Pull requests: flash-algo/flash-sparse-attention

1 Open 38 Closed

Author

Filter by author

Loading

Label

Filter by label

Loading

Projects

Filter by project

Loading

Milestones

Filter by milestone

Loading

Reviews

Filter by reviews

No reviews Review required Approved review Changes requested

Assignee

Filter by who’s assigned

Assigned to nobody Loading

Sort

Sort by

Newest Oldest Most commented Least commented Recently updated Least recently updated Best match

Most reactions

👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Optimize triton version: GQA, mask/bias broadcasting, skip inactive tiles, and stability fixes

#200 by LoserCheems was merged Nov 7, 2025 Loading…

2 of 5 tasks

24

[FEATURE SUPPORT] Convert attention mask storage from float to bool

#166 by LoserCheems was merged Sep 12, 2025 Loading…

5 tasks done

13

[BUG FIX] Prevent mask/bias materialization; avoid OOB for irregular seqlen

#168 by LoserCheems was merged Sep 17, 2025 Loading…

5 tasks done

11

[FEATURE SUPPORT] Triton special compact dynamic-mask attention: 1.6× faster fwd+bwd, numerically equivalent

#206 by LoserCheems was merged Nov 7, 2025 Loading…

4 of 5 tasks

10

[FEATURE SUPPORT] Variable-Length Attention with Padding-Free Execution

#188 by LoserCheems was merged Oct 11, 2025 Loading…

6 tasks done

8

[FEATURE SUPPORT] Centralize dynamic mask creation for FDMA

#197 by LoserCheems was merged Oct 23, 2025 Loading…

5 tasks done

1

6

[FEATURE SUPPORT] Optional mask/bias (3D & 4D)

#170 by LoserCheems was merged Sep 19, 2025 Loading…

5 of 7 tasks

6

[BUG FIX] Unify masking utilities and improve performance

#209 by LoserCheems was merged Nov 13, 2025 Loading…

4 tasks

6

[FEATURE SUPPORT] Broadcastable 4D mask/bias, 128‑rounded key length, stride‑0 broadcasting, and dbias reductions

#190 by LoserCheems was merged Oct 12, 2025 Loading…

5 tasks done

6

Implement variable-length attention with mask and bias support

#185 by LoserCheems was merged Oct 9, 2025 Loading…

6

Add issue/PR templates

#186 by LoserCheems was merged Oct 10, 2025 Loading…

6

[BUG FIX] Fix mask/bias memory access and vectorization issues in kernels

#182 by LoserCheems was merged Oct 1, 2025 Loading…

7 tasks done

6

Refactor attention block smoothing for consistency

#205 by LoserCheems was merged Nov 6, 2025 Loading…

4 tasks

5

Chore/sync after move

#208 by LoserCheems was merged Nov 11, 2025 Loading…

5

Add tensor operation utilities and performance enhancements

#165 by LoserCheems was merged Sep 11, 2025 Loading…

5

Update issue templates and add auto-assignment for PRs docs Improvements or additions to documentation

#158 by LoserCheems was merged Sep 9, 2025 Loading…

5

Enhance bias gradient accumulation in backward pass

#193 by LoserCheems was merged Oct 16, 2025 Loading…

5

Add block-wise smoothing to attention mask

#201 by LoserCheems was merged Oct 30, 2025 Loading…

5

Update documentation to use mask utility in examples

#198 by LoserCheems was merged Oct 23, 2025 Loading…

4

[FEATURE SUPPORT] Robust dBias accumulation for seqlen_q_bias == 1

#194 by LoserCheems was merged Oct 22, 2025 Loading…

4 of 5 tasks

4

Fix documentation and references for Flash Sparse Attention

#207 by LoserCheems was merged Nov 9, 2025 Loading…

4 tasks

4

Refactor attention mask and bias handling for efficiency

#177 by LoserCheems was merged Sep 21, 2025 Loading…

3

[FEATURE SUPPORT] Flexible head dims for mask/bias with in-kernel conversion path

#167 by LoserCheems was merged Sep 13, 2025 Loading…

4 of 5 tasks

3

[FEATURE SUPPORT] Move scaling out of streaming loops, bias-initialized acc_s, and fix dQ double-scaling

#203 by LoserCheems was merged Nov 4, 2025 Loading…

5 tasks done

1

3

Add selectable masking strategies for attention

#204 by LoserCheems was merged Nov 6, 2025 Loading…

2