COMMITS
September 9, 2026
C
Fix SM90 bwd crash for head_dim in (128, 192] (#2482)
cora-codes committed
P
Fix `seqlen_k_loaded` treating a window bound of 0 as unbounded (sliding window) (#2624)
Prathamesh Jadhav committed
C
cute: don't widen intentionally-empty offset windows to full attention (#2490)
cora-codes committed
J
[CuTe] Speed up SM100 hdim-64 bwd: dedicated P/dS TMEM slots + per-warp signaling (#2804)
jack-carlisle-2025 committed
September 8, 2026
K
[CuTe,Bwd,Sm100] Support head_dim not a multiple of 32 in backward (#2698)
Kaijing Ma committed
M
[CuTe] Include callable defaults in JIT cache hashes (#2773)
Mingfei Guo committed
Y
[CuTe, SM80/SM120] Guard invalid varlen forward tiles (#2763)
Yiming Zhang committed
August 28, 2026
D
[CuTe] Speed up scalar SM100 mask compilation (#2819)
Driss Guessous committed
August 15, 2026
D
[CuTe] Relax CUTLASS DSL requirement (#2798)
Driss Guessous committed
August 11, 2026
J
Fix removed Quack packed subtraction API (#2787)
Jiaxuan Bai committed
August 9, 2026
J
Fix CLC fuzz scheduler expectations (#2766)
Jiaxuan Bai committed
August 7, 2026
D
[CuTe] Fix forward dynamic-shape correctness (#2745)
Driss Guessous committed
Y
[CuTe, Fwd] Stabilize tensor max_seqlen compile key (#2762)
Yiming Zhang committed
August 6, 2026
J
[CuTe, SM100] Fix deadlock in varlen + block-sparse + SplitKV forward (#2761)
Jiaxuan Bai committed
H
Fix duplicated word in layer norm comment (#2744)
Haijie Zhi committed
D
[CuTe, FA4] Preserve first-tile flag during scheduler reconstruction (#2705)
dongxiao committed
August 5, 2026
H
[Cute, bwd, sm90/100/110] Support learnable sink in backward (#2706)
Henry Tsang committed
August 4, 2026
L
[CuTe, SM100] Sparse MLA bwd: don't scatter dK/dV at -1 sentinel indices (#2755)
Lequn Chen committed
H
[ROCm] Fix CK varlen_fwd binding argument mismatch (#2742)
Hosang Yoon committed
August 3, 2026
J
[AI] Add doc on debug methodology (#2753)
jayhshah committed
August 2, 2026
R
[CuTe,Sm100] Varlen Dynamic Persistent scheduler and metadata (#2559)
Reuben Stern committed
July 29, 2026
A
Remove SM100 Functions from Hopper (#2746)
ankutalev committed
July 28, 2026
D
Numeric tweaks to fp8 (#2731)
Driss Guessous committed
July 27, 2026
L
add linearize scheduling to combine kernel for full cudagraph (#2692)
liangel-02 committed
July 24, 2026
J
Expand FLASHATTENTION_DISABLE_DROPOUT to not bring in unneeded headers (#2669)
Jane (Yuan) Xu committed
July 23, 2026
R
[CuTe, Flex] Allow score mod use in varlen backward (#2547)
Reuben Stern committed
July 19, 2026
July 16, 2026
Y
Add paged-KV block_table bounds check in mha_fwd_kvcache (#2711)
Yunwei Li committed
July 14, 2026
D
Enable 2CTA for SM100 block-sparse backward (#2661)
Driss Guessous committed