[ROCm][CI] Fix cuda graph mem profile issue - #48764
Merged
AndreasKaratzas merged 2 commits intoJul 16, 2026
Merged
Conversation
Signed-off-by: charlifu <charlifu@amd.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
AndreasKaratzas
added
the
ready
AndreasKaratzas
approved these changes
Jul 15, 2026
AndreasKaratzas
left a comment
AndreasKaratzas
left a comment
Member
There was a problem hiding this comment.
LGTM
Thanks for fixing this.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Root cause
During
determine_available_memory→profile_cudagraph_memory, capture runs insidetorch.cuda.graph(). PyTorch requires a non-default stream (cuda_stream != 0).On ROCm, the profiling path was using:
GraphCaptureContext(torch.cuda.current_stream(self.device))
Before vLLM initializes its dedicated compute stream, torch.cuda.current_stream() is the per-thread default stream (cuda_stream=0). That object is not torch.cuda.default_stream() by identity, but PyTorch still rejects it for graph capture — which matches your error:
RuntimeError: CUDA graphs must be captured on a non-default stream.
Fix
Use vLLM's current_stream() instead. It creates a dedicated non-default compute stream (the same one used for normal inference), which:
Satisfies PyTorch's cudagraph capture requirement
Preserves the ROCm intent of profiling on the compute stream, not a throwaway side stream from graph_capture()