Skip to content

Pull requests: Dao-AILab/flash-attention

New pull request New
231 Open 713 Closed
231 Open 713 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
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

Add determnisitc var len support
#2795 opened Aug 14, 2026 by drisspg Collaborator Loading…
2
[CuTe, SM100] Support varlen block-sparse backward
#2794 opened Aug 14, 2026 by drisspg Collaborator Loading…
2
feat: support per-head rotary frequencies
#2792 opened Aug 12, 2026 by zupengwang Loading…
1
Mask the output store in the Triton linear forward kernel
#2791 opened Aug 12, 2026 by nileshpatil6 Loading…
1
[Triton] Apply the computed bounds mask in kernel_fwd's output store
#2790 opened Aug 11, 2026 by truong-v Loading…
1
publish.yml: trigger the release workflow on tag pushes
#2788 opened Aug 11, 2026 by aryanputta Contributor Loading…
1
1
[CuTe, SM120] Enable custom score-mod backward
#2786 opened Aug 11, 2026 by guoriyue • Draft
[Generation] Track EOS per sequence in batched decoding
#2784 opened Aug 11, 2026 by guoriyue • Draft
[CuTe] Add packed QKV and KV public APIs
#2785 opened Aug 11, 2026 by guoriyue • Draft
fix: mark FA2 auxiliary outputs non-differentiable
#2781 opened Aug 9, 2026 by zupengwang Loading…
1
[Pretrained] Load safetensors-only models from the Hub
#2778 opened Aug 9, 2026 by guoriyue • Draft
[FA2] Fix backward for non-contiguous Q/K/V inputs
#2777 opened Aug 9, 2026 by guoriyue • Draft
[CuTe, Bwd] Return zero gradients for empty Q/K workloads
#2776 opened Aug 9, 2026 by guoriyue • Draft
[CuTe, Fwd] Fix PackGQA predication for padded head dimensions
#2775 opened Aug 9, 2026 by guoriyue • Draft
[CuTe] Include callable defaults in JIT cache hashes
#2773 opened Aug 9, 2026 by guoriyue • Draft
[Build] Fix CUDA gencode selection for older toolkits
#2772 opened Aug 9, 2026 by guoriyue • Draft
[CuTe, SM120] Optimize backward on consumer Blackwell
#2771 opened Aug 9, 2026 by guoriyue • Draft
[CuTe] Pass Fwd/Bwd args as standardized NamedTuples
#2770 opened Aug 9, 2026 by reubenconducts Contributor Loading…
6
Add an Attention Sink to Flash MLA for SM100
#2768 opened Aug 7, 2026 by floatingtrees Loading…
1
[CuTe, SM80/SM120] Guard invalid varlen forward tiles
#2763 opened Aug 7, 2026 by eamonn-zh Contributor Loading…
2
Add varlen blocksparse backward hopper
#2759 opened Aug 6, 2026 by drisspg Collaborator • Draft
[CuTe, SM120] Support automatic num_splits
#2758 opened Aug 5, 2026 by bvolpato Loading…
[CuTe, SM90] Support KV128 block-sparse backward for hdim192
#2757 opened Aug 5, 2026 by drisspg Collaborator Loading…
1
2
[CuTe,Fwd,Sm90] add dynamic persistent scheduler to sm90
#2754 opened Aug 3, 2026 by reubenconducts Contributor Loading…
[CuTe, SM80/SM120] Support learnable_sink (attention sinks) on Ampere/Ada/consumer-Blackwell
#2752 opened Jul 31, 2026 by arbi-dev Loading…
2
ProTip! Add no:assignee to see everything that’s not assigned.