All pull requests

Pull requests

• [NPU] Enable native target and DSpark draft Attention graphs

  #545·jiaran-king opened Nov 5, 2026Collaborator·9 tasks done·

  ·
  2

• [NPU] Add DeepSeek-V4.1 flash AFD role split

  #534·wangyao-i opened Nov 4, 2026·

  ·
  3

• [Skill] deploy-afd-k8s: simplify readiness, timeouts, and recipe parsing

  #533·ronenkat opened Nov 4, 2026Collaborator·

  ·
  3

• feat(npu): add MiniMax-M3 AFD role split

  #530·wangyao-i opened Nov 3, 2026·

  ·
  3

• feat(elastic): AFD elastic scaling for Ascend NPU

  #523·yzchang-plus opened Nov 2, 2026·9 tasks done·

  ·
  13

• [GPU]: Replay microbatched CUDA graphs for under-filled DP ranks

  #519·ronenkat opened Nov 1, 2026Collaborator·9 tasks done·

  ·
  7

• feat(npu): carry a pre-quantized activation over the sync A2E transport

  #506·Pingzii opened Oct 29, 2026Contributor·

  ·

• feat(npu): transport BF16 Attention-side routing to the FFN role

  #503·Pingzii opened Oct 28, 2026Contributor·

  ·
  25

• feat(gpu): run Qwen3 MoE on FastAFDAFDConnector

  #492·ischencheng opened Oct 23, 2026Contributor·9 tasks done·

  ·

• feat(npu): serve fused BF16 DeepSeek-V2 through FFN FaaS

  #491·jiaran-king opened Oct 23, 2026Collaborator·1 of 3 tasks·

  ·
  2

• feat(npu): run standalone V4 FFN with native expert ownership

  #490·jiaran-king opened Oct 23, 2026Collaborator·

  ·
  12

• feat(gpu): add FastAFD M2N connector for SM100

  ready
  #489·ischencheng opened Oct 22, 2026Contributor·9 tasks done·

  ·
  8

• test(gpu): add a real-checkpoint MoE reference probe for the FFN role

  #482·ischencheng opened Oct 22, 2026Contributor·9 tasks done·

  ·

• [Feat]: Add runtime FULL and LIGHTWEIGHT profiling for Async CAM

  #468·lirx-pd opened Oct 19, 2026Contributor·8 of 9 tasks·

  ·
  6

• [WIP] feat(npu): support MXFP8 datatype for A2E operator

  Ascendenhancement
  #444·wangyibo1005 opened Oct 12, 2026Contributor·9 tasks·

  ·

• Feat/dsv4 afd phase1 delivery

  Ascendenhancement
  #408·wenhow opened Sep 29, 2026·9 tasks done·

  ·
  5

• refactor(moe): share remote MoE runners for GPU P2P

  NVIDIA
  #406·lirx-pd opened Sep 28, 2026Contributor·9 tasks done·

  ·
  5

• [Feat]:Refactor attention remote moe

  AscendenhancementNVIDIA
  #401·lirx-pd opened Sep 28, 2026Contributor·8 of 9 tasks·

  ·
  1