All pull requests

Pull requests

• Optimize K2 recurrent-state dataflow for SM103/B300

  #35·ICtuvk123 opened Sep 14, 2026·

  ·

• Support SM80 (Ampere) with dual-arch build

  #33·Leslie360 opened Aug 30, 2026·

  ·

• Handle all-empty varlen batches

  #32·0z5a opened Aug 17, 2026·

  ·

• Add reusable workspace support to the Python API

  #31·0z5a opened Aug 17, 2026·

  ·
  1

• Add intermediate states and exact CP transition prototype

  #30·0z5a opened Aug 14, 2026·

  ·

• Reuse tile prefix in recurrence kernel

  #29·0z5a opened Aug 14, 2026·

  ·

• Add CUDA training kernels (fwd+bwd) for KDA

  #28·xy200303 opened Aug 8, 2026·

  ·
  5

• Add differentiable CPU backend for KDA research

  #27·xy200303 opened Jul 29, 2026·

  ·
  1

• Report CUTLASS setup and CUDA launch-configuration failures

  #19·morluto opened Jul 27, 2026·

  ·
  1

• Validate the independent FLA 0.5 compatibility path

  #18·morluto opened Jul 27, 2026·

  ·

• Parallelize partial output-tile stores across the store warp

  #17·morluto opened Jul 27, 2026·

  ·
  1

• Tolerate one-ULP BF16 variation in correctness tests

  #16·morluto opened Jul 27, 2026·

  ·