All pull requests

Pull requests

• fwd: precompute flashmask n_block list, drop the per-tile scan

  #176·baoqiwen opened Sep 10, 2026·

  ·

• [Optim] SWA related performance optimization (part 3/3)

  #163·Enigmatisms opened Jul 7, 2026·

  ·

• [Sm100] FlashMask V4 fwd support head dim 256 via Split-D (q_stage=1, d==dv)

  #152·wangxudong10 opened May 22, 2026·

  ·
  3

• [WIP] Fa4 d256 varlen zero seq

  #149·umiswing opened May 19, 2026·

  ·

• [WIP] support cp case use_varlen

  #147·umiswing opened May 18, 2026·

  ·

• Fix CUDA 13.2 flash attention build compatibility

  #141·gouzil opened Apr 28, 2026·

  ·
  11

• feat: implement max_logits support for flashmask

  #139·xxyux opened Apr 22, 2026·

  ·

• [Feat] CP-balance formal incorporation as flash_mask sub-module

  #127·Enigmatisms opened Apr 9, 2026·

  ·
  4

• Add rrattn estimate func and interface

  #117·LLSGYN opened Mar 13, 2026·

  ·
  4

• Support Global Sliding Window (num_vec == 4) on FM4 BWD

  #111·umiswing opened Mar 3, 2026·

  ·

• adapt to torch version flashmaskv4

  #103·clouds1238 opened Jan 28, 2026·

  ·
  1

• add_flashmask_cpbalance

  #99·starcrown001 opened Dec 30, 2025·

  ·
  1

• add flashmask v2 torch flash_api.cpp flashmask_interface.py setup.py

  #98·clouds1238 opened Dec 23, 2025·

  ·
  10

• fine-tuned tile size & regitser for fwd_hdim64

  #92·xxyux opened Nov 14, 2025·

  ·

• Removed redundant templates and related compile-time/runtime code

  #91·Enigmatisms opened Nov 14, 2025·

  ·

• fix fa2 flashmask oob read

  #67·umiswing opened Jun 26, 2025·

  ·

• [WIP] fa3 varlen fix int32 overflow

  #65·umiswing opened Jun 19, 2025·

  ·

• scan from right to left and skip masked block for each row at kernel begin

  #55·GuoxiaWang opened Sep 23, 2024·

  ·

• optimize skip block calculate in bwd

  #49·GuoxiaWang opened Aug 28, 2024·

  ·

• [BugFix] fix_mask error using unpadding api

  #41·wwbitejotunn opened Apr 23, 2024·

  ·
  1

• Fix unpadding input with padding mask compute error

  #38·wwbitejotunn opened Apr 15, 2024·

  ·
  1

• Fa cmake extends op

  #31·AnnaTrainingG opened Dec 14, 2023·

  ·
  1

• Fa cmake

  #29·AnnaTrainingG opened Dec 6, 2023·

  ·
  31

• [WIP]Sparse seqparallel

  #9·zkh2016 opened Jun 8, 2023·

  ·
  1

• add block sparse api

  #7·kuizhiqing opened May 27, 2023·

  ·
  1