RuntimeError: GET was unable to find an engine to execute this computation

CUDA版本是11.1，虚拟环境中安装的torch是2.0.0，首先调用torch出现问题ImportError: libcupti.so.11.7:cannot open shared object file: No such file or directory该问题的解决办法如下：

• 检查是否存在该文件

  locate libcupti.so.11.7
  

  

• 在该虚拟环境中重新安装torch，还是没生成nvidia这个文件。

• 从其他的文件中将文件拷贝到虚拟环境的对应包下。
  

解决上述问题后又出现问题RuntimeError: GET was unable to find an engine to execute this computation该问题解决方法：

• 检测torch和cuda是否能用

  import torch
  print(torch.__version__)
  print(torch.cuda.is_available())
  

  

• 测试

  import torch
  
  print(torch.cuda.is_available())
  num_gpu =1 
  # Decide which device to run on
  device = torch.device("cuda:0" if (torch.cuda.is_available() and num_gpu > 0) else "cpu")
  print(device)
  print(torch.cuda.get_device_name(0))
  print(torch.rand(3,3).cuda())
  

  

• 检查cudnn是否可用

  print(torch.backends.cudnn.version())
  

  

• 官网下载cuda
  CUDA官网下载
  

• cuda安装
  

• 下载解压cudnn
  Local Installers for Windows and Linux, Ubuntu(x86_64, armsbsa)
  
  

• 安装cudnn
  

• 设置环境

   export PATH=/opt/xxx/soft/cuda-11.7/bin:$PATH
   export LD_LIBRARY_PATH=LD_LIBRARY_PATH:/opt/xxx/soft/cuda-11.7/lib64
   export CUDA_HOME=/opt/xxx/soft/cuda-11.7
  
  

• 运行代码，问题解决

参考资料
RuntimeError: GET was unable to find an engine to execute this computation