at the
This postcard
Oh great,
Can I
Hmm, this is
I understand your
Oh, look at
B
I'm so
Whoa, this
W
Oh, this
Sometimes
This piece of
What the
That was a
Hey, do
Oh no
Whoa, what
I feel like I
The documentary
Hello
Hello little one
I understand that my
Eww, that
Wow, an
Yes! Finally,
Although the physical location
Whoever is watching
That movie
I remember wondering about
Hey there, little
Who's
Hello, who
Hello everyone! Thank
Hello, can
That's too
Hey, just wanted
Hey there, I
Saying good
Hey there!
Who is there?
Oh my good
I am very
Oh no, what
Wow, thank
I was promised
Hi, is
Hey, I'
Guys, the
Oh no, that
Who is there
Hello, this
That movie really touched
If you have something
The documentary was
I'm starting
Are you kidd
That movie really
Hey everyone,
Thank you for considering
I didn'
Yes! I
Can you
Oh my god
Hey, whoever
That melody really
Thank you, little
Hello, may I
Look
Wow, we
It looks
What do these
Oh wow
I apologize
What are you all
It's such
It's clear
Hey, I was
Hey friend,
I can only
The weather outside is
Eww, this
I miss you
Wow
Aww,
Hi, is there
This artwork
Okay,
Oh well,
This
I'
Say
Hey there little gu
Hmm,
Whoa, who
I am thr
Oh man
Okay, stay calm
I'm happy
Oh, this cur
Oh man,
I'm sorry
Hello? Who
What?! That
This piece
Hey everyone
That's so
Are you okay?
What happened? Where
Hi there
The
Who the hell entered
I can
Guys,
What's
What in
It's important
I'm
I'm coming
It'
Yes! Finally
Wait, what
Wow, reading
I'm surprised
Hey, did
Hey,
Okay, let
I understand that you
Who the hell threw
Eww, who
Thank you for thinking
Who is this?\"
I am deeply
Thank you for including
Oh no, an
It looks like you
Aww
I'm confused
Wow, it
That poem really
Yes
Hey there, is
Hey, what'
Thank you for remember
To
This is
Thank you for making
I can'
That mel
Wow, they
I feel like
Although the
Who are you
Love
If
What the hell are
I am so sad
Oh, I found
Thank you
It looks like
Well, life is
I appreciate that
The artist's
Whoa, that
It's never

503
examples/cvector-generator/cvector-generator.cpp Normal file
View file

@ -0,0 +1,503 @@
#include "common.h"
#include "llama.h"
#include "ggml.h"
#include "pca.hpp"
#include "mean.hpp"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#include <cstdio>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#include <climits>
//////////////////////////////////////////////////
// utils
template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
std::string ret;
for (; begin != end; ++begin) {
ret += llama_token_to_piece(ctx, *begin);
}
return ret;
}
static void print_usage(int argc, char ** argv, const gpt_params & params) {
gpt_params_print_usage(argc, argv, params);
printf("\nexample usage:\n");
printf("\n CPU only: %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
printf("\n with GPU: %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
printf("\n advanced: %s -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100\n", argv[0]);
printf("\n using mean: %s -m ./llama-3.Q4_K_M.gguf --method mean\n", argv[0]);
printf("\n");
}
//////////////////////////////////////////////////
// cb_eval is reused for each pair of positive - negative prompt
struct callback_data {
ggml_context * ctx_ggml = nullptr; // holds v_pos, v_neg, v_diff_filtered
int n_layers = 0;
int n_tokens = 0;
bool is_eval_pos = true;
// each element of the vector correspond to one layer
std::vector<struct ggml_tensor *> v_pos; // vector of matrices of size [n_embd, n_tokens]
std::vector<struct ggml_tensor *> v_neg; // vector of matrices of size [n_embd, n_tokens]
std::vector<struct ggml_tensor *> v_diff_filtered; // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer
// save a tensor into either v_pos or v_neg (decided by is_eval_pos)
void save_tensor_for_layer(struct ggml_tensor * t) {
GGML_ASSERT(t->type == GGML_TYPE_F32);
if (ctx_ggml == nullptr) {
// alloc a new ctx_ggml if needed
struct ggml_init_params params_ggml = {
/*.mem_size =*/ ggml_tensor_overhead() * n_layers * 3u,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx_ggml = ggml_init(params_ggml);
}
// copy tensor data
auto n_bytes = ggml_nbytes(t);
struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]);
t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow
ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes);
ggml_set_name(t_layer, ggml_get_name(t));
//print_debug_tensor(t_layer);
if (is_eval_pos) {
v_pos.push_back(t_layer);
} else {
v_neg.push_back(t_layer);
}
}
// calculate diff (v_pos - v_neg) and place the result back to v_pos
// all zero rows in the diff tensor will also be removed
// NOTE: final layer is ignored. we only have (n_layers - 1) to process
std::vector<struct ggml_tensor *> calc_diff() {
for (float il = 0; il < v_pos.size(); il++) {
float * a = (float *) v_pos[il]->data;
float * b = (float *) v_neg[il]->data;
size_t n_elem = ggml_nelements(v_pos[il]);
for (size_t j = 0; j < n_elem; j++) {
a[j] -= b[j];
}
//print_debug_tensor(v_pos[i]);
auto diff_filtered = filter_nonzero_rows(v_pos[il]);
v_diff_filtered.push_back(diff_filtered);
}
return v_diff_filtered; // for convinient, we return the result std::vector
}
// delete zero rows from a given 2D tensor
struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) {
//printf("filter_nonzero_rows\n");
auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool {
// check if given row containing all zero elements
int n_cols = t->ne[0]; // hint: should be equal to n_embd
for (int col = 0; col < n_cols; ++col) {
if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) {
return false;
}
}
return true;
};
std::vector<int> rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered)
for (int i_row = 0; i_row < a->ne[1]; i_row++) {
if (!is_row_all_zeros(a, i_row, 1e-6)) {
rows_to_copy.push_back(i_row);
}
}
// get "n_nonzero_rows" for the output "diff_filtered"
int n_nonzero_rows = rows_to_copy.size();
//printf("n_nonzero_rows: %d\n", n_nonzero_rows);
int n_embd = a->ne[0];
GGML_ASSERT(n_nonzero_rows > 0);
// diff_filtered: [n_embd, n_nonzero_rows]
struct ggml_tensor * diff_filtered = ggml_new_tensor_2d(
ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows);
ggml_format_name(diff_filtered, "diff_filtered_%s", a->name);
diff_filtered->data = malloc(ggml_nbytes(diff_filtered));
// copy non-zero rows
for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) {
int src_row = rows_to_copy[dest_row];
for (int i = 0; i < n_embd; i++) {
float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0);
ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem);
}
}
//print_debug_tensor(diff_filtered);
return diff_filtered;
}
// we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
void reset() {
for (auto ptr : v_pos) free(ptr->data);
for (auto ptr : v_neg) free(ptr->data);
for (auto ptr : v_diff_filtered) free(ptr->data);
v_pos.clear();
v_neg.clear();
v_diff_filtered.clear();
if (ctx_ggml) {
ggml_free(ctx_ggml);
}
ctx_ggml = nullptr;
}
};
/**
* process_ctx is used to store the ggml context for pre-post processing the diff vectors
* in short, input => v_diff and output => v_final
*/
struct train_context {
ggml_context * ctx_ggml;
int n_embd;
int n_layers;
/* pair of prompts to be used for generating final vector */
std::vector<std::string> positive_entries;
std::vector<std::string> negative_entries;
// each element of the vector correspond to one layer
// NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here
// NOTE (2): v_diff is transposed from v_diff_tmp
std::vector<struct ggml_tensor *> v_diff; // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows)
std::vector<struct ggml_tensor *> v_final; // vector of vectors of size [n_embd] to be written to file
// to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor
// v_diff_tmp will get converted unto v_diff later on
std::vector<std::vector<uint8_t>> v_diff_tmp;
train_context(int n_embd_, int n_layers_) {
n_embd = n_embd_;
n_layers = n_layers_;
struct ggml_init_params params_ggml = {
/*.mem_size =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx_ggml = ggml_init(params_ggml);
for (int il = 0; il < n_layers - 1; il++) {
std::vector<uint8_t> empty;
v_diff_tmp.push_back(empty);
auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd);
t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible
v_final.push_back(t);
}
}
// add new rows into existing tensor in v_diff_tmp
void concat_diff_tmp(const std::vector<struct ggml_tensor *> & diff_filtered) {
GGML_ASSERT((int) diff_filtered.size() == n_layers - 1);
for (int il = 0; il < n_layers - 1; il++) {
auto t = diff_filtered[il];
auto & diff_tmp = v_diff_tmp[il];
size_t curr_size = diff_tmp.size();
diff_tmp.resize(curr_size + ggml_nbytes(t));
memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t));
}
}
// build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
// TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
void build_v_diff(bool transpose) {
printf("build_v_diff\n");
for (int il = 0; il < n_layers - 1; il++) {
auto & diff_tmp = v_diff_tmp[il];
int n_elem = diff_tmp.size() / sizeof(float);
GGML_ASSERT(n_elem % n_embd == 0);
int n_rows = n_elem / n_embd;
struct ggml_tensor * diff = transpose
? ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd)
: ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_embd, n_rows);
ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
if (transpose) {
// copy data & transpose
float * arr = (float *) diff_tmp.data();
for (int ir = 0; ir < n_rows; ++ir) {
for (int ic = 0; ic < n_embd; ++ic) {
float f = arr[ir*n_embd + ic];
ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
}
}
} else {
// only copy
memcpy(diff->data, diff_tmp.data(), ggml_nbytes(diff));
}
v_diff.push_back(diff);
print_debug_tensor(diff);
// free memory of diff_tmp
diff_tmp.resize(0);
}
}
~train_context() {
for (auto ptr : v_final) free(ptr->data);
for (auto ptr : v_diff) free(ptr->data);
// no need to free v_diff_tmp, since we didn't use malloc
ggml_free(ctx_ggml);
}
};
struct tokenized_prompt {
std::vector<llama_token> tokens_pos;
std::vector<llama_token> tokens_neg;
size_t max_seq_len;
tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
padding_seq(ctx, tokens_pos, max_seq_len);
padding_seq(ctx, tokens_neg, max_seq_len);
}
void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
// TODO: customize padding token
std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
llama_token pad_tok = pad_tokens.back();
while (tokens.size() < len) {
tokens.push_back(pad_tok);
}
}
};
//////////////////////////////////////////////////
template <typename T>
static std::string to_string(const T & val) {
std::stringstream ss;
ss << val;
return ss.str();
}
static std::vector<std::string> ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) {
std::vector<std::string> output;
std::ifstream file(path);
if (!file.is_open()) {
fprintf(stderr, "error: unable to open file: %s\n", path.c_str());
exit(1);
}
std::string line;
while (std::getline(file, line)) {
bool is_skip = skip_empty_lines && line.empty();
if (!is_skip) {
string_process_escapes(line);
output.push_back(line);
}
}
file.close();
return output;
}
//////////////////////////////////////////////////
static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
auto * cb_data = (callback_data *) user_data;
static const char * l_out_name = "l_out";
const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
if (ask) {
return is_l_out;
}
if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
return true;
}
// save the tensor to current context
cb_data->save_tensor_for_layer(t);
return true;
}
static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
llama_kv_cache_clear(ctx);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}
return true;
}
static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const std::string fname, const std::string model_hint) {
struct gguf_context * ctx = gguf_init_empty();
const std::string arch = "controlvector";
gguf_set_val_str(ctx, "general.architecture", arch.c_str());
gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str());
gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size());
for (size_t i = 0; i < v_ctrl.size(); ++i) {
gguf_add_tensor(ctx, v_ctrl[i]);
print_debug_tensor(v_ctrl[i]);
printf("Added tensor: %s\n", v_ctrl[i]->name);
}
printf("%s: writing file...\n", __func__);
gguf_write_to_file(ctx, fname.c_str(), false);
printf("%s: wrote file '%s'\n", __func__, fname.c_str());
gguf_free(ctx);
}
/**
* Load prompt files and completion file.
* Then format each pair of prompt + completion to make an entry.
*/
static int prepare_entries(gpt_params & params, train_context & ctx_train) {
// load prompts
std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
if (positive_prompts.size() != negative_prompts.size()) {
fprintf(stderr, "number of positive and negative prompts must be equal\n");
return 1;
}
if (positive_prompts.empty()) {
fprintf(stderr, "must provide at least one prompt pair\n");
return 1;
}
ctx_train.positive_entries = positive_prompts;
ctx_train.negative_entries = negative_prompts;
return 0;
}
int main(int argc, char ** argv) {
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
print_usage(argc, argv, params);
return 1;
}
if (params.n_pca_iterations % params.n_pca_batch != 0) {
fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n");
return 1;
}
callback_data cb_data;
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
params.cb_eval = cb_eval;
params.cb_eval_user_data = &cb_data;
params.warmup = false;
print_build_info();
llama_backend_init();
llama_numa_init(params.numa);
// load the model to get hparams
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
// int n_ctx = llama_n_ctx(ctx);
int n_layers = llama_n_layer(model);
int n_embd = llama_n_embd(model);
// get model hint param (a.k.a model arch name)
char model_hint[128];
llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
// init train_context
train_context ctx_train(n_embd, n_layers);
// load and prepare entries for training
prepare_entries(params, ctx_train);
// we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped
std::vector<tokenized_prompt> tokenized_prompts;
size_t n_total_tokens = 0;
for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]);
n_total_tokens += 2 * t.max_seq_len;
tokenized_prompts.push_back(std::move(t));
}
std::cout << "n_total_tokens: " << n_total_tokens << std::endl;
for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
bool success = false;
tokenized_prompt t = tokenized_prompts[i];
cb_data.n_layers = n_layers;
cb_data.n_tokens = t.max_seq_len;
printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
(int) i+1, (int) ctx_train.positive_entries.size(),
tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(),
tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
(int) t.max_seq_len);
cb_data.is_eval_pos = true;
success = get_hidden_layers(ctx, t.tokens_pos);
if (!success) break;
cb_data.is_eval_pos = false;
success = get_hidden_layers(ctx, t.tokens_neg);
if (!success) break;
// calculate diff and remove all zero rows
auto v_diff_filtered = cb_data.calc_diff();
// save & concat the filtered v_diff to ctx_train
ctx_train.concat_diff_tmp(v_diff_filtered);
// reset for next iteration
cb_data.reset();
}
// done with the model, we can now free it to make gain some memory
printf("Done evaluate prompts, unload model...\n");
llama_free(ctx);
llama_free_model(model);
bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
// prepare ctx_train for PCA
ctx_train.build_v_diff(use_pca);
if (use_pca) {
// run PCA
PCA::pca_params pca_params;
pca_params.n_threads = params.n_threads;
pca_params.n_batch = params.n_pca_batch;
pca_params.n_iterations = params.n_pca_iterations;
PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
} else {
// run mean
mean::run(ctx_train.v_diff, ctx_train.v_final);
}
// write output vectors to gguf
export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);
llama_backend_free();
return 0;
}

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