All pull requests

Pull requests

• feat: force expert balance

  #33·zongshuai1 opened Oct 27, 2026Contributor·

  ·

• build(npu): build PyTorch from source without mimalloc

  #32·saherby opened Oct 26, 2026·

  ·
  2

• refactor(moe): unify token dispatch and fix routing gradients

  #29·dazhuangzhuang1024 opened Oct 23, 2026Collaborator·

  ·
  2

• feat: support tp and etp for fsdp2 backend

  #28·dazhuangzhuang1024 opened Oct 23, 2026Collaborator·

  ·

• feat(npu): upgrade flash-linear-attention-npu-a3 to 26.9.1

  #25·ysg063 opened Oct 22, 2026Contributor·

  ·
  4

• [WIP] feat(ops): add opt-in AscendC chunk_kda provider

  #21·ChenxiQ opened Oct 21, 2026·

  ·

• Perf/packed workload balance

  #20·zongshuai1 opened Oct 21, 2026Contributor·

  ·
  1

• add mega_moe

  #19·zongshuai1 opened Oct 21, 2026Contributor·

  ·
  1

• feat(engine): add --tt_profiling_config with_shape and with_stack knobs

  #18·ChenxiQ opened Oct 21, 2026·

  ·

• feat(deepseek-v4): port model and NPU operator updates

  #17·i-zhen opened Oct 21, 2026Collaborator·

  ·

• build(image): compile fla and fla_npu from source in-image

  #16·ChenxiQ opened Oct 20, 2026·

  ·

• feat(kcp): add FLA context parallelism for Bailing KDA

  #15·Cruise20 opened Oct 20, 2026Collaborator·

  ·

• feat(dcp): route SavePlan metadata exchange over a gloo control-plane…

  #10·junejanejin opened Oct 19, 2026·

  ·
  1

• reorder tokens by torch_npu.npu_moe_re_routing

  #8·zongshuai1 opened Oct 18, 2026Contributor·

  ·
  4

• fix: guard all_tied_weights_keys access for FSDP2-wrapped models

  #7·junejanejin opened Oct 17, 2026·

  ·
  1

• feat: add Kimi K3 (Kimi Linear) model support

  #6·junejanejin opened Oct 17, 2026·

  ·
  1

  3