Repository navigation
[CI][ROCm][Bugfix] DeepSeek V4.1: don't pass expert intermediate padding to AITER - #61960
Merged
AndreasKaratzas merged 5 commits intoOct 17, 2026
Merged
Conversation
…to AITER AITER 0.1.25's FlyDSL MoE (ROCm/aiter#6125 configs) leaves the padded intermediate columns unwritten when told about the padding, so stage 2 reads stale memory and small decode batches (M <= 4 tokens) can come out NaN. This hits DeepSeek V4.1 at TP4, where the per-rank expert intermediate size is padded 576 -> 640. Pass intermediate_pad=0 for V4.1's routed experts so AITER computes over the zero padding instead of skipping it, which is exact. DeepSeek V4 never pads the expert intermediate size and is unchanged. Revert once vLLM pins an AITER with the stage-1 grid fix. Signed-off-by: Micah Williamson <micah.williamson@amd.com> Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Micah Williamson <micah.williamson@amd.com> Co-authored-by: Cursor <cursoragent@cursor.com>
micah-wil
requested review from
dllehr-amd,
hongxiayang,
tjtanaa and
zyongye
as code owners
October 16, 2026 05:08
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Collaborator
Author
|
/ci run |
|
✅ Triggered Buildkite CI #95332 for commit |
AndreasKaratzas
approved these changes
Oct 16, 2026
AndreasKaratzas
enabled auto-merge (squash)
October 16, 2026 07:11
Collaborator
Author
|
/ci retry |
|
✅ The previous CI build is still running: https://buildkite.com/vllm/ci/builds/95332 |
Collaborator
Author
|
/ci run |
|
✅ Triggered Buildkite CI #95382 for commit |
Member
|
/ci retry |
|
✅ Triggered Buildkite CI #95522 for commit |
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
2 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.
AITER 0.1.25's FlyDSL MoE (ROCm/aiter#6125 configs) leaves the padded intermediate columns unwritten when told about the padding, so stage 2 reads stale memory and small decode batches (M <= 4 tokens) can come out NaN. This hits DeepSeek V4.1 at TP4, where the per-rank expert intermediate size is padded 576 -> 640.
Pass intermediate_pad=0 for V4.1's routed experts so AITER computes over the zero padding instead of skipping it, which is exact. DeepSeek V4 never pads the expert intermediate size and is unchanged. Revert once vLLM pins an AITER with the fix ROCm/aiter#6598.
This hotfix was developed to resolve the flakiness in the newly added test group
(MI355) DeepSeek-V4.1 Mono Decode Numericslm-eval results: