diff --git a/Dockerfile b/Dockerfile
index cf5009c..97cfbf9 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -11,11 +11,11 @@ ENV USE_CUDA=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 \
-    USE_CUDA=1 USE_DISTRIBUTED=0 BUILD_TEST=0 TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0"
+    TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0"
 
 RUN apt-get update && apt-get install -y --no-install-recommends \
       python3 python3-dev python3-pip python3-distutils git cmake ninja-build \
@@ -30,7 +30,9 @@ RUN git clone --recursive -b v2.7.1 https://github.com/pytorch/pytorch.git
 
 WORKDIR /opt/pytorch
 ENV MAX_JOBS=${MAX_JOBS}
-RUN python3 setup.py bdist_wheel        # ≈50‒60 min 首编
+RUN echo "Building PyTorch with USE_DISTRIBUTED=$USE_DISTRIBUTED" && \
+    python3 setup.py bdist_wheel && \
+    python3 -c "from torch.distributed import Backend; print('✅ Build success. GLOO =', Backend.GLOO)"
 
 ###############################################################################
 # Stage 1 ─ builder-extras：用自编 Torch 装 TV / flashinfer / sglang，并收集轮子
@@ -105,7 +107,12 @@ COPY --from=builder-extras /usr/local/cuda/lib64/libcupti.so /usr/lib/x86_64-lin
 RUN ldconfig
 
 COPY --from=builder-extras /wheels /tmp/wheels
-RUN python3 -m pip install --no-cache-dir /tmp/wheels/* && rm -rf /tmp/wheels
+#RUN python3 -m pip install --no-cache-dir /tmp/wheels/* && rm -rf /tmp/wheels
+# ✅ 优先装你自编的 torch，避免被 PyPI 上的覆盖
+RUN python3 -m pip install --no-cache-dir /tmp/wheels/torch*.whl && \
+    python3 -m pip install --no-cache-dir /tmp/wheels/* && \
+    python3 -c "from torch.distributed import Backend; print('✅ Runtime torch distributed OK, GLOO =', Backend.GLOO)" && \
+    rm -rf /tmp/wheels
 
 # 安装运行时漏掉的依赖
 RUN python3 -m pip install --no-cache-dir pydantic orjson psutil pyzmq pynvml transformers==4.48.3 uvicorn fastapi IPython aiohttp setproctitle uvloop sentencepiece triton