All pull requests

Pull requests

• fix: retry checkpoint writes after transient failures

  #1917·w3lld1 opened Oct 2, 2026·

  ·
  1

• RF moments use_mask behind a flag

  #1915·polycrit opened Sep 30, 2026Contributor·

  ·

• RF float32 moments, behavior version 32

  #1914·polycrit opened Sep 30, 2026Contributor·

  ·

• Add tool to export models to AOTInductor packages via torch.export

  #1901·larissakl opened Sep 29, 2026·

  ·

• PT CTC FSA cache bypassed under CUDA graph capture

  #1899·polycrit opened Sep 29, 2026Contributor·

  ·

• PT native packed CTC mask lower bound

  #1898·polycrit opened Sep 29, 2026Contributor·

  ·

• RF packed batch norm over extra static axes

  #1892·polycrit opened Sep 27, 2026Contributor·

  ·

• RF packed batch norm statistics in float32, unused rows zeroed

  #1891·polycrit opened Sep 27, 2026Contributor·

  ·

• RF packed cu_seqlens on one device

  #1890·polycrit opened Sep 27, 2026Contributor·

  ·

• RF chunked rel-pos self-attention op

  #1874·polycrit opened Sep 24, 2026Contributor·

  ·

• RF rel-pos self-attention left context and lookahead band

  #1873·polycrit opened Sep 24, 2026Contributor·

  ·

• RF dot_attention group ranges with a Triton kernel

  #1872·polycrit opened Sep 24, 2026Contributor·

  ·

• PT Triton chunked rel-pos attention kernel

  #1871·polycrit opened Sep 24, 2026Contributor·

  ·

• RF dot_attention query and key groups

  #1870·polycrit opened Sep 24, 2026Contributor·

  ·

• RF packed monotonic RNN-T lattice and loss

  #1869·polycrit opened Sep 24, 2026Contributor·

  ·

• RF search_sorted along a dynamic axis

  #1868·polycrit opened Sep 24, 2026Contributor·

  ·

• RF packed keep and tighten static buffer sizes

  #1867·polycrit opened Sep 24, 2026Contributor·

  ·

• RF packed gather and scatter across packings

  #1866·polycrit opened Sep 24, 2026Contributor·

  ·

• RF SpecAugment mask from a running sum

  #1865·polycrit opened Sep 24, 2026Contributor·

  ·

• RF packed regap of empty sequences, capture-safe restore

  #1864·polycrit opened Sep 24, 2026Contributor·

  ·

• PT graph capture segmented mode

  #1862·polycrit opened Sep 24, 2026Contributor·

  ·

• PT engine report train steps one step late

  #1861·polycrit opened Sep 24, 2026Contributor·

  ·

• RF Loss inv norm factor cached per step

  #1860·polycrit opened Sep 24, 2026Contributor·

  ·

• PT graph capture compiled step for torch module models

  #1859·polycrit opened Sep 24, 2026Contributor·

  ·

• Native packed CTC per-sequence kernels

  #1858·polycrit opened Sep 24, 2026Contributor·

  ·
  2