Need Suggestions for upgrading older version pytorch

Hi guys, my model can only work with pytorch <=1.1.0, unfortunately pytorch 1.1.0 can’t work with CUDA 11, which is needed in new GPU card like Nvidia A100 and 30xx series.

When I tried to use pytorch 1.7 + CUDA 11, my trained model will produce strange noise in output images. Model trained on pytorch 1.1.0 is OK, models trained on pytorch 1.1.0 can work in a forward loop in pytorch 1.7+CUDA11 environment.

My question is,

  1. If I want to upgrade to pytorch 1.7, is there any guideline how to upgrade from pytorch 1.1.0 to pytorch 1.7, since there are many upgraded functions between versions. I don’t know which function I should pay attention to.
  2. If I want to use pytorch 1.1.0 with CUDA 11 in new GPU card, is possible, and how to?

Thanks for the attention~

  1. You could skim through the release notes for each released PyTorch version and check which changes might affect your training. I would probably start with backwards compatibility breaking changes and see, if fixing these might help.

  2. You would have to search for all CUDA 11.x related PRs and try to backport them to the older PyTorch release, which might be cumbersome depending how old the PyTorch release is (and might not be trivial at all).

1 Like

For qestion 1, there might be a huge amount of work to do. I’ll think about it. Maybe some debugging skills can simplify this work? I totally agree with you to start with backwards part. :smiley:
For qestion 2, in fact I just need pytorch 1.1.0, is it possible to build from source for this version using CUDA 11?
Many thanks for the helpful reply.

I haven’t tried to build PyTorch 1.1.0 with any CUDA 11 release, but since 1.1.0 was released in April 2019, I’m pretty sure it won’t work without some backports.