From 0992c81029db86e2dcc769acddadff3fea6e67bf Mon Sep 17 00:00:00 2001
From: hailin <hailin@gdzx.xyz>
Date: Thu, 18 Sep 2025 19:40:46 +0800
Subject: [PATCH] .

---
 Dockerfile | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/Dockerfile b/Dockerfile
index 17f6a5a36..59be356a0 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -41,7 +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 TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;8.9"
 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 \
@@ -73,7 +73,7 @@ WORKDIR /opt/flashinfer
 
 
 # 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"
+ENV FLASHINFER_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9"
 
 # 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 pip install --no-cache-dir numpy requests build "cuda-python>=12.0,<13" "nvidia-nvshmem-cu12" ninja pynvml && \
@@ -134,11 +134,16 @@ RUN pip wheel "gradio==5.38.2" requests -w /wheels
 RUN pip wheel pybase64==1.3.2 -w /wheels
 
 
+
+
 # 导出轮子的独立阶段
 FROM scratch AS wheelhouse
 COPY --from=builder-extras /wheels /
 
 
+
+
+
 # 从宿主机目录 _wheelhouse/ 安装轮子的 runtime
 ARG CUDA_VERSION=12.6.1
 FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04 AS runtime-prebuilt
@@ -179,6 +184,10 @@ RUN echo "/usr/local/cuda/extras/CUPTI/lib64" > /etc/ld.so.conf.d/cupti.conf &&
 # 保险起见，再加一行环境变量（有些基础镜像不把 extras 加入 ld.so.conf）：
 ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:${LD_LIBRARY_PATH}
 
+
+
+
+
 ###############################################################################
 # Stage 2 ─ runtime：极简运行镜像，仅离线安装 wheel
 ###############################################################################