State of affairs for NestedTensor (NJT) inference?
Open
Nobody has claimed this yet.
triaged
- Dominant language
- C++
- Stars
- 13.4k
- Forks
- 2.4k
- Avg merge
- 5d 3h
- Merged PRs (30d)
- 2
Description
PyTorch now has some support for representing varlen sequences. It is supported to some extent by HF:
- https://medium.com/pytorch/bettertransformer-out-of-the-box-performance-for-huggingface-transformers-3fbe27d50ab2
- https://github.com/pytorch/pytorch/issues/126025#issuecomment-2420369702
This is useful e.g. for saving compute on padding tokens for BERT inference. Does TRT has kernels for such NJT sdpa ops? (and can they be executed via CUDA graphs?) If so, how to benefit from it? Is there an example?
Thank you!
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
Start by reviewing the linked PyTorch and Hugging Face references, then investigate whether TensorRT supports NestedTensor SDPA inference and CUDA graph execution. Done means documenting the current support status, explaining how to use it if supported, and providing an example or clearly stating that no example is available.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- huggingface, pytorch
- Domain
- ai, machine-learning, performance
- Issue type
- Documentation
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100