导入torch时libcublas.so.10版本缺失问题求助

本地安装Torch以运行最新ONNX Runtime测试时，触发以下导入错误：

ImportError: /usr/local/cuda-10.0/lib64/libcublas.so.10: version `libcublas.so.10' not found (required by /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)

已执行的排查步骤： #

1. 确认文件存在，执行命令：

ll /usr/local/cuda-10.0/lib64/libcublas.so.10

输出结果显示该软链接有效：

lrwxrwxrwx 1 root root 44 Dec 20 11:46 /usr/local/cuda-10.0/lib64/libcublas.so.10 -> /usr/local/cuda-10.0/lib64/libcublas.so.10.0

2. 检查LD_LIBRARY_PATH，确认/usr/local/cuda-10.0/lib64/已包含在路径中，尝试将其移至路径前端仍无效。

3. 使用ldd查看库关联，输出如下：

ldd /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so

输出：

/home/users/username/.local/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so: /usr/local/cuda-10.0/lib64/libcublas.so.10: version `libcublas.so.10' not found (required by /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so)
        linux-vdso.so.1 =>  (0x00007ffffb33e000)
        libc10_cuda.so => /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libc10_cuda.so (0x00007fedb38ae000)
        libcudart-80664282.so.10.2 => /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libcudart-80664282.so.10.2 (0x00007fedb362d000)
        libnvToolsExt-3965bdd0.so.1 => /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libnvToolsExt-3965bdd0.so.1 (0x00007fedb3423000)
        libpthread.so.0 => /usr/lib64/libpthread.so.0 (0x00007fedb3207000)
        libc10.so => /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libc10.so (0x00007fedb2f77000)
        libtorch_cpu.so => /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libtorch_cpu.so (0x00007fed9a500000)
        libm.so.6 => /usr/lib64/libm.so.6 (0x00007fed9a1fe000)
        libdl.so.2 => /usr/lib64/libdl.so.2 (0x00007fed99ffa000)
        libcublas.so.10 => /usr/local/cuda-10.0/lib64/libcublas.so.10 (0x00007fed95a64000)
        librt.so.1 => /usr/lib64/librt.so.1 (0x00007fed9585c000)
        libstdc++.so.6 => /usr/local/gcc-5.4.0/lib64/libstdc++.so.6 (0x00007fed954e1000)
        libgcc_s.so.1 => /usr/local/gcc-5.4.0/lib64/libgcc_s.so.1 (0x00007fed952ca000)
        libc.so.6 => /usr/lib64/libc.so.6 (0x00007fed94efc000)
        /lib64/ld-linux-x86-64.so.2 (0x00007fede3d1e000)
        libgomp-a34b3233.so.1 => /home/users/username/.local/lib/python3.8/site-packages/torch/lib/libgomp-a34b3233.so.1 (0x00007fed94cd2000)

• 匹配CUDA与Torch版本：从ldd输出可见，Torch依赖的libcudart是10.2版本，但系统安装的是CUDA 10.0，版本不兼容导致库版本校验失败。解决思路二选一：

  • 升级系统CUDA到10.2版本，确保与Torch依赖一致；
  • 卸载当前Torch，安装适配CUDA 10.0的版本，示例命令：

    pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html
    

• 临时强制加载指定库：若不想修改版本，可在启动脚本前用LD_PRELOAD指定加载具体的库文件：

  LD_PRELOAD=/usr/local/cuda-10.0/lib64/libcublas.so.10.0 python your_test_script.py
  

• 检查并更新CUDA库小版本：部分Torch版本要求libcublas.so.10对应的实际文件是更高小版本（如10.0.130），可查看CUDA 10.0目录下是否有该版本文件，若有，重新创建软链接：

  rm /usr/local/cuda-10.0/lib64/libcublas.so.10
  ln -s /usr/local/cuda-10.0/lib64/libcublas.so.10.0.130 /usr/local/cuda-10.0/lib64/libcublas.so.10
  

内容的提问来源于stack exchange，提问作者Johnzy