fix(grpo): preserve prompt groups across rollout paths - #4013
Closed
ananthsub wants to merge 2 commits into
Closed
Conversation
Group harness rollouts by their source prompt rather than model-visible prompt tokens, and carry those identities through dynamic sampling, replay, the data plane, and checkpoint restore. Signed-off-by: Ananth Subramaniam <ansubramania@nvidia.com>
|
Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Signed-off-by: Ananth Subramaniam <ansubramania@nvidia.com>
Contributor
Author
|
closing in favor of #3882 |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
GRPO currently identifies a prompt group by tokenizing the prompt returned from the rollout harness. A harness can add rollout-specific content, such as a unique workspace path, so sibling generations become singleton groups and receive zero group-relative advantages.
{group_id}_g{index}sample IDs instead of harness-visible prompt tokens.PPO remains unchanged because its supported GAE and raw-reward estimators do not compute group-relative baselines from prompt IDs.
fyi @cmunley1