ad-executor - analyze multimodal metadata induced synchronize
Open
@galagam is already working on this.
Since May 4, 2026.
Multimodal
- Dominant language
- Python
- Stars
- 14.7k
- Forks
- 2.8k
- Avg merge
- 2d 23h
- Merged PRs (30d)
- 489
Description
Follow up from #12861
https://github.com/NVIDIA/TensorRT-LLM/pull/12861#discussion_r3077973131
in ad-executor, we have:
```
# Multimodal prefill can leave image/context work in flight on the execution
# stream after model.forward returns. Synchronize before the overlap scheduler
# advances to the next step so later batch-state reuse does not race that work
# and surface as a delayed CUDA illegal memory access.
torch.cuda.current_stream().synchronize()
This may not be not ideal perf-wise. Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.