Fix BatchSamplerShard tail sampling - #9907
Open
taking-lying-flat wants to merge 1 commit into
Open
Conversation
taking-lying-flat
marked this pull request as ready for review
August 14, 2026 00:39
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
BatchSamplerSharddrop_last=Trueand repeat-pad fordrop_last=FalseRoot cause
BatchSamplerShardpreviously floored the dataset size by the DP world size in its constructor. The non-length-grouped paths then generated indices only for that shortened prefix, so tail samples were permanently excluded even whendrop_last=False. For packed datasets, the omitted item could be a full packed bin.Impact
All original samples now participate when
drop_last=False, while every DP rank still receives the same number of samples. Whendrop_last=True, indices are truncated after the epoch-specific global order is generated.