mirror of
https://gitee.com/trusted-list/sglang.git
synced 2026-10-28 08:34:56 +08:00
62 lines
2.4 KiB
Docker
62 lines
2.4 KiB
Docker
ARG BASE_IMAGE=pytorch/manylinux2_28-builder
|
|
ARG CUDA_VERSION=13.4
|
|
|
|
FROM ${BASE_IMAGE}:cuda${CUDA_VERSION}
|
|
|
|
ARG ARCHITECTURE=x86_64
|
|
ARG CUDA_TAG=cu134
|
|
ARG CUDA_VERSION=13.4
|
|
ARG NCCL_VERSION=2.32.3
|
|
ARG PYTHON_TAG=cp312-cp312
|
|
ARG TORCH_CUDA_TAG=cu130
|
|
ARG TORCH_VERSION=2.14.1
|
|
|
|
ENV CUDA_HOME=/usr/local/cuda
|
|
ENV LD_LIBRARY_PATH=/usr/local/lib:/usr/local/lib64:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
|
|
ENV PATH=/opt/python/${PYTHON_TAG}/bin:${PATH}
|
|
ENV PYTHON_BIN=/opt/python/${PYTHON_TAG}/bin/python
|
|
|
|
# DeepEP 2.5 has no NVSHMEM or RDMA build dependency. DeepJIT needs C++20 with <format>
|
|
# (the builder images put gcc-toolset-13 first on PATH) and the elfutils headers; libdw
|
|
# itself is only dlopen'ed at runtime.
|
|
RUN yum install -y --nogpgcheck \
|
|
curl \
|
|
elfutils-devel \
|
|
git \
|
|
make \
|
|
patchelf \
|
|
pkgconfig \
|
|
wget \
|
|
&& yum clean all \
|
|
&& rm -rf /var/cache/yum
|
|
|
|
RUN set -eux; \
|
|
if [ "${ARCHITECTURE}" = aarch64 ]; then cuda_target=sbsa; else cuda_target="${ARCHITECTURE}"; fi; \
|
|
cuda_stub="/usr/local/cuda-${CUDA_VERSION}/targets/${cuda_target}-linux/lib/stubs/libcuda.so"; \
|
|
test -f "${cuda_stub}"; \
|
|
mkdir -p /usr/lib64 "/usr/lib/${ARCHITECTURE}-linux-gnu"; \
|
|
ln -sf "${cuda_stub}" /usr/lib64/libcuda.so; \
|
|
ln -sf "${cuda_stub}" "/usr/lib/${ARCHITECTURE}-linux-gnu/libcuda.so"
|
|
|
|
# PyTorch pins an older nvidia-nccl-cu13, so install the NCCL that DeepEP 2.5 needs over it.
|
|
RUN --mount=type=cache,id=sgl-deep-ep-v25-pip-${CUDA_TAG}-${PYTHON_TAG}-${ARCHITECTURE},target=/root/.cache/pip \
|
|
set -eux; \
|
|
"${PYTHON_BIN}" -m pip uninstall -y deep-ep sgl-deep-ep sgl-deep-ep-v25 || true; \
|
|
"${PYTHON_BIN}" -m pip install --upgrade pip; \
|
|
"${PYTHON_BIN}" -m pip install --force-reinstall \
|
|
"torch==${TORCH_VERSION}" \
|
|
--index-url "https://download.pytorch.org/whl/${TORCH_CUDA_TAG}"; \
|
|
"${PYTHON_BIN}" -m pip install --force-reinstall --no-deps \
|
|
"nvidia-nccl-cu13==${NCCL_VERSION}"; \
|
|
"${PYTHON_BIN}" -m pip install \
|
|
"auditwheel>=6.0" \
|
|
build \
|
|
ninja \
|
|
packaging \
|
|
setuptools \
|
|
wheel; \
|
|
TORCH_VERSION="${TORCH_VERSION}" "${PYTHON_BIN}" -c \
|
|
'import os, torch; assert torch.__version__.startswith(os.environ["TORCH_VERSION"]); print(torch.__version__, torch.version.cuda)'; \
|
|
g++ --version | head -1; \
|
|
nvcc --version | tail -1
|