Repository navigation
Use the actual global token count in distributed Sinkhorn - #9
Merged
Merged
Conversation
The Sinkhorn-Knopp denominator assumed every rank holds the same number of rows (local_rows * world_size). iBOT block masking gives ranks different masked-token counts, so the prototype targets were biased. Sum the real local row counts across ranks instead; equal-count inputs (e.g. CLS logits) are bit-identical. A rank may now hold zero rows; a globally empty batch raises. Adds single-process and two-rank gloo tests: equal-count parity with the historical implementation, unequal-count parity with a plain Sinkhorn over all tokens, and empty-rank handling.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The Sinkhorn-Knopp denominator assumed every rank holds the same number of rows (local_rows * world_size). iBOT block masking gives ranks different masked-token counts, so the prototype targets were biased. Sum the real local row counts across ranks instead; equal-count inputs (e.g. CLS logits) are bit-identical. A rank may now hold zero rows; a globally empty batch raises.
Adds single-process and two-rank gloo tests: equal-count parity with the historical implementation, unequal-count parity with a plain Sinkhorn over all tokens, and empty-rank handling.