All pull requests

Pull requests

• 优化 FlashMLA 64 头融合注意力内核的 Prefill 与 FP8 Decode 性能

  #227·cb521 opened Sep 24, 2026·

  ·

• fix: MSVC portability - uint4 transport for shared-memory 128-bit load/store

  #226·PerryLink opened Sep 16, 2026Contributor·

  ·

• fix: MSVC portability - standard C++ dependent-type init + drop GNU always_inline

  #225·PerryLink opened Sep 15, 2026Contributor·

  ·

• [SM90] Support dynamic top-k for V3.2 sparse decode

  #216·Livinfly opened Aug 30, 2026·

  ·

• docs: use consistent TFlops unit in README

  #215·pengfaming opened Aug 28, 2026·

  ·

• Fix README decode example to match the current flash_mla_with_kvcache signature

  #214·shoemoney opened Aug 24, 2026·

  ·

• Fix duplicate TRAP_ONLY_DEVICE_ASSERT definition

  #209·tianchongchong opened Aug 7, 2026·

  ·

• [Draft]SM100 persistent sparse attention for V4 CSA/HCA

  #207·uchihatmtkinu opened Aug 4, 2026Collaborator·

  ·

• Mark auxiliary LSE output non-differentiable

  #203·Mr-Neutr0n opened Aug 3, 2026·

  ·

• Preserve input validation under Python optimization

  #202·Mr-Neutr0n opened Jul 31, 2026·

  ·

• Fix scheduler state after backend failure

  #201·Mr-Neutr0n opened Jul 31, 2026·

  ·

• [SM90] Add packed selected-KV sparse decode

  #200·xyz2606 opened Jul 26, 2026·

  ·

• Add sparse MLA backward (DSA top-k), sliding-window attention + attention sink, and block-sparse MLA forward/backward for SM100

  #198·kmswin1 opened Jul 16, 2026·

  ·

• sparse_prefill_fwd (sm100): skip trailing fully-invalid top-k blocks when topk_length is not provided

  #196·zkyue opened Jul 10, 2026·

  ·

• Add pyproject.toml with build-system metadata

  #195·Functionhx opened Jul 9, 2026·

  ·

• Become ABI Stable with LibTorch 2.11+

  #193·janeyx99 opened Jul 2, 2026·

  ·
  1

  1

• fix typo

  #191·wangtianxia-sjtu opened Jun 25, 2026·

  ·

• Dsv4 sparse prefill optimization by skipping odd tail block

  #187·Njuapp opened Jun 9, 2026·

  ·

• Raise on unsupported head_dim in SM100 dense FMHA dispatch instead of silently returning uninitialized output

  #185·toffee-desuwa opened May 30, 2026·

  ·
  3

• Fix SM100 V32 sparse decode FP8 handling

  #184·jianglan89 opened May 10, 2026·

  ·

• Add SM80 (Ampere/A100) dense MLA decode kernel

  #183·bzantium opened May 1, 2026·

  ·

• docs: fix sign of online softmax rescaling factors

  #177·tangpanyu opened Apr 23, 2026·

  ·

• Install CUDAGuard before Arch() in api entry points (fixes #158)

  #176·yurekami opened Apr 19, 2026·

  ·
  1

• fix: tcgen05.mma.ws assert typo

  #175·KevinZeng08 opened Apr 16, 2026·

  ·

• Fix potential memory leak in sparse_attn_decode_interface

  #174·NJX-njx opened Apr 14, 2026·

  ·
  2