OpenImagingLab / OpenImagingLab/FlashVSR

About topk in sparse attention

Open
#42 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
1.9k
Forks
152
PR merge metrics
No merged PRs in 30d

Description

Why do the topk selection across all query-key pairs together but not topk selection across the row for every query respectively. There may be only a part of tokens are updated in the attention module.

https://github.com/OpenImagingLab/FlashVSR/blob/914dcd4b3b1155d8a10028177b7473d4f42d47da/diffsynth/models/wan_video_dit.py#L140-L149

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Read diffsynth/models/wan_video_dit.py around lines 140-149 and inspect how topk is applied across query-key pairs. Compare global selection with per-query row selection, then determine whether the attention module intentionally updates only part of the tokens. Done means the intended behavior and its rationale are documented or the required change is clearly specified.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.