How to set TensorRT to support Flash Attention 3
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 13.4k
- Forks
- 2.4k
- Avg merge
- 5d 3h
- Merged PRs (30d)
- 2
Description
HI，
How does TensorRT control whether to load flash attention? I tried to install flash attention 3 (which includes the right swipe module hopper for H100) from the source code, but the FTT and throughput running on TensorRT on H100 before and after installation were basically the same, and it seems that I did not use the flash attention I installed. Please provide some ideas and suggestions, thank you very much.
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.
Research direction
The issue mentions TensorRT, Flash Attention 3, and H100 performance comparisons, but names no source files, tests, or entry points. Start by identifying where TensorRT selects attention implementations and reproduce the before-and-after throughput results; the issue needs a defined configuration path and measurable acceptance criteria before work can be considered done.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- machine-learning, performance
- Issue type
- Documentation
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 15/100