All pull requests

Pull requests

• feat(wan): fused RMSNorm+RoPE+transpose Pallas producer kernel, Fixed-M optimizations, and serving stabilization

  #488·Perseus14 opened Sep 22, 2026·4/4·

  ·
  8

• perf(wan2.2): Optimize attention kernels, rematerialization, and Ulysses parallelism for WAN 2.2 Training on TPU 7x [Stacked on #470]

  #487·Toshi-31 opened Sep 21, 2026·

  ·
  6

• Fix optimizer state restoration and multi-host sharding in WAN trainer

  #484·ninatu opened Sep 16, 2026·

  ·
  13

• Run FLUX text encoders under Torchax instead of Flax

  #483·olehtika opened Sep 16, 2026·

  ·
  3

• Fix three FLUX regressions from the Flux.2-Klein onboarding

  #482·olehtika opened Sep 15, 2026·

  ·
  5

• Fix nightly and stable Docker image build pipeline

  #481·prishajain1 opened Sep 15, 2026·

  ·
  1

• Add Sparse VideoGen attention for Wan on TPUs

  #480·Ravi-VK opened Sep 15, 2026·

  ·
  27

• feat(wan): fast serving with persistent AOT caching and tuned inference recipe

  pull ready
  #479·Perseus14 opened Sep 13, 2026·3/4·

  ·
  20

  

• feat(attention): 2D Ulysses+Ring attention with Global Virtual K-Centering

  #478·Perseus14 opened Sep 13, 2026·2/4·

  ·
  16

  

• feat(attention): fixed-m splash attention kernel with dynamic bounds and safety fallbacks

  pull ready
  #477·Perseus14 opened Sep 13, 2026·1/4·

  ·
  33

  

• fix(deps): skip nvidia-cuda-cccl on macOS (no wheels available)

  pull ready
  #476·harkanwalbedi opened Sep 11, 2026·

  ·
  1

  3

• fix: prevent silent double-application of LoRA weights

  #473·WangXukang-cypher opened Sep 4, 2026·

  ·
  12

• Feat: internal 1d and 2d ring attention kernel

  #472·eltsai opened Sep 3, 2026·

  ·
  4

  

• Minor improvement to Flux models

  #457·Perseus14 opened Aug 6, 2026·

  ·
  3

• Add heads_per_tile (multi-head-per-tile) support to Ulysses ring attention

  pull ready
  #444·edgexyz opened Jul 12, 2026·

  ·
  9

• Tf records from gcs

  #213·entrpn opened Jul 28, 2025·

  ·
  1