Reimplement batched_rewards to fix compatibility, broken in #2698
#2876
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2876
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
@krammnic if this is ready for testing I could give it a shot :) Let me know. |
| tokenizer: Union[ModelTokenizer, HuggingFaceModelTokenizer], | ||
| completions: torch.Tensor, | ||
| answers: list[list[str]], | ||
| device: torch.device, |
There was a problem hiding this comment.
You will need to pass the device parameter of batched_rewards in https://github.com/pytorch/torchtune/blob/main/recipes/dev/grpo_full_finetune_distributed.py#L649
But even then an error occurs regarding how the answers per batch per roll out are handled.
[rank0]: Traceback (most recent call last):
[rank0]: File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 948, in <module>
[rank0]: sys.exit(recipe_main())
[rank0]: ^^^^^^^^^^^^^
[rank0]: File "/tmp/torchtune/torchtune/config/_parse.py", line 99, in wrapper
[rank0]: sys.exit(recipe_main(conf))
[rank0]: ^^^^^^^^^^^^^^^^^
[rank0]: File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 943, in recipe_main
[rank0]: recipe.train()
[rank0]: File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 812, in train
[rank0]: trajectory = self.generate_trajectory_batched(tokens, answers)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 703, in generate_trajectory_batched
[rank0]: self.generate_trajectory(batch_input_ids, batch_answers)
[rank0]: File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 649, in generate_trajectory
[rank0]: rewards, successes = batched_rewards(self._tokenizer, responses, answers, self._device)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/tmp/torchtune/torchtune/dev/rl/rewards.py", line 251, in batched_rewards
[rank0]: answers_list.append(answers[b][g])
[rank0]: ~~~~~~~~~~^^^
[rank0]: IndexError: string index out of range
|
Still need few updates and it will be ready |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Context
What is the purpose of this PR? Is it to
Please link to any issues this PR addresses.
Changelog
What are the changes made in this PR?
With introduction of async GRPO, the structure of reward modeling was changed leading to some breaks in common
dev/grpo_full_finetune_distributedwhich is used and not deprecated. The idea is to not just revert some old code, but to reimplement it to make the working recipe.Test plan
Please make sure to do each of the following if applicable to your PR. If you're unsure about any one of these just ask and we will happily help. We also have a contributing page for some guidance on contributing.
pre-commit install)pytest testspytest tests -m integration_testUX
If your function changed a public API, please add a dummy example of what the user experience will look like when calling it.
Here is a docstring example
and a tutorial example