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

13 Open 98 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

[FEATURE] Support total_k-head broadcastable mask/bias in varlen batch inference

#187 opened Oct 10, 2025 by Copilot AI • Draft

1

Fix INF issue in bf16 backward pass with safer value clamping

#181 opened Sep 26, 2025 by Copilot AI • Draft

1

Implement Unified Sparse Mask Strategy with Block-Level Skipping

#164 opened Sep 11, 2025 by Copilot AI • Draft

1

Add optional mask & bias inputs with adaptive computation skipping

#162 opened Sep 10, 2025 by Copilot AI • Draft

1

Fix installation error with improved backend fallback and error messages

#151 opened Sep 6, 2025 by Copilot AI • Draft

1

[PERFORMANCE] Implement Linear KV Cache optimization for inference acceleration

#143 opened Sep 1, 2025 by Copilot AI • Draft

1

[FEATURE] Reduce compute bubbles in backward skip path

#133 opened Aug 29, 2025 by Copilot AI • Draft

1

Fix NaN/Inf values in dV gradients during backward pass

#122 opened Aug 24, 2025 by Copilot AI • Draft

1

Add comprehensive documentation for efficient attention mask handling in long sequences

#120 opened Aug 23, 2025 by Copilot AI • Draft

1

2

Document query-agnostic design and implications for associative recall tasks

#119 opened Aug 23, 2025 by Copilot AI • Draft

1

[PERFORMANCE] Implement Backward Launch Template Optimization for SM 8.9 and SM 9.0

#112 opened Aug 17, 2025 by Copilot AI • Draft

1

Optimize sparse logic with block-level Tensor Core utilization

#110 opened Aug 15, 2025 by Copilot AI • Draft

1

1

Implement dBias gradient computation in CUDA backward kernel

#73 opened Jul 22, 2025 by Copilot AI • Draft

1