fix(kernel): exact single-launch triton top-k/top-p sampling - #345
Open
jason-fxz wants to merge 1 commit into
Open
fix(kernel): exact single-launch triton top-k/top-p sampling#345jason-fxz wants to merge 1 commit into
jason-fxz wants to merge 1 commit into
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 Triton sampling fallback (used when flashinfer is not installed) estimated the top-k / top-p threshold from a fixed-size outlier buffer. Whenever one token holds more than ~40% of the probability the estimate collapses and top-k keeps the whole vocabulary; top-p was off by a few percent everywhere.
This replaces the estimator with an exact histogram search over the fp32 bit pattern and fuses each op into a single kernel launch.
Verified against a tie-aware torch reference (1.9M rows, 0 mismatches) and against flashinfer (identical kept sets).
Supersedes #329 — thanks @taking-lying-flat for catching this.