Revert "fix(rollout): reject incomplete group-estimator batches (#109)" - #111
Merged
Merged
Conversation
This reverts commit 34fd910. A per-response drop (empty_tokens / vlm_media_unresolved / vlm_truncation / logprob_misalign / no_action_tokens) marks ONE rollout as unusable; #109 made a group estimator discard the prompt's healthy siblings as well and backfill, wasting 7 rollouts of generation per incident on the long-context / multi-image episodes that are most expensive. The motivations do not hold on main: every group estimator already handles short groups (mean over the remaining rollouts, grpo std guard, rloo n-1), and with dynamic filtering off batch divisibility is handled by balance_experiences dropping the trailing remainder, so there is no NCCL hang on this path. Plain over-length rollouts are truncated and kept. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This reverts commit 34fd910.
A per-response drop (empty_tokens / vlm_media_unresolved / vlm_truncation / logprob_misalign / no_action_tokens) marks ONE rollout as unusable; #109 made a group estimator discard the prompt's healthy siblings as well and backfill, wasting 7 rollouts of generation per incident on the long-context / multi-image episodes that are most expensive. The motivations do not hold on main: every group estimator already handles short groups (mean over the remaining rollouts, grpo std guard, rloo n-1), and with dynamic filtering off batch divisibility is handled by balance_experiences dropping the trailing remainder, so there is no NCCL hang on this path. Plain over-length rollouts are truncated and kept.