[GRPO] Fix loss normalization #2881
Merged
+1
−1
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
qgallouedec
approved these changes
Feb 17, 2025
|
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. |
|
As seen on Twitter, some discussion about this change: https://x.com/danielhanchen/status/1900844864134410695 |
|
should we make it configurable? |
This was referenced Mar 23, 2025
yxliu-TAMU
pushed a commit
to mincheolseong/ECEN743-GRPO-Project-Proposal
that referenced
this pull request
Apr 20, 2025
* fix GRPO loss normalization * fix sum dim * fix loss= repeated
tyler-griggs
added a commit
to NovaSky-AI/SkyRL
that referenced
this pull request
Jul 15, 2025
## What does this PR do? Adds support for token-level loss (ie, `token_mean` loss reduction type) as introduced by DAPO. With `token_mean` loss reduction, all tokens in all sequences contribute equally to loss. The loss reduction type is configurable via `trainer.algorithm.loss_reduction`, but the default is updated to be `token_mean`, as opposed to our previous implementation (`sequence_mean`). This loss reduction is what the community is standardizing on as default (TRL's [default](huggingface/trl#2881), verl's [default](https://github.com/volcengine/verl/blob/517cc23c9dbb0da5c2cd2b012466790e29cb781a/verl/trainer/config/actor/actor.yaml#L63)) Wandb report of comparing `token_mean` vs `sequence_mean`: https://wandb.ai/sky-posttraining-uc-berkeley/gsm8k/reports/Token-level-loss-token_mean---VmlldzoxMzYwMDc4MQ The only plot with a notable difference is `policy_loss`, which is much larger for `token_mean` than it is for `sequence_mean`: <img width="312" height="274" alt="Screenshot 2025-07-15 at 9 52 57 AM" src="https://github.com/user-attachments/assets/40f94cb6-c5e5-47f6-9b09-a076811746a0" /> However, this `policy_loss` matches the same magnitude of `pg_loss` we observe in verl: <img width="980" height="611" alt="Screenshot 2025-07-15 at 9 54 39 AM" src="https://github.com/user-attachments/assets/53714573-2b21-4e67-b30a-dd3648279438" /> --------- Co-authored-by: Sumanth R Hegde <[email protected]>
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.
What does this PR do?
The current GRPO implementation uses
per-sequence normalization, this PR corrects this to beglobal normalizationDetails:
In Causal Language Modelling, we typically use
global normalizationto scale the loss, so that each unmasked token's loss provides the same contribution to the total loss. Example from transformers codebase: https://github.com/huggingface/transformers/blob/fae0f3dde83b7a54441f7a5bb0fc45d354fe81ce/src/transformers/loss/loss_utils.py#L24-L29