diff --git a/Dockerfile b/Dockerfile
index b5c569c92..1d2fcf42b 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -25,6 +25,8 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
       libjpeg-dev libpng-dev ca-certificates && \
     python3 -m pip install --no-cache-dir --upgrade pip wheel setuptools sympy pyyaml typing-extensions numpy
 
+RUN python3 -m pip install --no-cache-dir numpy requests packaging build
+
 WORKDIR /opt
 RUN git clone --recursive -b v2.7.1 https://github.com/pytorch/pytorch.git
 
@@ -39,6 +41,7 @@ RUN echo "Building PyTorch with USE_DISTRIBUTED=$USE_DISTRIBUTED" && \
 ARG CUDA_VERSION=12.6.1
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS builder-extras
 
+ENV TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;8.9;9.0"
 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 \
@@ -74,8 +77,8 @@ WORKDIR /opt/flashinfer
 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 && \
+RUN python3 -m pip install --no-cache-dir numpy requests build && \
+    python3 -m flashinfer.aot && \
     python3 -m build --no-isolation --wheel && \
     ls -lh dist/