💔 Decouple loss computing and generation in GRPO #2762
Merged
+73
−54
Conversation
This file contains 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
qgallouedec
changed the title
|
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
commented
Feb 4, 2025
| attention_mask = torch.cat([prompt_mask_repeated, completion_mask], dim=1) # (B*G, P+C) | ||
|
|
||
| # Get the per-token log probabilities for the completions for the model and the reference model | ||
| def get_per_token_logps(model, input_ids, attention_mask, logits_to_keep): |
qgallouedec
commented
Feb 4, 2025
| ) | ||
|
|
||
| # Compute the KL divergence between the model and the reference model | ||
| per_token_kl = torch.exp(ref_per_token_logps - per_token_logps) - (ref_per_token_logps - per_token_logps) - 1 |
There was a problem hiding this comment.
this is later computed in compute_loss
qgallouedec
commented
Feb 4, 2025
Comment on lines
-524
to
-531
| # x - x.detach() allows for preserving gradients from x | ||
| per_token_loss = torch.exp(per_token_logps - per_token_logps.detach()) * advantages.unsqueeze(1) | ||
| per_token_loss = -(per_token_loss - self.beta * per_token_kl) | ||
| loss = ((per_token_loss * completion_mask).sum(dim=1) / completion_mask.sum(dim=1)).mean() | ||
|
|
||
| # Log the metrics | ||
| completion_length = self.accelerator.gather_for_metrics(completion_mask.sum(1)).float().mean().item() | ||
| self._metrics["completion_length"].append(completion_length) |
There was a problem hiding this comment.
this is later computed in compute_loss
qgallouedec
commented
Feb 4, 2025
| @@ -366,32 +366,41 @@ def _set_signature_columns_if_needed(self): | |||
| if self._signature_columns is None: | |||
| self._signature_columns = ["prompt"] | |||
|
|
|||
| # Trainer "prepares" the inputs before calling `compute_loss`. It converts to tensor and move to device. | |||
| # Since we preprocess the data in `compute_loss`, we need to override this method to skip this step. | |||
| def _prepare_inputs(self, inputs: dict[str, Union[torch.Tensor, Any]]) -> dict[str, Union[torch.Tensor, Any]]: | |||
There was a problem hiding this comment.
in this method, we now:
- generate
- compute reward
- compute ref log probs
edbeeching
approved these changes
Feb 4, 2025
5 tasks
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.
The motivation behind this PR is to decouple the whole part linked to the generation and calculation of rewards and ref log probs on the one hand, and to the calculation of loss on the other. It's a preparatory PR for the implementation of:
1: minibatching within a same group (reduce memory requirement)
2: the possibility of multiple optimization steps.