[GH-ISSUE #62] Exception: Capture cuda graph failed: CUDA error: capturing stream has unjoined work #60

Closed
opened 2026-10-26 09:53:32 +00:00 by bsakaguchi · 2 comments
Owner

Originally created by @liz-badada on GitHub (Mar 10, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/62

Hi,

I tried with cuda-graph capturing feature when intergrating DeepEP, but got "CUDA error: capturing stream has unjoined work".

Any comments?

Thanks!

Originally created by @liz-badada on GitHub (Mar 10, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/62 Hi, I tried with cuda-graph capturing feature when intergrating DeepEP, but got "CUDA error: capturing stream has unjoined work". Any comments? Thanks!
Author
Owner

@LyricZhao commented on GitHub (Mar 10, 2025):

Which series of kernels are you capturing, low-latency kernels? (Normal kernels doesn't support CUDA graph due to CPU sync)

I guess this is not related to DeepEP, but some capturing issues, e.g. https://stackoverflow.com/questions/79496131/trying-to-end-stream-capture-fails-due-to-unjoined-work-but-synchronizing-fai.

For low-latency kernels, the recommended way to use is a single stream with hook.

<!-- gh-comment-id:2709821152 --> @LyricZhao commented on GitHub (Mar 10, 2025): Which series of kernels are you capturing, low-latency kernels? (Normal kernels doesn't support CUDA graph due to CPU sync) I guess this is not related to DeepEP, but some capturing issues, e.g. https://stackoverflow.com/questions/79496131/trying-to-end-stream-capture-fails-due-to-unjoined-work-but-synchronizing-fai. For low-latency kernels, the recommended way to use is a single stream with hook.
Author
Owner

@liz-badada commented on GitHub (Mar 10, 2025):

Yes, Normal Kernels currently, thanks for the information!

<!-- gh-comment-id:2710051439 --> @liz-badada commented on GitHub (Mar 10, 2025): Yes, Normal Kernels currently, thanks for the information!
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepEP#60
No description provided.