mirror of
https://github.com/archlinux/aur.git
synced 2026-03-14 23:16:48 +01:00
llama.cpp all in one!
This commit is contained in:
Orion
commit
3a4943ba09
2 changed files with 199 additions and 0 deletions
51
.SRCINFO
Normal file
51
.SRCINFO
Normal file
|
|
@ -0,0 +1,51 @@
|
|||
pkgbase = llama.cpp-aio
|
||||
pkgdesc = Port of Facebook's LLaMA model in C/C++ (All in one: CUDA + ROCm + Vulkan)
|
||||
pkgver = b8037
|
||||
pkgrel = 1
|
||||
url = https://github.com/ggml-org/llama.cpp
|
||||
arch = x86_64
|
||||
arch = armv7h
|
||||
arch = aarch64
|
||||
license = MIT
|
||||
makedepends = cmake
|
||||
makedepends = git
|
||||
makedepends = rocm-hip-sdk
|
||||
makedepends = shaderc
|
||||
makedepends = vulkan-headers
|
||||
depends = curl
|
||||
depends = gcc-libs
|
||||
depends = glibc
|
||||
depends = python
|
||||
depends = cuda
|
||||
depends = nvidia-utils
|
||||
depends = hip-runtime-amd
|
||||
depends = hipblas
|
||||
depends = openmp
|
||||
depends = rocblas
|
||||
depends = rocwmma
|
||||
depends = vulkan-icd-loader
|
||||
optdepends = python-numpy: needed for convert_hf_to_gguf.py
|
||||
optdepends = python-safetensors: needed for convert_hf_to_gguf.py
|
||||
optdepends = python-sentencepiece: needed for convert_hf_to_gguf.py
|
||||
optdepends = python-pytorch: needed for convert_hf_to_gguf.py
|
||||
optdepends = python-transformers: needed for convert_hf_to_gguf.py
|
||||
optdepends = python-gguf: needed for convert_hf_to_gguf.py
|
||||
provides = llama.cpp
|
||||
provides = llama.cpp-cuda
|
||||
provides = llama.cpp-hip
|
||||
provides = llama.cpp-vulkan
|
||||
conflicts = llama.cpp
|
||||
conflicts = libggml
|
||||
conflicts = ggml
|
||||
conflicts = stable-diffusion.cpp
|
||||
options = lto
|
||||
options = !debug
|
||||
backup = etc/conf.d/llama.cpp
|
||||
source = llama.cpp-aio-b8037.tar.gz::https://github.com/ggml-org/llama.cpp/archive/refs/tags/b8037.tar.gz
|
||||
source = https://raw.githubusercontent.com/Orion-zhen/aur-packages/refs/heads/main/assets/llama.cpp/llama.cpp.service
|
||||
source = https://raw.githubusercontent.com/Orion-zhen/aur-packages/refs/heads/main/assets/llama.cpp/llama.cpp.conf
|
||||
sha256sums = 84a8535ab935e03ae1e4812f6b7826828f8939d93c92cc53b66f277f4a4ede47
|
||||
sha256sums = 0377d08a07bda056785981d3352ccd2dbc0387c4836f91fb73e6b790d836620d
|
||||
sha256sums = e4856f186f69cd5dbfcc4edec9f6b6bd08e923bceedd8622eeae1a2595beb2ec
|
||||
|
||||
pkgname = llama.cpp-aio
|
||||
148
PKGBUILD
Normal file
148
PKGBUILD
Normal file
|
|
@ -0,0 +1,148 @@
|
|||
# # Maintainer: Orion-zhen <https://github.com/Orion-zhen>
|
||||
# Contributor: txtsd <aur.archlinux@ihavea.quest>
|
||||
|
||||
pkgname=llama.cpp-aio
|
||||
_pkgname=${pkgname%%-aio}
|
||||
pkgver=b8037
|
||||
pkgrel=1
|
||||
pkgdesc="Port of Facebook's LLaMA model in C/C++ (All in one: CUDA + ROCm + Vulkan)"
|
||||
arch=(x86_64 armv7h aarch64)
|
||||
url='https://github.com/ggml-org/llama.cpp'
|
||||
license=('MIT')
|
||||
depends=(
|
||||
curl
|
||||
gcc-libs
|
||||
glibc
|
||||
python
|
||||
cuda
|
||||
nvidia-utils
|
||||
hip-runtime-amd
|
||||
hipblas
|
||||
openmp
|
||||
rocblas
|
||||
rocwmma
|
||||
vulkan-icd-loader
|
||||
)
|
||||
makedepends=(
|
||||
cmake
|
||||
git
|
||||
rocm-hip-sdk
|
||||
shaderc
|
||||
vulkan-headers
|
||||
)
|
||||
optdepends=(
|
||||
'python-numpy: needed for convert_hf_to_gguf.py'
|
||||
'python-safetensors: needed for convert_hf_to_gguf.py'
|
||||
'python-sentencepiece: needed for convert_hf_to_gguf.py'
|
||||
'python-pytorch: needed for convert_hf_to_gguf.py'
|
||||
'python-transformers: needed for convert_hf_to_gguf.py'
|
||||
'python-gguf: needed for convert_hf_to_gguf.py'
|
||||
)
|
||||
provides=(${_pkgname} ${_pkgname}-cuda ${_pkgname}-hip ${_pkgname}-vulkan)
|
||||
conflicts=(${_pkgname} libggml ggml stable-diffusion.cpp)
|
||||
options=(lto !debug)
|
||||
backup=("etc/conf.d/llama.cpp")
|
||||
source=(
|
||||
"${pkgname}-${pkgver}.tar.gz::https://github.com/ggml-org/llama.cpp/archive/refs/tags/${pkgver}.tar.gz"
|
||||
"https://raw.githubusercontent.com/Orion-zhen/aur-packages/refs/heads/main/assets/llama.cpp/llama.cpp.service"
|
||||
"https://raw.githubusercontent.com/Orion-zhen/aur-packages/refs/heads/main/assets/llama.cpp/llama.cpp.conf"
|
||||
)
|
||||
sha256sums=('84a8535ab935e03ae1e4812f6b7826828f8939d93c92cc53b66f277f4a4ede47'
|
||||
'0377d08a07bda056785981d3352ccd2dbc0387c4836f91fb73e6b790d836620d'
|
||||
'e4856f186f69cd5dbfcc4edec9f6b6bd08e923bceedd8622eeae1a2595beb2ec')
|
||||
|
||||
prepare() {
|
||||
ln -sf "${_pkgname}-${pkgver}" llama.cpp
|
||||
}
|
||||
|
||||
build() {
|
||||
# 配置环境
|
||||
if [[ -z "${NVCC_CCBIN}" ]]; then
|
||||
source /etc/profile
|
||||
fi
|
||||
if [[ -z "${ROCM_PATH}" ]]; then
|
||||
source /etc/profile
|
||||
fi
|
||||
export HIP_PATH="$(hipconfig -R)"
|
||||
export HIPCXX="$(hipconfig -l)/clang"
|
||||
export HIP_PLATFORM=amd
|
||||
|
||||
local _cmake_options=(
|
||||
-B build
|
||||
-S "${_pkgname}"
|
||||
-DCMAKE_BUILD_TYPE=Release
|
||||
-DCMAKE_INSTALL_PREFIX='/usr'
|
||||
-DBUILD_SHARED_LIBS=ON
|
||||
-DLLAMA_BUILD_TESTS=OFF
|
||||
-DLLAMA_USE_SYSTEM_GGML=OFF
|
||||
-DGGML_ALL_WARNINGS=OFF
|
||||
-DGGML_ALL_WARNINGS_3RD_PARTY=OFF
|
||||
-DGGML_BUILD_EXAMPLES=OFF
|
||||
-DGGML_BUILD_TESTS=OFF
|
||||
-DGGML_LTO=ON
|
||||
-DGGML_RPC=ON
|
||||
# CUDA part
|
||||
-DGGML_CUDA=ON
|
||||
-DGGML_CUDA_GRAPHS=ON
|
||||
# ROCm part
|
||||
-DGGML_HIP=ON
|
||||
-DGGML_HIP_GRAPHS=ON
|
||||
-DGGML_HIP_ROCWMMA_FATTN=ON
|
||||
-DHIP_PLATFORM=amd # 手动指定 AMD 平台, 防止因 rocm-nightly 禁用自动检测而报错
|
||||
# Vulkan part
|
||||
-DGGML_VULKAN=ON
|
||||
|
||||
-DGGML_CUDA_FA_ALL_QUANTS=ON
|
||||
-DLLAMA_BUILD_NUMBER="${pkgver#b}" # 修正版本号
|
||||
-Wno-dev
|
||||
)
|
||||
|
||||
# 检查是否在 CI 环境中构建
|
||||
if [ -n "$CI" ] && [ "$CI" != 0 ]; then
|
||||
msg2 "CI = $CI detected, building universal package"
|
||||
# 启用通用构建
|
||||
_cmake_options+=(
|
||||
-DGGML_BACKEND_DL=ON
|
||||
-DGGML_CPU_ALL_VARIANTS=ON
|
||||
-DGGML_NATIVE=OFF
|
||||
|
||||
# https://developer.nvidia.com/cuda-gpus
|
||||
-DCMAKE_CUDA_ARCHITECTURES="75;86;89;120;121"
|
||||
|
||||
# https://llvm.org/docs/AMDGPUUsage.html
|
||||
# gfx906: MI 50/60, Radeon VII
|
||||
# gfx101x: RX 5000 Series
|
||||
# gfx103x: RX 6000 Series
|
||||
# gfx110x: RX 7000 Series
|
||||
# gfx1151: Strix Halo
|
||||
# gfx120x: RX 9000 Series
|
||||
-DAMDGPU_TARGETS="gfx906;gfx1010;gfx1030;gfx1031;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201"
|
||||
-DGGML_ZENDNN=ON
|
||||
)
|
||||
else
|
||||
# 本地构建, 针对当前设备优化
|
||||
_cmake_options+=(
|
||||
-DGGML_NATIVE=ON
|
||||
)
|
||||
fi
|
||||
|
||||
# 允许用户自定义构建选项
|
||||
if [[ -n "$LLAMA_BUILD_EXTRA_ARGS" ]]; then
|
||||
msg2 "Applied custom CMake build args: $LLAMA_BUILD_EXTRA_ARGS"
|
||||
_cmake_options+=($LLAMA_BUILD_EXTRA_ARGS)
|
||||
fi
|
||||
|
||||
cmake "${_cmake_options[@]}"
|
||||
cmake --build build -- -j $(nproc)
|
||||
}
|
||||
|
||||
package() {
|
||||
DESTDIR="${pkgdir}" cmake --install build
|
||||
|
||||
install -Dm644 "${_pkgname}/LICENSE" "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE"
|
||||
install -Dm644 "llama.cpp.conf" "${pkgdir}/etc/conf.d/llama.cpp"
|
||||
install -Dm644 "llama.cpp.service" "${pkgdir}/usr/lib/systemd/system/llama.cpp.service"
|
||||
|
||||
msg2 "llama.cpp.service is now available"
|
||||
msg2 "llama-server arguments are in /etc/conf.d/llama.cpp"
|
||||
}
|
||||
Loading…
Add table
Reference in a new issue