All pull requests

Pull requests

• fix(dspine): correct training objective and add online example

  #1038·A-Gentle-Cat opened Oct 28, 2026·

  ·

• feat: add three-step MTP training

  #1037·Wuu0224 opened Oct 28, 2026·4 tasks done·

  ·

• feat(capture): detect the server capture backend by default (server_capture: auto)

  #1036·maocheng23 opened Oct 28, 2026Collaborator·3 tasks done·

  ·

• MoE drafts: the load-balance loss (stacked on #1034)

  #1035·yushengsu-thu opened Oct 28, 2026Contributor·

  ·
  1

• MoE drafts: the expert bias (stacked on #1033)

  #1034·yushengsu-thu opened Oct 28, 2026Contributor·

  ·
  1

• MoE FFN for DFlash-family drafts: the module

  #1033·yushengsu-thu opened Oct 28, 2026Contributor·

  ·
  1

• MoE FFN for DFlash-family drafts, staged: module, selection bias, aux loss (final)

  #1031·yushengsu-thu opened Oct 27, 2026Contributor·

  ·
  2

• [6/6] feat(attention): DSA and QSA draft attention training

  #1030·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [5/6] feat(attention): sparse attention and SWA interfaces

  #1029·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [4/6] feat(attention): unify visibility across dense attention backends

  #1028·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [3/6] feat(draft): residual hooks with Qwen GR and DeepSeek-V4.1 mHC

  #1027·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [2/6] refactor(draft): extract GQA/MLA attention into reusable layers

  #1026·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [1/6] refactor(draft): move draft layer components under modeling/draft/layers

  #1025·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• feat(eagle4): add config-gated EAGLE-4 draft modules

  #1024·yudigege86 opened Oct 27, 2026·2 of 4 tasks·

  ·

• Refactor draft layers and add GR/mHC residuals

  #1023·jiapingW opened Oct 27, 2026Collaborator·

  ·

• [Exp.] DSpark v2 MoE draft (stacked on #1003)

  #1022·yushengsu-thu opened Oct 26, 2026Contributor·4 tasks·

  ·
  1

• feat(dpal): add D-PAL training objectives for DFlash

  #1020·Lucas-TY opened Oct 25, 2026·

  ·

• serving: SGLang draft classes for the MoE DSpark/DFlash drafts (stacked on #1003)

  #1016·yushengsu-thu opened Oct 24, 2026Contributor·4 of 6 tasks·

  ·

• perf(mtp): run lm_head and cross-entropy only on supervised positions

  #1015·lijiayu0627 opened Oct 24, 2026·2 of 3 tasks·

  ·

• fix(mtp): Qwen3.5 MTP draft attention layout and RoPE mismatches

  #1014·lijiayu0627 opened Oct 23, 2026·2 of 3 tasks·

  ·

• tpu: online spec capture for sglang-torchtpu via sglang plugin

  #1013·junyanxu opened Oct 23, 2026·

  ·
  3

• tpu: specforge_tpu plugin for offline torch_tpu training

  #1012·junyanxu opened Oct 23, 2026·

  ·

• cli: load out-of-tree platform plugins

  #1011·junyanxu opened Oct 23, 2026·

  ·

• config: add gpt-oss-120B DSpark draft config

  #1010·junyanxu opened Oct 23, 2026·

  ·
  1

• [Feature] Add LongSpark training support for Qwen3-8B

  #1009·PFLiu1 opened Oct 22, 2026·3 of 4 tasks·

  ·