All pull requests

Pull requests

• [OPD] Add the Qwen3.6-35B-A3B multi-teacher OPD launcher (reasoning_gym, B200/GB300)

  #4493·maocheng23 opened Oct 28, 2026Collaborator·

  ·

• Add --rollout-data-yaml-config: a validated list of training datasets to mix

  #4492·Shi-Dong opened Oct 28, 2026Collaborator·

  ·
  5

• [rollout] Let a sample lower its own response budget via Sample.max_response_len

  #4491·Shi-Dong opened Oct 28, 2026Collaborator·

  ·
  1

• [rm_hub] Let a dataset row choose its reward function via metadata.custom_rm_path

  #4490·Shi-Dong opened Oct 28, 2026Collaborator·

  ·
  5

• fix(score-centering): give empty CP ranks the fp32 log-prob dtype

  #4489·Xiuyu-Li opened Oct 28, 2026Contributor·

  ·

• fix(dsv4.1): keep THD-packed samples apart in the attention and Engram

  #4488·Xiuyu-Li opened Oct 28, 2026Contributor·

  ·

• docs: add Simplified Chinese README (README.zh.md)

  #4487·JasonYeYuhe opened Oct 28, 2026·

  ·
  3

• fix(metrics): report raw reward before dynamic sampling filters

  #4486·guapisolo opened Oct 28, 2026Collaborator·

  ·
  8

• [CI only] Run LoRA MoE tests against sglang#46894

  #4484·HJSang opened Oct 27, 2026Collaborator·

  ·
  4

• Accumulate synchronized multi-LoRA gradients per request

  #4480·yueming-yuan opened Oct 27, 2026Collaborator·1/3·

  ·

• Let torchtitan take --dump-details: dump train data from the shared torch-native loop

  #4477·HJSang opened Oct 27, 2026Collaborator·

  ·
  8

• docs: add multi-LoRA training service overview

  #4476·yueming-yuan opened Oct 27, 2026Collaborator·

  ·
  2

• fix(command_utils): make same-second run ids not collide

  #4472·HJSang opened Oct 27, 2026Collaborator·

  ·
  9

• fix(megatron): reject --allgather-cp when an attention layer assumes zigzag CP

  run-ci-megatron
  #4471·HJSang opened Oct 27, 2026Collaborator·

  ·
  1

  8

• [OPD] Allow --skip-actor-forward-only with sampled-token OPD

  #4468·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [Megatron] Split TE FusedAdam multi-tensor calls under the NVTETensor handle cap

  #4467·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [OPD] Keep a scalar reward: store teacher log-probs when scoring returns (supersedes #3266)

  #4466·maocheng23 opened Oct 27, 2026Collaborator·

  ·
  1

• fix(megatron): pass FA4 unbounded window sides as None from TE (wrong causal attention on Blackwell)

  #4465·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• [Feature Request] Sampler KL for score centering

  #4462·zhoutong-hai opened Oct 27, 2026·

  ·
  1

• feat(weight-update): add WeightBridge weight transfer mode

  #4461·SwiftSeal03 opened Oct 27, 2026·7 tasks done·

  ·
  1

• [OPD] Cut wasted teacher-scoring work: pooled session, response-only teacher logprobs, skip stale fully-async groups

  #4459·maocheng23 opened Oct 27, 2026Collaborator·

  ·

• Support disaggregated-mode in native LoRA and PP>1 in native + bridge lora

  #4458·yueming-yuan opened Oct 27, 2026Collaborator·

  ·

• weight update: colocated p2p through the local pull transport

  #4452·yueming-yuan opened Oct 27, 2026Collaborator·

  ·

• p2p: query each engine rank's published rank-info record

  #4451·yueming-yuan opened Oct 27, 2026Collaborator·

  ·

• weight update: exchange expert units only with the ranks that load them

  #4450·yueming-yuan opened Oct 27, 2026Collaborator·

  ·