Add CISPO loss option and documentation #4298
Open
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
Add CISPO (truncated importance-sampling REINFORCE) as a new loss_type in GRPOConfig and document the objective, usage, and citation. Implement docs describing the CISPO formulation, the cispo_clip_max parameter (default 5.0), and how to enable it via loss_type="cispo". Add related logging metrics (cispo/importance_ratio/{mean,truncated_mean,max,max_truncated} and cispo/clip_fraction) so importance ratios and truncation behavior can be monitored. Update GRPOConfig help text to mention CISPO and explain its behavior. This enables using the ScaleRL-inspired variant and improves observability during training.
pramodith
requested changes
Oct 19, 2025
| }, | ||
| ) | ||
|
|
||
| cispo_clip_max: float = field( |
There was a problem hiding this comment.
I'd recommend re-using the epsilon-high config instead of creating a new one here.
| if self.loss_type == "cispo": | ||
| cispo_cap = torch.full_like(coef_1, self.cispo_clip_max) | ||
| cispo_truncated_weights = torch.minimum(coef_1, cispo_cap) | ||
| cispo_clipped_mask = coef_1 > cispo_cap |
There was a problem hiding this comment.
Can be simplified by:
cispo_weights = torch.where(coef_1 < self.epsilon_high, coef_1, self.epsilon_high).detach()
|
@gustavorubim just wanted to check if you've had the chance to make the requested changes. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Introduce the CISPO loss type in GRPOConfig, enhancing the training process with truncated importance-sampling REINFORCE. Update documentation to explain the objective, usage, and related parameters, while adding logging metrics for better observability during training. This implementation supports the ScaleRL-inspired variant and improves the overall training experience.