Updating to latest CuDNN without building pytorch from source

1

I want to update to CuDNN 8.9.6; the version included in a pip build of torch 2.1.1 is 8.9.2. Nowadays, CuDNN is installed as part of a torch pip install using a pip dependency - with torch 2.1.1, the dependency is nvidia-cudnn-cu12==8.9.2.26. There’s an updated version of this package with the version I want - nvidia-cudnn-cu12==8.9.6.50. Can I just install this version and expect that it’ll work properly? Or does the version need to exactly match the version that pytorch was built with?

I have tried updating it, and nothing seems broken, but I don’t know how to know if it’s actually using the new version or not.

2

Updating libs manually by overwriting the files is not supported and might break. If you want to update specific libs, source builds are the recommended approach.