All pull requests

Pull requests

• Fix eagle3 offline loss masking

  #1049·cih9088 opened Oct 29, 2026Contributor·2 of 6 tasks·

  ·

• Extract the Qwen3.5 native MTP checkpoint adapter

  #1048·jiapingW opened Oct 29, 2026Collaborator·

  ·

• Preserve capture layers for unchanged DFlash depth

  #1047·jiapingW opened Oct 29, 2026Collaborator·

  ·

• [5/5] Migrate draft settings and native recipe names

  #1046·jiapingW opened Oct 29, 2026Collaborator·

  ·

• [4/5] Add canonical draft settings and serving serialization

  #1045·jiapingW opened Oct 29, 2026Collaborator·

  ·

• [3/5] Organize draft models and reusable components

  #1044·jiapingW opened Oct 29, 2026Collaborator·

  ·

• [2/5] Define typed draft model options

  #1043·jiapingW opened Oct 29, 2026Collaborator·

  ·

• [1/5] Unify draft config loading

  #1042·jiapingW opened Oct 29, 2026Collaborator·

  ·

• Add the MiniMax-M3 DSpark training recipe

  #1039·subhajitdchow opened Oct 28, 2026·2 of 6 tasks·

  ·

• 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·

  ·

• 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·

  ·
  12

• 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·

  ·