All pull requests

Pull requests

• [Fix] Align GLM-5.3-Flash SFT supervision at role boundaries and wire clear_thinking into the VL path

  #2217·ShilohYu opened Oct 23, 2026Collaborator·

  ·

• [Fix] Register async all-to-all tensors across streams

  #2214·YanhuiDua opened Oct 22, 2026Collaborator·

  ·

• [Docs] Fix LLaVA dataset guide links in zh-CN config READMEs

  #2212·pratikgx opened Oct 21, 2026·

  ·

• [Feature] Add opt-in SFT rank memory monitoring

  #2211·YanhuiDua opened Oct 20, 2026Collaborator·

  ·
  5

• [Feature] Share GLM/Qwen lazy images, SP decode and async prefetch

  #2210·xs1997zju opened Oct 20, 2026Collaborator·

  ·

• Fix NUMA binding using GPU locality and available CPU resources

  #2209·YanhuiDua opened Oct 20, 2026Collaborator·

  ·
  15

• Avoid host synchronization in packed KDA and loss paths

  #2204·jayhenry opened Oct 18, 2026Collaborator·

  ·

• GLM-5.3-Flash: training performance and memory optimizations

  #2203·jayhenry opened Oct 17, 2026Collaborator·

  ·

• [Fix] Bound GLM53 original-image residency with per-image preprocessing

  #2202·xs1997zju opened Oct 17, 2026Collaborator·

  ·

• [Feature] Add opt-in GLM53 lazy media and SP-group decode

  #2201·xs1997zju opened Oct 17, 2026Collaborator·

  ·

• [WIP] Diagnose SP media duplication and prototype pack sharing

  #2200·xs1997zju opened Oct 17, 2026Collaborator·

  ·

• [feat] support GLM5.3-flash RL

  #2197·PengchengShi00 opened Oct 16, 2026Collaborator·

  ·
  4

• [Fix] Render GLM-5.3-Flash assistant media as placeholder markers

  #2193·ShilohYu opened Oct 16, 2026Collaborator·

  ·
  4

• [feat] Support value head and critic loss

  #2188·YanhuiDua opened Oct 16, 2026Collaborator·3/4·

  ·
  16

• [Feature] Support PPO with shared actor and critic packing

  #2186·YanhuiDua opened Oct 16, 2026Collaborator·4/4·

  ·
  25

• [Refactor] Add index-based data packing and worker-side materialization

  #2185·YanhuiDua opened Oct 16, 2026Collaborator·2/4·

  ·
  7

• [Fix] Align GLM-5.3-Flash chat rendering and SFT loss masks

  #2184·gulixin0922 opened Oct 16, 2026Contributor·

  ·

• [Feature] Add opt-in training progress diagnostics

  #2183·xs1997zju opened Oct 16, 2026Collaborator·

  ·

• feat(tools): add model normalization tooling with reference-aligned output

  #2182·caixianzhang opened Oct 16, 2026Collaborator·

  ·

• [Fix] Restore SwapAdamW state and parameter aliases after checkpoint loading

  #2175·matrix72c opened Oct 15, 2026Contributor·

  ·
  11

• [Fix] Preserve sandbox failures in evaluation results

  #2173·matrix72c opened Oct 15, 2026Contributor·

  ·
  11

• [WIP] Restore env vars after colocate/disagg weight-update tests

  #2165·PengchengShi00 opened Oct 13, 2026Collaborator·

  ·

• [Fix] Remesh FP32 ignored params onto world-wide replicas

  #2162·YanhuiDua opened Oct 12, 2026Collaborator·

  ·
  9

• [Refactor] Unify Muon matrix layout for Newton-Schulz and lr_ratio

  #2161·YanhuiDua opened Oct 12, 2026Collaborator·

  ·
  8

• test: add trainer-free AutoModel to existing HF/XT parity test

  #2152·xs1997zju opened Oct 10, 2026Collaborator·

  ·