x

docker/Dockerfile

@@ -28,7 +28,7 @@ ENV DEBIAN_FRONTEND=noninteractive \
     FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE \
     MINERU_MODEL_SOURCE=huggingface \
     TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 \
-    HSA_ENABLE_SDMA=1 \
+    HSA_ENABLE_SDMA=1 \               # 启用系统 DMA 加速 GPU 显存传输
     VLLM_TARGET_DEVICE=rocm
 
 WORKDIR /opt
@@ -159,6 +159,13 @@ RUN set -ex && \
     cd /opt/vllm && \
     sed -i '109,121s/^/\/\/ /' csrc/mamba/mamba_ssm/selective_scan.h && \
     echo "vllm mamba operator+ patch applied." && \
+    # cmake 别名兜底：ROCm 7.2 可能缺少 hiprand/hipblas cmake target
+    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 \
@@ -177,13 +184,14 @@ RUN set -ex && \
     cd /opt/vllm && ${VENV}/bin/pip install --no-cache-dir -e . --no-build-isolation && \
     # 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)" && \
-    # 清理可能的 CUDA triton 残余
-    ${VENV}/bin/pip uninstall -y triton triton-rocm 2>/dev/null; \
+    # 先重装 ROCm PyTorch 覆盖可能的 CUDA 版，再清理 CUDA triton 元数据
+    # 顺序重要：pytorch-triton-rocm 和 triton 共享 triton/ 物理目录，必须先重装后卸载
     ${VENV}/bin/pip install --no-cache-dir --force-reinstall \
         torch==2.11.0+rocm7.2 torchvision pytorch-triton-rocm \
         --index-url ${TORCH_INDEX} && \
-    # 最终验证 vllm 平台检测
-    ${VENV}/bin/python -c "from vllm.platforms import current_platform; print('Platform:', type(current_platform).__name__); print('is_rocm:', current_platform.is_rocm()); print('device_type:', current_platform.device_type); assert current_platform.is_rocm(), 'vllm ROCm detection failed!'; print('vllm OK')" && \
+    ${VENV}/bin/pip uninstall -y triton triton-rocm 2>/dev/null; \
+    # vllm 平台模块导入验证（GPU 检测只能在运行时，容器构建时无 GPU 设备）
+    ${VENV}/bin/python -c "from vllm.platforms import current_platform; print('Platform module:', type(current_platform).__name__); print('vllm import OK')" && \
     # 清理构建目录（减小镜像体积，约 3-5GB）
     rm -rf /opt/vllm_build