Fixing image segmentation with inference mode.

[Narsil]

What does this PR do?

It seems detr models are modifying tensors inplace, which is not allowed
by inference_mode context manager, effectively breaking image-segmentation for
torch > 1.9 users.

This PR proposes to override the context manager only in image-segmentation to no_grad
so the pipeline works again, without checking underlying model. Using a method to do the
switch.

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@mishig25 @LysandreJik

[patrickvonplaten]

Fine with this solution, but a better one might actually to disable all those:

        if torch.isinf(hidden_states.clone()).any() or torch.isnan(hidden_states.clone()).any():
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) 

for inference to make inference mode possible. Those "value" clamping checks are IMO only really useful to allow fp16 training (not so much for fp16 inference) and thus we could disable them with an if self.training statement.

@patil-suraj and I have somewhat unsuccesfully added lots of those statements to T5 in the hope of enabling fp16 training, which turned out to not work very well...with bfloat16 being available more for PyTorch those statements might not serve a good purpose anymore anyways.

=> So I would be fine with hiding them behind a self.training.

Pinging @stas00 here as well to hear his opinion :-)

[Narsil]

Agreed with @patrickvonplaten it would be better if we could keep inference_mode and work we the other fix.

[sgugger]

This fix is fine for a quick patch (patch release is going to be made in the next couple of hours), and we can make a better one afterward. So I would merge this now :-)

[LysandreJik]

Thank you @Narsil!