All pull requests

Pull requests

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

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

  ·

• feat(mooncake): enable Ascend pinned receive and loader-side H2D

  #1071·curnane-lab opened Nov 3, 2026Collaborator·6 tasks·

  ·

• fix(capture): Ascend NPU host publication hardening

  #1070·curnane-lab opened Nov 3, 2026Collaborator·6 tasks·

  ·

• data: DeepSeek-V4 thinking mode for regeneration and training

  #1069·yushengsu-thu opened Nov 2, 2026Contributor·

  ·
  1

• residual: DeepSeek-V4 mHC via pre_from_prev and a learned head

  #1068·yushengsu-thu opened Nov 2, 2026Contributor·

  ·
  1

• Add plugin-based DMTP capture and training for Qwen3.6

  #1067·jiapingW opened Nov 2, 2026Collaborator·3 of 4 tasks·

  ·

• [4/4] build: align training with the pinned SGLang revision

  #1066·jiapingW opened Nov 1, 2026Collaborator·

  ·
  2

• [3/4] capture: remove legacy SGLang patch support

  #1065·jiapingW opened Nov 1, 2026Collaborator·

  ·

• [2/4] offline: prepare target features through the capture plugin

  #1064·jiapingW opened Nov 1, 2026Collaborator·

  ·
  1

• [7/10] Define per-layer specs and component contracts

  #1061·jiapingW opened Nov 1, 2026Collaborator·

  ·

• [6/10] Extract reusable draft components

  #1060·jiapingW opened Nov 1, 2026Collaborator·

  ·

• [4/10] Require explicit draft configuration when loading weights

  #1059·jiapingW opened Nov 1, 2026Collaborator·

  ·

• [3/10] Initialize draft models from embedded YAML settings

  #1058·jiapingW opened Nov 1, 2026Collaborator·

  ·

• MoE drafts: DeepSeek-V4's own DSpark draft architecture (stacked on #1056)

  #1057·yushengsu-thu opened Oct 31, 2026Contributor·

  ·
  2

• MoE drafts: the DeepSeek-V4 family (stacked on #1035)

  #1056·yushengsu-thu opened Oct 31, 2026Contributor·

  ·
  1

• fix(checkpoint): reject weight_map shard names that leave the checkpoint directory

  #1054·li-lizhe opened Oct 30, 2026Contributor·3 tasks done·

  ·
  1

• Run CPU config tests on stacked pull requests

  #1053·jiapingW opened Oct 30, 2026Collaborator·

  ·

• [8/10] Build draft models from per-layer specs

  #1052·jiapingW opened Oct 30, 2026Collaborator·

  ·

• docs(landing): name the Ascend vendor explicitly

  #1051·Devin-DC-Huang opened Oct 30, 2026·3 of 6 tasks·

  ·
  1

• Reproduce each target's served logit transform in teacher and head logits

  #1050·maocheng23 opened Oct 29, 2026Collaborator·3 of 4 tasks·

  ·

• 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·

  ·

• [10/10] Consolidate draft settings in training recipes

  #1046·jiapingW opened Oct 29, 2026Collaborator·

  ·

• [9/10] Version draft settings and separate serving serialization

  #1045·jiapingW opened Oct 29, 2026Collaborator·

  ·
  1