源码编译构建

获取SGLang源码，从源码构建Python wheel和CUDA扩展wheel后安装构建产物。

获取SGLang 0.5.10源码。

git clone --branch v0.5.10 --depth 1 https://github.com/sgl-project/sglang.git

从源码构建Python wheel。

cd sglang
SETUPTOOLS_SCM_PRETEND_VERSION=0.5.10 \
python3 -m pip wheel --no-build-isolation --no-deps \
  --wheel-dir /tmp/sglang-build/python ./python

从源码构建CUDA扩展wheel。

export CUDA_HOME=/usr/local/cuda-13.0 
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:${LD_LIBRARY_PATH:-}
export TORCH_CUDA_ARCH_LIST="8.0"
export MAX_JOBS=8 
export NVCC_THREADS=1 
export CMAKE_BUILD_PARALLEL_LEVEL=8 
export CMAKE_ARGS="-DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_CUDA_ARCHITECTURES=80"
export Torch_DIR=$(python3 - <<'PY'
import pathlib 
import torch  

print(pathlib.Path(torch.__file__).resolve().parent / "share" / "cmake" / "Torch") 
PY 
) 
export CMAKE_PREFIX_PATH=$(python3 - <<'PY'
import torch.utils  

print(torch.utils.cmake_prefix_path) 
PY 
)  

python3 -m pip wheel --no-build-isolation --no-deps \
  --wheel-dir /tmp/sglang-build/kernel ./sgl-kernel

安装本地构建产物。

python3 -m pip install \
  typing_extensions filelock fsspec jinja2 networkx sympy mpmath numpy packaging \
  tqdm requests aiohttp fastapi uvicorn pydantic orjson pyzmq psutil pillow \
  sentencepiece safetensors tokenizers transformers==5.3.0 pybase64 setproctitle \
  partial_json_parser diskcache msgspec interegular lark pycountry pycryptodomex \
  gguf openai anthropic prometheus-client uvloop watchfiles soundfile tiktoken IPython
python3 -m pip install --force-reinstall --no-deps /tmp/sglang-build/kernel/*.whl
python3 -m pip install --force-reinstall --no-deps /tmp/sglang-build/python/*.whl

安装本地构建产物时使用--no-deps，用于保留前面已确认的CUDA版本PyTorch，避免依赖解析过程替换PyTorch版本。安装完成后，应再次确认torch.backends.cuda.is_built()和torch.cuda.is_available()均为True。

构建产物示例如下：

sglang-0.5.10-py3-none-any.whl 
sglang_kernel-0.4.1-*.whl

父主题： 开发指南