Eval self.left_padding whenever it is updated in BatchRotatingKVCache#960
Open
rltakashige wants to merge 1 commit intoml-explore:mainfrom
Open
Eval self.left_padding whenever it is updated in BatchRotatingKVCache#960rltakashige wants to merge 1 commit intoml-explore:mainfrom
rltakashige wants to merge 1 commit intoml-explore:mainfrom
Conversation
rltakashige
added a commit
to exo-explore/exo
that referenced
this pull request
Mar 7, 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation:
I was running into
RuntimeError: [metal::malloc] Resource limit (499000) exceeded.when using batching for GPT OSS. (see the attached log.txt). Upon investigation, this happened for any model with rotating KV cache.Steps to Reproduce:
Run my attached reproduce_batch_kvcache_leak.py with any model that uses sliding window attention with
python reproduce_batch_kvcache_leak.py --model <model path> --crash. This runs the model in a batch generator with two requests for 50000 steps together. I have been using GPT OSS 120B MXFP4 Q8 primarily.--add-eval adds an eval to the left padding, which prevents this from occurring.
Issue and Proposed changes
I think the issue is caused by the left padding never being evaluated, meaning buffers are accumulated in an unbounded fashion.
I am not sure whether you'd prefer moving the evals outside this function. However, it is only necessary to evaluate the left padding when it is updated (from testing).