sglang_v0.5.2/Dockerfile

184 lines
8.6 KiB
Docker
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

###############################################################################
# Stage 0 ─ builder-torch编译 PyTorch 2.7.1 (+cu126)
###############################################################################
ARG CUDA_VERSION=12.6.1
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS builder-torch
ENV USE_CUDA=1 \
USE_DISTRIBUTED=1 \
USE_MPI=1 \
USE_GLOO=1 \
USE_NCCL=1 \
USE_SYSTEM_NCCL=1 \
BUILD_TEST=0
ARG MAX_JOBS=90
ENV DEBIAN_FRONTEND=noninteractive \
PYTHONUNBUFFERED=1 LANG=C.UTF-8 LC_ALL=C.UTF-8 \
TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;8.9;9.0"
RUN apt-get update && apt-get install -y --no-install-recommends \
python3 python3-dev python3-pip python3-distutils git cmake ninja-build \
libopenblas-dev libopenmpi-dev \
libnccl2=2.22.3-1+cuda12.6 \
libnccl-dev=2.22.3-1+cuda12.6 \
libjpeg-dev libpng-dev ca-certificates && \
python3 -m pip install --no-cache-dir --upgrade pip wheel setuptools sympy pyyaml typing-extensions numpy
WORKDIR /opt
RUN git clone --recursive -b v2.7.1 https://github.com/pytorch/pytorch.git
WORKDIR /opt/pytorch
ENV MAX_JOBS=${MAX_JOBS}
RUN echo "Building PyTorch with USE_DISTRIBUTED=$USE_DISTRIBUTED" && \
python3 setup.py bdist_wheel
###############################################################################
# Stage 1 ─ builder-extras用自编 Torch 装 TV / flashinfer / sglang并收集轮子
###############################################################################
ARG CUDA_VERSION=12.6.1
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS builder-extras
ENV DEBIAN_FRONTEND=noninteractive PYTHONUNBUFFERED=1 LANG=C.UTF-8 LC_ALL=C.UTF-8
RUN apt-get update && apt-get install -y --no-install-recommends \
python3 python3-pip python3-distutils python3.10-dev git build-essential \
cmake ninja-build libjpeg-dev libpng-dev ca-certificates \
libopenmpi-dev libopenblas-dev\
libnccl2=2.22.3-1+cuda12.6 \
libnccl-dev=2.22.3-1+cuda12.6 && \
python3 -m pip install --no-cache-dir --upgrade pip wheel setuptools
# ── 安装自编 torch 轮子 ──────────────────────────────────────────────────────
COPY --from=builder-torch /opt/pytorch/dist /tmp/torch_dist
RUN set -e && \
echo "==> Files in /tmp/torch_dist:" && ls -lh /tmp/torch_dist && \
find /tmp/torch_dist -name 'torch-*.whl' -print | xargs -r python3 -m pip install --no-cache-dir
# ── 编译 torchvision 0.22.1 (依赖本地 torch) ────────────────────────────────
WORKDIR /opt
RUN git clone -b v0.22.1 https://github.com/pytorch/vision.git
WORKDIR /opt/vision
RUN python3 setup.py bdist_wheel
# ── 编译 flashinfer (主分支支持 torch 2.7 / cu126) ─────────────────────────
WORKDIR /opt
RUN git clone https://github.com/flashinfer-ai/flashinfer.git
WORKDIR /opt/flashinfer
RUN pip install . && \
python3 -m pip wheel . --no-deps -w dist/
# ── 下载 vllm 预编译 wheel避免编译 flash-attn ───────────────────────────────
WORKDIR /opt
RUN pip download --only-binary=:all: --no-deps vllm==0.9.1 -d /tmp/vllm_wheels
# ── 编译你本地 sglang 源码并打 wheel ───────────────────────────────────────
COPY ./sglang /sgl/sglang
WORKDIR /sgl/sglang/python
RUN python3 -m pip install ".[srt,openai]" --no-build-isolation && \
python3 -m pip wheel ".[srt,openai]" --no-deps -w /tmp/sg_wheels
# ── 🔄 下载 sgl-kernel与 sglang 同步)───────────────────────────────────────
RUN pip download --only-binary=:all: --no-deps sgl-kernel==0.3.7 -d /tmp/sgl_kernel_wheels
# ── 收集所有 wheel 到 /wheels ──────────────────────────────────────────────
RUN mkdir -p /wheels && \
cp /tmp/torch_dist/torch*.whl /wheels/ && \
cp /opt/vision/dist/torchvision-*.whl /wheels/ && \
cp /opt/flashinfer/dist/flashinfer_python-*.whl /wheels/ && \
cp /tmp/vllm_wheels/vllm-*.whl /wheels/ && \
cp /tmp/sg_wheels/sglang-*.whl /wheels/ && \
cp /tmp/sgl_kernel_wheels/sgl_kernel-*.whl /wheels/ && \
pip wheel filelock typing-extensions sympy fsspec jinja2 networkx -w /wheels
# ── ✅ 再打包 runtime 阶段必需依赖 ────────────────────────────────────────────
RUN pip wheel \
pydantic orjson psutil pyzmq pynvml \
transformers==4.52.0 uvicorn fastapi IPython aiohttp \
setproctitle uvloop sentencepiece triton pillow cachetools msgspec blake3 cloudpickle compressed-tensors einops openai py-cpuinfo dill partial_json_parser python-multipart torchao \
-w /wheels
# ── ✅ 打包 gradio UI 所需依赖 ────────────────────────────────────────────────
RUN pip wheel "gradio==5.38.2" requests -w /wheels
###############################################################################
# Stage 2 ─ runtime极简运行镜像仅离线安装 wheel
###############################################################################
ARG CUDA_VERSION=12.6.1
FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04
ENV DEBIAN_FRONTEND=noninteractive PYTHONUNBUFFERED=1 LANG=C.UTF-8 LC_ALL=C.UTF-8
RUN apt-get update && apt-get install -y --no-install-recommends gcc g++ build-essential ninja-build cuda-compiler-12-6\
python3 python3-dev python3-pip python3-distutils curl ca-certificates \
libopenblas-dev libgomp1 libcupti-dev libnuma1 libopenmpi-dev openmpi-bin libnuma-dev libpng16-16 libjpeg8 && \
rm -rf /var/lib/apt/lists/* && \
python3 -m pip install --no-cache-dir --upgrade pip \
&& python3 -m pip install --no-cache-dir xgrammar
# 👉 拷贝 cupti 动态库(避免写死版本号)
COPY --from=builder-extras /usr/local/cuda/lib64/libcupti.so.12 /usr/lib/x86_64-linux-gnu/
COPY --from=builder-extras /usr/local/cuda/lib64/libcupti.so /usr/lib/x86_64-linux-gnu/
# 👇建议在后面补上
RUN ldconfig
COPY --from=builder-extras /wheels /tmp/wheels
# ✅ 优先装你自编的 torch避免被 PyPI 上的覆盖
RUN ls -lh /tmp/wheels && \
rm -f /tmp/wheels/torch-2.7.1a0+*.whl && \
rm -f /tmp/wheels/huggingface_hub-0.33.4*.whl && \
python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/torch*.whl && \
python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/vllm-*.whl && \
python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/sgl_kernel-*.whl && \
python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/gradio-5.38.2*.whl && \
python3 -m pip install --no-cache-dir --no-deps $(ls /tmp/wheels | grep -v '^gradio-' | sed 's|^|/tmp/wheels/|') && \
python3 -m pip install --no-cache-dir --no-deps $(find /tmp/wheels -maxdepth 1 -type f -name '*.whl' ! -name 'gradio-*') && \
python3 -c "import gradio, sys; print('✅ Gradio version =', gradio.__version__)" && \
rm -rf /tmp/wheels
# ✅ 安装 Prometheus client
RUN python3 -m pip install --no-cache-dir prometheus_client
# ✅ 设置多进程 metrics 收集目录(用于 MultiProcessCollector
ENV PROMETHEUS_MULTIPROC_DIR=/tmp/prometheus
# ✅ 确保目录存在
RUN mkdir -p /tmp/prometheus
# ---- 拷贝预调优的 MoE Triton kernel config ----------------------------
COPY moe_kernels /usr/local/lib/python3.10/dist-packages/sglang/srt/layers/moe/fused_moe_triton/configs
# ✅ 添加 Tini推荐
ENV TINI_VERSION=v0.19.0
ADD https://github.com/krallin/tini/releases/download/${TINI_VERSION}/tini /tini
RUN chmod +x /tini
ENTRYPOINT ["/tini", "--"]
# ---- 拷贝模型(路径可换) ----
# COPY ./Alibaba/Qwen3-30B-A3B /root/.cradle/Alibaba/Qwen3-30B-A3B
HEALTHCHECK --interval=30s --timeout=2s --start-period=300s --retries=5 CMD curl -fs http://localhost:30000/health || exit 1
# ---- 暴露端口 ----
EXPOSE 30000 30001
# 安装 supervisor
RUN apt-get update && apt-get install -y supervisor && \
mkdir -p /etc/supervisor/conf.d
# 拷贝 supervisord 配置文件和 UI 脚本
COPY ./meta_ui.py /app/meta_ui.py
COPY ./supervisord.conf /etc/supervisor/supervisord.conf
# 作为容器主进程运行 supervisor
CMD ["/usr/bin/supervisord", "-c", "/etc/supervisor/supervisord.conf"]