All pull requests

Pull requests

• feat(opd): support multi-node teacher and rollout engines in MultiEngineManager

  #450·DrRyanHuang opened Oct 17, 2026Collaborator·4 of 11 tasks·

  ·
  5

• fix(training): correct optimizer, export and resume handling

  #449·Yangruipis opened Oct 17, 2026Collaborator·3 of 11 tasks·

  ·
  6

• feat(sync): publish Codex example, image SFT, and streaming fix

  #448·SigureMo opened Oct 17, 2026Member·6 of 11 tasks·

  ·
  8

• perf(megatron): reduce SFT loss backward and actor snapshot overhead

  #447·DrRyanHuang opened Oct 16, 2026Collaborator·

  ·
  7

• [Draft]feat(gagar): agentic grader

  #446·kkyyxhll opened Oct 16, 2026Contributor·6 of 11 tasks·

  ·
  6

• docs(skills): require minimal-change reviews

  #443·Yangruipis opened Oct 16, 2026Collaborator·3 of 11 tasks·

  ·
  6

• feat(algorithms): add score centering policy loss

  #440·ldemon2333 opened Oct 15, 2026Contributor·6 of 11 tasks·

  ·
  16

• feat(dpo): support Qwen3.6 CP/PP training

  #438·Baiiiii126 opened Oct 13, 2026Contributor·

  ·
  23

• fix(rollout): clean up routers in their owning processes

  #434·ldemon2333 opened Oct 13, 2026Contributor·3 of 11 tasks·

  ·
  1

  16

• fix(mini-swe-agent): prevent false-positive R2E rewards from test

  #425·Echo-Nie opened Oct 9, 2026Contributor·1 of 7 tasks·

  ·
  18

  

• 【No.11】feat(straggler): observe-only Megatron slow-rank analysis

  Hackathon 2nd
  #423·hongwei-2026 opened Oct 8, 2026·3 of 7 tasks·

  ·
  3

  

• docs: focus READMEs on using Relax

  #407·SigureMo opened Oct 3, 2026Member·5 tasks done·

  ·
  6

• 【No.5】fix(metrics): deduplicate spec decoding metrics

  Hackathon 2nd
  #406·AllureCurtain opened Oct 3, 2026·11 of 15 tasks·

  ·
  12

  

• feat(docs): add bilingual blog

  #404·SigureMo opened Oct 2, 2026Member·6 of 11 tasks·

  ·
  28

• docs(readme): document Relax downstream projects

  #399·ShigureNyako opened Oct 1, 2026Contributor·11 tasks·

  ·
  3

• 【No.5】fix(agentic): aggregate speculative decoding metrics

  Hackathon 2nd
  #391·Pinoeer-kingxi opened Sep 29, 2026·9 of 13 tasks·

  ·
  25

  

• 【No.7】feat(lora): immutable version publication, session binding and safe reclamation

  Hackathon 2nd
  #382·ying123ww opened Sep 27, 2026·6 of 12 tasks·

  ·
  14

  

• 【No.3】feat(inference): Unified Inference Service

  Hackathon 2nd
  #381·leelingrui opened Sep 26, 2026·

  ·
  64

  

• 【No.11】Megatron 慢卡诊断与平台上报

  Hackathon 2nd
  #378·shanyulu opened Sep 25, 2026·

  ·
  

• 【No.7】feat(lora): implement immutable adapter publication and session pinnin

  Hackathon 2nd
  #377·treasevenzk opened Sep 25, 2026·

  ·
  13

  

• 【No.4】GenRM 弹性扩缩容与自动伸缩

  Hackathon 2nd
  #370·shanyulu opened Sep 24, 2026·

  ·
  124

  

• 【No.3】feat(inference): Unified Inference Service

  Hackathon 2nd
  #368·oliver-0424 opened Sep 24, 2026Contributor·

  ·
  30

  

• 【No.11】feat(straggler): low-overhead per-rank straggler profiler for the Megatron backend

  Hackathon 2nd
  #363·Mukayi opened Sep 23, 2026·6 tasks done·

  ·
  32

  

• 【No.3】feat(inference): unify Rollout, GenRM, and Teacher service

  Hackathon 2nd
  #356·Arazyu opened Sep 22, 2026·

  ·
  35

  

• 【No.5】fix(agentic): deduplicate and aggregate speculative decoding metrics

  Hackathon 2nd
  #349·wenhaolai opened Sep 21, 2026·6 of 11 tasks·

  ·
  36