Skip to content

Conversation

@LucasWilkinson
Copy link
Collaborator

@LucasWilkinson LucasWilkinson commented Dec 2, 2025

FIX: #29852

Make sure we only call build_for_cudagraph_capture if we are capturing cudagraphs not for DP dummy runs; this behavior; NOTE prior to #28579 we used always called build_for_cudagraph_capture for all dummy runs including DP dummy batches (this was a bit messy), this shouldn't be the case anymore since we build with padding which can cause asserts in build_for_cudagraph_capture

Signed-off-by: Lucas Wilkinson [email protected]

Signed-off-by: Lucas Wilkinson <[email protected]>
@LucasWilkinson LucasWilkinson changed the title [BugFix] Potentially fix `build_for_cudagraph_capture [BugFix] Potentially fix build_for_cudagraph_capture Dec 2, 2025
@LucasWilkinson LucasWilkinson marked this pull request as ready for review December 2, 2025 16:04
@LucasWilkinson
Copy link
Collaborator Author

Confirmed fixed by @varun-sundar-rabindranath

@varun-sundar-rabindranath
Copy link
Contributor

I was hitting this assert #29852 on DP/EP on the Decode nodes in a PD setup. I no longer hit this assert with this PR.

@LucasWilkinson LucasWilkinson added the ready ONLY add when PR is ready to merge/full CI is needed label Dec 2, 2025
@LucasWilkinson LucasWilkinson changed the title [BugFix] Potentially fix build_for_cudagraph_capture [BugFix] Fix assert in build_for_cudagraph_capture Dec 2, 2025
@github-project-automation github-project-automation bot moved this to In review in NVIDIA Dec 2, 2025
@tlrmchlsmth tlrmchlsmth enabled auto-merge (squash) December 2, 2025 16:07
@tlrmchlsmth tlrmchlsmth added this to the v0.12.0 milestone Dec 2, 2025
@simon-mo simon-mo disabled auto-merge December 3, 2025 00:56
@simon-mo simon-mo merged commit 5cdd664 into vllm-project:main Dec 3, 2025
17 of 19 checks passed
@github-project-automation github-project-automation bot moved this from In review to Done in NVIDIA Dec 3, 2025
khluu pushed a commit that referenced this pull request Dec 3, 2025
Signed-off-by: Lucas Wilkinson <[email protected]>
(cherry picked from commit 5cdd664)
minosfuture added a commit to minosfuture/vllm that referenced this pull request Dec 4, 2025
charlotte12l pushed a commit to charlotte12l/vllm that referenced this pull request Dec 5, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

nvidia ready ONLY add when PR is ready to merge/full CI is needed v1

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

[Bug]: DSR1 NVFP4 DEP cannot run because MLA only supports decode-only full CUDAGraph capture

4 participants