Skip to content

Pull requests: Dao-AILab/flash-attention

New pull request New
212 Open 694 Closed
212 Open 694 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

[CuTe] Fix forward dynamic-shape correctness
#2745 opened Jul 29, 2026 by drisspg Collaborator Loading…
4
Fix duplicated word in layer norm comment
#2744 opened Jul 29, 2026 by cupkk Loading…
1
[CuTe] Add explicit forward config autotuning
#2743 opened Jul 28, 2026 by drisspg Collaborator • Draft
[ROCm] Fix CK varlen_fwd binding argument mismatch
#2742 opened Jul 28, 2026 by hyoon1 Contributor Loading…
1
[CuTe] Tune B200 forward scheduling
#2741 opened Jul 27, 2026 by drisspg Collaborator • Draft
Pass loc and ip to llvm.inline_asm
#2740 opened Jul 27, 2026 by pchen7e2 Loading…
[CuTe] Tune GB300 forward scheduling
#2738 opened Jul 25, 2026 by drisspg Collaborator • Draft
[CuTe] Add explicit forward config benchmark
#2735 opened Jul 25, 2026 by drisspg Collaborator • Draft
[CuTe] Route forward through typed configs
#2733 opened Jul 25, 2026 by drisspg Collaborator • Draft
[CuTe] Add typed forward configurations
#2732 opened Jul 25, 2026 by drisspg Collaborator • Draft
2
tests: add GQA decode and serving pattern tests
#2730 opened Jul 25, 2026 by priyanka25aug Loading…
perf(cute): optimize SM100/SM103 hd256 long-sequence causal varlen prefill
#2726 opened Jul 23, 2026 by muse-coder Loading…
1
[FA3 ring attn support] Ring-attention primitives that keep partials & gradients in fp32, eliminating the per-hop bf16↔fp32 round-trips a naive ring pays (faster + more accurate)
#2723 opened Jul 22, 2026 by AntonotnaWang Loading…
Add two-section causal backward for shared K/V
#2722 opened Jul 22, 2026 by foraxe Loading…
1
Different numerics
#2721 opened Jul 21, 2026 by drisspg Collaborator Loading…
4
[AMD]: add FLASH_ATTENTION_USE_SYSTEM_AITER env to use pre-installed aiter
#2720 opened Jul 21, 2026 by ZiguanWang Loading…
PR: [CuTe, Bwd, SM100] Remove unneeded proxy fences in postprocess kernel
#2719 opened Jul 20, 2026 by Johnsonms Collaborator Loading…
[CuTe, SM100, bwd] fix: SM100 backward for padded head dimensions
#2715 opened Jul 17, 2026 by Luosuu Contributor Loading…
9
hopper: expose seqused_q for KV-cache decode
#2714 opened Jul 17, 2026 by sj0618 Loading…
[CuTe,Sm100,Fwd] fp8: bound lazy-rescale threshold by e4m3 range (fixes silent O corruption on peaked softmax rows)
#2713 opened Jul 17, 2026 by james-p-xu Loading…
[Cute, bwd, sm90/100/110] Support learnable sink in backward
#2706 opened Jul 13, 2026 by henrylhtsang Contributor Loading…
17
[CuTe, FA4] Preserve first-tile flag during scheduler reconstruction
#2705 opened Jul 13, 2026 by dongxiao92 Loading…
1
Migrate deprecated cutlass.utils spellings to canonical namespaces
#2704 opened Jul 13, 2026 by yiwangchunyu • Draft
Fa4 return lse consistency
#2700 opened Jul 9, 2026 by yijianggit Loading…
[CuTe,Bwd,Sm100] Support head_dim not a multiple of 32 in backward
#2698 opened Jul 9, 2026 by KaijingOfficial Loading…
3 of 4 tasks
1
ProTip! Add no:assignee to see everything that’s not assigned.