👮 Deprecate policy in favor of model in PPOTrainer

[qgallouedec]

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? 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.
• 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.

[qgallouedec]

model replaced by policy_and_value
policy replaced by model

[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.

[lewtun]

If I'm not mistaken, won't this breaking change be introduced in the next TRL release (i.e. v0.13.0 instead of 0.150)? In other words, don't we technically need to support both sets of args until v0.15?

[qgallouedec]

The decorator allows for supporting both args, and when the old one is used, raises a warning:

>>> from datasets import load_dataset
>>> from transformers import AutoModelForCausalLM, AutoModelForSequenceClassification, AutoTokenizer
>>> from trl import PPOConfig, PPOTrainer
>>> args = PPOConfig("my_out_dir")
>>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-1b-deduped")
>>> model = AutoModelForCausalLM.from_pretrained("EleutherAI/pythia-1b-deduped")
>>> ref_model = AutoModelForCausalLM.from_pretrained("EleutherAI/pythia-1b-deduped")
>>> reward_model = AutoModelForSequenceClassification.from_pretrained("EleutherAI/pythia-1b-deduped", num_labels=1)
>>> value_model = AutoModelForSequenceClassification.from_pretrained("EleutherAI/pythia-1b-deduped", num_labels=1)
>>> dataset = load_dataset("trl-internal-testing/descriptiveness-sentiment-trl-style", split="descriptiveness")
>>> trainer = PPOTrainer(
...      config=args,
...      processing_class=tokenizer,
...      policy=model,  # old arg here
...      ref_model=ref_model,
...      reward_model=reward_model,
...      value_model=value_model,
...      train_dataset=dataset,
...  )
/fsx/qgallouedec/transformers/src/transformers/utils/deprecation.py:165: FutureWarning: `policy` is deprecated and removed starting from version 0.15.0 for `PPOTrainer.__init__`. Use `model` instead.
  return func(*args, **kwargs)

Did I misunderstand your point?

[lewtun]

Ah I was mistaken indeed. This looks great then!

[lewtun]

Overall LGTM, but I have a concern this will break the trainer in the next release instead of the desired v0.15. I realise this trainer isn't used much, so it's probably OK to break across revisions in the interest of reaching stability long-term