vllm源码编译

参考：如何从源码编译安装vllm - 知乎

官方推荐使用 uv 创建 venv

虚拟环境中先安装 pyproject.toml 中 requires 依赖，安装 requirements 目录下 common.txt  cuda.txt 依赖

在 setup.py 文件中可以指定cmake每次使用固定目录，默认每次随机不同的临时目录

class cmake_build_ext(build_ext):
    def finalize_options(self):
        super().finalize_options()
        self.build_temp = os.path.join(os.getcwd(), "build")

提前下载依赖代码并切换版本

mkdir -p .deps

git clone https://github.com/NVIDIA/cutlass.git .deps/cutlass-src
git clone https://github.com/vllm-project/flash-attention.git .deps/vllm-flash-attn-src
git clone https://github.com/vllm-project/FlashMLA.git .deps/flashmla-src
git clone https://github.com/deepseek-ai/DeepGEMM.git .deps/deepgemm-src
git clone https://github.com/IST-DASLab/qutlass.git .deps/qutlass-src
git clone https://github.com/triton-lang/triton.git .deps/triton_kernels-src

grep -RInA8 -B4 "GIT_REPOSITORY" CMakeLists.txt cmake/external_projects

grep -RInA8 -B4 "GIT_REPOSITORY" CMakeLists.txt cmake/external_projects
git -C .deps/cutlass-src checkout <tag-or-commit>

安装：

cd /root/code/vllm

export VLLM_TARGET_DEVICE=cuda
export CMAKE_BUILD_TYPE=RelWithDebInfo
# export CMAKE_BUILD_TYPE=Debug # 调试版本
# export CMAKE_ARGS="-DCMAKE_CUDA_FLAGS_DEBUG='-G -lineinfo'" # 指定kernel调试

# 指定 blackwell 架构
export TORCH_CUDA_ARCH_LIST="12.0"

# 指定并发数，防止编译时oom
export MAX_JOBS=2
export NVCC_THREADS=2

export VERBOSE=1

uv pip install -e . --no-build-isolation -v