.

Dockerfile

@@ -65,11 +65,24 @@ 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
+RUN git clone  --recursive -b v0.3.1 https://github.com/flashinfer-ai/flashinfer.git
 WORKDIR /opt/flashinfer
 
-RUN pip install . && \
-    python3 -m pip wheel . --no-deps -w dist/
+
+
+# 覆盖你的目标算力：3090=8.6，4090=8.9，H100=9.0a；可按需增/减
+ENV FLASHINFER_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0a"
+
+# 先做 AOT 预编译，再直接打 wheel（不隔离，使用同一份自编 torch）
+RUN python3 -m flashinfer.aot && \
+    python3 -m pip install --no-cache-dir build && \
+    python3 -m build --no-isolation --wheel && \
+    ls -lh dist/
+
+
+
+# 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
@@ -103,6 +116,10 @@ RUN pip wheel \
     setproctitle uvloop sentencepiece triton pillow cachetools msgspec blake3 cloudpickle compressed-tensors einops openai py-cpuinfo dill partial_json_parser python-multipart torchao \
     -w /wheels
 
+
+# 产出 openai-harmony 的离线 wheel
+RUN pip wheel --no-deps openai-harmony==0.0.4 -w /wheels    
+
 # ── ✅ 打包 gradio UI 所需依赖 ────────────────────────────────────────────────
 RUN pip wheel "gradio==5.38.2" requests -w /wheels
 
@@ -185,9 +202,9 @@ RUN ls -lh /tmp/wheels && \
     python3 -c "import gradio, sys; print('✅ Gradio version =', gradio.__version__)" && \
     rm -rf /tmp/wheels
 
-RUN PIP_NO_INDEX= PIP_FIND_LINKS= python3 -m pip install --no-cache-dir --no-deps \
-    openai-harmony==0.0.4 \
-    flashinfer-python==0.3.1
+# RUN PIP_NO_INDEX= PIP_FIND_LINKS= python3 -m pip install --no-cache-dir --no-deps \
+#     openai-harmony==0.0.4 \
+#     flashinfer-python==0.3.1
 
 # ✅ 安装 Prometheus client
 RUN python3 -m pip install --no-cache-dir prometheus_client