fa4-v4.0.0.beta27

What's Changed

• [CuTe] Relax CUTLASS DSL requirement by @drisspg in #2798

Full Changelog: fa4-v4.0.0.beta26...fa4-v4.0.0.beta27

fa4-v4.0.0.beta26

What's Changed

• [CuTe,Bwd,Sm90] Fix: wait for bwd_preprocess on the block-sparse path, matching the dense path by @Fugoes in #2756
• [Cute, bwd, sm90/100/110] Support learnable sink in backward by @henrylhtsang in #2706
• [CuTe, FA4] Preserve first-tile flag during scheduler reconstruction by @dongxiao92 in #2705
• Fix duplicated word in layer norm comment by @cupkk in #2744
• [CuTe, SM100] Fix deadlock in varlen + block-sparse + SplitKV forward by @JiaxuanBai in #2761
• [CuTe, Fwd] Fix forward compile key churn when max_seqlen is a tensor by @eamonn-zh in #2762
• [CuTe] Fix forward dynamic-shape correctness by @drisspg in #2745
• Fix CLC fuzz scheduler expectations by @JiaxuanBai in #2766
• Fix removed Quack packed subtraction API by @JiaxuanBai in #2787

New Contributors

• @Fugoes made their first contribution in #2756
• @dongxiao92 made their first contribution in #2705
• @cupkk made their first contribution in #2744
• @JiaxuanBai made their first contribution in #2761
• @eamonn-zh made their first contribution in #2762

Full Changelog: fa4-v4.0.0.beta25...fa4-v4.0.0.beta26

fa4-v4.0.0.beta25

What's Changed

• Remove SM100 Functions from Hopper Flash Attention 3 by @ankutalev in #2746
• [CuTe,Sm100] Varlen Dynamic Persistent scheduler and metadata by @reubenconducts in #2559
• [AI] Add doc on debug methodology by @jayhshah in #2753
• [ROCm] Fix CK varlen_fwd binding argument mismatch by @hyoon1 in #2742
• [CuTe, SM100] Sparse MLA bwd: don't scatter dK/dV at -1 sentinel indices by @abcdabcd987 in #2755

New Contributors

• @hyoon1 made their first contribution in #2742
• @abcdabcd987 made their first contribution in #2755

Full Changelog: fa4-v4.0.0.beta24...fa4-v4.0.0.beta25

fa4-v4.0.0.beta24

What's Changed

• [CuTe, Flex] Allow score mod use in varlen backward by @reubenconducts in #2547
• Expand FLASHATTENTION_DISABLE_DROPOUT to not bring in unneeded headers by @janeyx99 in #2669
• add linearize scheduling to combine kernel for full cudagraph by @liangel-02 in #2692
• Numeric tweaks to fp8 by @drisspg in #2731

Full Changelog: fa4-v4.0.0.beta23...fa4-v4.0.0.beta24

fa4-v4.0.0.beta23

What's Changed

• Add paged-KV block_table bounds check in mha_fwd_kvcache by @yunweili3 in #2711
• [CuTe, SM100] Fix FP8 e4m3 accuracy: make max_offset dtype-aware to avoid P saturation by @yunweili3 in #2717

Full Changelog: fa4-v4.0.0.beta22...fa4-v4.0.0.beta23

fa4-v4.0.0.beta22

What's Changed

• add tcgen.ld.red support to sm103a arch by @drisspg in #2696
• [CuTe, Bwd] Fix backward compile key churn due to pickling, max_seqlen is a tensor by @michaelxu-msft in #2507
• hopper/setup.py: harden tarfile extraction against path traversal and symlink escape by @aryanputta in #2702
• Enable 2CTA for SM100 block-sparse backward by @drisspg in #2661

New Contributors

• @michaelxu-msft made their first contribution in #2507

Full Changelog: fa4-v4.0.0.beta21...fa4-v4.0.0.beta22

fa4-v4.0.0.beta21

What's Changed

• [FA3] uv installation support by @bbuschkaemper in #2458
• [AMD ROCm] Enable RDNA backward and adopt CK unified workspace by @rocking5566 in #2675
• Fix CuTe SM120 compile-time argument handling by @Kevin-Li-2025 in #2671
• [NVIDIA][CuTe,Fwd,sm120] Implement Pack-GQA on SM120 (+ graceful SplitKV fallback) by @johnnynunez in #2656

New Contributors

• @bbuschkaemper made their first contribution in #2458
• @Kevin-Li-2025 made their first contribution in #2671

Full Changelog: fa4-v4.0.0.beta20...fa4-v4.0.0.beta21

fa4-v4.0.0.beta20

What's Changed

• fix: sync callers with new _flash_attn_fwd 4-tuple return signature by @hhy3 in #2674
• Fix compatibility issues with CuTe DSL 4.6.0+ by @anakinxc in #2648
• [CuTe] Pass tmem scalar fields as .ptr to TmemAllocator on SM100 by @pashu-cohere in #2679
• Add FLASHATTENTION_DISABLE_SPLIT_ALIGNMENT by @janeyx99 in #2680
• ci: rebake cu130 image for cutlass-dsl 4.6.0.dev0 floor by @Johnsonms in #2684
• Update FA4 cute quack compatibility by @Luosuu in #2676
• ci: install cutlass-dsl/quack at runtime to decouple from the baked image by @Johnsonms in #2685
• [Cute] Assume 16B stride divisibility for SM100 backward LSE/dPsum bulk-copy inputs by @Johnsonms in #2686
• fix(hd256/sm100): forward reads actual input strides, drop .contiguous() patch by @oattia in #2670
• ci: add MLA absorbed coverage to FA4 CI by @Johnsonms in #2690
• Parallelize splitkv alignment templated kernels, remove flag by @janeyx99 in #2683

New Contributors

• @hhy3 made their first contribution in #2674
• @pashu-cohere made their first contribution in #2679
• @oattia made their first contribution in #2670

Full Changelog: fa4-v4.0.0.beta19...fa4-v4.0.0.beta20

fa4-v4.0.0.beta19

What's Changed

• fix(hd256/sm100): make q/k/v contiguous before dedicated hd256 kernel by @yunweili3 in #2666
• [Cute,Bwd,Sm100] add sparse MLA (Deepseek v4) backward kernels by @jayhshah in #2621

New Contributors

• @yunweili3 made their first contribution in #2666

Full Changelog: fa4-v4.0.0.beta18...fa4-v4.0.0.beta19

fa4-v4.0.0.beta18

What's Changed

• Fix SM100 FP8 fwd with cutlass-dsl >=4.5.2 (MmaF8F6F4Op) by @Johnsonms in #2640
• [cute] Fix int32 overflow in SM100 LPT tile scheduler for long context by @sryap in #2662
• [Fwd,Sm100] Tune FP8 causal hd128 ex2_emu_freq (8 vs inherited 16) by @Johnsonms in #2642
• Make q_subtile_factor default to identity by @drisspg in #2660

Full Changelog: fa4-v4.0.0.beta17...fa4-v4.0.0.beta18