Fix Laguna KV ring commits for oversized prefills#614
Open
christophsturm wants to merge 1 commit into
Open
Conversation
Laguna prefill attention stages all KV rows before committing them to its 512-row sliding-window ring. When a chunk exceeds that capacity, the existing kernel dispatches every token and races multiple writers onto each ring row, so older rows can survive and corrupt subsequent decoding. Commit only the newest cache-capacity rows, cover a 1024-token prefill with a bit-exact Metal regression test, and bump the checkpoint payload ABI because earlier Laguna checkpoints may persist raced ring contents. Tested on an Apple M4 Max with the Metal backend and Poolside Laguna S 2.1 Q4_K_M: make clean && make; ./ds4_test --metal-kernels; and ./ds4_test --server. The regression test reported 30,912 key and 30,464 value mismatches on the old kernel and zero mismatches with this fix.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Laguna prefill attention stages all KV rows before committing them to its 512-row sliding-window ring. When a chunk exceeds that capacity, the existing kernel dispatches every token and races multiple writers onto each ring row, so older rows can survive and corrupt subsequent decoding.
Commit only the newest cache-capacity rows, cover a 1024-token prefill with a bit-exact Metal regression test, and bump the checkpoint payload ABI because earlier Laguna checkpoints may persist raced ring contents.
Tested on an Apple M4 Max with the Metal backend and Poolside Laguna S 2.1 Q4_K_M: make clean && make; ./ds4_test --metal-kernels; and ./ds4_test --server. The regression test reported 30,912 key and 30,464 value mismatches on the old kernel and zero mismatches with this fix.