diff --git a/Dockerfile b/Dockerfile
index 2f1a17b00..2e6ef4896 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -49,8 +49,9 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
       cmake ninja-build libjpeg-dev libpng-dev ca-certificates \
       libopenmpi-dev libopenblas-dev\
       libnccl2=2.22.3-1+cuda12.6 \
-      libnccl-dev=2.22.3-1+cuda12.6 && \
-    python3 -m pip install --no-cache-dir --upgrade pip wheel setuptools
+      libnccl-dev=2.22.3-1+cuda12.6 \
+      curl xz-utils \
+    && python3 -m pip install --no-cache-dir --upgrade pip wheel setuptools
 
 # ── 安装自编 torch 轮子 ──────────────────────────────────────────────────────
 COPY --from=builder-torch /opt/pytorch/dist /tmp/torch_dist
@@ -97,18 +98,38 @@ 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==0.3.9.post2 -d /tmp/sgl_kernel_wheels
 
+ENV PATH=/usr/local/cuda/bin:${PATH}
+
 # ── 用你本地源码编 sgl-kernel==0.3.9.post2（与自编 torch 完全 ABI 对齐） ──────
 WORKDIR /sgl/sglang/sgl-kernel
+
+# 覆盖安装 ptxas 12.8（保留 nvcc 12.6），并打印版本确认
+RUN bash -lc '\
+  set -euo pipefail; \
+  NVCC_ARCHIVE_VERSION=12.8.93; \
+  T=cuda_nvcc-linux-x86_64-${NVCC_ARCHIVE_VERSION}-archive; \
+  curl -fsSLO https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/linux-x86_64/${T}.tar.xz && \
+  tar -xf ${T}.tar.xz && \
+  install -m 0755 ${T}/bin/ptxas /usr/local/cuda/bin/ptxas && \
+  /usr/local/cuda/bin/ptxas --version \
+'
+
+# 限制构建并行；避免 ptxas 多线程崩溃
+ENV CMAKE_BUILD_PARALLEL_LEVEL=8
+ENV SGL_KERNEL_COMPILE_THREADS=1
+
+# 构建 sgl-kernel（保持 FA3；去掉无效的关 90a 标志）
 RUN python3 -m pip install --no-cache-dir "cmake>=3.27,<4.0" scikit-build-core==0.11.6 pybind11[global] packaging && \
     bash -lc '\
       export CMAKE_PREFIX_PATH="$(python3 -c "import torch; print(torch.utils.cmake_prefix_path)")" && \
-      # 这些卡就够了：A100(80), 3090(86), 4090(89)。别碰 90a。
       export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9" && \
       export CUDAARCHS="80;86;89" && \
       export CMAKE_CUDA_ARCHITECTURES="$CUDAARCHS" && \
-      # 同时把它塞进 CMAKE_ARGS；不认识的开关会被忽略，不会报错
-      export CMAKE_ARGS="-DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_CUDA_ARCHITECTURES=$CUDAARCHS -DFA3_WITH_SM90A=OFF -DSGLK_DISABLE_SM90A=ON" && \
-      python3 -m pip wheel . --no-deps --no-build-isolation -w /tmp/sgl_kernel_wheels'
+      # 这里保留常规参数；如果项目支持，也把内核编译线程设为 1（未知项将被忽略，不会报错）
+      export CMAKE_ARGS="-DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_CUDA_ARCHITECTURES=$CUDAARCHS -Ccmake.define.SGL_KERNEL_COMPILE_THREADS=1" && \
+      python3 -m pip wheel . --no-deps --no-build-isolation -w /tmp/sgl_kernel_wheels \
+    '
+
 
 
 # ── 收集所有 wheel 到 /wheels ──────────────────────────────────────────────