Revert 45777#45942
Conversation
| # The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True. | ||
| # We set it to `False` for CI. See https://github.com/pytorch/pytorch/issues/157274#issuecomment-3090791615 | ||
| enable_tf32(False) | ||
| # # torch.backends.fp32_precision does not cascade to torch.backends.cudnn.conv.fp32_precision and torch.backends.cudnn.rnn.fp32_precision |
There was a problem hiding this comment.
Could we check the torch version for this block where the fix exists?
There was a problem hiding this comment.
@vasqu It is already taken care inside the enable_tf32 method that I added.
There was a problem hiding this comment.
Took care of it in new fix. :)
|
cc @ydshieh @khushali9, what version of Torch do we need in the CI before we can remove those lines again? |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
it's torch 2.12, not released yet I think |
|
@Rocketknight1 I checked it is not 2.12 but the next release 2.13 on June 8. I will keep a check and create my PR again to have it merge once Pytorch version is released. cc @ydshieh Thanks! |
|
Oh yes, sorry indeed 2.13. Thank you for handling it. You can use |
|
Sure @ydshieh thanks |
|
@ydshieh I checked pytorch 2.13 release is not out yet, I am keeping an eye, I will make this changes as soon as they have release out. |
#45777 removed some lines in the conftest which work around a torch issue, but the Torch fix hasn't propagated through to our CI version of Torch yet, so we need to put them back for now!
cc @vasqu