Scale loss before backward

[qgallouedec]

What does this PR do?

Fixes huggingface/trl#2456

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?

[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.

[muellerzr]

This one feels more like the right solution. Do the slow tests pass under this? (There is one in there for grad accum)

[muellerzr]

Yes because in reality this is exactly the right solution: since we are no longer relying on accelerate to div the loss, we need to do so before backward() if we don't know our num items in the batch

[SunMarc]

LGTM !

[qgallouedec]

Do the slow tests pass under this?

They pass locally. I get errors but I think they aren't related:

============================== short test summary info ==============================
FAILED tests/trainer/test_trainer.py::TrainerIntegrationTest::test_auto_batch_size_finder - AttributeError: 'DownloadConfig' object has no attribute 'use_auth_token'
FAILED tests/trainer/test_trainer.py::TrainerIntegrationTest::test_auto_batch_size_with_deepspeed - ModuleNotFoundError: No module named 'mpi4py'
FAILED tests/trainer/test_trainer.py::TrainerIntegrationTest::test_use_liger_kernel_patching - AssertionError: False is not true
FAILED tests/trainer/test_trainer_seq2seq.py::Seq2seqTrainerTester::test_finetune_bert2bert - TypeError: BertModel.forward() got an unexpected keyword argument 'num_items_in_batch'
======== 4 failed, 250 passed, 82 skipped, 89 warnings in 203.93s (0:03:23) =========

[qgallouedec]

@ArthurZucker @muellerzr @SunMarc I'll let you merge when you think it's good.