diff --git a/DeepSpeed/install_deepspeed_src_mamba_single_gpu.sh b/DeepSpeed/install_deepspeed_src_mamba_single_gpu.sh
index 8b61681..0558872 100644
--- a/DeepSpeed/install_deepspeed_src_mamba_single_gpu.sh
+++ b/DeepSpeed/install_deepspeed_src_mamba_single_gpu.sh
@@ -10,7 +10,7 @@ print("Torch CUDA tag:", torch.version.cuda)
 print("CUDA available:", torch.cuda.is_available())
 PY
 
-# 尝试用 nvidia 官方 11.8 频道；失败则走 conda-forge 逐组件
+# 安装 CUDA 11.8 工具链
 echo "==> Installing CUDA 11.8 toolchain into current env..."
 if mamba install -y -c "nvidia/label/cuda-11.8.0" cuda-toolkit; then
   echo "Installed cuda-toolkit (nvidia channel)."
@@ -24,23 +24,28 @@ else
     cuda-profiler-api=11.8
 fi
 
-# 强制当前会话使用 env 里的 11.8 nvcc/库
+# 强制用当前环境里的 CUDA 11.8
 export CUDA_HOME="$CONDA_PREFIX"
+export CUDA_PATH="$CUDA_HOME"
+export TORCH_CUDA_HOME="$CUDA_HOME"
+export CUDACXX="$CUDA_HOME/bin/nvcc"
 export PATH="$CUDA_HOME/bin:$PATH"
 export LD_LIBRARY_PATH="$CUDA_HOME/lib:$CUDA_HOME/lib64:${LD_LIBRARY_PATH:-}"
+hash -r  # 刷新 shell 命令缓存
 
 echo "==> nvcc should now be 11.8:"
+which nvcc
 nvcc --version
 
 # 架构：3090 = sm_86
 export TORCH_CUDA_ARCH_LIST="8.6"
 
-# 编译依赖（用 mamba 提速）
+# 编译依赖
 echo "==> Build deps"
 mamba install -y -c conda-forge cmake ninja pybind11 libaio git
 pip install -U pip setuptools wheel
 
-# 获取 DeepSpeed 源码（固定较稳 tag）
+# 获取 DeepSpeed 源码
 echo "==> Clone DeepSpeed (if not exists)"
 cd "${HOME}/train/new"
 [ -d DeepSpeed ] || git clone --branch v0.14.3 https://github.com/microsoft/DeepSpeed.git
@@ -48,13 +53,14 @@ cd DeepSpeed
 
 # 清理旧安装
 pip uninstall -y deepspeed || true
+rm -rf build
 
-# 仅启用训练相关内核
+# 启用训练相关内核
 export DS_BUILD_OPS=1
 export DS_BUILD_AIO=1
 export DS_BUILD_FUSED_ADAM=1
 export DS_BUILD_CPU_ADAM=1
-# 推理/transformer 内核先关，减少兼容风险
+# 推理/transformer 内核先关
 # export DS_BUILD_TRANSFORMER=1
 
 echo "==> Build & install DeepSpeed"