NVIDIA / NVIDIA/TransformerEngine
no attention backend support for softmax = learnable and cp_comm_type = p2p
Open
Nobody has claimed this yet.
bug
- Dominant language
- Python
- Stars
- 3.5k
- Forks
- 831
- Avg merge
- 3d 11h
- Merged PRs (30d)
- 65
Description
When using context parallel to fine-tune gpt-oss, no attention backend is supported for this configuration. I have to change cp_comm_type to a2a to enable FusedAttention. But this is potentially less efficient than a2a at large context length (say 128k).
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 tracing attention backend selection for context-parallel configurations, focusing on FusedAttention with softmax=learnable and cp_comm_type=p2p. Compare this with the currently working cp_comm_type=a2a path; done means the reported configuration has a supported attention backend and is covered by an appropriate validation or regression test.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- backend, machine-learning, performance
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Needs clarification
- Newbie friendliness
- 42/100