All pull requests

Pull requests

• fix(cutedsl): support CUTLASS DSL 4.4.2 and 4.6.2

  #383·akoumpa opened Oct 14, 2026Contributor·

  ·
  2

• [WIP] feat(ffa): Add SM100 inner-loop-K backward kernel for full mask

  #380·lijinnn opened Oct 8, 2026Collaborator·

  ·

• [WIP]: feat(ffa): native softcap in the cutedsl backend

  #379·cherhh opened Oct 5, 2026Collaborator·6/6·

  ·

• [WIP]: fix(ext): pad the unique_consecutive_pairs tail and launch on the current stream

  #377·cherhh opened Oct 4, 2026Collaborator·

  ·

• [WIP]: feat(ffa): cat_gqa backward and packed atomic forward in the cutedsl backend

  #376·cherhh opened Oct 4, 2026Collaborator·5/6·

  ·

• [WIP]: feat(ffa): return max logits in the cutedsl backend

  #375·cherhh opened Oct 4, 2026Collaborator·4/6·

  ·

• [WIP]: feat(ffa): sm_margin and dynamic tile scheduling in the cutedsl backend

  #374·cherhh opened Oct 3, 2026Collaborator·3/6·

  ·

• [WIP] Adapt ffa_fa4 for sm120

  #369·lijinnn opened Sep 22, 2026Collaborator·

  ·

• [WIP]: feat(ffa): learnable attention sink in the cutedsl backend

  #367·cherhh opened Sep 21, 2026Collaborator·2/6·

  ·
  1

• feat: add cuDNN-FlexAttention as high-perf backend

  #353·NVIDIA-JerryChen opened Sep 14, 2026Collaborator·

  ·
  1

• [WIP]: test(sparse): add deterministic and overlap-safe tests for BlockSparse/IndexSparse

  #335·cennn opened Jul 17, 2026Collaborator·2 tasks·

  ·

• [WIP]: feat(cutedsl-sm100): BWD LoopK with warp-specialized pipeline + IndexSparse/BlockSparse support

  #334·cennn opened Jul 13, 2026Collaborator·

  ·