Trainer: Pass num_items_in_batch to compute_loss in prediction_step

[pramodith]

What does this PR do?

Ensures that num_items_in_batch is passed to the compute_loss function in the prediction_step to ensure that loss is calculated the same way both at train and eval time.

Fixes #41108

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.

@SunMarc

[pramodith]

I think the test case could be better open to hearing suggestions to add tests (for multi-gpu?) or modify the test.

[SunMarc]

LGTM ! Thanks for this nice PR !

[SunMarc]

Hmmm these tests are still failing: tests/trainer/test_trainer.py::TrainerIntegrationTest::test_evaluate_with_jit, can you quickly check why ? I guess the simplest solution would be to check if we have self.args.jit_mode_eval or not. Btw, we should probably deprecate this arg also

[HuggingFaceDocBuilderDev]

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.

[pramodith]

Hmmm these tests are still failing: tests/trainer/test_trainer.py::TrainerIntegrationTest::test_evaluate_with_jit, can you quickly check why ? I guess the simplest solution would be to check if we have self.args.jit_mode_eval or not. Btw, we should probably deprecate this arg also

Will take a look in a few hours once I'm off work.

[pramodith]

The issue was with RegressionPretrainedModel's signature not having num_items_in_batch as an argument, which leads to jit compile mode not knowing about this argument. I've fixed those test cases in the least intrusive way possible by setting trainer.model_accepts_loss_kwargs=False. A bit hacky, let me know if that's not okay and if we should pass an argument to get_regression_trainer instead.