deepspeedai / deepspeedai/DeepSpeed
Can Deepspeed-Sparse-Attention work with auto-regressive attention (left triangle attention)?
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Hi ,
Recently, I have been working on the NLG(Natural Language Generation) model whose attention is a left triangle attention, like following picuture:
But from the code I have read from deepspeed, it only support block sparse (block_size > 1) which can't satisfy the condition : "a word can only look back", e.g. the word "1" can watch word "2" in the following picture:
So my question is can I use DeepSpeed in the NLG model?
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 DeepSpeed's sparse-attention implementation and its block-size constraints, then compare them with the left-triangle mask shown in the issue. Confirm whether autoregressive masking is supported; if not, define the required scope and validation before changing code.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 15/100