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

Merge branch 'master' into dev-refactoring

This commit is contained in:
hongruichen 2025-02-24 14:39:12 +08:00
parent a822d00753 7ad0779f5d
commit 84328ffbda
194 changed files with 9118 additions and 3583 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.devops/cuda.Dockerfile
View File

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG CUDA_VERSION=12.6.0
ARG CUDA_VERSION=12.4.0
# Target the CUDA build image
ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}

4
.devops/llama-cpp-cuda.srpm.spec
View File

@ -17,10 +17,10 @@ Version: %( date "+%%Y%%m%%d" )
Release: 1%{?dist}
Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
License: MIT
Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
Source0: https://github.com/ggml-org/llama.cpp/archive/refs/heads/master.tar.gz
BuildRequires: coreutils make gcc-c++ git cuda-toolkit
Requires: cuda-toolkit
URL: https://github.com/ggerganov/llama.cpp
URL: https://github.com/ggml-org/llama.cpp
%define debug_package %{nil}
%define source_date_epoch_from_changelog 0

4
.devops/llama-cpp.srpm.spec
View File

@ -18,10 +18,10 @@ Version: %( date "+%%Y%%m%%d" )
Release: 1%{?dist}
Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
License: MIT
Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
Source0: https://github.com/ggml-org/llama.cpp/archive/refs/heads/master.tar.gz
BuildRequires: coreutils make gcc-c++ git libstdc++-devel
Requires: libstdc++
URL: https://github.com/ggerganov/llama.cpp
URL: https://github.com/ggml-org/llama.cpp
%define debug_package %{nil}
%define source_date_epoch_from_changelog 0

2
.devops/musa.Dockerfile
View File

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG MUSA_VERSION=rc3.1.0
ARG MUSA_VERSION=rc3.1.1
# Target the MUSA build image
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}

6
.devops/nix/package.nix
View File

@ -133,12 +133,12 @@ effectiveStdenv.mkDerivation (finalAttrs: {
--replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
'';
# With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
# With PR#6015 https://github.com/ggml-org/llama.cpp/pull/6015,
# `default.metallib` may be compiled with Metal compiler from XCode
# and we need to escape sandbox on MacOS to access Metal compiler.
# `xcrun` is used find the path of the Metal compiler, which is varible
# and not on $PATH
# see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
# see https://github.com/ggml-org/llama.cpp/pull/6118 for discussion
__noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
nativeBuildInputs =
@ -220,7 +220,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
broken = (useMetalKit && !effectiveStdenv.isDarwin);
description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
homepage = "https://github.com/ggerganov/llama.cpp/";
homepage = "https://github.com/ggml-org/llama.cpp/";
license = lib.licenses.mit;
# Accommodates `nix run` and `lib.getExe`

2
.devops/rocm.Dockerfile
View File

@ -11,7 +11,7 @@ ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-co
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
# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
# This is mostly tied to rocBLAS supported archs.
# gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
# gfx906 is deprecated

6
.github/ISSUE_TEMPLATE/020-enhancement.yml vendored
View File

@ -6,7 +6,7 @@ body:
- type: markdown
attributes:
value: |
[Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas)
[Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggml-org/llama.cpp/discussions/categories/ideas)
- type: checkboxes
id: prerequisites
@ -16,11 +16,11 @@ body:
options:
- label: I am running the latest code. Mention the version if possible as well.
required: true
- label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
- label: I carefully followed the [README.md](https://github.com/ggml-org/llama.cpp/blob/master/README.md).
required: true
- label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
required: true
- label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share.
- label: I reviewed the [Discussions](https://github.com/ggml-org/llama.cpp/discussions), and have a new and useful enhancement to share.
required: true
- type: textarea

2
.github/ISSUE_TEMPLATE/030-research.yml vendored
View File

@ -6,7 +6,7 @@ body:
- type: markdown
attributes:
value: |
Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
Don't forget to check for any [duplicate research issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
- type: checkboxes
id: research-stage

4
.github/ISSUE_TEMPLATE/040-refactor.yml vendored
View File

@ -6,8 +6,8 @@ body:
- type: markdown
attributes:
value: |
Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
Don't forget to [check for existing refactor issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
Also you may want to check [Pull request refactor label as well](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
- type: textarea
id: background-description

6
.github/ISSUE_TEMPLATE/config.yml vendored
View File

@ -1,11 +1,11 @@
blank_issues_enabled: true
contact_links:
- name: Got an idea?
url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas
url: https://github.com/ggml-org/llama.cpp/discussions/categories/ideas
about: Pop it there. It may then become an enhancement ticket.
- name: Got a question?
url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a
url: https://github.com/ggml-org/llama.cpp/discussions/categories/q-a
about: Ask a question there!
- name: Want to contribute?
url: https://github.com/ggerganov/llama.cpp/wiki/contribute
url: https://github.com/ggml-org/llama.cpp/wiki/contribute
about: Head to the contribution guide page of the wiki for areas you can help with

2
.github/pull_request_template.md vendored
View File

@ -1 +1 @@
*Make sure to read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*
*Make sure to read the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*

12
.github/workflows/bench.yml.disabled vendored
View File

@ -1,5 +1,5 @@
# TODO: there have been some issues with the workflow, so disabling for now
# https://github.com/ggerganov/llama.cpp/issues/7893
# https://github.com/ggml-org/llama.cpp/issues/7893
#
# Benchmark
name: Benchmark
@ -57,17 +57,7 @@ jobs:
if: |
inputs.gpu-series == 'Standard_NC4as_T4_v3'
|| (
github.event_name == 'schedule'
&& github.ref_name == 'master'
&& github.repository_owner == 'ggerganov'
)
|| github.event_name == 'pull_request_target'
|| (
github.event_name == 'push'
&& github.event.ref == 'refs/heads/master'
&& github.repository_owner == 'ggerganov'
)
steps:
- name: Clone
id: checkout

54
.github/workflows/build.yml vendored
View File

@ -129,7 +129,7 @@ jobs:
run: |
sysctl -a
# 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
# https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build \
-DCMAKE_BUILD_RPATH="@loader_path" \
-DLLAMA_FATAL_WARNINGS=ON \
@ -173,7 +173,15 @@ jobs:
name: llama-bin-macos-x64.zip
ubuntu-cpu-cmake:
runs-on: ubuntu-22.04
strategy:
matrix:
include:
- build: 'x64'
os: ubuntu-22.04
- build: 'arm64'
os: ubuntu-22.04-arm
runs-on: ${{ matrix.os }}
steps:
- name: Clone
@ -239,14 +247,14 @@ jobs:
run: |
cp LICENSE ./build/bin/
cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/*
zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*
- name: Upload artifacts
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
uses: actions/upload-artifact@v4
with:
path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
name: llama-bin-ubuntu-x64.zip
path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
name: llama-bin-ubuntu-${{ matrix.build }}.zip
ubuntu-latest-cmake-sanitizer:
runs-on: ubuntu-latest
@ -374,6 +382,8 @@ jobs:
- name: Clone
id: checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: ccache
uses: hendrikmuhs/ccache-action@v1.2.16
@ -401,7 +411,35 @@ jobs:
run: |
cd build
# This is using llvmpipe and runs slower than other backends
ctest -L main --verbose --timeout 1800
ctest -L main --verbose --timeout 2700
- name: Determine tag name
id: tag
shell: bash
run: |
BUILD_NUMBER="$(git rev-list --count HEAD)"
SHORT_HASH="$(git rev-parse --short=7 HEAD)"
if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
else
SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
fi
- name: Pack artifacts
id: pack_artifacts
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
run: |
cp LICENSE ./build/bin/
cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
- name: Upload artifacts
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
uses: actions/upload-artifact@v4
with:
path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
name: llama-bin-ubuntu-vulkan-x64.zip
ubuntu-22-cmake-hip:
runs-on: ubuntu-22.04
@ -443,7 +481,7 @@ jobs:
ubuntu-22-cmake-musa:
runs-on: ubuntu-22.04
container: mthreads/musa:rc3.1.0-devel-ubuntu22.04
container: mthreads/musa:rc3.1.1-devel-ubuntu22.04
steps:
- name: Clone
@ -1345,8 +1383,10 @@ jobs:
needs:
- ubuntu-cpu-cmake
- ubuntu-22-cmake-vulkan
- windows-latest-cmake
- windows-2019-cmake-cuda
- windows-latest-cmake-sycl
- windows-latest-cmake-hip-release
- macOS-latest-cmake-arm64
- macOS-latest-cmake-x64

2
.github/workflows/docker.yml vendored
View File

@ -51,6 +51,8 @@ jobs:
- name: Set up QEMU
uses: docker/setup-qemu-action@v3
with:
image: tonistiigi/binfmt:qemu-v7.0.0-28
- name: Set up Docker Buildx
uses: docker/setup-buildx-action@v3

2
.github/workflows/labeler.yml vendored
View File

@ -11,7 +11,7 @@ jobs:
steps:
- uses: actions/checkout@v4
with:
repository: "ggerganov/llama.cpp"
repository: "ggml-org/llama.cpp"
- uses: actions/labeler@v5
with:
configuration-path: '.github/labeler.yml'

1
.gitignore vendored
View File

@ -98,6 +98,7 @@ examples/server/*.css.hpp
examples/server/*.html.hpp
examples/server/*.js.hpp
examples/server/*.mjs.hpp
examples/server/*.gz.hpp
!build_64.sh
!examples/*.bat
!examples/*/*.kts

10
CONTRIBUTING.md
View File

@ -1,10 +1,12 @@
# Pull requests (for contributors)
- llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier
- Test your changes:
- Execute [the full CI locally on your machine](ci/README.md) before publishing
- Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
- If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
- If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
- Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
@ -12,7 +14,7 @@
- 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
- Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
- Consider adding yourself to [CODEOWNERS](CODEOWNERS)
# Coding guidelines
@ -40,14 +42,14 @@
- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` to format the added code
- For anything not covered in the current guidelines, refer to the [C++ Core Guidelines](https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines)
- 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 $C^T = A B^T \Leftrightarrow C = B A^T.$
- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggml-org/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
![matmul](media/matmul.png)
# Naming guidelines
- Use `snake_case` for function, variable and type names
- Naming usually optimizes for longest common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
- Naming usually optimizes for longest common prefix (see https://github.com/ggml-org/ggml/pull/302#discussion_r1243240963)
```cpp
// not OK
@ -122,4 +124,4 @@
The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
https://github.com/ggerganov/llama.cpp/projects
https://github.com/ggml-org/llama.cpp/projects

24
Makefile
View File

@ -1,5 +1,5 @@
ifndef LLAMA_MAKEFILE
$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
endif
# Define the default target now so that it is always the first target
@ -463,7 +463,7 @@ endif
ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
# https://github.com/ggerganov/llama.cpp/issues/2922
# https://github.com/ggml-org/llama.cpp/issues/2922
MK_CFLAGS += -Xassembler -muse-unaligned-vector-move
MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
@ -680,6 +680,10 @@ ifdef GGML_CUDA_CCBIN
MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN)
endif # GGML_CUDA_CCBIN
ifdef GGML_CUDA_NO_FA
MK_NVCCFLAGS += -DGGML_CUDA_NO_FA
endif # GGML_CUDA_NO_FA
ifdef GGML_CUDA_FA_ALL_QUANTS
MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
endif # GGML_CUDA_FA_ALL_QUANTS
@ -800,6 +804,10 @@ ifdef GGML_CUDA_NO_PEER_COPY
HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
endif # GGML_CUDA_NO_PEER_COPY
ifdef GGML_CUDA_NO_FA
HIPFLAGS += -DGGML_CUDA_NO_FA
endif # GGML_CUDA_NO_FA
OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
@ -847,7 +855,7 @@ ifdef GGML_MUSA
CXX := $(MUSA_PATH)/bin/clang++
MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc
MUSAFLAGS = -x musa -mtgpu
MUSAFLAGS = -fsigned-char -x musa -mtgpu
MUSAFLAGS += $(foreach arch,$(subst ;, ,$(MUSA_ARCHITECTURES)),--cuda-gpu-arch=mp_$(arch))
ifdef GGML_CUDA_FORCE_MMQ
@ -876,6 +884,10 @@ ifdef GGML_CUDA_NO_PEER_COPY
MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY
endif # GGML_CUDA_NO_PEER_COPY
ifdef GGML_CUDA_NO_FA
MUSAFLAGS += -DGGML_CUDA_NO_FA
endif # GGML_CUDA_NO_FA
ifdef GGML_CUDA_FA_ALL_QUANTS
MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
endif # GGML_CUDA_FA_ALL_QUANTS
@ -1078,8 +1090,8 @@ endif
ifdef REMOVE_WARNING
$(info !!! REMOVAL WARNING !!!)
$(info The following LLAMA_ options have been removed and are no longer supported)
$(info - LLAMA_DISABLE_LOGS (https://github.com/ggerganov/llama.cpp/pull/9418))
$(info - LLAMA_SERVER_VERBOSE (https://github.com/ggerganov/llama.cpp/pull/9418))
$(info - LLAMA_DISABLE_LOGS (https://github.com/ggml-org/llama.cpp/pull/9418))
$(info - LLAMA_SERVER_VERBOSE (https://github.com/ggml-org/llama.cpp/pull/9418))
$(info )
endif
@ -1364,7 +1376,7 @@ llama-server: \
examples/server/index.html.hpp \
examples/server/loading.html.hpp \
common/chat.cpp \
common/chat.hpp \
common/chat.h \
common/chat-template.hpp \
common/json.hpp \
common/minja.hpp \

74
README.md
View File

@ -3,26 +3,33 @@
![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)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
[![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
[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)
[Roadmap](https://github.com/users/ggml-org/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
> [!IMPORTANT]
> New `llama.cpp` package location: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/pkgs/container/llama.cpp)
>
> Update your container URLs to: `ghcr.io/ggml-org/llama.cpp`
>
> More info: https://github.com/ggml-org/llama.cpp/discussions/11801
## Recent API changes
- [Changelog for `libllama` API](https://github.com/ggerganov/llama.cpp/issues/9289)
- [Changelog for `llama-server` REST API](https://github.com/ggerganov/llama.cpp/issues/9291)
- [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)
- [Changelog for `llama-server` REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
## Hot topics
- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggerganov/llama.cpp/pull/11427
- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
- **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
- Universal tool call support in `llama-server`: https://github.com/ggerganov/llama.cpp/pull/9639
- Universal tool call support in `llama-server`: https://github.com/ggml-org/llama.cpp/pull/9639
- Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
- Introducing GGUF-my-LoRA https://github.com/ggerganov/llama.cpp/discussions/10123
- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
- Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
----
@ -39,7 +46,7 @@ range of hardware - locally and in the cloud.
- Vulkan and SYCL backend support
- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggml-org/ggml) library.
<details>
<summary>Models</summary>
@ -59,23 +66,23 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [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] [BERT](https://github.com/ggml-org/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)
- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
- [X] [Starcoder models](https://github.com/ggml-org/llama.cpp/pull/3187)
- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
- [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
- [X] [MPT](https://github.com/ggml-org/llama.cpp/pull/3417)
- [X] [Bloom](https://github.com/ggml-org/llama.cpp/pull/3553)
- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
- [X] [StableLM models](https://huggingface.co/stabilityai)
- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
- [x] [PLaMo-13B](https://github.com/ggml-org/llama.cpp/pull/3557)
- [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
- [x] [PhiMoE](https://github.com/ggerganov/llama.cpp/pull/11003)
- [x] [PhiMoE](https://github.com/ggml-org/llama.cpp/pull/11003)
- [x] [GPT-2](https://huggingface.co/gpt2)
- [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
- [x] [Orion 14B](https://github.com/ggml-org/llama.cpp/pull/5118)
- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
- [x] [CodeShell](https://github.com/WisdomShell/codeshell)
- [x] [Gemma](https://ai.google.dev/gemma)
@ -146,7 +153,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
- Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
- 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)
- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggml-org/llama.cpp/pull/6326)
- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
@ -245,7 +252,7 @@ The project also includes many example programs and tools using the `llama` libr
- Clone this repository and build locally, see [how to build](docs/build.md)
- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
- Use a Docker image, see [documentation for Docker](docs/docker.md)
- Download pre-built binaries from [releases](https://github.com/ggerganov/llama.cpp/releases)
- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases)
## Obtaining and quantizing models
@ -258,14 +265,14 @@ You can either manually download the GGUF file or directly use any `llama.cpp`-c
After downloading a model, use the CLI tools to run it locally - see below.
`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggml-org/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`:
- Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes
- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggerganov/llama.cpp/discussions/10123)
- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggerganov/llama.cpp/discussions/9268)
- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggerganov/llama.cpp/discussions/9669)
- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggml-org/llama.cpp/discussions/10123)
- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268)
- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669)
To learn more about model quantization, [read this documentation](examples/quantize/README.md)
@ -488,9 +495,9 @@ To learn more about model quantization, [read this documentation](examples/quant
- 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, PRs and projects 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
- See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
- Make sure to read this: [Inference at the edge](https://github.com/ggml-org/llama.cpp/discussions/205)
- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
## Other documentation
@ -505,7 +512,7 @@ To learn more about model quantization, [read this documentation](examples/quant
- [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)
- [GGML tips & tricks](https://github.com/ggml-org/llama.cpp/wiki/GGML-Tips-&-Tricks)
#### Seminal papers and background on the models
@ -519,5 +526,18 @@ If your issue is with model generation quality, then please at least scan the fo
- [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
#### References
## Completions
Command-line completion is available for some environments.
#### Bash Completion
```bash
$ build/bin/llama-cli --completion-bash > ~/.llama-completion.bash
$ source ~/.llama-completion.bash
```
Optionally this can be added to your `.bashrc` or `.bash_profile` to load it
automatically. For example:
```console
$ echo "source ~/.llama-completion.bash" >> ~/.bashrc
```
## References

2
SECURITY.md
View File

@ -62,6 +62,6 @@ Beware that none of the topics under [Using llama.cpp securely](#using-llamacpp-
<!-- normal version -->
However, If you have discovered a security vulnerability in this project, please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.
Please disclose it as a private [security advisory](https://github.com/ggerganov/llama.cpp/security/advisories/new).
Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new).
A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.

4
ci/README.md
View File

@ -1,11 +1,11 @@
# CI
In addition to [Github Actions](https://github.com/ggerganov/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
In addition to [Github Actions](https://github.com/ggml-org/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
https://github.com/ggml-org/ci
It monitors the `master` branch for new commits and runs the
[ci/run.sh](https://github.com/ggerganov/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
[ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
to cover various hardware architectures, including GPU and Apple Silicon instances.

29
common/CMakeLists.txt
View File

@ -57,8 +57,7 @@ add_library(${TARGET} STATIC
arg.h
base64.hpp
chat.cpp
chat.hpp
chat-template.hpp
chat.h
common.cpp
common.h
console.cpp
@ -68,7 +67,8 @@ add_library(${TARGET} STATIC
llguidance.cpp
log.cpp
log.h
minja.hpp
minja/chat-template.hpp
minja/minja.hpp
ngram-cache.cpp
ngram-cache.h
sampling.cpp
@ -96,6 +96,22 @@ if (LLAMA_LLGUIDANCE)
include(ExternalProject)
set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
# Set the correct library file extension based on platform
if (WIN32)
set(LLGUIDANCE_LIB_NAME "llguidance.lib")
# Add Windows-specific libraries
set(LLGUIDANCE_PLATFORM_LIBS
ws2_32 # Windows Sockets API
userenv # For GetUserProfileDirectoryW
ntdll # For NT functions
bcrypt # For BCryptGenRandom
)
else()
set(LLGUIDANCE_LIB_NAME "libllguidance.a")
set(LLGUIDANCE_PLATFORM_LIBS "")
endif()
ExternalProject_Add(llguidance_ext
GIT_REPOSITORY https://github.com/guidance-ai/llguidance
# v0.6.12:
@ -106,17 +122,18 @@ if (LLAMA_LLGUIDANCE)
CONFIGURE_COMMAND ""
BUILD_COMMAND cargo build --release
INSTALL_COMMAND ""
BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/libllguidance.a ${LLGUIDANCE_PATH}/llguidance.h
BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
UPDATE_COMMAND ""
)
target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_LLGUIDANCE)
add_library(llguidance STATIC IMPORTED)
set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/libllguidance.a)
set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME})
add_dependencies(llguidance llguidance_ext)
target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance)
# Add platform libraries to the main target
set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
endif ()
target_include_directories(${TARGET} PUBLIC .)

188
common/arg.cpp
View File

@ -2,6 +2,7 @@
#include "log.h"
#include "sampling.h"
#include "chat.h"
#include <algorithm>
#include <climits>
@ -365,6 +366,112 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
print_options(specific_options);
}
static void common_params_print_completion(common_params_context & ctx_arg) {
std::vector<common_arg *> common_options;
std::vector<common_arg *> sparam_options;
std::vector<common_arg *> specific_options;
for (auto & opt : ctx_arg.options) {
if (opt.is_sparam) {
sparam_options.push_back(&opt);
} else if (opt.in_example(ctx_arg.ex)) {
specific_options.push_back(&opt);
} else {
common_options.push_back(&opt);
}
}
printf("_llama_completions() {\n");
printf(" local cur prev opts\n");
printf(" COMPREPLY=()\n");
printf(" cur=\"${COMP_WORDS[COMP_CWORD]}\"\n");
printf(" prev=\"${COMP_WORDS[COMP_CWORD-1]}\"\n\n");
printf(" opts=\"");
auto print_options = [](const std::vector<common_arg *> & options) {
for (const common_arg * opt : options) {
for (const char * arg : opt->args) {
printf("%s ", arg);
}
}
};
print_options(common_options);
print_options(sparam_options);
print_options(specific_options);
printf("\"\n\n");
printf(" case \"$prev\" in\n");
printf(" --model)\n");
printf(" COMPREPLY=( $(compgen -f -X '!*.gguf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
printf(" return 0\n");
printf(" ;;\n");
printf(" --grammar-file)\n");
printf(" COMPREPLY=( $(compgen -f -X '!*.gbnf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
printf(" return 0\n");
printf(" ;;\n");
printf(" --chat-template-file)\n");
printf(" COMPREPLY=( $(compgen -f -X '!*.jinja' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
printf(" return 0\n");
printf(" ;;\n");
printf(" *)\n");
printf(" COMPREPLY=( $(compgen -W \"${opts}\" -- \"$cur\") )\n");
printf(" return 0\n");
printf(" ;;\n");
printf(" esac\n");
printf("}\n\n");
std::set<std::string> executables = {
"llama-batched",
"llama-batched-bench",
"llama-bench",
"llama-cli",
"llama-convert-llama2c-to-ggml",
"llama-cvector-generator",
"llama-embedding",
"llama-eval-callback",
"llama-export-lora",
"llama-gbnf-validator",
"llama-gen-docs",
"llama-gguf",
"llama-gguf-hash",
"llama-gguf-split",
"llama-gritlm",
"llama-imatrix",
"llama-infill",
"llama-llava-cli",
"llama-llava-clip-quantize-cli",
"llama-lookahead",
"llama-lookup",
"llama-lookup-create",
"llama-lookup-merge",
"llama-lookup-stats",
"llama-minicpmv-cli",
"llama-parallel",
"llama-passkey",
"llama-perplexity",
"llama-q8dot",
"llama-quantize",
"llama-quantize-stats",
"llama-qwen2vl-cli",
"llama-retrieval",
"llama-run",
"llama-save-load-state",
"llama-server",
"llama-simple",
"llama-simple-chat",
"llama-speculative",
"llama-speculative-simple",
"llama-tokenize",
"llama-tts",
"llama-vdot"
};
for (const auto& exe : executables) {
printf("complete -F _llama_completions %s\n", exe.c_str());
}
}
static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & value) {
std::vector<ggml_backend_dev_t> devices;
auto dev_names = string_split<std::string>(value, ',');
@ -426,6 +533,10 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
}
exit(0);
}
if (ctx_arg.params.completion) {
common_params_print_completion(ctx_arg);
exit(0);
}
} catch (const std::invalid_argument & ex) {
fprintf(stderr, "%s\n", ex.what());
ctx_arg.params = params_org;
@ -494,6 +605,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
exit(0);
}
));
add_opt(common_arg(
{"--completion-bash"},
"print source-able bash completion script for llama.cpp",
[](common_params & params) {
params.completion = true;
}
));
add_opt(common_arg(
{"--verbose-prompt"},
string_format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"),
@ -946,6 +1064,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.sampling.min_p = std::stof(value);
}
).set_sparam());
add_opt(common_arg(
{"--top-nsigma"}, "N",
string_format("top-n-sigma sampling (default: %.1f, -1.0 = disabled)", params.sampling.top_n_sigma),
[](common_params & params, const std::string & value) {
params.sampling.top_n_sigma = std::stof(value);
}
).set_examples({LLAMA_EXAMPLE_MAIN}).set_sparam());
add_opt(common_arg(
{"--xtc-probability"}, "N",
string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
@ -1445,7 +1570,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
"- isolate: only spawn threads on CPUs on the node that execution started on\n"
"- numactl: use the CPU map provided by numactl\n"
"if run without this previously, it is recommended to drop the system page cache before using this\n"
"see https://github.com/ggerganov/llama.cpp/issues/1437",
"see https://github.com/ggml-org/llama.cpp/issues/1437",
[](common_params & params, const std::string & value) {
/**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
@ -1975,6 +2100,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.use_jinja = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
add_opt(common_arg(
{"--reasoning-format"}, "FORMAT",
"reasoning format (default: deepseek; allowed values: deepseek, none)\n"
"controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).\n"
"only supported for non-streamed responses",
[](common_params & params, const std::string & value) {
/**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
else if (value == "none") { params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
else { std::invalid_argument("invalid value"); }
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
add_opt(common_arg(
{"--chat-template"}, "JINJA_TEMPLATE",
string_format(
@ -2112,7 +2248,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
).set_env("LLAMA_LOG_VERBOSITY"));
add_opt(common_arg(
{"--log-prefix"},
"Enable prefx in log messages",
"Enable prefix in log messages",
[](common_params &) {
common_log_set_prefix(common_log_main(), true);
}
@ -2366,5 +2502,53 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--fim-qwen-1.5b-default"},
string_format("use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
params.n_cache_reuse = 256;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--fim-qwen-3b-default"},
string_format("use default Qwen 2.5 Coder 3B (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
params.n_cache_reuse = 256;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--fim-qwen-7b-default"},
string_format("use default Qwen 2.5 Coder 7B (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
params.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
params.n_cache_reuse = 256;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
return ctx_arg;
}

1033
common/chat.cpp
View File

File diff suppressed because it is too large Load Diff

134
common/chat.h Normal file
View File

@ -0,0 +1,134 @@
// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers.
#pragma once
#include "common.h"
#include <string>
#include <vector>
struct common_chat_templates;
struct common_chat_tool_call {
std::string name;
std::string arguments;
std::string id;
};
struct common_chat_msg_content_part {
std::string type;
std::string text;
};
struct common_chat_msg {
std::string role;
std::string content;
std::vector<common_chat_msg_content_part> content_parts = {};
std::vector<common_chat_tool_call> tool_calls = {};
std::string reasoning_content;
std::string tool_name;
std::string tool_call_id;
};
struct common_chat_tool {
std::string name;
std::string description;
std::string parameters;
};
enum common_chat_tool_choice {
COMMON_CHAT_TOOL_CHOICE_AUTO,
COMMON_CHAT_TOOL_CHOICE_REQUIRED,
COMMON_CHAT_TOOL_CHOICE_NONE,
};
enum common_chat_format {
COMMON_CHAT_FORMAT_CONTENT_ONLY,
COMMON_CHAT_FORMAT_GENERIC,
COMMON_CHAT_FORMAT_MISTRAL_NEMO,
COMMON_CHAT_FORMAT_LLAMA_3_X,
COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
COMMON_CHAT_FORMAT_DEEPSEEK_R1,
COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING,
COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
COMMON_CHAT_FORMAT_HERMES_2_PRO,
COMMON_CHAT_FORMAT_COMMAND_R7B,
COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING,
COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
};
struct common_chat_templates_inputs {
std::vector<common_chat_msg> messages;
std::string grammar;
std::string json_schema;
bool add_generation_prompt = true;
bool use_jinja = true;
// Parameters below only supported when use_jinja is true
std::vector<common_chat_tool> tools;
common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
bool parallel_tool_calls = false;
bool extract_reasoning = true;
};
struct common_chat_params {
common_chat_format format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
std::string prompt;
std::string grammar;
bool grammar_lazy = false;
std::vector<common_grammar_trigger> grammar_triggers;
std::vector<std::string> preserved_tokens;
std::vector<std::string> additional_stops;
};
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
void common_chat_templates_free(struct common_chat_templates * tmpls);
struct common_chat_templates_deleter { void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); } };
typedef std::unique_ptr<struct common_chat_templates, common_chat_templates_deleter> common_chat_templates_ptr;
common_chat_templates_ptr common_chat_templates_init(
const struct llama_model * model,
const std::string & chat_template_override,
const std::string & bos_token_override = "",
const std::string & eos_token_override = "");
bool common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr);
struct common_chat_params common_chat_templates_apply(
const struct common_chat_templates * tmpls,
const struct common_chat_templates_inputs & inputs);
// Format single message, while taking into account the position of that message in chat history
std::string common_chat_format_single(
const struct common_chat_templates * tmpls,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass,
bool use_jinja);
// Returns an example of formatted chat
std::string common_chat_format_example(
const struct common_chat_templates * tmpls,
bool use_jinja);
std::string common_chat_format_name(common_chat_format format);
common_chat_msg common_chat_parse( const std::string & input, common_chat_format format);
common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
// Parses a JSON array of messages in OpenAI's chat completion API format.
// T can be std::string containing JSON or nlohmann::ordered_json
template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);
template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
// Parses a JSON array of tools in OpenAI's chat completion tool call API format.
// T can be std::string containing JSON or nlohmann::ordered_json
template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);

52
common/chat.hpp
View File

@ -1,52 +0,0 @@
// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers.
#pragma once
#include "common.h"
#include <json.hpp>
#include <optional>
#include <string>
#include <vector>
using json = nlohmann::ordered_json;
struct common_chat_inputs {
json messages;
json tools;
json tool_choice;
json json_schema;
bool parallel_tool_calls;
bool stream;
std::string grammar;
bool add_generation_prompt = true;
};
enum common_chat_format {
COMMON_CHAT_FORMAT_CONTENT_ONLY,
COMMON_CHAT_FORMAT_GENERIC,
COMMON_CHAT_FORMAT_MISTRAL_NEMO,
COMMON_CHAT_FORMAT_LLAMA_3_X,
COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
COMMON_CHAT_FORMAT_DEEPSEEK_R1,
COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
COMMON_CHAT_FORMAT_HERMES_2_PRO,
COMMON_CHAT_FORMAT_COMMAND_R7B,
COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
};
struct common_chat_params {
common_chat_format format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
json prompt;
std::string grammar;
bool grammar_lazy = false;
std::vector<common_grammar_trigger> grammar_triggers;
std::vector<std::string> preserved_tokens;
std::vector<std::string> additional_stops;
};
struct common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & params);
std::string common_chat_format_name(common_chat_format format);
common_chat_msg common_chat_parse( const std::string & input, common_chat_format format);

170
common/common.cpp
View File

@ -12,8 +12,6 @@
#include "json.hpp"
#include "json-schema-to-grammar.h"
#include "llama.h"
#include "chat.hpp"
#include "chat-template.hpp"
#include <algorithm>
#include <cinttypes>
@ -1768,174 +1766,6 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
return text;
}
//
// Chat template utils
//
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
if (use_jinja) {
try {
auto chat_template = common_chat_template(tmpl, "<s>", "</s>");
common_chat_inputs inputs;
inputs.messages = json::array({{
{"role", "user"},
{"content", "test"},
}});
common_chat_params_init(chat_template, inputs);
return true;
} catch (const std::exception & e) {
LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what());
return false;
}
}
llama_chat_message chat[] = {{"user", "test"}};
const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0);
return res >= 0;
}
std::string common_chat_apply_template(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & msgs,
bool add_ass,
bool use_jinja) {
if (use_jinja) {
auto messages = json::array();
for (const auto & msg : msgs) {
messages.push_back({{"role", msg.role}, {"content", msg.content}});
}
common_chat_inputs inputs;
inputs.messages = messages;
inputs.add_generation_prompt = add_ass;
return common_chat_params_init(tmpl, inputs).prompt;
}
int alloc_size = 0;
std::vector<llama_chat_message> chat;
for (const auto & msg : msgs) {
chat.push_back({msg.role.c_str(), msg.content.c_str()});
alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
}
std::vector<char> buf(alloc_size);
// run the first time to get the total output length
int32_t res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size());
// error: chat template is not supported
if (res < 0) {
// if the custom "tmpl" is not supported, we throw an error
// this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
throw std::runtime_error("this custom template is not supported");
}
// if it turns out that our buffer is too small, we resize it
if ((size_t) res > buf.size()) {
buf.resize(res);
res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size());
}
std::string formatted_chat(buf.data(), res);
return formatted_chat;
}
std::string common_chat_format_single(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass,
bool use_jinja) {
std::ostringstream ss;
auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(tmpl, past_msg, false, use_jinja);
std::vector<common_chat_msg> chat_new(past_msg);
// if the past_msg ends with a newline, we must preserve it in the formatted version
if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
ss << "\n";
};
// format chat with new_msg
chat_new.push_back(new_msg);
auto fmt_new_msg = common_chat_apply_template(tmpl, chat_new, add_ass, use_jinja);
// get the diff part
ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
return ss.str();
}
std::string common_chat_format_example(const common_chat_template & tmpl, bool use_jinja) {
std::vector<common_chat_msg> msgs = {
{"system", "You are a helpful assistant", {}},
{"user", "Hello", {}},
{"assistant", "Hi there", {}},
{"user", "How are you?", {}},
};
return common_chat_apply_template(tmpl, msgs, true, use_jinja);
}
#define CHATML_TEMPLATE_SRC \
"{%- for message in messages -%}\n" \
" {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' -}}\n" \
"{%- endfor -%}\n" \
"{%- if add_generation_prompt -%}\n" \
" {{- '<|im_start|>assistant\n' -}}\n" \
"{%- endif -%}"
common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override)
{
std::string default_template_src;
std::string template_tool_use_src;
bool has_explicit_template = !chat_template_override.empty();
if (chat_template_override.empty()) {
auto str = llama_model_chat_template(model, /* name */ nullptr);
if (str) {
default_template_src = str;
has_explicit_template = true;
}
str = llama_model_chat_template(model, /* name */ "tool_use");
if (str) {
template_tool_use_src = str;
has_explicit_template = true;
}
} else {
default_template_src = chat_template_override;
}
if (default_template_src.empty() || default_template_src == "chatml") {
if (!template_tool_use_src.empty()) {
default_template_src = template_tool_use_src;
} else {
default_template_src = CHATML_TEMPLATE_SRC;
}
}
auto vocab = llama_model_get_vocab(model);
const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {
if (token == LLAMA_TOKEN_NULL) {
if (default_template_src.find(jinja_variable_name) != std::string::npos
|| template_tool_use_src.find(jinja_variable_name) != std::string::npos) {
LOG_WRN("%s: warning: vocab does not have a %s token, jinja template won't work as intended.\n", __func__, name);
}
return std::string();
} else {
return common_token_to_piece(vocab, token, true);
}
};
auto token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");
auto token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");
try {
return {
has_explicit_template,
std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos),
template_tool_use_src.empty()
? nullptr
: std::make_unique<minja::chat_template>(template_tool_use_src, token_bos, token_eos),
};
} catch (const std::exception & e) {
LOG_ERR("%s: failed to parse chat template: %s\n", __func__, e.what());
return {
has_explicit_template,
std::make_unique<minja::chat_template>(CHATML_TEMPLATE_SRC, token_bos, token_eos),
nullptr,
};
}
}
//
// KV cache utils
//

68
common/common.h
View File

@ -140,6 +140,7 @@ struct common_params_sampling {
int32_t dry_allowed_length = 2; // tokens extending repetitions beyond this receive penalty
int32_t dry_penalty_last_n = -1; // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float top_n_sigma = -1.00f;// -1.0 = disabled
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool ignore_eos = false;
@ -177,10 +178,10 @@ struct common_params_speculative {
int32_t n_ctx = 0; // draft context size
int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
int32_t n_min = 5; // minimum number of draft tokens to use for speculative decoding
int32_t n_min = 0; // minimum number of draft tokens to use for speculative decoding
int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
float p_split = 0.1f; // speculative decoding split probability
float p_min = 0.9f; // minimum speculative decoding probability (greedy)
float p_min = 0.75f; // minimum speculative decoding probability (greedy)
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
@ -202,6 +203,11 @@ struct common_params_vocoder {
bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy // NOLINT
};
enum common_reasoning_format {
COMMON_REASONING_FORMAT_NONE,
COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
};
struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 4096; // context size
@ -292,6 +298,7 @@ struct common_params {
bool kl_divergence = false; // compute KL divergence
bool usage = false; // print usage
bool completion = false; // print source-able completion script
bool use_color = false; // use color to distinguish generations and inputs
bool special = false; // enable special token output
bool interactive = false; // interactive mode
@ -346,6 +353,7 @@ struct common_params {
std::string chat_template = ""; // NOLINT
bool use_jinja = false; // NOLINT
bool enable_chat_template = true;
common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
std::vector<std::string> api_keys;
@ -608,62 +616,6 @@ std::string common_detokenize(
const std::vector<llama_token> & tokens,
bool special = true);
//
// Chat template utils
//
struct common_tool_call {
std::string name;
std::string arguments;
std::string id;
};
// same with llama_chat_message, but uses std::string
struct common_chat_msg {
std::string role;
std::string content;
std::vector<common_tool_call> tool_calls;
std::string tool_plan = "";
};
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
namespace minja {
class chat_template;
}
typedef minja::chat_template common_chat_template;
struct common_chat_templates {
bool has_explicit_template; // Model had builtin template or template overridde was specified.
std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
std::unique_ptr<common_chat_template> template_tool_use;
};
// 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 common_chat_apply_template(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & chat,
bool add_ass,
bool use_jinja);
// Format single message, while taking into account the position of that message in chat history
std::string common_chat_format_single(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass,
bool use_jinja);
// Returns an example of formatted chat
std::string common_chat_format_example(
const common_chat_template & tmpl, bool use_jinja);
common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override);
//
// KV cache utils
//

0
common/chat-template.hpp → common/minja/chat-template.hpp
View File

0
common/minja.hpp → common/minja/minja.hpp
View File

98
common/sampling.cpp
View File

@ -134,11 +134,11 @@ std::string common_params_sampling::print() const {
snprintf(result, sizeof(result),
"\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
"\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
"\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
"\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, top_n_sigma = %.3f, temp = %.3f\n"
"\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, top_n_sigma, temp,
mirostat, mirostat_eta, mirostat_tau);
return std::string(result);
@ -151,12 +151,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
lparams.no_perf = params.no_perf;
std::vector<const char *> trigger_words;
trigger_words.reserve(params.grammar_trigger_words.size());
for (const auto & str : params.grammar_trigger_words) {
trigger_words.push_back(str.word.c_str());
}
struct llama_sampler * grmr;
if (params.grammar.compare(0, 11, "%llguidance") == 0) {
#ifdef LLAMA_USE_LLGUIDANCE
@ -165,6 +159,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
#endif // LLAMA_USE_LLGUIDANCE
} else {
std::vector<const char *> trigger_words;
trigger_words.reserve(params.grammar_trigger_words.size());
for (const auto & str : params.grammar_trigger_words) {
trigger_words.push_back(str.word.c_str());
}
grmr = params.grammar_lazy
? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
trigger_words.data(), trigger_words.size(),
@ -188,45 +188,51 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
params.logit_bias.data()));
if (params.mirostat == 0) {
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY:
{
std::vector<const char *> c_breakers;
c_breakers.reserve(params.dry_sequence_breakers.size());
for (const auto & str : params.dry_sequence_breakers) {
c_breakers.push_back(str.c_str());
}
if (params.top_n_sigma >= 0) {
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
llama_sampler_chain_add(result->chain, llama_sampler_init_temp (params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
} else {
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY:
{
std::vector<const char *> c_breakers;
c_breakers.reserve(params.dry_sequence_breakers.size());
for (const auto & str : params.dry_sequence_breakers) {
c_breakers.push_back(str.c_str());
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
break;
case COMMON_SAMPLER_TYPE_TOP_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_MIN_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_XTC:
llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
break;
case COMMON_SAMPLER_TYPE_TYPICAL_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TEMPERATURE:
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab));
break;
case COMMON_SAMPLER_TYPE_PENALTIES:
llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
break;
case COMMON_SAMPLER_TYPE_TOP_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_MIN_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_XTC:
llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
break;
case COMMON_SAMPLER_TYPE_TYPICAL_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TEMPERATURE:
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab));
break;
case COMMON_SAMPLER_TYPE_PENALTIES:
llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
}
}
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));

10
common/speculative.cpp
View File

@ -252,11 +252,6 @@ llama_tokens common_speculative_gen_draft(
// add drafted token for each sequence
const llama_token id = cur_p->data[0].id;
// only collect very high-confidence draft tokens
if (cur_p->data[0].p < params.p_min) {
break;
}
common_sampler_accept(smpl, id, true);
result.push_back(id);
@ -265,6 +260,11 @@ llama_tokens common_speculative_gen_draft(
break;
}
// only collect very high-confidence draft tokens
if (cur_p->data[0].p < params.p_min) {
break;
}
common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
// evaluate the drafted tokens on the draft model

2
common/speculative.h
View File

@ -9,7 +9,7 @@ struct common_speculative_params {
int n_draft = 16; // max drafted tokens
int n_reuse = 256;
float p_min = 0.9f; // min probability required to accept a token in the draft
float p_min = 0.75f; // min probability required to accept a token in the draft
};
struct common_speculative * common_speculative_init(struct llama_context * ctx_dft);

6
convert_hf_to_gguf.py
View File

@ -558,7 +558,7 @@ class Model:
# NOTE: this function is generated by convert_hf_to_gguf_update.py
# do not modify it manually!
# ref: https://github.com/ggerganov/llama.cpp/pull/6920
# ref: https://github.com/ggml-org/llama.cpp/pull/6920
# Marker: Start get_vocab_base_pre
def get_vocab_base_pre(self, tokenizer) -> str:
# encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
@ -708,7 +708,7 @@ class Model:
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("** ref: https://github.com/ggerganov/llama.cpp/pull/6920")
logger.warning("** ref: https://github.com/ggml-org/llama.cpp/pull/6920")
logger.warning("**")
logger.warning(f"** chkhsh: {chkhsh}")
logger.warning("**************************************************************************************")
@ -2835,7 +2835,7 @@ class InternLM2Model(Model):
if chat_eos_token_id is not None:
# For the chat model, we replace the eos with '<|im_end|>'.
# TODO: this is a hack, should be fixed
# https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
# https://github.com/ggml-org/llama.cpp/pull/6745#issuecomment-2067687048
special_vocab.special_token_ids["eos"] = chat_eos_token_id
logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}"
" in chat mode so that the conversation can end normally.")

4
convert_hf_to_gguf_update.py
View File

@ -8,7 +8,7 @@
# provide the necessary information to llama.cpp via the GGUF header in order to implement
# the same pre-tokenizer.
#
# ref: https://github.com/ggerganov/llama.cpp/pull/6920
# ref: https://github.com/ggml-org/llama.cpp/pull/6920
#
# Instructions:
#
@ -246,7 +246,7 @@ src_func = f"""
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("** ref: https://github.com/ggerganov/llama.cpp/pull/6920")
logger.warning("** ref: https://github.com/ggml-org/llama.cpp/pull/6920")
logger.warning("**")
logger.warning(f"** chkhsh: {{chkhsh}}")
logger.warning("**************************************************************************************")

4
convert_lora_to_gguf.py
View File

@ -395,7 +395,7 @@ if __name__ == '__main__':
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
logger.error("Please refer to https://github.com/ggml-org/llama.cpp/pull/9948")
sys.exit(1)
if base_name in tensor_map:
@ -419,7 +419,7 @@ if __name__ == '__main__':
# some archs may have the same tensor for lm_head and output (tie word embeddings)
# in this case, adapters targeting lm_head will fail when using llama-export-lora
# therefore, we ignore them for now
# see: https://github.com/ggerganov/llama.cpp/issues/9065
# see: https://github.com/ggml-org/llama.cpp/issues/9065
if name == "lm_head.weight" and len(dest) == 0:
raise ValueError("lm_head is present in adapter, but is ignored in base model")
for dest_name, dest_data in dest:

2
docs/android.md
View File

@ -12,7 +12,7 @@ $ apt update && apt upgrade -y
$ apt install git cmake
```
Then, follow the [build instructions](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md), specifically for CMake.
Then, follow the [build instructions](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md), specifically for CMake.
Once the binaries are built, download your model of choice (e.g., from Hugging Face). It's recommended to place it in the `~/` directory for best performance:

4
docs/backend/OPENCL.md
View File

@ -122,7 +122,7 @@ cp libOpenCL.so ~/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x
```sh
cd ~/dev/llm
git clone https://github.com/ggerganov/llama.cpp && \
git clone https://github.com/ggml-org/llama.cpp && \
cd llama.cpp && \
mkdir build-android && cd build-android
@ -182,7 +182,7 @@ cmake --build . --target install
mkdir -p ~/dev/llm
cd ~/dev/llm
git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
mkdir build && cd build
cmake .. -G Ninja `

6
docs/backend/SYCL.md
View File

@ -36,8 +36,8 @@ The following release is verified with good quality:
|Commit ID|Tag|Release|Verified Platform| Update date|
|-|-|-|-|-|
|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
|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||
|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/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
@ -58,7 +58,7 @@ The following release is verified with good quality:
- 2024.3
- Release binary files of Windows.
- A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
- New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
- New base line is ready: [tag b2437](https://github.com/ggml-org/llama.cpp/tree/b2437).
- Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
- Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
- Support detecting all GPUs with level-zero and same top **Max compute units**.

12
docs/build.md
View File

@ -3,7 +3,7 @@
**To get the Code:**
```bash
git clone https://github.com/ggerganov/llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
```
@ -46,7 +46,7 @@ cmake --build build --config Release
```
- 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,...):
- Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/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
@ -206,6 +206,14 @@ This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GP
cmake --build build --config Release
```
For static build:
```bash
cmake -B build -DGGML_MUSA=ON \
-DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON
cmake --build build --config Release
```
The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) 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.

2
docs/cuda-fedora.md
View File

@ -248,7 +248,7 @@ You have successfully set up CUDA on Fedora within a toolbox environment using t
- **Building `llama.cpp`:**
- With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support.
- With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support.
- Ensure that any CUDA-specific build flags or paths are correctly set in your build configuration.
- **Using the Toolbox Environment:**

20
docs/development/HOWTO-add-model.md
View File

@ -104,16 +104,16 @@ Note: to debug the inference graph: you can use [llama-eval-callback](/examples/
## GGUF specification
https://github.com/ggerganov/ggml/blob/master/docs/gguf.md
https://github.com/ggml-org/ggml/blob/master/docs/gguf.md
## Resources
- YaRN RoPE scaling https://github.com/ggerganov/llama.cpp/pull/2268
- support Baichuan serial models https://github.com/ggerganov/llama.cpp/pull/3009
- support attention bias https://github.com/ggerganov/llama.cpp/pull/4283
- Mixtral support https://github.com/ggerganov/llama.cpp/pull/4406
- BERT embeddings https://github.com/ggerganov/llama.cpp/pull/5423
- Grok-1 support https://github.com/ggerganov/llama.cpp/pull/6204
- Command R Plus support https://github.com/ggerganov/llama.cpp/pull/6491
- support arch DBRX https://github.com/ggerganov/llama.cpp/pull/6515
- How to convert HuggingFace model to GGUF format https://github.com/ggerganov/llama.cpp/discussions/2948
- YaRN RoPE scaling https://github.com/ggml-org/llama.cpp/pull/2268
- support Baichuan serial models https://github.com/ggml-org/llama.cpp/pull/3009
- support attention bias https://github.com/ggml-org/llama.cpp/pull/4283
- Mixtral support https://github.com/ggml-org/llama.cpp/pull/4406
- BERT embeddings https://github.com/ggml-org/llama.cpp/pull/5423
- Grok-1 support https://github.com/ggml-org/llama.cpp/pull/6204
- Command R Plus support https://github.com/ggml-org/llama.cpp/pull/6491
- support arch DBRX https://github.com/ggml-org/llama.cpp/pull/6515
- How to convert HuggingFace model to GGUF format https://github.com/ggml-org/llama.cpp/discussions/2948

36
docs/docker.md
View File

@ -7,21 +7,21 @@
## 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`)
1. `ghcr.io/ggml-org/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/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
3. `ghcr.io/ggml-org/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`)
- `ghcr.io/ggerganov/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
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, ROCm or MUSA library, you'll need to build the images locally for now).
@ -32,25 +32,25 @@ The easiest way to download the models, convert them to ggml and optimize them i
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
docker run -v /path/to/models:/models ghcr.io/ggml-org/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
docker run -v /path/to/models:/models ghcr.io/ggml-org/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
docker run -v /path/to/models:/models ghcr.io/ggml-org/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 run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggml-org/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
```
## Docker With CUDA
@ -69,7 +69,7 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
The defaults are:
- `CUDA_VERSION` set to `12.6.0`
- `CUDA_VERSION` set to `12.4.0`
- `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
The resulting images, are essentially the same as the non-CUDA images:
@ -104,7 +104,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
The defaults are:
- `MUSA_VERSION` set to `rc3.1.0`
- `MUSA_VERSION` set to `rc3.1.1`
The resulting images, are essentially the same as the non-MUSA images:

2
docs/install.md
View File

@ -7,7 +7,7 @@ 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
The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
## Nix

4
docs/llguidance.md
View File

@ -13,13 +13,15 @@ cmake -B build -DLLAMA_LLGUIDANCE=ON
make -C build -j
```
For Windows use `cmake --build build --config Release` instead of `make`.
This requires the Rust compiler and the `cargo` tool to be [installed](https://www.rust-lang.org/tools/install).
## Interface
There are no new command-line arguments or modifications to `common_params`. When enabled, grammars starting with `%llguidance` are passed to LLGuidance instead of the [current](../grammars/README.md) llama.cpp grammars. Additionally, JSON Schema requests (e.g., using the `-j` argument in `llama-cli`) are also passed to LLGuidance.
For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/scripts/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format.
For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/python/llguidance/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format.
## Performance

6
examples/cvector-generator/README.md
View File

@ -3,9 +3,9 @@
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)
- [Add support for control vectors](https://github.com/ggml-org/llama.cpp/pull/5970)
- (Issue) [Generate control vector using llama.cpp](https://github.com/ggml-org/llama.cpp/issues/6880)
- [Add cvector-generator example](https://github.com/ggml-org/llama.cpp/pull/7514)
## Examples

4
examples/imatrix/README.md
View File

@ -1,7 +1,7 @@
# llama.cpp/examples/imatrix
Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models.
More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861
Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models.
More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861
## Usage

2
examples/imatrix/imatrix.cpp
View File

@ -100,7 +100,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
// this has been adapted to the new format of storing merged experts in a single 3d tensor
// ref: https://github.com/ggerganov/llama.cpp/pull/6387
// ref: https://github.com/ggml-org/llama.cpp/pull/6387
if (t->op == GGML_OP_MUL_MAT_ID) {
// ids -> [n_experts_used, n_tokens]
// src1 -> [cols, n_expert_used, n_tokens]

11
examples/llama-bench/llama-bench.cpp
View File

@ -876,8 +876,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
struct test {
static const std::string build_commit;
static const int build_number;
static const std::string cpu_info;
static const std::string gpu_info;
const std::string cpu_info;
const std::string gpu_info;
std::string model_filename;
std::string model_type;
uint64_t model_size;
@ -903,7 +903,10 @@ struct test {
std::string test_time;
std::vector<uint64_t> samples_ns;
test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) :
cpu_info(get_cpu_info()),
gpu_info(get_gpu_info()) {
model_filename = inst.model;
char buf[128];
llama_model_desc(lmodel, buf, sizeof(buf));
@ -1058,8 +1061,6 @@ struct test {
const std::string test::build_commit = LLAMA_COMMIT;
const int test::build_number = LLAMA_BUILD_NUMBER;
const std::string test::cpu_info = get_cpu_info();
const std::string test::gpu_info = get_gpu_info();
struct printer {
virtual ~printer() {}

2
examples/llama.android/llama/src/main/cpp/CMakeLists.txt
View File

@ -14,7 +14,7 @@ project("llama-android")
#include(FetchContent)
#FetchContent_Declare(
# llama
# GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
# GIT_REPOSITORY https://github.com/ggml-org/llama.cpp
# GIT_TAG master
#)

4
examples/llama.swiftui/README.md
View File

@ -3,9 +3,9 @@
Local inference of llama.cpp on an iPhone. This is a sample app that can be used as a starting
point for more advanced projects.
For usage instructions and performance stats, check the following discussion: https://github.com/ggerganov/llama.cpp/discussions/4508
For usage instructions and performance stats, check the following discussion: https://github.com/ggml-org/llama.cpp/discussions/4508
![image](https://github.com/ggerganov/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
![image](https://github.com/ggml-org/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
Video demonstration:

25
examples/llama.swiftui/llama.swiftui/UI/ContentView.swift
View File

@ -124,15 +124,26 @@ struct ContentView: View {
}
}
}.sheet(isPresented: $showingHelp) { // Sheet for help modal
VStack(alignment: .leading) {
NavigationView {
VStack(alignment: .leading) {
Text("1. Make sure the model is in GGUF Format")
.padding()
Text("2. Copy the download link of the quantized model")
.padding()
VStack(alignment: .leading) {
Text("1. Make sure the model is in GGUF Format")
.padding()
Text("2. Copy the download link of the quantized model")
.padding()
}
Spacer()
}
Spacer()
}
.navigationTitle("Help")
.navigationBarTitleDisplayMode(.inline)
.toolbar {
ToolbarItem(placement: .navigationBarTrailing) {
Button("Done") {
showingHelp = false
}
}
}
}
}
}
}

2
examples/llama.vim
View File

@ -39,7 +39,7 @@
"
" :call llama#init()
"
" more info: https://github.com/ggerganov/llama.cpp/pull/9787
" more info: https://github.com/ggml-org/llama.cpp/pull/9787
"
" colors (adjust to your liking)

2
examples/llava/README-minicpmo2.6.md
View File

@ -26,7 +26,7 @@ python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
```
Build llama.cpp using `CMake`:
https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md
https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md
```bash
cmake -B build

2
examples/llava/README-minicpmv2.5.md
View File

@ -6,7 +6,7 @@ Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-
Clone llama.cpp:
```bash
git clone https://github.com/ggerganov/llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
```

14
examples/llava/clip.cpp
View File

@ -1729,11 +1729,11 @@ void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) {
}
}
static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img) {
img->nx = nx;
img->ny = ny;
img->buf.resize(3 * nx * ny);
memcpy(img->buf.data(), data, img->buf.size());
memcpy(img->buf.data(), rgb_pixels, img->buf.size());
}
bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
@ -1743,7 +1743,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
LOG_ERR("%s: failed to load image '%s'\n", __func__, fname);
return false;
}
build_clip_img_from_data(data, nx, ny, img);
clip_build_img_from_pixels(data, nx, ny, img);
stbi_image_free(data);
return true;
}
@ -1755,7 +1755,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
LOG_ERR("%s: failed to decode image bytes\n", __func__);
return false;
}
build_clip_img_from_data(data, nx, ny, img);
clip_build_img_from_pixels(data, nx, ny, img);
stbi_image_free(data);
return true;
}
@ -2712,9 +2712,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
if (!ctx->has_glm_projector) {
struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
// The patches vector is used to get rows to index into the embeds with;
// we should skip dim 0 only if we have CLS to avoid going out of bounds
// when retrieving the rows.
int patch_offset = ctx->has_class_embedding ? 1 : 0;
int* patches_data = (int*)malloc(ggml_nbytes(patches));
for (int i = 0; i < num_patches; i++) {
patches_data[i] = i + 1;
patches_data[i] = i + patch_offset;
}
ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
free(patches_data);

3
examples/llava/clip.h
View File

@ -73,6 +73,9 @@ CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch * batch);
CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
/** build image from pixels decoded by other libraries instead of stb_image.h for better performance. The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes */
CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img);
CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
/** interpret bytes as an image file with length bytes_length, and use the result to populate img */

2
examples/lookahead/README.md
View File

@ -4,4 +4,4 @@ Demonstration of lookahead decoding technique:
https://lmsys.org/blog/2023-11-21-lookahead-decoding/
More info: https://github.com/ggerganov/llama.cpp/pull/4207
More info: https://github.com/ggml-org/llama.cpp/pull/4207

4
examples/lookup/README.md
View File

@ -8,5 +8,5 @@ The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft
More info:
https://github.com/ggerganov/llama.cpp/pull/4484
https://github.com/ggerganov/llama.cpp/issues/4226
https://github.com/ggml-org/llama.cpp/pull/4484
https://github.com/ggml-org/llama.cpp/issues/4226

12
examples/main/README.md
View File

@ -1,6 +1,6 @@
# llama.cpp/examples/main
This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggml-org/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
## Table of Contents
@ -121,7 +121,7 @@ When --in-prefix or --in-suffix options are enabled the chat template ( --chat-t
### Chat templates
`--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name. Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
`--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name. Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
Example usage: `--chat-template gemma`
@ -265,6 +265,14 @@ Being experimental and unique, XTC is disabled by default. The recommended combi
Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
### Top-nσ Sampling
- `--top-nsigma N`: Limit the next token selection to a subset of tokens with pre-softmax logits that are within n * σ less than the max logit (default: -1, -1 = disabled).
Top-nσ sampling is a text generation method that selects tokens based on a statistical threshold in pre-softmax logits. It works by only sampling from tokens with logits that are within n * σ of the maximum logit. This method helps maintain a stable sampling space regardless of temperature scaling, allowing it to perform well on reasoning tasks even in high temperatures. Without complex probability manipulation, it efficiently filters tokens directly on the pre-softmax logits. A higher value for top-nsigma (e.g., 5) will take more noisy tokens into consideration, while a lower value (e.g., 1) will focous on the more informative region of the sampling space.
Example usage: `--top-nsigma 1`
### Logit Bias
- `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.

27
examples/main/main.cpp
View File

@ -4,7 +4,7 @@
#include "log.h"
#include "sampling.h"
#include "llama.h"
#include "chat-template.hpp"
#include "chat.h"
#include <cstdio>
#include <cstring>
@ -158,7 +158,7 @@ int main(int argc, char ** argv) {
}
const llama_vocab * vocab = llama_model_get_vocab(model);
auto chat_templates = common_chat_templates_from_model(model, params.chat_template);
auto chat_templates = common_chat_templates_init(model, params.chat_template);
LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
@ -201,7 +201,7 @@ int main(int argc, char ** argv) {
}
// auto enable conversation mode if chat template is available
const bool has_chat_template = chat_templates.has_explicit_template && chat_templates.template_default;
const bool has_chat_template = common_chat_templates_was_explicit(chat_templates.get());
if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {
if (has_chat_template) {
LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);
@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
// print chat template example in conversation mode
if (params.conversation_mode) {
if (params.enable_chat_template) {
LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(*chat_templates.template_default, params.use_jinja).c_str());
LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja).c_str());
} else {
LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
}
@ -264,9 +264,11 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd_inp;
auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {
common_chat_msg new_msg{role, content, {}};
auto formatted = common_chat_format_single(*chat_templates.template_default, chat_msgs, new_msg, role == "user", g_params->use_jinja);
chat_msgs.push_back({role, content, {}});
common_chat_msg new_msg;
new_msg.role = role;
new_msg.content = content;
auto formatted = common_chat_format_single(chat_templates.get(), chat_msgs, new_msg, role == "user", g_params->use_jinja);
chat_msgs.push_back(new_msg);
LOG_DBG("formatted: '%s'\n", formatted.c_str());
return formatted;
};
@ -755,11 +757,14 @@ int main(int argc, char ** argv) {
// check for reverse prompt using special tokens
llama_token last_token = common_sampler_last(smpl);
if (std::find(antiprompt_token.begin(), antiprompt_token.end(), last_token) != antiprompt_token.end()) {
if (params.interactive) {
is_interacting = true;
for (auto token : antiprompt_token) {
if (token == last_token) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
break;
}
is_antiprompt = true;
}
if (is_antiprompt) {

4
examples/passkey/README.md
View File

@ -5,8 +5,8 @@ models ability to recall information from long contexts.
See the following PRs for more info:
- https://github.com/ggerganov/llama.cpp/pull/3856
- https://github.com/ggerganov/llama.cpp/pull/4810
- https://github.com/ggml-org/llama.cpp/pull/3856
- https://github.com/ggml-org/llama.cpp/pull/4810
### Usage

2
examples/pydantic_models_to_grammar_examples.py
View File

@ -23,7 +23,7 @@ def create_completion(host, prompt, gbnf_grammar):
"""Calls the /completion API on llama-server.
See
https://github.com/ggerganov/llama.cpp/tree/HEAD/examples/server#api-endpoints
https://github.com/ggml-org/llama.cpp/tree/HEAD/examples/server#api-endpoints
"""
print(f" Request:\n Grammar:\n{textwrap.indent(gbnf_grammar, ' ')}\n Prompt:\n{textwrap.indent(prompt.rstrip(), ' ')}")
headers = {"Content-Type": "application/json"}

30
examples/quantize/README.md
View File

@ -69,22 +69,22 @@ Several quantization methods are supported. They differ in the resulting model d
| 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)
- [k-quants](https://github.com/ggml-org/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-quants 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)
- [#2707](https://github.com/ggml-org/llama.cpp/pull/2707)
- [#2807](https://github.com/ggml-org/llama.cpp/pull/2807)
- [#4773 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4773)
- [#4856 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4856)
- [#4861 - importance matrix](https://github.com/ggml-org/llama.cpp/pull/4861)
- [#4872 - MoE models](https://github.com/ggml-org/llama.cpp/pull/4872)
- [#4897 - 2-bit quantization](https://github.com/ggml-org/llama.cpp/pull/4897)
- [#4930 - imatrix for all k-quants](https://github.com/ggml-org/llama.cpp/pull/4930)
- [#4951 - imatrix on the GPU](https://github.com/ggml-org/llama.cpp/pull/4957)
- [#4969 - imatrix for legacy quants](https://github.com/ggml-org/llama.cpp/pull/4969)
- [#4996 - k-quants tuning](https://github.com/ggml-org/llama.cpp/pull/4996)
- [#5060 - Q3_K_XS](https://github.com/ggml-org/llama.cpp/pull/5060)
- [#5196 - 3-bit i-quants](https://github.com/ggml-org/llama.cpp/pull/5196)
- [quantization tuning](https://github.com/ggml-org/llama.cpp/pull/5320), [another one](https://github.com/ggml-org/llama.cpp/pull/5334), and [another one](https://github.com/ggml-org/llama.cpp/pull/5361)
**Llama 2 7B**

2
examples/retrieval/README.md
View File

@ -3,7 +3,7 @@
Demonstration of simple retrieval technique based on cosine similarity
More info:
https://github.com/ggerganov/llama.cpp/pull/6193
https://github.com/ggml-org/llama.cpp/pull/6193
### How to use

173
examples/run/run.cpp
View File

@ -24,7 +24,7 @@
#include <string>
#include <vector>
#include "chat-template.hpp"
#include "chat.h"
#include "common.h"
#include "json.hpp"
#include "linenoise.cpp/linenoise.h"
@ -113,6 +113,7 @@ class Opt {
llama_context_params ctx_params;
llama_model_params model_params;
std::string model_;
std::string chat_template_file;
std::string user;
bool use_jinja = false;
int context_size = -1, ngl = -1;
@ -148,6 +149,16 @@ class Opt {
return 0;
}
int handle_option_with_value(int argc, const char ** argv, int & i, std::string & option_value) {
if (i + 1 >= argc) {
return 1;
}
option_value = argv[++i];
return 0;
}
int parse(int argc, const char ** argv) {
bool options_parsing = true;
for (int i = 1, positional_args_i = 0; i < argc; ++i) {
@ -169,6 +180,11 @@ class Opt {
verbose = true;
} else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
use_jinja = true;
} else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0){
if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
return 1;
}
use_jinja = true;
} else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
help = true;
return 0;
@ -207,6 +223,11 @@ class Opt {
"Options:\n"
" -c, --context-size <value>\n"
" Context size (default: %d)\n"
" --chat-template-file <path>\n"
" Path to the file containing the chat template to use with the model.\n"
" Only supports jinja templates and implicitly sets the --jinja flag.\n"
" --jinja\n"
" Use jinja templating for the chat template of the model\n"
" -n, -ngl, --ngl <value>\n"
" Number of GPU layers (default: %d)\n"
" --temp <value>\n"
@ -261,13 +282,12 @@ static int get_terminal_width() {
#endif
}
#ifdef LLAMA_USE_CURL
class File {
public:
FILE * file = nullptr;
FILE * open(const std::string & filename, const char * mode) {
file = fopen(filename.c_str(), mode);
file = ggml_fopen(filename.c_str(), mode);
return file;
}
@ -303,6 +323,20 @@ class File {
return 0;
}
std::string to_string() {
fseek(file, 0, SEEK_END);
const size_t size = ftell(file);
fseek(file, 0, SEEK_SET);
std::string out;
out.resize(size);
const size_t read_size = fread(&out[0], 1, size, file);
if (read_size != size) {
printe("Error reading file: %s", strerror(errno));
}
return out;
}
~File() {
if (fd >= 0) {
# ifdef _WIN32
@ -327,6 +361,7 @@ class File {
# endif
};
#ifdef LLAMA_USE_CURL
class HttpClient {
public:
int init(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
@ -557,7 +592,7 @@ class LlamaData {
llama_model_ptr model;
llama_sampler_ptr sampler;
llama_context_ptr context;
std::vector<llama_chat_message> messages;
std::vector<llama_chat_message> messages; // TODO: switch to common_chat_msg
std::list<std::string> msg_strs;
std::vector<char> fmtted;
@ -834,44 +869,23 @@ static void add_message(const char * role, const std::string & text, LlamaData &
}
// Function to apply the chat template and resize `formatted` if needed
static int apply_chat_template(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, bool use_jinja) {
if (use_jinja) {
json messages = json::array();
for (const auto & msg : llama_data.messages) {
messages.push_back({
{"role", msg.role},
{"content", msg.content},
});
}
try {
minja::chat_template_inputs tmpl_inputs;
tmpl_inputs.messages = messages;
tmpl_inputs.add_generation_prompt = append;
minja::chat_template_options tmpl_opts;
tmpl_opts.use_bos_token = false;
tmpl_opts.use_eos_token = false;
auto result = tmpl.apply(tmpl_inputs, tmpl_opts);
llama_data.fmtted.resize(result.size() + 1);
memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1);
return result.size();
} catch (const std::exception & e) {
printe("failed to render the chat template: %s\n", e.what());
return -1;
}
}
int result = llama_chat_apply_template(
tmpl.source().c_str(), llama_data.messages.data(), llama_data.messages.size(), append,
append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0);
if (append && result > static_cast<int>(llama_data.fmtted.size())) {
llama_data.fmtted.resize(result);
result = llama_chat_apply_template(tmpl.source().c_str(), llama_data.messages.data(),
llama_data.messages.size(), append, llama_data.fmtted.data(),
llama_data.fmtted.size());
static int apply_chat_template(const struct common_chat_templates * tmpls, LlamaData & llama_data, const bool append, bool use_jinja) {
common_chat_templates_inputs inputs;
for (const auto & msg : llama_data.messages) {
common_chat_msg cmsg;
cmsg.role = msg.role;
cmsg.content = msg.content;
inputs.messages.push_back(cmsg);
}
inputs.add_generation_prompt = append;
inputs.use_jinja = use_jinja;
return result;
auto chat_params = common_chat_templates_apply(tmpls, inputs);
// TODO: use other params for tool calls.
auto result = chat_params.prompt;
llama_data.fmtted.resize(result.size() + 1);
memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1);
return result.size();
}
// Function to tokenize the prompt
@ -963,7 +977,8 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
}
static int read_user_input(std::string & user_input) {
static const char * prompt_prefix = "> ";
static const char * prompt_prefix_env = std::getenv("LLAMA_PROMPT_PREFIX");
static const char * prompt_prefix = prompt_prefix_env ? prompt_prefix_env : "> ";
#ifdef WIN32
printf("\r" LOG_CLR_TO_EOL LOG_COL_DEFAULT "%s", prompt_prefix);
@ -1015,8 +1030,8 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt,
}
// Helper function to apply the chat template and handle errors
static int apply_chat_template_with_error_handling(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) {
const int new_len = apply_chat_template(tmpl, llama_data, append, use_jinja);
static int apply_chat_template_with_error_handling(const common_chat_templates * tmpls, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) {
const int new_len = apply_chat_template(tmpls, llama_data, append, use_jinja);
if (new_len < 0) {
printe("failed to apply the chat template\n");
return -1;
@ -1074,40 +1089,68 @@ static int get_user_input(std::string & user_input, const std::string & user) {
return 0;
}
// Reads a chat template file to be used
static std::string read_chat_template_file(const std::string & chat_template_file) {
File file;
if (!file.open(chat_template_file, "r")) {
printe("Error opening chat template file '%s': %s", chat_template_file.c_str(), strerror(errno));
return "";
}
return file.to_string();
}
static int process_user_message(const Opt & opt, const std::string & user_input, LlamaData & llama_data,
const common_chat_templates_ptr & chat_templates, int & prev_len,
const bool stdout_a_terminal) {
add_message("user", opt.user.empty() ? user_input : opt.user, llama_data);
int new_len;
if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, true, new_len, opt.use_jinja) < 0) {
return 1;
}
std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len);
std::string response;
if (generate_response(llama_data, prompt, response, stdout_a_terminal)) {
return 1;
}
if (!opt.user.empty()) {
return 2;
}
add_message("assistant", response, llama_data);
if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, false, prev_len, opt.use_jinja) < 0) {
return 1;
}
return 0;
}
// Main chat loop function
static int chat_loop(LlamaData & llama_data, const std::string & user, bool use_jinja) {
static int chat_loop(LlamaData & llama_data, const Opt & opt) {
int prev_len = 0;
llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
auto chat_templates = common_chat_templates_from_model(llama_data.model.get(), "");
GGML_ASSERT(chat_templates.template_default);
std::string chat_template;
if (!opt.chat_template_file.empty()) {
chat_template = read_chat_template_file(opt.chat_template_file);
}
common_chat_templates_ptr chat_templates = common_chat_templates_init(llama_data.model.get(), chat_template);
static const bool stdout_a_terminal = is_stdout_a_terminal();
while (true) {
// Get user input
std::string user_input;
if (get_user_input(user_input, user) == 1) {
if (get_user_input(user_input, opt.user) == 1) {
return 0;
}
add_message("user", user.empty() ? user_input : user, llama_data);
int new_len;
if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, true, new_len, use_jinja) < 0) {
const int ret = process_user_message(opt, user_input, llama_data, chat_templates, prev_len, stdout_a_terminal);
if (ret == 1) {
return 1;
}
std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len);
std::string response;
if (generate_response(llama_data, prompt, response, stdout_a_terminal)) {
return 1;
}
if (!user.empty()) {
} else if (ret == 2) {
break;
}
add_message("assistant", response, llama_data);
if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, false, prev_len, use_jinja) < 0) {
return 1;
}
}
return 0;
@ -1165,7 +1208,7 @@ int main(int argc, const char ** argv) {
return 1;
}
if (chat_loop(llama_data, opt.user, opt.use_jinja)) {
if (chat_loop(llama_data, opt)) {
return 1;
}

2
examples/server/CMakeLists.txt
View File

@ -5,7 +5,7 @@ option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
if (MINGW)
# fix: https://github.com/ggerganov/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006
# fix: https://github.com/ggml-org/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006
add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
endif()

331
examples/server/README.md
View File

@ -7,14 +7,14 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
**Features:**
* LLM inference of F16 and quantized models on GPU and CPU
* [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
* Reranking endoint (WIP: https://github.com/ggerganov/llama.cpp/pull/9510)
* Reranking endoint (WIP: https://github.com/ggml-org/llama.cpp/pull/9510)
* Parallel decoding with multi-user support
* Continuous batching
* Multimodal (wip)
* Monitoring endpoints
* Schema-constrained JSON response format
The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggml-org/llama.cpp/issues/4216).
## Usage
@ -65,7 +65,7 @@ The project is under active development, and we are [looking for feedback and co
| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
| `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
| `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
| `--list-devices` | print list of available devices and exit |
| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
@ -127,6 +127,7 @@ The project is under active development, and we are [looking for feedback and co
| `--grammar-file FNAME` | file to read grammar from |
| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
| `--jinja` | Enable experimental Jinja templating engine (required for tool use) |
| `--reasoning-format FORMAT` | Controls extraction of model thinking traces and the format / field in which they are returned (default: `deepseek`; allowed values: `deepseek`, `none`; requires `--jinja`). `none` will leave thinking traces inline in `message.content` in a model-specific format, while `deepseek` will return them separately under `message.reasoning_content` |
**Example-specific params**
@ -177,7 +178,7 @@ Example usage of docker compose with environment variables:
```yml
services:
llamacpp-server:
image: ghcr.io/ggerganov/llama.cpp:server
image: ghcr.io/ggml-org/llama.cpp:server
ports:
- 8080:8080
volumes:
@ -272,10 +273,10 @@ You can consume the endpoints with Postman or NodeJS with axios library. You can
### Docker
```bash
docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
# or, with CUDA:
docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggml-org/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
```
## Testing with CURL
@ -1065,7 +1066,7 @@ print(completion.choices[0].text)
### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
*Options:*
@ -1119,7 +1120,7 @@ curl http://localhost:8080/v1/chat/completions \
*Tool call support*
[Function calling](https://platform.openai.com/docs/guides/function-calling) is supported for all models (see https://github.com/ggerganov/llama.cpp/pull/9639):
[Function calling](https://platform.openai.com/docs/guides/function-calling) is supported for all models (see https://github.com/ggml-org/llama.cpp/pull/9639):
- Requires `--jinja` flag
- Native tool call formats supported:
@ -1136,61 +1137,252 @@ curl http://localhost:8080/v1/chat/completions \
| Template | Format |
|----------|--------|
| CohereForAI-c4ai-command-r-plus-default.jinja | generic tool calls |
| CohereForAI-c4ai-command-r-plus-rag.jinja | generic tool calls |
| CohereForAI-c4ai-command-r-plus-tool_use.jinja | generic tool calls |
| MiniMaxAI-MiniMax-Text-01.jinja | generic tool calls |
| NexaAIDev-Octopus-v2.jinja | generic tool calls |
| NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | generic tool calls |
| NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | hermes 2 pro tool calls |
| NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | generic tool calls |
| NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | hermes 2 pro tool calls |
| NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | generic tool calls |
| NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | hermes 2 pro tool calls |
| OrionStarAI-Orion-14B-Chat.jinja | generic tool calls |
| Qwen-QwQ-32B-Preview.jinja | hermes 2 pro tool calls |
| Qwen-Qwen2-7B-Instruct.jinja | generic tool calls |
| Qwen-Qwen2-VL-7B-Instruct.jinja | generic tool calls |
| Qwen-Qwen2.5-7B-Instruct.jinja | hermes 2 pro tool calls |
| Qwen-Qwen2.5-Math-7B-Instruct.jinja | hermes 2 pro tool calls |
| TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | generic tool calls |
| abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | generic tool calls |
| bofenghuang-vigogne-2-70b-chat.jinja | generic tool calls |
| databricks-dbrx-instruct.jinja | generic tool calls |
| deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | generic tool calls |
| deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | deepseek r1 tool calls |
| deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | deepseek r1 tool calls |
| deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | deepseek r1 tool calls |
| deepseek-ai-DeepSeek-V2.5.jinja | deepseek r1 tool calls |
| deepseek-ai-deepseek-coder-33b-instruct.jinja | generic tool calls |
| google-gemma-2-2b-it.jinja | generic tool calls |
| google-gemma-7b-it.jinja | generic tool calls |
| indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | generic tool calls |
| mattshumer-Reflection-Llama-3.1-70B.jinja | generic tool calls |
| meetkai-functionary-medium-v3.2.jinja | functionary v3.2 tool calls |
| meta-llama-Llama-3.1-8B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) |
| meta-llama-Llama-3.2-3B-Instruct.jinja | llama 3.x tool calls |
| meta-llama-Llama-3.3-70B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) |
| meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) |
| microsoft-Phi-3-medium-4k-instruct.jinja | generic tool calls |
| microsoft-Phi-3-mini-4k-instruct.jinja | generic tool calls |
| microsoft-Phi-3-small-8k-instruct.jinja | generic tool calls |
| microsoft-Phi-3.5-mini-instruct.jinja | generic tool calls |
| microsoft-Phi-3.5-vision-instruct.jinja | generic tool calls |
| mistralai-Mistral-7B-Instruct-v0.2.jinja | generic tool calls |
| mistralai-Mistral-Large-Instruct-2407.jinja | mistral nemo tool calls |
| mistralai-Mistral-Large-Instruct-2411.jinja | generic tool calls |
| mistralai-Mistral-Nemo-Instruct-2407.jinja | mistral nemo tool calls |
| mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | generic tool calls |
| mlabonne-AlphaMonarch-7B.jinja | generic tool calls |
| nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | llama 3.x tool calls (w/ builtin tools) |
| openchat-openchat-3.5-0106.jinja | generic tool calls |
| teknium-OpenHermes-2.5-Mistral-7B.jinja | generic tool calls |
| Almawave-Velvet-14B.jinja | Hermes 2 Pro |
| AtlaAI-Selene-1-Mini-Llama-3.1-8B.jinja | Llama 3.x |
| CohereForAI-aya-expanse-8b.jinja | Generic |
| CohereForAI-c4ai-command-r-plus-default.jinja | Generic |
| CohereForAI-c4ai-command-r-plus-rag.jinja | Generic |
| CohereForAI-c4ai-command-r-plus-tool_use.jinja | Generic |
| CohereForAI-c4ai-command-r7b-12-2024-default.jinja | Command R7B (extract reasoning) |
| CohereForAI-c4ai-command-r7b-12-2024-rag.jinja | Command R7B (extract reasoning) |
| CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja | Command R7B (extract reasoning) |
| CohereForAI-c4ai-command-r7b-12-2024.jinja | Generic |
| DavieLion-Llama-3.2-1B-SPIN-iter3.jinja | Generic |
| Delta-Vector-Rei-12B.jinja | Mistral Nemo |
| EpistemeAI-Mistral-Nemo-Instruct-12B-Philosophy-Math.jinja | Mistral Nemo |
| FlofloB-83k_continued_pretraining_Qwen2.5-0.5B-Instruct_Unsloth_merged_16bit.jinja | Hermes 2 Pro |
| FlofloB-test_continued_pretraining_Phi-3-mini-4k-instruct_Unsloth_merged_16bit.jinja | Generic |
| HelpingAI-HAI-SER.jinja | Generic |
| HuggingFaceTB-SmolLM2-1.7B-Instruct.jinja | Generic |
| HuggingFaceTB-SmolLM2-135M-Instruct.jinja | Generic |
| HuggingFaceTB-SmolLM2-360M-Instruct.jinja | Generic |
| INSAIT-Institute-BgGPT-Gemma-2-27B-IT-v1.0.jinja | Generic |
| Ihor-Text2Graph-R1-Qwen2.5-0.5b.jinja | Hermes 2 Pro |
| Infinigence-Megrez-3B-Instruct.jinja | Generic |
| Josephgflowers-TinyLlama_v1.1_math_code-world-test-1.jinja | Generic |
| LGAI-EXAONE-EXAONE-3.5-2.4B-Instruct.jinja | Generic |
| LGAI-EXAONE-EXAONE-3.5-7.8B-Instruct.jinja | Generic |
| LatitudeGames-Wayfarer-12B.jinja | Generic |
| Magpie-Align-Llama-3-8B-Magpie-Align-v0.1.jinja | Generic |
| Magpie-Align-Llama-3.1-8B-Magpie-Align-v0.1.jinja | Generic |
| MaziyarPanahi-calme-3.2-instruct-78b.jinja | Generic |
| MiniMaxAI-MiniMax-Text-01.jinja | Generic |
| MiniMaxAI-MiniMax-VL-01.jinja | Generic |
| NaniDAO-deepseek-r1-qwen-2.5-32B-ablated.jinja | DeepSeek R1 (extract reasoning) |
| NexaAIDev-Octopus-v2.jinja | Generic |
| NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | Generic |
| NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | Hermes 2 Pro |
| NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | Generic |
| NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | Hermes 2 Pro |
| NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | Generic |
| NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | Hermes 2 Pro |
| NovaSky-AI-Sky-T1-32B-Flash.jinja | Hermes 2 Pro |
| NovaSky-AI-Sky-T1-32B-Preview.jinja | Hermes 2 Pro |
| OnlyCheeini-greesychat-turbo.jinja | Generic |
| Orenguteng-Llama-3.1-8B-Lexi-Uncensored-V2.jinja | Llama 3.x |
| OrionStarAI-Orion-14B-Chat.jinja | Generic |
| PowerInfer-SmallThinker-3B-Preview.jinja | Generic |
| PrimeIntellect-INTELLECT-1-Instruct.jinja | Generic |
| Qwen-QVQ-72B-Preview.jinja | Generic |
| Qwen-QwQ-32B-Preview.jinja | Hermes 2 Pro |
| Qwen-Qwen1.5-7B-Chat.jinja | Generic |
| Qwen-Qwen2-7B-Instruct.jinja | Generic |
| Qwen-Qwen2-VL-72B-Instruct.jinja | Generic |
| Qwen-Qwen2-VL-7B-Instruct.jinja | Generic |
| Qwen-Qwen2.5-0.5B.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-1.5B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-14B-Instruct-1M.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-14B.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-32B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-32B.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-3B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-72B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-7B-Instruct-1M.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-7B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-7B.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-Coder-32B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-Coder-7B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-Math-1.5B.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-Math-7B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-VL-3B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-VL-72B-Instruct.jinja | Hermes 2 Pro |
| Qwen-Qwen2.5-VL-7B-Instruct.jinja | Hermes 2 Pro |
| RWKV-Red-Team-ARWKV-7B-Preview-0.1.jinja | Hermes 2 Pro |
| SakanaAI-TinySwallow-1.5B-Instruct.jinja | Hermes 2 Pro |
| SakanaAI-TinySwallow-1.5B.jinja | Hermes 2 Pro |
| Sao10K-70B-L3.3-Cirrus-x1.jinja | Llama 3.x |
| SentientAGI-Dobby-Mini-Leashed-Llama-3.1-8B.jinja | Llama 3.x |
| SentientAGI-Dobby-Mini-Unhinged-Llama-3.1-8B.jinja | Llama 3.x |
| Steelskull-L3.3-Damascus-R1.jinja | Llama 3.x |
| Steelskull-L3.3-MS-Nevoria-70b.jinja | Llama 3.x |
| Steelskull-L3.3-Nevoria-R1-70b.jinja | Llama 3.x |
| THUDM-glm-4-9b-chat.jinja | Generic |
| THUDM-glm-edge-1.5b-chat.jinja | Generic |
| Tarek07-Progenitor-V1.1-LLaMa-70B.jinja | Llama 3.x |
| TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | Generic |
| TinyLlama-TinyLlama-1.1B-Chat-v1.0.jinja | Generic |
| UCLA-AGI-Mistral7B-PairRM-SPPO-Iter3.jinja | Generic |
| ValiantLabs-Llama3.1-8B-Enigma.jinja | Llama 3.x |
| abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | Generic |
| ai21labs-AI21-Jamba-1.5-Large.jinja | Generic |
| allenai-Llama-3.1-Tulu-3-405B-SFT.jinja | Generic |
| allenai-Llama-3.1-Tulu-3-405B.jinja | Generic |
| allenai-Llama-3.1-Tulu-3-8B.jinja | Generic |
| arcee-ai-Virtuoso-Lite.jinja | Hermes 2 Pro |
| arcee-ai-Virtuoso-Medium-v2.jinja | Hermes 2 Pro |
| arcee-ai-Virtuoso-Small-v2.jinja | Hermes 2 Pro |
| avemio-GRAG-NEMO-12B-ORPO-HESSIAN-AI.jinja | Generic |
| bespokelabs-Bespoke-Stratos-7B.jinja | Hermes 2 Pro |
| bfuzzy1-acheron-m1a-llama.jinja | Generic |
| bofenghuang-vigogne-2-70b-chat.jinja | Generic |
| bytedance-research-UI-TARS-72B-DPO.jinja | Generic |
| bytedance-research-UI-TARS-7B-DPO.jinja | Generic |
| bytedance-research-UI-TARS-7B-SFT.jinja | Generic |
| carsenk-phi3.5_mini_exp_825_uncensored.jinja | Generic |
| cyberagent-DeepSeek-R1-Distill-Qwen-14B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
| cyberagent-DeepSeek-R1-Distill-Qwen-32B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
| databricks-dbrx-instruct.jinja | Generic |
| deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | Generic |
| deepseek-ai-DeepSeek-Coder-V2-Lite-Base.jinja | Generic |
| deepseek-ai-DeepSeek-Coder-V2-Lite-Instruct.jinja | Generic |
| deepseek-ai-DeepSeek-R1-Distill-Llama-70B.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1-Distill-Qwen-14B.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1-Zero.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-V2-Lite.jinja | Generic |
| deepseek-ai-DeepSeek-V2.5.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-DeepSeek-V3.jinja | DeepSeek R1 (extract reasoning) |
| deepseek-ai-deepseek-coder-33b-instruct.jinja | Generic |
| deepseek-ai-deepseek-coder-6.7b-instruct.jinja | Generic |
| deepseek-ai-deepseek-coder-7b-instruct-v1.5.jinja | Generic |
| deepseek-ai-deepseek-llm-67b-chat.jinja | Generic |
| deepseek-ai-deepseek-llm-7b-chat.jinja | Generic |
| dicta-il-dictalm2.0-instruct.jinja | Generic |
| ehristoforu-Falcon3-8B-Franken-Basestruct.jinja | Hermes 2 Pro |
| fireworks-ai-llama-3-firefunction-v2.jinja | FireFunction v2 |
| godlikehhd-alpaca_data_sampled_ifd_new_5200.jinja | Hermes 2 Pro |
| godlikehhd-alpaca_data_score_max_0.7_2600.jinja | Hermes 2 Pro |
| google-gemma-2-27b-it.jinja | Generic |
| google-gemma-2-2b-it.jinja | Generic |
| google-gemma-2-2b-jpn-it.jinja | Generic |
| google-gemma-7b-it.jinja | Generic |
| huihui-ai-DeepSeek-R1-Distill-Llama-70B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
| huihui-ai-DeepSeek-R1-Distill-Llama-8B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
| huihui-ai-DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
| huihui-ai-DeepSeek-R1-Distill-Qwen-32B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
| huihui-ai-DeepSeek-R1-Distill-Qwen-7B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
| huihui-ai-Qwen2.5-14B-Instruct-1M-abliterated.jinja | Hermes 2 Pro |
| ibm-granite-granite-3.1-8b-instruct.jinja | Generic |
| indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | Generic |
| inflatebot-MN-12B-Mag-Mell-R1.jinja | Generic |
| jinaai-ReaderLM-v2.jinja | Generic |
| kms7530-chemeng_qwen-math-7b_24_1_100_1_nonmath.jinja | Hermes 2 Pro |
| knifeayumu-Cydonia-v1.3-Magnum-v4-22B.jinja | Mistral Nemo |
| langgptai-qwen1.5-7b-chat-sa-v0.1.jinja | Generic |
| lightblue-DeepSeek-R1-Distill-Qwen-7B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
| mattshumer-Reflection-Llama-3.1-70B.jinja | Generic |
| meetkai-functionary-medium-v3.1.jinja | Functionary v3.1 Llama 3.1 |
| meetkai-functionary-medium-v3.2.jinja | Functionary v3.2 |
| meta-llama-Llama-2-7b-chat-hf.jinja | Generic |
| meta-llama-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
| meta-llama-Llama-3.2-11B-Vision-Instruct.jinja | Llama 3.x |
| meta-llama-Llama-3.2-1B-Instruct.jinja | Llama 3.x |
| meta-llama-Llama-3.2-3B-Instruct.jinja | Llama 3.x |
| meta-llama-Llama-3.3-70B-Instruct.jinja | Llama 3.x |
| meta-llama-Meta-Llama-3-8B-Instruct.jinja | Generic |
| meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
| microsoft-Phi-3-medium-4k-instruct.jinja | Generic |
| microsoft-Phi-3-mini-4k-instruct.jinja | Generic |
| microsoft-Phi-3-small-8k-instruct.jinja | Generic |
| microsoft-Phi-3.5-mini-instruct.jinja | Generic |
| microsoft-Phi-3.5-vision-instruct.jinja | Generic |
| microsoft-phi-4.jinja | Generic |
| migtissera-Tess-3-Mistral-Nemo-12B.jinja | Generic |
| ministral-Ministral-3b-instruct.jinja | Generic |
| mistralai-Codestral-22B-v0.1.jinja | Generic |
| mistralai-Mistral-7B-Instruct-v0.1.jinja | Generic |
| mistralai-Mistral-7B-Instruct-v0.2.jinja | Generic |
| mistralai-Mistral-7B-Instruct-v0.3.jinja | Mistral Nemo |
| mistralai-Mistral-Large-Instruct-2407.jinja | Mistral Nemo |
| mistralai-Mistral-Large-Instruct-2411.jinja | Generic |
| mistralai-Mistral-Nemo-Instruct-2407.jinja | Mistral Nemo |
| mistralai-Mistral-Small-24B-Instruct-2501.jinja | Generic |
| mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | Generic |
| mkurman-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
| mlabonne-AlphaMonarch-7B.jinja | Generic |
| mlx-community-Josiefied-Qwen2.5-0.5B-Instruct-abliterated-v1-float32.jinja | Hermes 2 Pro |
| mlx-community-Qwen2.5-VL-7B-Instruct-8bit.jinja | Hermes 2 Pro |
| mobiuslabsgmbh-DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1.jinja | DeepSeek R1 (extract reasoning) |
| netcat420-MFANNv0.20.jinja | Generic |
| netcat420-MFANNv0.24.jinja | Generic |
| netease-youdao-Confucius-o1-14B.jinja | Hermes 2 Pro |
| nvidia-AceMath-7B-RM.jinja | Hermes 2 Pro |
| nvidia-Eagle2-1B.jinja | Hermes 2 Pro |
| nvidia-Eagle2-9B.jinja | Hermes 2 Pro |
| nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | Llama 3.x |
| onnx-community-DeepSeek-R1-Distill-Qwen-1.5B-ONNX.jinja | DeepSeek R1 (extract reasoning) |
| open-thoughts-OpenThinker-7B.jinja | Hermes 2 Pro |
| openchat-openchat-3.5-0106.jinja | Generic |
| pankajmathur-orca_mini_v6_8b.jinja | Generic |
| princeton-nlp-Mistral-7B-Base-SFT-RDPO.jinja | Generic |
| princeton-nlp-Mistral-7B-Instruct-DPO.jinja | Generic |
| princeton-nlp-Mistral-7B-Instruct-RDPO.jinja | Generic |
| prithivMLmods-Bellatrix-Tiny-1.5B-R1.jinja | Hermes 2 Pro |
| prithivMLmods-Bellatrix-Tiny-1B-R1.jinja | Llama 3.x |
| prithivMLmods-Bellatrix-Tiny-1B-v3.jinja | Generic |
| prithivMLmods-Bellatrix-Tiny-3B-R1.jinja | Llama 3.x |
| prithivMLmods-Blaze-14B-xElite.jinja | Generic |
| prithivMLmods-Calcium-Opus-14B-Elite2-R1.jinja | Hermes 2 Pro |
| prithivMLmods-Calme-Ties-78B.jinja | Generic |
| prithivMLmods-Calme-Ties2-78B.jinja | Generic |
| prithivMLmods-Calme-Ties3-78B.jinja | Generic |
| prithivMLmods-ChemQwen2-vL.jinja | Generic |
| prithivMLmods-GWQ2b.jinja | Generic |
| prithivMLmods-LatexMind-2B-Codec.jinja | Generic |
| prithivMLmods-Llama-3.2-6B-AlgoCode.jinja | Llama 3.x |
| prithivMLmods-Megatron-Opus-14B-Exp.jinja | Hermes 2 Pro |
| prithivMLmods-Megatron-Opus-14B-Stock.jinja | Hermes 2 Pro |
| prithivMLmods-Megatron-Opus-7B-Exp.jinja | Hermes 2 Pro |
| prithivMLmods-Omni-Reasoner-Merged.jinja | Hermes 2 Pro |
| prithivMLmods-Omni-Reasoner4-Merged.jinja | Hermes 2 Pro |
| prithivMLmods-Primal-Opus-14B-Optimus-v1.jinja | Hermes 2 Pro |
| prithivMLmods-QwQ-Math-IO-500M.jinja | Hermes 2 Pro |
| prithivMLmods-Qwen-7B-Distill-Reasoner.jinja | DeepSeek R1 (extract reasoning) |
| prithivMLmods-Qwen2.5-1.5B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
| prithivMLmods-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
| prithivMLmods-Qwen2.5-32B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
| prithivMLmods-Qwen2.5-7B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
| prithivMLmods-Triangulum-v2-10B.jinja | Hermes 2 Pro |
| qingy2024-Falcon3-2x10B-MoE-Instruct.jinja | Hermes 2 Pro |
| rubenroy-Zurich-14B-GCv2-5m.jinja | Hermes 2 Pro |
| rubenroy-Zurich-7B-GCv2-5m.jinja | Hermes 2 Pro |
| silma-ai-SILMA-Kashif-2B-Instruct-v1.0.jinja | Generic |
| simplescaling-s1-32B.jinja | Hermes 2 Pro |
| sometimesanotion-Lamarck-14B-v0.7.jinja | Hermes 2 Pro |
| sonthenguyen-zephyr-sft-bnb-4bit-DPO-mtbr-180steps.jinja | Generic |
| sthenno-tempesthenno-icy-0130.jinja | Generic |
| sumink-qwft.jinja | Hermes 2 Pro |
| teknium-OpenHermes-2.5-Mistral-7B.jinja | Generic |
| thirdeyeai-elevate360m.jinja | Generic |
| tiiuae-Falcon3-10B-Instruct.jinja | Hermes 2 Pro |
| unsloth-DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit.jinja | DeepSeek R1 (extract reasoning) |
| unsloth-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
| unsloth-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
| unsloth-Mistral-Small-24B-Instruct-2501-unsloth-bnb-4bit.jinja | Generic |
| upstage-solar-pro-preview-instruct.jinja | Generic |
| whyhow-ai-PatientSeek.jinja | Generic |
| xwen-team-Xwen-72B-Chat.jinja | Hermes 2 Pro |
| xwen-team-Xwen-7B-Chat.jinja | Hermes 2 Pro |
This table can be generated with:
```bash
./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null
```
</details>
@ -1202,11 +1394,20 @@ curl http://localhost:8080/v1/chat/completions \
```shell
# Native support:
llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L
llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M
llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
# Native support for DeepSeek R1 works best w/ our own template (official template buggy)
llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \
--chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \
--chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
# Native support requires the right template for these GGUFs:
llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \
@ -1236,17 +1437,17 @@ curl http://localhost:8080/v1/chat/completions \
{
"type":"function",
"function":{
"name":"get_current_weather",
"description":"Get the current weather in a given location",
"name":"python",
"description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
"parameters":{
"type":"object",
"properties":{
"location":{
"code":{
"type":"string",
"description":"The city and state, e.g. San Francisco, CA"
"description":"The code to run in the ipython interpreter."
}
},
"required":["location"]
"required":["code"]
}
}
}
@ -1254,7 +1455,7 @@ curl http://localhost:8080/v1/chat/completions \
"messages": [
{
"role": "user",
"content": "What is the weather like in Istanbul?."
"content": "Print a hello world message with python."
}
]
}'
@ -1398,7 +1599,7 @@ Apart from error types supported by OAI, we also have custom types that are spec
### Legacy completion web UI
A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggml-org/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
For example:

669
examples/server/httplib.h
View File

File diff suppressed because it is too large Load Diff

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

Binary file not shown.

111
examples/server/server.cpp
View File

@ -42,7 +42,7 @@ enum stop_type {
STOP_TYPE_LIMIT,
};
// state diagram: https://github.com/ggerganov/llama.cpp/pull/9283
// state diagram: https://github.com/ggml-org/llama.cpp/pull/9283
enum slot_state {
SLOT_STATE_IDLE,
SLOT_STATE_STARTED, // TODO: this state is only used for setting up the initial prompt processing; maybe merge it with launch_slot_with_task in the future
@ -173,6 +173,7 @@ struct slot_params {
{"grammar_trigger_words", grammar_trigger_words},
{"grammar_trigger_tokens", sampling.grammar_trigger_tokens},
{"preserved_tokens", sampling.preserved_tokens},
{"chat_format", common_chat_format_name(oaicompat_chat_format)},
{"samplers", samplers},
{"speculative.n_max", speculative.n_max},
{"speculative.n_min", speculative.n_min},
@ -273,7 +274,7 @@ struct server_task {
params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min);
params.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min);
params.speculative.n_min = std::max(params.speculative.n_min, 2);
params.speculative.n_min = std::max(params.speculative.n_min, 0);
params.speculative.n_max = std::max(params.speculative.n_max, 0);
// Use OpenAI API logprobs only if n_probs wasn't provided
@ -328,9 +329,6 @@ struct server_task {
}
// process "json_schema" and "grammar"
if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) {
throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both");
}
if (data.contains("json_schema") && !data.contains("grammar")) {
try {
auto schema = json_value(data, "json_schema", json::object());
@ -724,9 +722,19 @@ struct server_task_result_cmpl_final : server_task_result {
msg.content = content;
}
json tool_calls;
json message {
{"role", "assistant"},
};
if (!msg.reasoning_content.empty()) {
message["reasoning_content"] = msg.reasoning_content;
}
if (msg.content.empty() && !msg.tool_calls.empty()) {
message["content"] = json();
} else {
message["content"] = msg.content;
}
if (!msg.tool_calls.empty()) {
tool_calls = json::array();
auto tool_calls = json::array();
for (const auto & tc : msg.tool_calls) {
tool_calls.push_back({
{"type", "function"},
@ -737,15 +745,7 @@ struct server_task_result_cmpl_final : server_task_result {
{"id", tc.id},
});
}
}
json message {
{"content", msg.content},
{"tool_calls", tool_calls},
{"role", "assistant"},
};
if (!msg.tool_plan.empty()) {
message["tool_plan"] = msg.tool_plan;
message["tool_calls"] = tool_calls;
}
json choice {
@ -1804,7 +1804,7 @@ struct server_context {
// Necessary similarity of prompt for slot selection
float slot_prompt_similarity = 0.0f;
common_chat_templates chat_templates;
common_chat_templates_ptr chat_templates;
~server_context() {
// Clear any sampling context
@ -1888,45 +1888,17 @@ struct server_context {
llama_init_dft.context.reset();
}
if (params_base.chat_template.empty() && !validate_builtin_chat_template(params.use_jinja)) {
chat_templates = common_chat_templates_init(model, params_base.chat_template);
try {
common_chat_format_example(chat_templates.get(), params.use_jinja);
} catch (const std::exception & e) {
SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
chat_templates = common_chat_templates_from_model(model, "chatml");
} else {
chat_templates = common_chat_templates_from_model(model, params_base.chat_template);
chat_templates = common_chat_templates_init(model, "chatml");
}
GGML_ASSERT(chat_templates.template_default.get() != nullptr);
return true;
}
bool validate_builtin_chat_template(bool use_jinja) const {
llama_chat_message chat[] = {{"user", "test"}};
if (use_jinja) {
auto templates = common_chat_templates_from_model(model, "");
common_chat_inputs inputs;
inputs.messages = json::array({{
{"role", "user"},
{"content", "test"},
}});
GGML_ASSERT(templates.template_default);
try {
common_chat_params_init(*templates.template_default, inputs);
if (templates.template_tool_use) {
common_chat_params_init(*templates.template_tool_use, inputs);
}
return true;
} catch (const std::exception & e) {
SRV_ERR("failed to apply template: %s\n", e.what());
return false;
}
} else {
const char * tmpl = llama_model_chat_template(model, /* name */ nullptr);
const int32_t chat_res = llama_chat_apply_template(tmpl, chat, 1, true, nullptr, 0);
return chat_res > 0;
}
}
void init() {
const int32_t n_ctx_slot = n_ctx / params_base.n_parallel;
@ -2073,8 +2045,8 @@ struct server_context {
if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
// Might be better to reject the request with a 400 ?
SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.params.n_predict, slot.n_predict);
slot.params.n_predict = slot.n_predict;
SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict);
}
if (slot.params.ignore_eos && has_eos_token) {
@ -3653,7 +3625,7 @@ int main(int argc, char ** argv) {
}, {
{"name", "n_busy_slots_per_decode"},
{"help", "Average number of busy slots per llama_decode() call"},
{"value", (float) res_metrics->n_busy_slots_total / (float) res_metrics->n_decode_total}
{"value", (float) res_metrics->n_busy_slots_total / std::max((float) res_metrics->n_decode_total, 1.f)}
}}},
{"gauge", {{
{"name", "prompt_tokens_seconds"},
@ -3819,13 +3791,15 @@ int main(int argc, char ** argv) {
{ "default_generation_settings", ctx_server.default_generation_settings_for_props },
{ "total_slots", ctx_server.params_base.n_parallel },
{ "model_path", ctx_server.params_base.model },
{ "chat_template", ctx_server.chat_templates.template_default->source() },
{ "bos_token", ctx_server.chat_templates.template_default->bos_token() },
{ "eos_token", ctx_server.chat_templates.template_default->eos_token() },
{ "chat_template", common_chat_templates_source(ctx_server.chat_templates.get()) },
{ "bos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
{ "eos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
{ "build_info", build_info },
};
if (ctx_server.params_base.use_jinja && ctx_server.chat_templates.template_tool_use) {
data["chat_template_tool_use"] = ctx_server.chat_templates.template_tool_use->source();
if (ctx_server.params_base.use_jinja) {
if (auto tool_use_src = common_chat_templates_source(ctx_server.chat_templates.get(), "tool_use")) {
data["chat_template_tool_use"] = tool_use_src;
}
}
res_ok(res, data);
@ -4060,7 +4034,7 @@ int main(int argc, char ** argv) {
}
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get());
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
@ -4073,7 +4047,7 @@ int main(int argc, char ** argv) {
// same with handle_chat_completions, but without inference part
const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get());
res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
};
@ -4260,6 +4234,11 @@ int main(int argc, char ** argv) {
// return;
//}
// if true, use TEI API format, otherwise use Jina API format
// Jina: https://jina.ai/reranker/
// TEI: https://huggingface.github.io/text-embeddings-inference/#/Text%20Embeddings%20Inference/rerank
bool is_tei_format = body.contains("texts");
json query;
if (body.count("query") == 1) {
query = body.at("query");
@ -4272,7 +4251,8 @@ int main(int argc, char ** argv) {
return;
}
std::vector<std::string> documents = json_value(body, "documents", std::vector<std::string>());
std::vector<std::string> documents = json_value(body, "documents",
json_value(body, "texts", std::vector<std::string>()));
if (documents.empty()) {
res_error(res, format_error_response("\"documents\" must be a non-empty string array", ERROR_TYPE_INVALID_REQUEST));
return;
@ -4317,7 +4297,12 @@ int main(int argc, char ** argv) {
}
// write JSON response
json root = format_response_rerank(body, responses);
json root = format_response_rerank(
body,
responses,
is_tei_format,
documents);
res_ok(res, root);
};
@ -4479,8 +4464,8 @@ int main(int argc, char ** argv) {
// print sample chat example to make it clear which template is used
LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
ctx_server.chat_templates.template_default->source().c_str(),
common_chat_format_example(*ctx_server.chat_templates.template_default, ctx_server.params_base.use_jinja).c_str());
common_chat_templates_source(ctx_server.chat_templates.get()),
common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str());
ctx_server.queue_tasks.on_new_task([&ctx_server](const server_task & task) {
ctx_server.process_single_task(task);

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

@ -48,7 +48,7 @@ DEBUG=1 ./tests.sh -s -v -x
To run all the tests in a file:
```shell
./tests.sh unit/test_chat_completion.py.py -v -x
./tests.sh unit/test_chat_completion.py -v -x
```
To run a single test:

45
examples/server/tests/unit/test_chat_completion.py
View File

@ -21,6 +21,8 @@ def create_server():
(None, "Book", "What is the best book", 8, "^ blue", 23, 8, "length", True, "This is not a chat template, it is"),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", False, None),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True, None),
(None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", False, None),
(None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", True, None),
]
)
def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason, jinja, chat_template):
@ -44,7 +46,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
assert res.body["usage"]["completion_tokens"] == n_predicted
choice = res.body["choices"][0]
assert "assistant" == choice["message"]["role"]
assert match_regex(re_content, choice["message"]["content"])
assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}'
assert choice["finish_reason"] == finish_reason
@ -169,6 +171,47 @@ def test_completion_with_response_format(response_format: dict, n_predicted: int
assert "error" in res.body
@pytest.mark.parametrize("jinja,json_schema,n_predicted,re_content", [
(False, {"const": "42"}, 6, "\"42\""),
(True, {"const": "42"}, 6, "\"42\""),
])
def test_completion_with_json_schema(jinja: bool, json_schema: dict, n_predicted: int, re_content: str):
global server
server.jinja = jinja
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predicted,
"messages": [
{"role": "system", "content": "You are a coding assistant."},
{"role": "user", "content": "Write an example"},
],
"json_schema": json_schema,
})
assert res.status_code == 200, f'Expected 200, got {res.status_code}'
choice = res.body["choices"][0]
assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}'
@pytest.mark.parametrize("jinja,grammar,n_predicted,re_content", [
(False, 'root ::= "a"{5,5}', 6, "a{5,5}"),
(True, 'root ::= "a"{5,5}', 6, "a{5,5}"),
])
def test_completion_with_grammar(jinja: bool, grammar: str, n_predicted: int, re_content: str):
global server
server.jinja = jinja
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predicted,
"messages": [
{"role": "user", "content": "Does not matter what I say, does it?"},
],
"grammar": grammar,
})
assert res.status_code == 200, res.body
choice = res.body["choices"][0]
assert match_regex(re_content, choice["message"]["content"]), choice["message"]["content"]
@pytest.mark.parametrize("messages", [
None,
"string",

38
examples/server/tests/unit/test_rerank.py
View File

@ -10,17 +10,20 @@ def create_server():
server = ServerPreset.jina_reranker_tiny()
TEST_DOCUMENTS = [
"A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.",
"Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.",
"Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.",
"Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine."
]
def test_rerank():
global server
server.start()
res = server.make_request("POST", "/rerank", data={
"query": "Machine learning is",
"documents": [
"A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.",
"Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.",
"Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.",
"Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine."
]
"documents": TEST_DOCUMENTS,
})
assert res.status_code == 200
assert len(res.body["results"]) == 4
@ -38,6 +41,29 @@ def test_rerank():
assert least_relevant["index"] == 3
def test_rerank_tei_format():
global server
server.start()
res = server.make_request("POST", "/rerank", data={
"query": "Machine learning is",
"texts": TEST_DOCUMENTS,
})
assert res.status_code == 200
assert len(res.body) == 4
most_relevant = res.body[0]
least_relevant = res.body[0]
for doc in res.body:
if doc["score"] > most_relevant["score"]:
most_relevant = doc
if doc["score"] < least_relevant["score"]:
least_relevant = doc
assert most_relevant["score"] > least_relevant["score"]
assert most_relevant["index"] == 2
assert least_relevant["index"] == 3
@pytest.mark.parametrize("documents", [
[],
None,

177
examples/server/tests/unit/test_tool_call.py
View File

@ -92,6 +92,7 @@ def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, a
tool_calls = choice["message"].get("tool_calls")
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
tool_call = tool_calls[0]
assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
assert expected_function_name == tool_call["function"]["name"]
actual_arguments = tool_call["function"]["arguments"]
@ -155,11 +156,11 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict,
(TEST_TOOL, "success", "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(PYTHON_TOOL, "code", "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(PYTHON_TOOL, "code", "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
# (PYTHON_TOOL, "code", "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
(TEST_TOOL, "success", "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
(PYTHON_TOOL, "code", "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
(PYTHON_TOOL, "code", "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"),
# (PYTHON_TOOL, "code", "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"),
(TEST_TOOL, "success", "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
(PYTHON_TOOL, "code", "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
@ -175,7 +176,7 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict,
(TEST_TOOL, "success", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
(PYTHON_TOOL, "code", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
(PYTHON_TOOL, "code", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"),
# (PYTHON_TOOL, "code", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"),
# TODO: fix these
# (TEST_TOOL, "success", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# (PYTHON_TOOL, "code", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
@ -214,6 +215,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str
tool_calls = choice["message"].get("tool_calls")
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
tool_call = tool_calls[0]
assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
assert expected_function_name == tool_call["function"]["name"]
actual_arguments = tool_call["function"]["arguments"]
@ -273,7 +275,6 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t
@pytest.mark.slow
@pytest.mark.parametrize("hf_repo,template_override", [
("bartowski/c4ai-command-r7b-12-2024-GGUF:Q4_K_M", ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")),
("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
@ -298,13 +299,16 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t
("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"),
("bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L", ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")),
("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
("bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
# ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
# ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
])
def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None):
def test_weather(hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
n_predict = 512
server.n_slots = 1
@ -323,6 +327,7 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predict,
"messages": [
{"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things."},
{"role": "user", "content": "What is the weather in Istanbul?"},
],
"tools": [WEATHER_TOOL],
@ -332,6 +337,7 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
tool_calls = choice["message"].get("tool_calls")
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
tool_call = tool_calls[0]
assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"]
actual_arguments = json.loads(tool_call["function"]["arguments"])
assert 'location' in actual_arguments, f"location not found in {json.dumps(actual_arguments)}"
@ -340,22 +346,166 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
assert re.match('^Istanbul(, (TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}'
@pytest.mark.slow
@pytest.mark.parametrize("result_override,n_predict,hf_repo,template_override", [
(None, 128, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"),
(None, 128, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
(None, 128, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, 128, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(None, 128, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
(None, 128, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
(None, 128, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
(None, 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
(None, 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
(None, 128, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
# TODO: fix these (wrong results, either didn't respect decimal instruction or got wrong value)
("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
])
def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
# n_predict = 512
server.n_slots = 1
server.jinja = True
server.n_ctx = 8192 * 2
server.n_predict = n_predict
server.model_hf_repo = hf_repo
server.model_hf_file = None
if isinstance(template_override, tuple):
(template_hf_repo, template_variant) = template_override
server.chat_template_file = f"../../../models/templates/{template_hf_repo.replace('/', '-') + ('-' + template_variant if template_variant else '')}.jinja"
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
elif isinstance(template_override, str):
server.chat_template = template_override
server.start(timeout_seconds=TIMEOUT_SERVER_START)
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predict,
"messages": [
{"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things, and provide very concise answers. Do not explain your reasoning to the user. Provide any numerical values back to the user with at most two decimals."},
{"role": "user", "content": "What's the y coordinate of a point on the unit sphere at angle 30 degrees?"},
{
"role": "assistant",
"content": None,
"tool_calls": [
{
"id": "call_6789",
"type": "function",
"function": {
"name": "calculate",
"arguments": "{\"expression\":\"sin(30 * pi / 180)\"}"
}
}
]
},
{
"role": "tool",
"name": "calculate",
"content": "0.55644242476",
"tool_call_id": "call_6789"
}
],
"tools": [
{
"type":"function",
"function":{
"name":"calculate",
"description":"A calculator function that computes values of arithmetic expressions in the Python syntax",
"parameters":{
"type":"object",
"properties":{
"expression":{
"type":"string",
"description":"An arithmetic expression to compute the value of (Python syntad, assuming all floats)"
}
},
"required":["expression"]
}
}
}
]
}, timeout=TIMEOUT_HTTP_REQUEST)
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
choice = res.body["choices"][0]
tool_calls = choice["message"].get("tool_calls")
assert tool_calls is None, f'Expected no tool call in {choice["message"]}'
content = choice["message"].get("content")
assert content is not None, f'Expected content in {choice["message"]}'
if result_override is not None:
assert re.match(result_override, content), f'Expected {result_override}, got {content}'
else:
assert re.match('^[\\s\\S]*?The (y[ -])?coordinate [\\s\\S]*?is (approximately )?0\\.56\\b|^0\\.56$', content), \
f'Expected something like "The y coordinate is 0.56.", got {content}'
@pytest.mark.slow
@pytest.mark.parametrize("n_predict,reasoning_format,expect_content,expect_reasoning_content,hf_repo,template_override", [
(128, 'deepseek', "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(128, None, "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, 'deepseek', "To find the sum of.*", "I need to calculate the sum of 102 and 7.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'none', "^I need[\\s\\S]*?</think>\n?To find.*", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'deepseek', "To find the sum of.*", "First, I [\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
])
def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none'] | None, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
server.n_slots = 1
server.reasoning_format = reasoning_format
server.jinja = True
server.n_ctx = 8192 * 2
server.n_predict = n_predict
server.model_hf_repo = hf_repo
server.model_hf_file = None
if isinstance(template_override, tuple):
(template_hf_repo, template_variant) = template_override
server.chat_template_file = f"../../../models/templates/{template_hf_repo.replace('/', '-') + ('-' + template_variant if template_variant else '')}.jinja"
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
elif isinstance(template_override, str):
server.chat_template = template_override
server.start(timeout_seconds=TIMEOUT_SERVER_START)
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predict,
"messages": [
{"role": "user", "content": "What's the sum of 102 and 7?"},
]
}, timeout=TIMEOUT_HTTP_REQUEST)
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
choice = res.body["choices"][0]
assert choice["message"].get("tool_calls") is None, f'Expected no tool call in {choice["message"]}'
content = choice["message"].get("content")
if expect_content is None:
assert content is None, f'Expected no content in {choice["message"]}'
else:
assert re.match(expect_content, content), f'Expected {expect_content}, got {content}'
reasoning_content = choice["message"].get("reasoning_content")
if expect_reasoning_content is None:
assert reasoning_content is None, f'Expected no reasoning content in {choice["message"]}'
else:
assert re.match(expect_reasoning_content, reasoning_content), f'Expected {expect_reasoning_content}, got {reasoning_content}'
@pytest.mark.slow
@pytest.mark.parametrize("expected_arguments_override,hf_repo,template_override", [
(None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# (None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai-functionary-medium-v3.2", None)),
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"),
(None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
('{"code":"print("}', "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
('{"code":"print("}', "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
('{"code":"print("}', "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
(None, "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
(None, "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"),
('{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
('{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"),
@ -371,15 +521,13 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
# Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
(None, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
# (None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
])
def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
def test_hello_world(expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
server.n_slots = 1
server.jinja = True
server.n_ctx = 8192
server.n_predict = 128
server.n_predict = 512 # High because of DeepSeek R1
server.model_hf_repo = hf_repo
server.model_hf_file = None
if isinstance(template_override, tuple):
@ -406,6 +554,7 @@ def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo:
tool_calls = choice["message"].get("tool_calls")
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
tool_call = tool_calls[0]
assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
assert tool_call["function"]["name"] == PYTHON_TOOL["function"]["name"]
actual_arguments = tool_call["function"]["arguments"]
if expected_arguments_override is not None:

3
examples/server/tests/utils.py
View File

@ -78,6 +78,7 @@ class ServerProcess:
draft_max: int | None = None
no_webui: bool | None = None
jinja: bool | None = None
reasoning_format: Literal['deepseek', 'none'] | None = None
chat_template: str | None = None
chat_template_file: str | None = None
@ -172,6 +173,8 @@ class ServerProcess:
server_args.append("--no-webui")
if self.jinja:
server_args.append("--jinja")
if self.reasoning_format is not None:
server_args.extend(("--reasoning-format", self.reasoning_format))
if self.chat_template:
server_args.extend(["--chat-template", self.chat_template])
if self.chat_template_file:

190
examples/server/utils.hpp
View File

@ -7,14 +7,14 @@
// increase max payload length to allow use of larger context size
#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
// disable Nagle's algorithm
#define CPPHTTPLIB_TCP_NODELAY true
#include "httplib.h"
// Change JSON_ASSERT from assert() to GGML_ASSERT:
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
#include "minja.hpp"
#include "chat.hpp"
#include "chat-template.hpp"
#include "chat.h"
#include <random>
#include <sstream>
@ -347,41 +347,6 @@ static llama_tokens format_infill(
return embd_inp;
}
// Format given chat. If tmpl is empty, we take the template from model metadata
inline std::string format_chat(const common_chat_template & tmpl, const std::vector<json> & messages) {
std::vector<common_chat_msg> chat;
for (size_t i = 0; i < messages.size(); ++i) {
const auto & curr_msg = messages[i];
std::string role = json_value(curr_msg, "role", std::string(""));
std::string content;
if (curr_msg.contains("content")) {
if (curr_msg["content"].is_string()) {
content = curr_msg["content"].get<std::string>();
} else if (curr_msg["content"].is_array()) {
for (const auto & part : curr_msg["content"]) {
if (part.contains("text")) {
content += "\n" + part["text"].get<std::string>();
}
}
} else {
throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
}
} else {
throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
}
chat.push_back({role, content, /* tool_calls= */ {}});
}
const auto formatted_chat = common_chat_apply_template(tmpl, chat, true, /* use_jinja= */ false);
LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str());
return formatted_chat;
}
//
// base64 utils (TODO: move to common in the future)
//
@ -578,12 +543,10 @@ static json oaicompat_completion_params_parse(const json & body) {
static json oaicompat_completion_params_parse(
const json & body, /* openai api json semantics */
bool use_jinja,
const common_chat_templates & chat_templates)
common_reasoning_format reasoning_format,
const struct common_chat_templates * tmpls)
{
json llama_params;
const auto & tmpl = body.contains("tools") && chat_templates.template_tool_use
? *chat_templates.template_tool_use
: *chat_templates.template_default;
auto tools = json_value(body, "tools", json());
auto stream = json_value(body, "stream", false);
@ -609,61 +572,58 @@ static json oaicompat_completion_params_parse(
llama_params["stop"] = json_value(body, "stop", json::array());
}
auto json_schema = json_value(body, "json_schema", json());
auto grammar = json_value(body, "grammar", std::string());
if (!json_schema.is_null() && !grammar.empty()) {
throw std::runtime_error("Cannot use both json_schema and grammar");
}
// Handle "response_format" field
if (body.contains("response_format")) {
json response_format = json_value(body, "response_format", json::object());
std::string response_type = json_value(response_format, "type", std::string());
if (response_type == "json_object") {
llama_params["json_schema"] = json_value(response_format, "schema", json::object());
json_schema = json_value(response_format, "schema", json::object());
} else if (response_type == "json_schema") {
json json_schema = json_value(response_format, "json_schema", json::object());
llama_params["json_schema"] = json_value(json_schema, "schema", json::object());
json_schema = json_value(json_schema, "schema", json::object());
} else if (!response_type.empty() && response_type != "text") {
throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
}
}
// Apply chat template to the list of messages
if (use_jinja) {
auto tool_choice = json_value(body, "tool_choice", std::string("auto"));
if (tool_choice != "none" && tool_choice != "auto" && tool_choice != "required") {
throw std::runtime_error("Invalid tool_choice: " + tool_choice);
}
if (tool_choice != "none" && llama_params.contains("grammar")) {
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
}
common_chat_inputs inputs;
inputs.messages = body.at("messages");
inputs.tools = tools;
inputs.tool_choice = tool_choice;
inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) {
LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n");
inputs.parallel_tool_calls = false;
}
inputs.stream = stream;
// TODO: support mixing schema w/ tools beyond generic format.
inputs.json_schema = json_value(llama_params, "json_schema", json());
auto chat_params = common_chat_params_init(tmpl, inputs);
common_chat_templates_inputs inputs;
inputs.messages = common_chat_msgs_parse_oaicompat(body.at("messages"));
inputs.tools = common_chat_tools_parse_oaicompat(tools);
inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto")));
inputs.json_schema = json_schema.is_null() ? "" : json_schema.dump();
inputs.grammar = grammar;
inputs.add_generation_prompt = true;
inputs.use_jinja = use_jinja;
inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
inputs.extract_reasoning = reasoning_format != COMMON_REASONING_FORMAT_NONE;
if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) {
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
}
llama_params["chat_format"] = static_cast<int>(chat_params.format);
llama_params["prompt"] = chat_params.prompt;
llama_params["grammar"] = chat_params.grammar;
llama_params["grammar_lazy"] = chat_params.grammar_lazy;
auto grammar_triggers = json::array();
for (const auto & trigger : chat_params.grammar_triggers) {
grammar_triggers.push_back({
{"word", trigger.word},
{"at_start", trigger.at_start},
});
}
llama_params["grammar_triggers"] = grammar_triggers;
llama_params["preserved_tokens"] = chat_params.preserved_tokens;
for (const auto & stop : chat_params.additional_stops) {
llama_params["stop"].push_back(stop);
}
} else {
llama_params["prompt"] = format_chat(tmpl, body.at("messages"));
// Apply chat template to the list of messages
auto chat_params = common_chat_templates_apply(tmpls, inputs);
llama_params["chat_format"] = static_cast<int>(chat_params.format);
llama_params["prompt"] = chat_params.prompt;
llama_params["grammar"] = chat_params.grammar;
llama_params["grammar_lazy"] = chat_params.grammar_lazy;
auto grammar_triggers = json::array();
for (const auto & trigger : chat_params.grammar_triggers) {
grammar_triggers.push_back({
{"word", trigger.word},
{"at_start", trigger.at_start},
});
}
llama_params["grammar_triggers"] = grammar_triggers;
llama_params["preserved_tokens"] = chat_params.preserved_tokens;
for (const auto & stop : chat_params.additional_stops) {
llama_params["stop"].push_back(stop);
}
// Handle "n" field
@ -735,29 +695,51 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
return res;
}
static json format_response_rerank(const json & request, const json & ranks) {
json data = json::array();
int32_t n_tokens = 0;
int i = 0;
for (const auto & rank : ranks) {
data.push_back(json{
{"index", i++},
{"relevance_score", json_value(rank, "score", 0.0)},
});
static json format_response_rerank(
const json & request,
const json & ranks,
bool is_tei_format,
std::vector<std::string> & texts) {
json res;
if (is_tei_format) {
// TEI response format
res = json::array();
bool return_text = json_value(request, "return_text", false);
for (const auto & rank : ranks) {
int index = json_value(rank, "index", 0);
json elem = json{
{"index", index},
{"score", json_value(rank, "score", 0.0)},
};
if (return_text) {
elem["text"] = std::move(texts[index]);
}
res.push_back(elem);
}
} else {
// Jina response format
json results = json::array();
int32_t n_tokens = 0;
for (const auto & rank : ranks) {
results.push_back(json{
{"index", json_value(rank, "index", 0)},
{"relevance_score", json_value(rank, "score", 0.0)},
});
n_tokens += json_value(rank, "tokens_evaluated", 0);
n_tokens += json_value(rank, "tokens_evaluated", 0);
}
res = json{
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json{
{"prompt_tokens", n_tokens},
{"total_tokens", n_tokens}
}},
{"results", results}
};
}
json res = json {
{"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
{"object", "list"},
{"usage", json {
{"prompt_tokens", n_tokens},
{"total_tokens", n_tokens}
}},
{"results", data}
};
return res;
}

29
examples/server/webui/src/components/ChatMessage.tsx
View File

@ -159,6 +159,35 @@ export default function ChatMessage({
</div>
</details>
)}
{msg.extra && msg.extra.length > 0 && (
<details
className={classNames({
'collapse collapse-arrow mb-4 bg-base-200': true,
'bg-opacity-10': msg.role !== 'assistant',
})}
>
<summary className="collapse-title">
Extra content
</summary>
<div className="collapse-content">
{msg.extra.map(
(extra, i) =>
extra.type === 'textFile' ? (
<div key={extra.name}>
<b>{extra.name}</b>
<pre>{extra.content}</pre>
</div>
) : extra.type === 'context' ? (
<div key={i}>
<pre>{extra.content}</pre>
</div>
) : null // TODO: support other extra types
)}
</div>
</details>
)}
<MarkdownDisplay
content={content}
isGenerating={isPending}

27
examples/server/webui/src/components/ChatScreen.tsx
View File

@ -1,10 +1,11 @@
import { useEffect, useMemo, useState } from 'react';
import { useEffect, useMemo, useRef, useState } from 'react';
import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context';
import ChatMessage from './ChatMessage';
import { CanvasType, Message, PendingMessage } from '../utils/types';
import { classNames, throttle } from '../utils/misc';
import CanvasPyInterpreter from './CanvasPyInterpreter';
import StorageUtils from '../utils/storage';
import { useVSCodeContext } from '../utils/llama-vscode';
/**
* A message display is a message node with additional information for rendering.
@ -81,6 +82,14 @@ export default function ChatScreen() {
replaceMessageAndGenerate,
} = useAppContext();
const [inputMsg, setInputMsg] = useState('');
const inputRef = useRef<HTMLTextAreaElement>(null);
const { extraContext, clearExtraContext } = useVSCodeContext(
inputRef,
setInputMsg
);
// TODO: improve this when we have "upload file" feature
const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;
// keep track of leaf node for rendering
const [currNodeId, setCurrNodeId] = useState<number>(-1);
@ -115,10 +124,20 @@ export default function ChatScreen() {
setCurrNodeId(-1);
// get the last message node
const lastMsgNodeId = messages.at(-1)?.msg.id ?? null;
if (!(await sendMessage(currConvId, lastMsgNodeId, inputMsg, onChunk))) {
if (
!(await sendMessage(
currConvId,
lastMsgNodeId,
inputMsg,
currExtra,
onChunk
))
) {
// restore the input message if failed
setInputMsg(lastInpMsg);
}
// OK
clearExtraContext();
};
const handleEditMessage = async (msg: Message, content: string) => {
@ -129,6 +148,7 @@ export default function ChatScreen() {
viewingChat.conv.id,
msg.parent,
content,
msg.extra,
onChunk
);
setCurrNodeId(-1);
@ -143,6 +163,7 @@ export default function ChatScreen() {
viewingChat.conv.id,
msg.parent,
null,
msg.extra,
onChunk
);
setCurrNodeId(-1);
@ -203,9 +224,11 @@ export default function ChatScreen() {
<textarea
className="textarea textarea-bordered w-full"
placeholder="Type a message (Shift+Enter to add a new line)"
ref={inputRef}
value={inputMsg}
onChange={(e) => setInputMsg(e.target.value)}
onKeyDown={(e) => {
if (e.nativeEvent.isComposing || e.keyCode === 229) return;
if (e.key === 'Enter' && e.shiftKey) return;
if (e.key === 'Enter' && !e.shiftKey) {
e.preventDefault();

6
examples/server/webui/src/utils/app.context.tsx
View File

@ -25,6 +25,7 @@ interface AppContextValue {
convId: string | null,
leafNodeId: Message['id'] | null,
content: string,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
) => Promise<boolean>;
stopGenerating: (convId: string) => void;
@ -32,6 +33,7 @@ interface AppContextValue {
convId: string,
parentNodeId: Message['id'], // the parent node of the message to be replaced
content: string | null,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
) => Promise<void>;
@ -274,6 +276,7 @@ export const AppContextProvider = ({
convId: string | null,
leafNodeId: Message['id'] | null,
content: string,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
): Promise<boolean> => {
if (isGenerating(convId ?? '') || content.trim().length === 0) return false;
@ -298,6 +301,7 @@ export const AppContextProvider = ({
convId,
role: 'user',
content,
extra,
parent: leafNodeId,
children: [],
},
@ -324,6 +328,7 @@ export const AppContextProvider = ({
convId: string,
parentNodeId: Message['id'], // the parent node of the message to be replaced
content: string | null,
extra: Message['extra'],
onChunk: CallbackGeneratedChunk
) => {
if (isGenerating(convId)) return;
@ -339,6 +344,7 @@ export const AppContextProvider = ({
convId,
role: 'user',
content,
extra,
parent: parentNodeId,
children: [],
},

62
examples/server/webui/src/utils/llama-vscode.ts Normal file
View File

@ -0,0 +1,62 @@
import { useEffect, useState } from 'react';
import { MessageExtraContext } from './types';
// Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
// Ref: https://github.com/ggml-org/llama.cpp/pull/11940
interface SetTextEvData {
text: string;
context: string;
}
/**
* To test it:
* window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n return 123' }, '*');
*/
export const useVSCodeContext = (
inputRef: React.RefObject<HTMLTextAreaElement>,
setInputMsg: (text: string) => void
) => {
const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
null
);
// Accept setText message from a parent window and set inputMsg and extraContext
useEffect(() => {
const handleMessage = (event: MessageEvent) => {
if (event.data?.command === 'setText') {
const data: SetTextEvData = event.data;
setInputMsg(data?.text);
if (data?.context && data.context.length > 0) {
setExtraContext({
type: 'context',
content: data.context,
});
}
inputRef.current?.focus();
}
};
window.addEventListener('message', handleMessage);
return () => window.removeEventListener('message', handleMessage);
}, [inputRef, setInputMsg]);
// Add a keydown listener that sends the "escapePressed" message to the parent window
useEffect(() => {
const handleKeyDown = (event: KeyboardEvent) => {
if (event.key === 'Escape') {
window.parent.postMessage({ command: 'escapePressed' }, '*');
}
};
window.addEventListener('keydown', handleKeyDown);
return () => window.removeEventListener('keydown', handleKeyDown);
}, []);
return {
extraContext,
// call once the user message is sent, to clear the extra context
clearExtraContext: () => setExtraContext(null),
};
};

13
examples/server/webui/src/utils/misc.ts
View File

@ -53,12 +53,23 @@ export const copyStr = (textToCopy: string) => {
/**
* filter out redundant fields upon sending to API
* also format extra into text
*/
export function normalizeMsgsForAPI(messages: Readonly<Message[]>) {
return messages.map((msg) => {
let newContent = '';
for (const extra of msg.extra ?? []) {
if (extra.type === 'context') {
newContent += `${extra.content}\n\n`;
}
}
newContent += msg.content;
return {
role: msg.role,
content: msg.content,
content: newContent,
};
}) as APIMessage[];
}

14
examples/server/webui/src/utils/types.ts
View File

@ -42,11 +42,25 @@ export interface Message {
role: 'user' | 'assistant' | 'system';
content: string;
timings?: TimingReport;
extra?: MessageExtra[];
// node based system for branching
parent: Message['id'];
children: Message['id'][];
}
type MessageExtra = MessageExtraTextFile | MessageExtraContext; // TODO: will add more in the future
export interface MessageExtraTextFile {
type: 'textFile';
name: string;
content: string;
}
export interface MessageExtraContext {
type: 'context';
content: string;
}
export type APIMessage = Pick<Message, 'role' | 'content'>;
export interface Conversation {

4
examples/simple-cmake-pkg/README.md
View File

@ -1,6 +1,6 @@
# llama.cpp/example/simple-cmake-pkg
This program builds [simple](../simple) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
This program builds [simple](../simple) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggml-org/llama.cpp) in projects which live outside of the source tree.
## Building
@ -13,7 +13,7 @@ When hardware acceleration libraries are used (e.g. CUDA, Metal, Vulkan, etc.),
### Build llama.cpp and install to llama.cpp/inst
```sh
git clone https://github.com/ggerganov/llama.cpp
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -S . -B build
cmake --build build

6
examples/speculative/README.md
View File

@ -4,6 +4,6 @@ Demonstration of speculative decoding and tree-based speculative decoding techni
More info:
- https://github.com/ggerganov/llama.cpp/pull/2926
- https://github.com/ggerganov/llama.cpp/pull/3624
- https://github.com/ggerganov/llama.cpp/pull/5625
- https://github.com/ggml-org/llama.cpp/pull/2926
- https://github.com/ggml-org/llama.cpp/pull/3624
- https://github.com/ggml-org/llama.cpp/pull/5625

8
flake.nix
View File

@ -36,7 +36,7 @@
# ```
# nixConfig = {
# extra-substituters = [
# # Populated by the CI in ggerganov/llama.cpp
# # Populated by the CI in ggml-org/llama.cpp
# "https://llama-cpp.cachix.org"
#
# # A development cache for nixpkgs imported with `config.cudaSupport = true`.
@ -56,11 +56,11 @@
# };
# ```
# For inspection, use `nix flake show github:ggerganov/llama.cpp` or the nix repl:
# For inspection, use `nix flake show github:ggml-org/llama.cpp` or the nix repl:
#
# ```bash
# ❯ nix repl
# nix-repl> :lf github:ggerganov/llama.cpp
# nix-repl> :lf github:ggml-org/llama.cpp
# Added 13 variables.
# nix-repl> outputs.apps.x86_64-linux.quantize
# { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/llama-quantize"; type = "app"; }
@ -176,7 +176,7 @@
#
# We could test all outputs e.g. as `checks = confg.packages`.
#
# TODO: Build more once https://github.com/ggerganov/llama.cpp/issues/6346 has been addressed
# TODO: Build more once https://github.com/ggml-org/llama.cpp/issues/6346 has been addressed
checks = {
inherit (config.packages) default vulkan;
};

3
ggml/CMakeLists.txt
View File

@ -102,6 +102,7 @@ endif()
option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF)
option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
option(GGML_CPU_KLEIDIAI "ggml: use KleidiAI optimized kernels if applicable" OFF)
option(GGML_AVX "ggml: enable AVX" ${INS_ENB})
option(GGML_AVX_VNNI "ggml: enable AVX-VNNI" OFF)
option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB})
@ -121,6 +122,7 @@ endif()
option(GGML_LASX "ggml: enable lasx" ON)
option(GGML_LSX "ggml: enable lsx" ON)
option(GGML_RVV "ggml: enable rvv" ON)
option(GGML_VXE "ggml: enable vxe" ON)
option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
@ -150,6 +152,7 @@ set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
"ggml: max. batch size for using peer access")
option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF)
option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM" OFF)
option(GGML_CUDA_FA "ggml: compile ggml FlashAttention CUDA kernels" ON)
option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF)
option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT})

4
ggml/include/ggml-cpu.h
View File

@ -8,7 +8,7 @@ extern "C" {
#endif
// the compute plan that needs to be prepared for ggml_graph_compute()
// since https://github.com/ggerganov/ggml/issues/287
// since https://github.com/ggml-org/ggml/issues/287
struct ggml_cplan {
size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()`
uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
@ -95,9 +95,11 @@ extern "C" {
GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
GGML_BACKEND_API int ggml_cpu_has_sve (void);
GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes
GGML_BACKEND_API int ggml_cpu_has_sme (void);
// other
GGML_BACKEND_API int ggml_cpu_has_riscv_v (void);
GGML_BACKEND_API int ggml_cpu_has_vsx (void);
GGML_BACKEND_API int ggml_cpu_has_vxe (void);
GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void);
GGML_BACKEND_API int ggml_cpu_has_llamafile (void);

2
ggml/include/ggml-metal.h
View File

@ -45,7 +45,7 @@ GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_DEPRECATED(
GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
"obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
"obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);

123
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -111,14 +111,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
function(check_arm_feature tag code)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
check_cxx_source_runs(
"${code}"
GGML_MACHINE_SUPPORTS_${tag}
)
check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
if (GGML_MACHINE_SUPPORTS_${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
else()
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
if (GGML_MACHINE_SUPPORTS_no${tag})
set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
endif()
endif()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endfunction()
@ -126,6 +127,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
check_arm_feature(i8mm "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
check_arm_feature(sve "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
check_arm_feature(sme "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
else()
@ -150,7 +152,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
if (ARM_FEATURE_RESULT)
message(WARNING "Failed to get ARM features")
else()
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC)
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
if (NOT ${feature_pos} EQUAL -1)
message(STATUS "ARM feature ${feature} enabled")
@ -308,6 +310,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
if (GGML_RVV)
list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
message(STATUS "s390x detected")
file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
# TODO: Separation to determine activation of VX/VXE/VXE2
if (${S390X_M} MATCHES "8561|8562")
message(STATUS "z15 target")
list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
elseif (${S390X_M} MATCHES "3931")
message(STATUS "z16 target")
list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
else()
message(STATUS "Unknown target")
message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
list(APPEND ARCH_FLAGS -march=native -mtune=native)
endif()
if (GGML_VXE)
list(APPEND ARCH_FLAGS -mvx -mzvector)
endif()
else()
message(STATUS "Unknown architecture")
endif()
@ -316,6 +339,94 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
endif()
if (GGML_CPU_KLEIDIAI)
message(STATUS "Using KleidiAI optimized kernels if applicable")
# Disable the KleidiAI tests
set(KLEIDIAI_BUILD_TESTS OFF)
# Fetch KleidiAI sources:
include(FetchContent)
set(KLEIDIAI_COMMIT_TAG "v1.3.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
set(KLEIDIAI_ARCHIVE_MD5 "060bd2dc64642b091f461cc8dd7426d9")
if (POLICY CMP0135)
cmake_policy(SET CMP0135 NEW)
endif()
FetchContent_Declare(KleidiAI_Download
URL ${KLEIDIAI_DOWNLOAD_URL}
DOWNLOAD_EXTRACT_TIMESTAMP NEW
URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
FetchContent_MakeAvailable(KleidiAI_Download)
FetchContent_GetProperties(KleidiAI_Download
SOURCE_DIR KLEIDIAI_SRC
POPULATED KLEIDIAI_POPULATED)
if (NOT KLEIDIAI_POPULATED)
message(FATAL_ERROR "KleidiAI source downloaded failed.")
endif()
add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
# Remove kleidiai target after fetching it
if (TARGET kleidiai)
set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
endif()
list(APPEND GGML_CPU_SOURCES
ggml-cpu/kleidiai/kleidiai.cpp
ggml-cpu/kleidiai/kernels.cpp
ggml-cpu/kleidiai/kleidiai.h
ggml-cpu/kleidiai/kernels.h
)
# KleidiAI
include_directories(
${KLEIDIAI_SRC}/
${KLEIDIAI_SRC}/kai/
${KLEIDIAI_SRC}/kai/ukernels/
${KLEIDIAI_SRC}/kai/ukernels/matmul/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
if (NOT ARCH_FLAGS_TEMP)
string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}")
endif()
string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS})
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
if (NOT DOTPROD_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
endif()
if (NOT I8MM_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
endif()
if (NOT SME_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c)
list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c)
set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2")
endif()
set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
endif()
message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")
target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES})
target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})

151
ggml/src/ggml-cpu/ggml-cpu-impl.h
View File

@ -59,6 +59,15 @@ struct ggml_compute_params {
#endif
#endif
#if defined(__s390x__) && defined(__VEC__)
#ifndef __VXE__
#define __VXE__
#endif
#ifndef __VXE2__
#define __VXE2__
#endif
#endif
#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#include <sys/prctl.h>
@ -359,6 +368,148 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
#endif
#endif
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#define vec_neg(a) (-(a)) // Vector Negate
#define vec_add(a, b) ((a) + (b)) // Vector Add
#define vec_sub(a, b) ((a) - (b)) // Vector Subtract
#define vec_mul(a, b) ((a) * (b)) // Vector Multiply
#define vec_div(a, b) ((a) / (b)) // Vector Divide
#define vec_sl(a, b) ((a) << (b)) // Vector Shift Left
#define vec_sra(a, b) ((a) >> (b)) // Vector Shift Right
#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebraic
#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
#ifndef vec_and
#define vec_and(a, b) ((a) & (b)) // Vector AND
#endif
#ifndef vec_or
#define vec_or(a, b) ((a) | (b)) // Vector OR
#endif
#ifndef vec_xor
#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
#endif
typedef signed char char8x16_t __attribute__((vector_size(16)));
typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
typedef int8_t int8x16_t __attribute__((vector_size(16)));
typedef int16_t int16x8_t __attribute__((vector_size(16)));
typedef int32_t int32x4_t __attribute__((vector_size(16)));
typedef uint8_t uint8x16_t __attribute__((vector_size(16)));
typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
typedef float float32x4_t __attribute__((vector_size(16)));
typedef double double64x2_t __attribute((vector_size(16)));
typedef signed long long long64x2_t __attribute((vector_size(16)));
typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
res.val[2] = vec_xl(32, ptr);
res.val[3] = vec_xl(48, ptr);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
res.val[2] = vec_xl(32, ptr);
res.val[3] = vec_xl(48, ptr);
return res;
}
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
return res;
}
/*
! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
! or iq4_nl for example implementation.
*/
inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
const uchar8x16_t v_maske = { 0, 1, 4, 5, 8, 9, 12, 13,
16, 17, 20, 21, 24, 25, 28, 29 };
const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
const int16x8_t v_abe = vec_perm(a, b, v_maske);
return v_abo + v_abe;
}
inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
return acc + (vec_unpackh(p) + vec_unpackl(p));
}
#endif
#if defined(__loongarch_asx)
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(const float val) {

1094
ggml/src/ggml-cpu/ggml-cpu-quants.c
View File

File diff suppressed because it is too large Load Diff

128
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -112,7 +112,8 @@ struct ggml_arm_arch_features_type {
int has_i8mm;
int has_sve;
int sve_cnt;
} ggml_arm_arch_features = {-1, -1, -1, -1, 0};
int has_sme;
} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
#endif
@ -236,6 +237,8 @@ typedef pthread_t ggml_thread_t;
#else
#if defined(__POWER9_VECTOR__)
#define CACHE_LINE_SIZE 128
#elif defined(__VXE__) || defined(__VXE2__)
#define CACHE_LINE_SIZE 256
#else
#define CACHE_LINE_SIZE 64
#endif
@ -1210,6 +1213,87 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
#elif defined(__VXE__) || defined(__VXE2__)
#define GGML_SIMD
// F32 s390x
#define GGML_F32_STEP 32
#define GGML_F32_EPR 4
#define GGML_F32x4 __vector float
#define GGML_F32x4_ZERO vec_splats(0.0f)
#define GGML_F32x4_SET1 vec_splats
#define GGML_F32x4_LOAD(p) vec_xl(0, p)
#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p)
#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
#define GGML_F32x4_ADD vec_add
#define GGML_F32x4_MUL vec_mul
#define GGML_F32x4_REDUCE(res, x) \
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset + i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset + i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vec_add(x[i], x[offset + i]); \
} \
res = vec_extract(x[0], 0) + \
vec_extract(x[0], 1) + \
vec_extract(x[0], 2) + \
vec_extract(x[0], 3); \
}
#define GGML_F32_VEC GGML_F32x4
#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
#define GGML_F32_VEC_STORE GGML_F32x4_STORE
#define GGML_F32_VEC_FMA GGML_F32x4_FMA
#define GGML_F32_VEC_ADD GGML_F32x4_ADD
#define GGML_F32_VEC_MUL GGML_F32x4_MUL
#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
// F16 s390x
#define GGML_F16_STEP GGML_F32_STEP
#define GGML_F16_EPR GGML_F32_EPR
static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
float tmp[4];
for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]);
}
return vec_xl(0, tmp);
}
static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
float arr[4];
vec_xst(y, 0, arr);
for (int i = 0; i < 4; i++) {
x[i] = GGML_FP32_TO_FP16(arr[i]);
}
}
#define GGML_F16_VEC GGML_F32x4
#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO
#define GGML_F16_VEC_SET1 GGML_F32x4_SET1
#define GGML_F16_VEC_LOAD(p, i) __lzs_f16cx4_load(p)
#define GGML_F16_VEC_STORE(p, r, i) __lzs_f16cx4_store(p, r[i])
#define GGML_F16_VEC_FMA GGML_F32x4_FMA
#define GGML_F16_VEC_ADD GGML_F32x4_ADD
#define GGML_F16_VEC_MUL GGML_F32x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
#endif
// GGML_F32_ARR / GGML_F16_ARR
@ -1816,7 +1900,7 @@ inline static float ggml_silu_f32(float x) {
#if __FINITE_MATH_ONLY__
#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
#error "ref: https://github.com/ggerganov/llama.cpp/pull/7154#issuecomment-2143844461"
#error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461"
#endif
#if defined(__ARM_NEON) && defined(__aarch64__)
@ -2381,15 +2465,20 @@ bool ggml_is_numa(void) {
#define HWCAP2_I8MM (1 << 13)
#endif
#if !defined(HWCAP2_SME)
#define HWCAP2_SME (1 << 23)
#endif
static void ggml_init_arm_arch_features(void) {
#if defined(__linux__) && defined(__aarch64__)
uint32_t hwcap = getauxval(AT_HWCAP);
uint32_t hwcap2 = getauxval(AT_HWCAP2);
ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE);
ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE);
ggml_arm_arch_features.has_sme = !!(hwcap2 & HWCAP2_SME);
#if defined(__ARM_FEATURE_SVE)
ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
@ -2412,6 +2501,11 @@ static void ggml_init_arm_arch_features(void) {
}
ggml_arm_arch_features.has_i8mm = oldp;
if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
oldp = 0;
}
ggml_arm_arch_features.has_sme = oldp;
ggml_arm_arch_features.has_sve = 0;
ggml_arm_arch_features.sve_cnt = 0;
#else
@ -2435,6 +2529,12 @@ static void ggml_init_arm_arch_features(void) {
ggml_arm_arch_features.has_sve = 0;
ggml_arm_arch_features.sve_cnt = 0;
#endif
#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
ggml_arm_arch_features.has_sme = 1;
#else
ggml_arm_arch_features.has_sme = 0;
#endif
#endif
}
#endif
@ -7574,7 +7674,7 @@ UseGgmlGemm2:;
int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
// If the chunking is poor for the number of threads on this setup, scrap the whole plan. Re-chunk it by thread.
// Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggerganov/llama.cpp/pull/6915
// Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggml-org/llama.cpp/pull/6915
// In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that.
if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) {
// distribute the thread work across the inner or outer loop based on which one is larger
@ -14402,6 +14502,14 @@ int ggml_cpu_has_vsx(void) {
#endif
}
int ggml_cpu_has_vxe(void) {
#if defined(__VXE__) || defined(__VXE2__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_neon(void) {
#if defined(__ARM_ARCH) && defined(__ARM_NEON)
return ggml_arm_arch_features.has_neon;
@ -14442,6 +14550,14 @@ int ggml_cpu_get_sve_cnt(void) {
#endif
}
int ggml_cpu_has_sme(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
return ggml_arm_arch_features.has_sme;
#else
return 0;
#endif
}
void ggml_cpu_init(void) {
// needed to initialize f16 tables
{

19
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -14,6 +14,10 @@
#include "ggml-cpu-hbm.h"
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
#include "kleidiai/kleidiai.h"
#endif
#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
@ -39,6 +43,12 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
}
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
if (ggml_backend_cpu_kleidiai_buffer_type()) {
bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_AARCH64
if (ggml_backend_cpu_aarch64_buffer_type()) {
bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
@ -538,12 +548,18 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
}
if (ggml_cpu_has_sme()) {
features.push_back({ "SME", "1" });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
if (ggml_cpu_has_vsx()) {
features.push_back({ "VSX", "1" });
}
if (ggml_cpu_has_vxe()) {
features.push_back({ "VXE", "1" });
}
if (ggml_cpu_has_wasm_simd()) {
features.push_back({ "WASM_SIMD", "1" });
}
@ -559,6 +575,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
#ifdef GGML_USE_OPENMP
features.push_back({ "OPENMP", "1" });
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
features.push_back({ "KLEIDIAI", "1" });
#endif
#ifdef GGML_USE_CPU_AARCH64
features.push_back({ "AARCH64_REPACK", "1" });
#endif

259
ggml/src/ggml-cpu/kleidiai/kernels.cpp Normal file
View File

@ -0,0 +1,259 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
// KleidiAI micro-kernels
#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
#include "kai_common.h"
#include "kernels.h"
#define NELEMS(x) sizeof(x) / sizeof(*x)
static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
#if defined(__ARM_FEATURE_SME)
{
/* SME GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
},
/* SME GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
/* .require_aligned_m_idx = */ true,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
},
/* .required_cpu = */ CPU_FEATURE_SME,
},
#endif
#if defined(__APPLE__)
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
},
/* DOTPROD GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
/* .require_aligned_m_idx = */ false,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
},
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
},
/* i8mm GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
/* .require_aligned_m_idx = */ false,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
},
#endif
#else
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
},
/* i8mm GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
/* .require_aligned_m_idx = */ false,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
},
#endif
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
},
/* DOTPROD GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
/* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
/* .require_aligned_m_idx = */ false,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
},
#endif
#endif
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) {
ggml_kleidiai_kernels * kernels = nullptr;
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
kernels = &gemm_gemv_kernels[i];
break;
}
}
return kernels;
}

61
ggml/src/ggml-cpu/kleidiai/kernels.h Normal file
View File

@ -0,0 +1,61 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
enum cpu_feature {
CPU_FEATURE_NONE = 0,
CPU_FEATURE_DOTPROD = 1,
CPU_FEATURE_I8MM = 2,
CPU_FEATURE_SVE = 4,
CPU_FEATURE_SME = 8
};
inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
lhs = static_cast<cpu_feature>(lhs | rhs);
return lhs;
}
inline cpu_feature operator|(cpu_feature lhs, cpu_feature rhs) {
return static_cast<cpu_feature>(static_cast<int>(lhs) | static_cast<int>(rhs));
}
struct kernel_info {
size_t (*get_m_step)(void);
size_t (*get_n_step)(void);
size_t (*get_mr)(void);
size_t (*get_nr)(void);
size_t (*get_kr)(void);
size_t (*get_sr)(void);
size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl);
size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl);
size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
size_t (*get_dst_size)(size_t m, size_t n);
void (*run_kernel)(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max);
};
struct lhs_packing_info {
size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
size_t lhs_stride, void* lhs_packed);
bool require_aligned_m_idx;
};
struct rhs_packing_info {
size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
void (*pack_func)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params);
};
struct ggml_kleidiai_kernels {
kernel_info gemm;
kernel_info gemv;
lhs_packing_info lhs_info;
rhs_packing_info rhs_info;
cpu_feature required_cpu;
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features);

287
ggml/src/ggml-cpu/kleidiai/kleidiai.cpp Normal file
View File

@ -0,0 +1,287 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#include <arm_neon.h>
#include <assert.h>
#include <cfloat>
#include <stdint.h>
#include <string.h>
#if defined(__linux__)
#include <asm/hwcap.h>
#include <sys/auxv.h>
#elif defined(__APPLE__)
#include <string_view>
#include <sys/sysctl.h>
#include <sys/types.h>
#elif defined(_WIN32)
#include <windows.h>
#include <excpt.h>
#endif
#include "kleidiai.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-threading.h"
#include "ggml-cpu-traits.h"
#include "kernels.h"
#include "kai_common.h"
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
struct ggml_kleidiai_context {
ggml_kleidiai_kernels * kernels;
} static ctx = { NULL };
static void init_kleidiai_context(void) {
ggml_critical_section_start();
static bool initialized = false;
if (!initialized) {
initialized = true;
const char *env_var = getenv("GGML_KLEIDIAI_SME");
int sme_enabled = 0;
cpu_feature features = (ggml_cpu_has_dotprod() ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
(ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM : CPU_FEATURE_NONE) |
(ggml_cpu_has_sve() ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);
if (env_var) {
sme_enabled = atoi(env_var);
}
if (sme_enabled != 0) {
features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
}
ctx.kernels = ggml_kleidiai_select_kernels(features);
}
ggml_critical_section_end();
}
static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS);
return tensor->ne[dim];
}
namespace ggml::cpu::kleidiai {
class tensor_traits : public ggml::cpu::tensor_traits {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
GGML_ASSERT(ctx.kernels);
kernel_info * kernel = op->src[1]->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
size_t k = op->src[0]->ne[0];
size_t m = op->src[1]->ne[1];
size_t mr = kernel->get_mr();
size_t kr = kernel->get_kr();
size_t sr = kernel->get_sr();
size = ctx.kernels->lhs_info.packed_size(m, k, QK4_0, mr, kr, sr);
return true;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
if (dst->op == GGML_OP_MUL_MAT) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT(ctx.kernels);
kernel_info * kernel = src1->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
lhs_packing_info * lhs_info = &ctx.kernels->lhs_info;
GGML_ASSERT(kernel);
const int ith = params->ith;
const int nth = params->nth;
const size_t k = ne00;
const size_t m = ne11;
const size_t n = ne01;
const size_t n_step = kernel->get_n_step();
const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
const size_t n_start = ith * num_n_per_thread;
size_t n_to_process = num_n_per_thread;
if ((n_start + n_to_process) > n) {
n_to_process = n - n_start;
}
const uint8_t * lhs = static_cast<const uint8_t *>(src1->data);
uint8_t * lhs_packed = (uint8_t*)params->wdata;
const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
size_t mr = kernel->get_mr();
size_t kr = kernel->get_kr();
size_t sr = kernel->get_sr();
// Calculate number of columns to be processed per thread
const bool use_multithread = lhs_info->require_aligned_m_idx && m <= mr ? false : true;
const size_t num_m_per_thread = use_multithread ? kai_roundup(m, nth) / nth : m;
const size_t m_start = ith * num_m_per_thread;
size_t m_to_process = num_m_per_thread;
if ((m_start + m_to_process) > m) {
m_to_process = m - m_start;
}
if(m_start < m) {
// Transform LHS
const size_t src_stride = src1->nb[1];
const float * src_ptr = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(0, dst->src[1]->nb[1]));
const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr);
void * lhs_packed_ptr = static_cast<void *>(lhs_packed + lhs_packed_offset);
lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, m_start, src_ptr, src_stride, lhs_packed_ptr);
}
ggml_barrier(params->threadpool);
// Perform the operation
const size_t dst_stride = dst->nb[1];
const size_t lhs_packed_offset = lhs_info->get_packed_offset(0, k, QK4_0, mr, kr, sr);
const size_t rhs_packed_offset = kernel->get_rhs_packed_offset(n_start, k, QK4_0);
const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride);
const void * rhs_ptr = static_cast<const void *>(rhs_packed + rhs_packed_offset);
const void* lhs_ptr = (const void*)((const char *)lhs_packed + lhs_packed_offset);
float *dst_ptr = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
kernel->run_kernel(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr,
dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
return true;
}
return false;
}
public:
int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
GGML_ASSERT(ctx.kernels);
const size_t n = tensor->ne[1];
const size_t k = tensor->ne[0];
size_t nr = ctx.kernels->gemm.get_nr();
size_t kr = ctx.kernels->gemm.get_kr();
size_t sr = ctx.kernels->gemm.get_sr();
#ifndef NDEBUG
const size_t repacked_size = ctx.kernels->rhs_info.packed_size(n, k, nr, kr, QK4_0);
GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
#endif
struct kai_rhs_pack_qs4cxs1s0_param params;
params.lhs_zero_point = 1;
params.rhs_zero_point = 8;
ctx.kernels->rhs_info.pack_func(1, n, k, nr, kr, sr, QK4_0, (const uint8_t *)data, NULL, tensor->data, 0, &params);
return 0;
GGML_UNUSED(data_size);
}
};
static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
static tensor_traits traits;
return &traits;
}
} // namespace ggml::cpu::kleidiai
static void ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
auto tensor_traits = (ggml::cpu::kleidiai::tensor_traits *) tensor->extra;
auto OK = tensor_traits->repack(tensor, data, size);
GGML_ASSERT(OK == 0);
GGML_UNUSED(buffer);
}
static const char * ggml_backend_cpu_kleidiai_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_KLEIDIAI";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
if (buffer == nullptr) {
return nullptr;
}
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_kleidiai_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_kleidiai_buffer_set_tensor;
buffer->iface.get_tensor = nullptr;
buffer->iface.cpy_tensor = nullptr;
return buffer;
}
static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
namespace ggml::cpu::kleidiai {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
if ( op->op == GGML_OP_MUL_MAT &&
op->src[0]->type == GGML_TYPE_Q4_0 &&
op->src[0]->buffer &&
(ggml_n_dims(op->src[0]) == 2) &&
op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels
) {
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
if (op->src[1]->type == GGML_TYPE_F32 &&
ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) {
return true;
}
}
return false;
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT) {
if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
}
return nullptr;
}
};
} // namespace ggml::cpu::kleidiai
ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) {
static ggml::cpu::kleidiai::extra_buffer_type ctx;
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_kleidiai = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_kleidiai_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ &ctx,
};
init_kleidiai_context();
return &ggml_backend_cpu_buffer_type_kleidiai;
}

17
ggml/src/ggml-cpu/kleidiai/kleidiai.h Normal file
View File

@ -0,0 +1,17 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
#include "ggml-alloc.h"
#ifdef __cplusplus
extern "C" {
#endif
ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void);
#ifdef __cplusplus
}
#endif

17
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@ -280,14 +280,6 @@ template <> inline __m256bh load(const float *p) {
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// CONSTANTS
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// FLOATING POINT MATRIX MULTIPLICATION
@ -614,6 +606,14 @@ class tinyBLAS_Q0_AVX {
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
const int8_t kvalues_iq4nl[16] = {
-127, -104, -83, -65,
-49, -35, -22, -10,
1, 13, 25, 38,
53, 69, 89, 113
};
iq4nlt = _mm_loadu_si128((const __m128i *)kvalues_iq4nl);
}
void matmul(int64_t m, int64_t n) {
@ -1038,6 +1038,7 @@ class tinyBLAS_Q0_AVX {
const int64_t ldc;
const int ith;
const int nth;
__m128i iq4nlt;
};
#endif // __AVX__

10
ggml/src/ggml-cuda/CMakeLists.txt
View File

@ -7,7 +7,7 @@ if (CUDAToolkit_FOUND)
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# native == GPUs available at build time
# 52 == Maxwell, lowest CUDA 12 standard
# 50 == Maxwell, lowest CUDA 12 standard
# 60 == P100, FP16 CUDA intrinsics
# 61 == Pascal, __dp4a instruction (per-byte integer dot product)
# 70 == V100, FP16 tensor cores
@ -15,9 +15,9 @@ if (CUDAToolkit_FOUND)
if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
set(CMAKE_CUDA_ARCHITECTURES "native")
elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80")
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
set(CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80")
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@ -69,6 +69,10 @@ if (CUDAToolkit_FOUND)
add_compile_definitions(GGML_CUDA_NO_VMM)
endif()
if (NOT GGML_CUDA_FA)
add_compile_definitions(GGML_CUDA_NO_FA)
endif()
if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
add_compile_definitions(GGML_CUDA_F16)
endif()

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