.

Dockerfile

@@ -84,7 +84,7 @@ RUN python3 -m pip install ".[srt,openai]" --no-build-isolation && \
 
 
 # ── 🔄 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
+RUN pip download --only-binary=:all: --no-deps sgl-kernel==0.3.9.post2 -d /tmp/sgl_kernel_wheels
 
 # ── 收集所有 wheel 到 /wheels ──────────────────────────────────────────────
 RUN mkdir -p /wheels && \
@@ -109,11 +109,55 @@ RUN pip wheel "gradio==5.38.2" requests -w /wheels
 # 把运行时所需依赖也打包进入wheel ────────────────────────────────────────────────
 RUN pip wheel pybase64==1.3.2 -w /wheels
 
+
+# >>> ADD: 导出轮子的独立阶段
+FROM scratch AS wheelhouse
+COPY --from=builder-extras /wheels /
+# <<< END ADD
+
+
+# >>> ADD: 从宿主机目录 _wheelhouse/ 安装轮子的 runtime
+ARG CUDA_VERSION=12.6.1
+FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04 AS runtime-prebuilt
+
+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 \
+      libnccl2=2.22.3-1+cuda12.6 && \
+    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（与 autobuild 相同）
+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/
+RUN ldconfig
+
+# ★ 从宿主机构建上下文复制本地轮子（目录名固定：_wheelhouse/）
+COPY _wheelhouse/ /tmp/wheels/
+
+# 安装顺序与 runtime-autobuild 完全一致（优先 torch，再装其它）
+RUN ls -lh /tmp/wheels || true && \
+    rm -f /tmp/wheels/torch-2.7.1a0+*.whl && \
+    rm -f /tmp/wheels/huggingface_hub-0.34.4*.whl || true && \
+    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 || true && \
+    python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/sgl_kernel-*.whl || true && \
+    python3 -m pip install --no-cache-dir --no-deps /tmp/wheels/gradio-5.38.2*.whl || true && \
+    python3 -m pip install --no-cache-dir --no-deps $(find /tmp/wheels -maxdepth 1 -type f -name '*.whl' ! -name 'gradio-*' -printf "/tmp/wheels/%f ") && \
+    python3 -c "import gradio, sys; print('✅ Gradio version =', gradio.__version__)" && \
+    rm -rf /tmp/wheels
+
+
+
 ###############################################################################
 # Stage 2 ─ runtime：极简运行镜像，仅离线安装 wheel
 ###############################################################################
 ARG CUDA_VERSION=12.6.1
-FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04
+FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04 AS runtime-autobuild
 
 ENV DEBIAN_FRONTEND=noninteractive PYTHONUNBUFFERED=1 LANG=C.UTF-8 LC_ALL=C.UTF-8
 
@@ -142,7 +186,7 @@ RUN ls -lh /tmp/wheels && \
     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 $(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
@@ -185,3 +229,5 @@ 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"]
+
+