Cudnn is used when pytorch installed from pip?

1

4D1E6621519DC40F0FD85FE766B719C6
4D1E6621519DC40F0FD85FE766B719C61308×504 93.2 KB

I download pytorch from this link and get it installed in my server. It works fine in a test file. However,
LD_LIBRARY_PATH contains no cudnn. I wondering pytorch installed in this way get no utilization
of cudnn lib, leading to the slow training in DNNs.

  1. If I add to cudnn path to LD_LIBRARY_PATH, does pytorch automatically utilizes it ?
  2. Can I print the information of cudnn that pytorch is using?

Thank you!

2

You can print the version with:

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

The wheels come with some pre-built libraties (CUDA, cuDNN, etc.).