All pull requests

Pull requests

• [Transfer 6/6] Switch training to canonical configuration

  #1187·jiapingW opened Nov 14, 2026Collaborator·

  ·

• [Transfer 5/6] Migrate training recipes to canonical presets

  #1186·jiapingW opened Nov 14, 2026Collaborator·

  ·

• [Transfer 4/6] Migrate EAGLE cached attention and sequence parallelism

  #1185·jiapingW opened Nov 14, 2026Collaborator·

  ·

• [Transfer 3/6: 3/3] Remove legacy loss configuration

  #1184·jiapingW opened Nov 14, 2026Collaborator·

  ·
  11

• [Transfer 3/6: 2/3] Configure named loss terms and canonical resume

  #1183·jiapingW opened Nov 14, 2026Collaborator·

  ·
  13

• [Transfer 3/6: 1/3] Unify main and auxiliary loss execution

  #1182·jiapingW opened Nov 14, 2026Collaborator·

  ·
  13

• [Transfer 2/6] Consolidate shared numerical losses

  #1181·jiapingW opened Nov 14, 2026Collaborator·

  ·

• [MoE 3/3] Balance: The auxiliary load-balance loss

  #1180·yushengsu-thu opened Nov 14, 2026Contributor·5 tasks done·

  ·

• [MoE 2/3] Balance: Routing as a typed output and the expert-bias update

  #1179·yushengsu-thu opened Nov 13, 2026Contributor·7 tasks done·

  ·

• [MoE 1/3] MOE module: The MoE FFN module for canonical drafts

  #1178·yushengsu-thu opened Nov 13, 2026Contributor·9 tasks done·

  ·

• docs: training backend design (program, step runner, trainer) and target-execution ownership

  #1177·maocheng23 opened Nov 13, 2026Collaborator·2 of 3 tasks·

  ·
  2

• Checkpoint layout for fused MoE experts

  #1176·yushengsu-thu opened Nov 13, 2026Contributor·

  ·

• [Transfer 1/6] Migrate canonical training optimizations

  #1175·jiapingW opened Nov 13, 2026Collaborator·

  ·
  12

• [All] MoE FFN for draft models on the canonical path

  #1174·yushengsu-thu opened Nov 12, 2026Contributor·

  ·
  17

• fix(dspine): track replaced draft config

  #1173·1fanwang opened Nov 11, 2026·1 of 6 tasks·

  ·
  1

• feat: prepare the SpecForge remote development container

  #1172·mvanhorn opened Nov 11, 2026·3 of 6 tasks·

  ·
  1

• perf(dspark): normalize the loss once per optimizer window

  #1171·subhajitdchow opened Nov 11, 2026·4 of 6 tasks·

  ·
  2

• fix(mooncake): roll back keys written by a failed partial put

  #1169·Frank0415 opened Nov 10, 2026·2 of 6 tasks·

  ·
  1

  6

• fix(data): supervise an unterminated final Muse span

  #1167·subhajitdchow opened Nov 10, 2026·3 of 6 tasks·

  ·
  1

• perf: make chunked training vocabulary projections contiguous

  #1166·Lazycat1119 opened Nov 9, 2026·

  ·

• feat(training): save_on_epoch_end - pinned per-epoch checkpoints exempt from rotation

  #1081·subhajitdchow opened Nov 7, 2026·3 of 6 tasks·

  ·
  1

• ci: cache flash-attn wheels only after the ABI check, trust the workspace once

  #1080·curnane-lab opened Nov 7, 2026Collaborator·4 of 6 tasks·

  ·
  

• docs: fix dead links to plugin and offline READMEs breaking pages deploy

  #1077·curnane-lab opened Nov 7, 2026Collaborator·4 of 6 tasks·

  ·

• docs: restructure draft config design (HF-style components, single source of truth)

  #1076·maocheng23 opened Nov 6, 2026Collaborator·

  ·
  7

• MoE drafts: DeepSeek-V4.1's own DSpark draft (stacked on #1035)

  #1073·yushengsu-thu opened Nov 4, 2026Contributor·

  ·
  2