Skip to content

Pull requests: vllm-project/flash-attention

New pull request New
42 Open 137 Closed
42 Open 137 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

Optimize hdim 256
#180 opened Aug 17, 2026 by simon-veitner-redhat Loading…
hdim 256 mask_mod support
#179 opened Aug 12, 2026 by MatthewBonanni Member Loading…
Add SM90 FA4 Dense and MLA
#177 opened Aug 7, 2026 by simon-veitner-redhat Loading…
12
Apply split-K to varlen paged KV when max_seqlen_q > 1 (spec decode)
#176 opened Aug 6, 2026 by arbi-dev Loading…
1
Rope dim
#172 opened Jul 23, 2026 by JaredforReal Loading…
4
Sync upstream
#171 opened Jul 22, 2026 by MatthewBonanni Member Loading…
[WIP/RFC][SM120] Add bounded BF16 D256 paged-decode specialization
#170 opened Jul 22, 2026 by goodnight654 • Draft
1
Fix FA3 batch-dependent kernel specialization
#169 opened Jul 21, 2026 by yoyouwen • Draft
[Kernel] Support FP32 rotary tables with low-precision inputs
#167 opened Jul 19, 2026 by guoriyue • Draft
Refactor FA3 onto the PyTorch stable API
#163 opened Jul 14, 2026 by LucasWilkinson Collaborator • Draft
1
fix: align get_scheduler_metadata with mha_fwd in flash_api.cpp
#162 opened Jul 11, 2026 by elvircrn • Draft
1
fix: derive is_varlen in get_scheduler_metadata to match mha_fwd
#161 opened Jul 10, 2026 by elvircrn • Draft
1
Fa4 fp8 foldscale opt
#159 opened Jul 9, 2026 by qixiang-99 • Draft
[FA4] Support paged/decode compile specs in compile_flash_attn_varlen_func_from_specs
#158 opened Jul 9, 2026 by sfc-gh-goliaro Loading…
[Bugfix] Fix two SM80/SM120 forward kernel bugs: missing is_split_kv default, mDynamicCausal NameError
#156 opened Jun 30, 2026 by tgmerritt Loading…
feat(fa4): fuse per-group fp8 output
#151 opened Jun 18, 2026 by carlyou Loading…
4
[Fix] Mark kcache and vcache as mutable in fwd_kvcache
#148 opened Jun 15, 2026 by remi-or Loading…
feat(cute): add SM90 FP8 KV support with in-kernel dequantization
#147 opened Jun 12, 2026 by qixiang-99 • Draft
fix: handle suffix-less runtime arch (sm_103/GB300) in SM 10.x gate
#146 opened Jun 11, 2026 by aoshen02 Loading…
SM100 dynamic causal
#144 opened Jun 10, 2026 by MatthewBonanni Member Loading…
Fix illegal memory access in FA2 varlen SplitKV early-exit LSE write
#139 opened May 18, 2026 by wangyxbh Loading…
Reapply #122
#137 opened May 6, 2026 by MatthewBonanni Member Loading…
SM100 tile size 64
#132 opened Apr 9, 2026 by MatthewBonanni Member • Draft
[Perf] SM103 tcgen05.ld.red for fused TMEM load + row-max in softmax
#131 opened Apr 9, 2026 by LopezCastroRoberto Loading…
Combine kernel: increase pipeline depth from 4 to 8 stages
#124 opened Mar 4, 2026 by jmkuebler Loading…
1
ProTip! Follow long discussions with comments:>50.