[SFT] Log mean token accuracy from Liger kernel

[kashif]

What does this PR do?

We log the mean token accuracy from Liger kernel if it is available

Requires linkedin/Liger-Kernel#910

to test do:

pip install git+https://github.com/kashif/Liger-Kernel.git@mean_token_accuracy
pip install git+https://github.com/huggingface/trl.git@iiger-mean-acc

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? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes?
• 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.

[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.

[qgallouedec]

I prefer using an explicit condition, like this:

if self.args.use_liger_kernel and version("liger_kernel") >= "0.6.4":
    token_accuracy = self.accelerator.gather_for_metrics(outputs.token_accuracy).mean().item()
    self._metrics[mode]["mean_token_accuracy"].append(token_accuracy)

Having the explicit condition makes it much easier to spot when the check can be removed after we bump the minimum liger_kernel version. It’s clearer and more maintainable than hiding the logic elsewhere.

[qgallouedec]

the question is then, which version will include this feature

[kashif]

the issue with this is that I cannot test it since the dev version is the current version in their setup

[qgallouedec]

ah true