Setup loss_type in config at model init time

[ChanderG]

ensures no additional graph break introduced when torch.compile'ed

fixes #34615

What does this PR do?

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.

[ChanderG]

If this approach is fine - I will extend the PR to cover a wide set of models.

[qubvel]

Hi @ChanderG! Thanks for opening an issue and PR! Nice catch with the graph breaks!
Re the PR it is usually not the best decision to modify class attribute inplace, however this pattern is widely used across classification models losses

if self.num_labels == 1:
    self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
    self.config.problem_type = "single_label_classification"
else:
    self.config.problem_type = "multi_label_classification"

do you have other ideas on how to handle it without modifying the config attribute? Is there a way to fix re.findall instead of introducing the change for all models?

Alternatively we can have a model attribute, smth like

self.loss_type = "ForCausalLM" if self.config.loss_type is None else self.config.loss_type

but not feel strong re this :)

[ChanderG]

A more generic option may be to run the regex call at model init time, instead of hardcoding. I need to check, but it may still be efficient placed there (ie, won't cause a graph break from init).

[ArthurZucker]

Hey! We should rather do this in the call of super().__init__(config) as otherwise we have to monkey patch each and every model!

[ChanderG]

@ArthurZucker Updated to generic lookup. Tested with the repro here #34615 (comment) and no Graph Breaks with this latest version as well.

[ArthurZucker]

Waiting on this to put in the upcoming release !! 🤗

[ArthurZucker]

Okay this is failing tests because the loss_type is stored in the config, which is not reset in the tests.
It is actually be better to store the loss_type in the PreTrainedModel rather than in the config!

[ArthurZucker]

Pushed directly, hope it's not a problem 🤗 wanted to get this fix in fast!

[ArthurZucker]

Thanks for your patience! I thought this had been merged a while ago!