x

2026-06-05 13:35:14 +08:00
parent 270faac7b7
commit 9d0b4b4749
2 changed files with 31 additions and 100 deletions
@@ -1,15 +1,11 @@
 # =============================================================================
-# MinerU on ROCm 7.2.1 Docker Image
+# MinerU on ROCm 7.2.1 Docker Image（HK 云服务器 / 海外版）
 # 原生 Linux + Ubuntu 24.04 + ROCm 7.2.1 + PyTorch 2.11.0 + vllm + MinerU 3.2.0
-#
-# 国内网络优化版：Ubuntu/PyPI/GitHub 全部使用国内镜像
-# 构建前请根据你的 GPU 修改 ARCH 参数（默认 gfx1201 = RX 9070）
 # =============================================================================

 FROM ubuntu:24.04

 # -- 构建参数 ---------------------------------------------------------------
-# GPU 架构：gfx1201(RX 9070) gfx1200(RX 9060) gfx1100(RX 7900) gfx1101(RX 7800/7700) gfx1030(RX 6900/6800)
 ARG ARCH=gfx1201
 ARG PYTHON_VER=3.12
 ARG VENV=/opt/mineru_venv
@@ -18,8 +14,7 @@ ARG TORCH_INDEX=https://download.pytorch.org/whl/rocm7.2
 # -- GitHub 访问（国内无镜像，需代理）----------------------------------------
 ARG GIT_PROXY=http://127.0.0.1:8118

-# -- 国内镜像配置 -----------------------------------------------------------
-# PyPI 镜像
+# -- 国内镜像 ----------------------------------------------------------------
 ARG PIP_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple

 # -- 环境变量 ---------------------------------------------------------------
@@ -37,45 +32,37 @@ WORKDIR /opt
 # ===========================================================================
 # 阶段 1：换国内源 + 安装 ROCm 7.2.1
 # ===========================================================================
-# Ubuntu 24.04 使用 deb822 格式，默认源文件是 /etc/apt/sources.list.d/ubuntu.sources
 RUN sed -i 's|http://.*archive.ubuntu.com|http://mirrors.tuna.tsinghua.edu.cn|g' /etc/apt/sources.list.d/ubuntu.sources && \
    sed -i 's|http://.*security.ubuntu.com|http://mirrors.tuna.tsinghua.edu.cn|g' /etc/apt/sources.list.d/ubuntu.sources && \
    apt-get update && apt-get install -y --no-install-recommends \
    wget curl ca-certificates gnupg software-properties-common && \
-    # 添加 AMD ROCm 仓库（repo.radeon.com 通常国内可直连）
    wget -q https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
        gpg --dearmor | tee /etc/apt/trusted.gpg.d/rocm.gpg > /dev/null && \
    echo 'deb [arch=amd64] https://repo.radeon.com/rocm/apt/7.2.1 noble main' \
        > /etc/apt/sources.list.d/rocm.list && \
-    # apt pinning：AMD 仓库优先级高于 Ubuntu 自带（避免拿到旧版 rocminfo）
    printf 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600\n' \
        > /etc/apt/preferences.d/rocm-pin-600 && \
    apt-get update && \
-    # 安装 ROCm 基础组件（apt pinning 确保从 AMD 仓库拉）
    apt-get install -y --no-install-recommends \
        rocminfo rocm-device-libs hip-dev miopen-hip && \
-    # 清理
    apt-get clean && rm -rf /var/lib/apt/lists/*

 # ===========================================================================
-# 阶段 2：ROCm 头文件补丁（LLVM 22 兼容性修复）
-# 这些是 ROCm 7.2.1 在 24.04 上的已知问题，每次 apt 升级 ROCm 后需重新应用
+# 阶段 2：ROCm 头文件补丁
 # ===========================================================================
 RUN set -ex && \
-    # 补丁 1: hipcc/clang 符号链接（hipcc.pl 硬编码 clang-17，实际是 clang-22）
+    # 补丁 1: hipcc/clang 符号链接
    ln -sf /usr/bin/hipvars.pm /usr/share/perl5/hipvars.pm && \
-    # hipcc.pl 在 WSL2 存在，原生 Linux 不存在；不存在时保留 apt 安装的版本
    ([ -f /usr/bin/hipcc.pl ] && ln -sf /usr/bin/hipcc.pl /opt/rocm/bin/hipcc) || echo "hipcc.pl not found, keeping apt-installed hipcc" && \
    ln -sf /opt/rocm/llvm/bin/clang-22 /opt/rocm/llvm/bin/clang-17 && \
    ln -sf /opt/rocm/llvm/bin/clang++   /opt/rocm/llvm/bin/clang++-17 && \
    # 补丁 2: __hip_internal::conditional → std::conditional
    find /opt/rocm/include/hip -name "*.h" \
        -exec sed -i 's/__hip_internal::conditional/std::conditional/g' {} + && \
-    # 补丁 3: warpSize 常量（__AMDGCN_WAVEFRONT_SIZE 在 LLVM 22 未定义）
+    # 补丁 3: warpSize 常量
    find /opt/rocm/include/hip -name "amd_warp_functions.h" \
        -exec sed -i 's/static constexpr int warpSize = __AMDGCN_WAVEFRONT_SIZE;/constexpr int warpSize = 32;/g' {} + && \
    # 补丁 4: __activemask() → __builtin_amdgcn_read_exec()
-    # 注意：只改 amd_warp_sync_functions.h，不要动 amd_warp_functions.h（那是定义本身）
    sed -i 's/__activemask()/__builtin_amdgcn_read_exec()/g' \
        /opt/rocm/include/hip/amd_detail/amd_warp_sync_functions.h && \
    echo "ROCm 7.2.1 header patches applied."
@@ -87,16 +74,13 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential git ninja-build pkg-config \
    python${PYTHON_VER} python${PYTHON_VER}-venv python${PYTHON_VER}-dev \
    libnuma-dev libdrm2 libhwloc-dev libgl1 \
-    # vllm 运行时依赖
    libgomp1 libopenblas0 && \
    apt-get clean && rm -rf /var/lib/apt/lists/*

 # ===========================================================================
-# 阶段 4：CMake 4.0（vllm 要求 ≥ 4.0，Ubuntu 24.04 自带 3.28 不够）
+# 阶段 4：CMake 4.0
 # ===========================================================================
 RUN cd /tmp && \
-    # wget 通过代理下载 CMake（国内 GitHub 不通）
-    export http_proxy=${GIT_PROXY} https_proxy=${GIT_PROXY} && \
    wget -q https://github.com/Kitware/CMake/releases/download/v4.0.0/cmake-4.0.0-linux-x86_64.tar.gz && \
    tar -xzf cmake-4.0.0-linux-x86_64.tar.gz && \
    cp -r cmake-4.0.0-linux-x86_64/bin/* /usr/local/bin/ && \
@@ -108,22 +92,19 @@ RUN cd /tmp && \
 # 阶段 5：Python 虚拟环境 + PyTorch ROCm
 # ===========================================================================
 RUN python${PYTHON_VER} -m venv ${VENV} && \
-    # 配置 pip 国内镜像
    mkdir -p /root/.pip && \
    echo "[global]" > /root/.pip/pip.conf && \
    echo "index-url = ${PIP_INDEX}" >> /root/.pip/pip.conf && \
    ${VENV}/bin/pip install --no-cache-dir -U pip setuptools wheel && \
-    # 安装 PyTorch ROCm 版（指定 index-url 覆盖全局镜像）
    ${VENV}/bin/pip install --no-cache-dir --pre \
        torch==2.11.0+rocm7.2 \
        torchvision \
        pytorch-triton-rocm \
        --index-url ${TORCH_INDEX} && \
-    # 验证
-    ${VENV}/bin/python -c "import torch; print('PyTorch:', torch.__version__); print('ROCm:', torch.version.hip); assert torch.version.hip is not None"
+    ${VENV}/bin/python -c "import torch; print('PyTorch:', torch.__version__); assert torch.version.hip is not None"

 # ===========================================================================
-# 阶段 6：ROCm 开发包（vllm 编译必需）
+# 阶段 6：ROCm 开发包
 # ===========================================================================
 RUN apt-get update && apt-get install -y --no-install-recommends \
    hipblas-dev hiprand-dev hipsparse-dev hipsparselt-dev \
@@ -133,7 +114,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
    apt-get clean && rm -rf /var/lib/apt/lists/*

 # ===========================================================================
-# 阶段 6.5：安装 amdsmi（原生 Linux 必需，vllm 平台检测依赖）
+# 阶段 6.5：安装 amdsmi
 # ===========================================================================
 RUN if [ -d /opt/rocm/share/amd_smi ]; then \
        cp -r /opt/rocm/share/amd_smi /opt/amd_smi && \
@@ -144,35 +125,27 @@ RUN if [ -d /opt/rocm/share/amd_smi ]; then \
    fi

 # ===========================================================================
-# 阶段 7：amd-aiter + flash_attn（宿主机预克隆，COPY 进镜像）
+# 阶段 7：amd-aiter + flash_attn
 # ===========================================================================
-# 构建前请先在宿主机 docker/ 目录下克隆：
-#   git clone --depth 1 --recursive https://github.com/ROCm/aiter.git
-#   git clone --depth 1 https://github.com/Dao-AILab/flash-attention.git
-#   cd flash-attention && git checkout bba578d43974c1d3ba157ab597124dd0fe2ccdb4
-#   git submodule update --init --depth 1 && cd ..
-# ===========================================================================
-COPY aiter/ /opt/aiter/
-COPY flash-attention/ /opt/flash-attention/
 RUN set -ex && \
-    # aiter（AMD 优化的 attention 算子）
+    git config --global http.proxy ${GIT_PROXY} && \
+    git config --global https.proxy ${GIT_PROXY} && \
+    cd /opt && git clone --recursive --depth 1 https://github.com/ROCm/aiter.git && \
    ${VENV}/bin/pip install --no-cache-dir -e /opt/aiter && \
-    # flash_attn（Triton AMD 后端，锁定已验证的 commit）
+    cd /opt && git clone --recursive https://github.com/Dao-AILab/flash-attention.git && \
+    cd flash-attention && git checkout bba578d43974c1d3ba157ab597124dd0fe2ccdb4 && \
    ${VENV}/bin/pip install --no-cache-dir --no-build-isolation -e /opt/flash-attention && \
-    # 验证 PyTorch 没被覆盖
    ${VENV}/bin/python -c "import torch; v=torch.__version__; assert 'rocm' in v, f'PyTorch overwritten: {v}'; print('PyTorch OK:', v)"

 # ===========================================================================
-# 阶段 8a：编译 vllm（宿主机预克隆 + COPY，cmake + ninja，最耗时，单独一层缓存）
-# 构建前请在宿主机 docker/ 目录下克隆：
-#   git clone --depth 1 https://github.com/vllm-project/vllm.git
+# 阶段 8a：编译 vllm（cmake + ninja，最耗时，单独缓存）
 # ===========================================================================
-COPY vllm/ /opt/vllm/
 RUN set -ex && \
-    # setuptools 升级（PEP 639 兼容）
+    git config --global http.proxy ${GIT_PROXY} && \
+    git config --global https.proxy ${GIT_PROXY} && \
    ${VENV}/bin/pip install --no-cache-dir -U \
        "setuptools>=77.0.3" setuptools_scm setuptools_rust wheel && \
-    # 补丁 5：注释掉 vllm mamba 模块的 operator+ 定义（可能已被 vllm 移除）
+    cd /opt && git clone --depth 1 https://github.com/vllm-project/vllm.git && \
    cd /opt/vllm && \
    if [ -f csrc/mamba/mamba_ssm/selective_scan.h ]; then \
        sed -i '109,121s/^/\/\/ /' csrc/mamba/mamba_ssm/selective_scan.h && \
@@ -180,14 +153,12 @@ RUN set -ex && \
    else \
        echo "vllm mamba selective_scan.h not found (upstream removed), skipping patch."; \
    fi && \
-    # cmake 别名兜底
    mkdir -p /opt/rocm/lib/cmake/hiprand && \
    printf 'include(/opt/rocm/lib/cmake/rocrand/rocrand-config.cmake)\nif(TARGET roc::rocrand AND NOT TARGET hip::hiprand)\n  add_library(hip::hiprand ALIAS roc::rocrand)\nendif()\n' \
        > /opt/rocm/lib/cmake/hiprand/hiprand-config.cmake && \
    mkdir -p /opt/rocm/lib/cmake/hipblas && \
    printf 'include(/opt/rocm/lib/cmake/rocblas/rocblas-config.cmake)\nif(TARGET roc::rocblas AND NOT TARGET hip::hipblas)\n  add_library(hip::hipblas ALIAS roc::rocblas)\nendif()\n' \
        > /opt/rocm/lib/cmake/hipblas/hipblas-config.cmake && \
-    # cmake 配置
    mkdir -p /opt/vllm_build && \
    cmake -S /opt/vllm -B /opt/vllm_build -G Ninja \
        -DCMAKE_BUILD_TYPE=RelWithDebInfo \
@@ -200,36 +171,27 @@ RUN set -ex && \
        -DHIP_COMPILER=/opt/rocm/llvm/bin/clang++ \
        -DHIP_PATH=/opt/rocm \
        -DCMAKE_PREFIX_PATH="/opt/rocm;${VENV}/lib/python${PYTHON_VER}/site-packages/torch/share/cmake" && \
-    # ninja 编译（-j8，32GB 内存；若 < 16GB 请改为 -j2）
    cd /opt/vllm_build && ninja -j8 && \
-    # 安装 .so 到 vllm 源码目录
    cp /opt/vllm_build/*.abi3.so /opt/vllm/vllm/ && \
    echo "vllm C++ build complete (layer cached)."

 # ===========================================================================
-# 阶段 8b：安装 vllm + 验证 PyTorch（独立层，失败不影响 8a 缓存）
+# 阶段 8b：安装 vllm + 平台补丁 + 验证
 # ===========================================================================
 RUN set -ex && \
-    # 安装 vllm（让 pip 解析运行时依赖：xgrammar, regex, compressed_tensors 等）
    cd /opt/vllm && ${VENV}/bin/pip install --no-cache-dir -e . --no-build-isolation && \
-    # --no-build-isolation 可能漏装部分运行时依赖，显式补装
    ${VENV}/bin/pip install --no-cache-dir regex && \
-    # 应用 vllm 平台检测补丁（独立脚本，避免 Dockerfile 解析问题）
    ${VENV}/bin/python /opt/patch_vllm_platform.py && \
-    # 验证 PyTorch 没被 vllm 依赖覆盖
-    ${VENV}/bin/python -c "import torch; v=torch.__version__; assert 'rocm' in v, f'PyTorch overwritten by vllm deps: {v}'; print('PyTorch OK:', v)" && \
-    # 先重装 ROCm PyTorch 覆盖可能的 CUDA 版，再清理 CUDA triton 元数据
+    ${VENV}/bin/python -c "import torch; v=torch.__version__; assert 'rocm' in v, f'PyTorch overwritten: {v}'; print('PyTorch OK:', v)" && \
    ${VENV}/bin/pip install --no-cache-dir --force-reinstall \
        torch==2.11.0+rocm7.2 torchvision pytorch-triton-rocm \
        --index-url ${TORCH_INDEX} && \
    ${VENV}/bin/pip uninstall -y triton triton-rocm 2>/dev/null; \
-    # 确保 vllm 基本可导入（构建时无 GPU，仅验证 import）
    ${VENV}/bin/python -c "import vllm; print('vllm import OK:', vllm.__version__)" && \
-    # 清理构建目录（减小镜像体积，约 3-5GB）
    rm -rf /opt/vllm_build

 # ===========================================================================
-# 阶段 8.5：复制辅助脚本（必须在 MinerU 安装前就位）
+# 阶段 8.5：复制辅助脚本
 # ===========================================================================
 COPY scripts/apply_mineru_patches.py /opt/apply_mineru_patches.py
 COPY scripts/cache_warmer.py /opt/cache_warmer.py
@@ -239,20 +201,15 @@ COPY scripts/patch_vllm_platform.py /opt/patch_vllm_platform.py
 # 阶段 9：安装 MinerU + RDNA 适配补丁
 # ===========================================================================
 RUN set -ex && \
-    # pip 国内镜像已在阶段 5 全局配置
    ${VENV}/bin/pip install --no-cache-dir 'mineru[core]' && \
-    # 验证 PyTorch 没被覆盖
    ${VENV}/bin/python -c "import torch; v=torch.__version__; assert 'rocm' in v, f'PyTorch overwritten: {v}'; print('PyTorch OK:', v)" && \
-    # 应用 MinerU RDNA 适配补丁
    ${VENV}/bin/python /opt/apply_mineru_patches.py

 # ===========================================================================
 # 阶段 10：入口与最终验证
 # ===========================================================================
 RUN echo 'source /opt/mineru_venv/bin/activate' >> /etc/bash.bashrc && \
-    echo "MinerU Docker image built successfully." && \
-    ${VENV}/bin/python -c "import torch, vllm, mineru; print('='*50); print('MinerU ROCm Docker Image Ready'); print(f'  PyTorch : {torch.__version__}'); print(f'  ROCm    : {torch.version.hip}'); print(f'  vllm    : {vllm.__version__}'); print(f'  MinerU  : {mineru.__version__}'); print(f'  Arch    : ${ARCH}'); print('='*50)"
+    ${VENV}/bin/python -c "import torch, vllm, mineru; print('='*50); print('MinerU ROCm Docker Image Ready'); print(f'  PyTorch: {torch.__version__}'); print(f'  vllm:   {vllm.__version__}'); print(f'  MinerU:  {mineru.__version__}'); print('='*50)"

-# 容器入口：默认 bash，用户可 override
 ENTRYPOINT ["/bin/bash", "-c"]
 CMD ["bash"]
@@ -9,56 +9,30 @@ services:
    build:
      context: .
      dockerfile: Dockerfile
-      network: host                       # 容器共享宿主机网络（构建时需要）
+      network: host
      args:
-        # ---- 按你的 GPU 修改 ----
-        # gfx1201 = RX 9070 / 9070 XT / 9070 GRE
-        # gfx1200 = RX 9060 XT / 9060 XT LP
-        # gfx1100 = RX 7900 XTX / XT / GRE
-        # gfx1101 = RX 7800 XT / 7700 XT
-        # gfx1030 = RX 6950 / 6900 / 6800 系列
        ARCH: gfx1201
+        GIT_PROXY: http://127.0.0.1:8118
    container_name: mineru-rocm
    stdin_open: true
    tty: true
-    ipc: host                          # vllm 共享内存需要
-
-    # ---- GPU 设备透传（必需）----
+    ipc: host
    devices:
-      - /dev/kfd                       # ROCm KFD 内核驱动接口
-      - /dev/dri                       # GPU 渲染节点 (renderD*)
-
-    # ---- 安全配置（ROCm 需要）----
+      - /dev/kfd
+      - /dev/dri
    security_opt:
-      - seccomp=unconfined             # 允许 ROCm 系统调用
-
-    # ---- 用户权限（访问 GPU 设备）----
+      - seccomp=unconfined
    group_add:
-      - video                          # /dev/dri/render* 权限
-      - render                         # /dev/kfd 权限
-
-    # ---- 环境变量 ----
+      - video
+      - render
    environment:
      - MINERU_MODEL_SOURCE=${MINERU_MODEL_SOURCE:-huggingface}
      - HF_HUB_CACHE=${HF_HUB_CACHE:-/opt/models/huggingface}
      - MODELSCOPE_CACHE=${MODELSCOPE_CACHE:-/opt/models/modelscope}
-
-    # ---- 卷挂载 ----
    volumes:
-      # 输入/输出目录（按需修改）
      - ${INPUT_DIR:-./data/input}:/data/input:ro
      - ${OUTPUT_DIR:-./data/output}:/data/output
-      # 模型缓存（持久化，避免每次下载）
      - ${MODEL_DIR:-./data/models}:/opt/models
-      # MIOpen kernel 缓存（持久化，避免每次预热）
      - ${MIOPEN_CACHE:-./data/miopen}:/root/.cache/miopen
-
-    # ---- 启动命令（默认 bash，可改）----
    command: bash
-
-    # ---- 端口（WebUI / API 模式时启用）----
-    # ports:
-    #   - "7860:7860"   # WebUI
-    #   - "8000:8000"   # API
-
    restart: "no"