Repository navigation
[ROCm][DSv4.1] Lift the skinny-GEMM cap in the gfx942 draft lm_head test - #64532
Merged
alexm-redhat merged 3 commits intoOct 29, 2026
Merged
alexm-redhat merged 3 commits into
alexm-redhat merged 3 commits into
Conversation
The 315 MiB shard is over MI300 and MI355 L3, so the guard skips wvSplitK for n > 1 and the monokernel pad is no longer visible. Set the cap to this shard. The guard has its own tests. Signed-off-by: Frida Andersson <fanderss@amd.com>
mergify
Bot
added
DSv4.1
alexm-redhat
approved these changes
Oct 29, 2026
alexm-redhat
left a comment
alexm-redhat
left a comment
Contributor
There was a problem hiding this comment.
@frida-andersson thanks for the quick fix, LGTM!
6 tasks done
This test runs on gfx942. The shard is over the 256 MiB cap that skipped wvSplitK there. Signed-off-by: Frida Andersson <fanderss@amd.com>
Contributor
|
@frida-andersson verified this PR on my side, looks good. |
frida-andersson
marked this pull request as ready for review
October 29, 2026 15:31
frida-andersson
requested review from
WoosukKwon,
mgoin and
tlrmchlsmth
as code owners
October 29, 2026 15:31
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
frida-andersson
requested review from
AndreasKaratzas,
yewentao256 and
zyongye
as code owners
October 29, 2026 15:31
Contributor
Author
Thanks for testing it, I'll start CI |
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #98362 for commit ℹ️ This PR is 2 commits behind upstream |
Member
Contributor
Author
|
CI is green @alexm-redhat |
6 tasks done
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Overview
The gfx942 draft lm_head test still counts wvSplitK after #63179. The 315 MiB shard is over the skinny-GEMM cap, so 4- and 5-row calls leave wvSplitK and the monokernel pad is no longer visible.
Claims
test_draft_lm_head_takes_5_rows_off_wvsplitkagain sees wvSplitK for up to 5 rows, and no wvSplitK for the 5-row--monokernelpad.VLLM_SKINNY_GEMM_MAX_WEIGHT_BYTESto the shard and clears the cached cap on the way out.Validation
The failure is the three assertions on MI300 in CI #98321 and CI #98239. The same cases passed in CI #97749, before #63179. Ruff check and format pass on the file. The test skips off gfx942, so it was not rerun here.
Details
#63179 skips wvSplitK when the weight is over the memory-side cache and
1 < n <= 5. This shard is 32320×5120 bf16, 315 MiB, and the MI300 cap is 256 MiB. #60731's test still expects those calls. The cap in this test is the shard itself, so the row count still shows the pad. Noted on #63179 with no follow-up yet.Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.