Repository navigation
Remove Mega-Gate finite-score device trap - #11
Merged
ZJY0516 merged 1 commit intoSep 16, 2026
Merged
Conversation
Retain the selected expert ID clamp while allowing non-finite scores to propagate without the combined validation trap. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
vllm-project/vllm#56266
vLLM may pad tokens in a batch to the next CUDA graph capture size. These padded tokens may result in non-finite gate score, which trigger the device-side trap in Mega-Gate kernel.
This PR simply removes the check. A potential regression behavior is that non-padded tokens producing non-finite scores now won't be caught. But in practice, if such cases happen, we wouldn't want the server to crash anyway.