Merged
Conversation
skhorasganiTT
approved these changes
Feb 27, 2026
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
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Purpose
Fixing hanged benchmark pipeline if multiple similar images are used
Test Plan
I see an issue with this condition in vllm 1.0
https://github.com/tenstorrent/vllm/blob/dev/vllm/v1/core/sched/ascend_scheduler.py#L261
Extra condition
len(encoder_inputs_to_schedule) == 0: When a VLM request's encoder output is already cached (from a previous request with the same image),_try_schedule_encoder_inputscorrectly returns an empty list (nothing needs recomputation) while keepingnum_new_tokens > 0. The base Scheduler correctly schedules the request using cached encoder outputs. The AscendScheduler incorrectly treats "nothing to compute" as "can't schedule" and breaks.breakinstead ofcontinue: The base Scheduler uses continue to skip to the next request. The AscendScheduler uses break to stop scheduling entirely, blocking ALL subsequent requests in the queue forever.Result: when a VLM request at the head of the waiting queue has a cached encoder output (same image processed earlier), the scheduler permanently deadlocks -- it can never schedule that request or any request behind it.
Fix
Correcting the condition to
if num_new_tokens == 0: