All pull requests

Pull requests

• feat(examples): add Pixel-Reasoner image and video rollouts

  #1894·Tournesoly opened Nov 6, 2026Collaborator·5 of 9 tasks·

  ·
  1

• feat(engine): adapt CPU-staged Muon to WeftMoE

  #1892·yulangz opened Nov 6, 2026Collaborator·6 tasks done·

  ·
  1

• feat(engine): integrate WeftMoE training and inference with AdamW

  safe-to-test
  #1891·yulangz opened Nov 6, 2026Collaborator·6 tasks done·

  ·
  1

• fix(engine): map Qwen VL FSDP weights for SGLang AWEX

  #1888·77malong opened Nov 5, 2026·4 of 7 tasks·

  ·
  1

  1

• feat(engine): support Megatron score centering with TP/CP/PP

  #1884·ActuallyEdward opened Nov 4, 2026Contributor·10 of 11 tasks·

  ·
  6

• feat(data): improve offline data prep with validation, tutorials, image util tests

  #1875·DoubleCore opened Nov 3, 2026·

  ·

• fix: stabilize Arena sessions and AWEX rollout handover

  #1873·yulangz opened Nov 3, 2026Collaborator·4 of 8 tasks·

  ·
  1

• fix(reward): backport Geometry3K expression parsing to ascend

  #1869·ActuallyEdward opened Nov 2, 2026Contributor·5 of 15 tasks·

  ·
  3

• fix(infra): drop rollout results the controller stopped waiting for

  #1864·omgwht opened Oct 31, 2026Contributor·5 of 9 tasks·

  ·
  1

  1

• feat: Support Qwen35 dense fp8 RL

  #1861·sitabulaixizawaluduo opened Oct 31, 2026Collaborator·15 tasks·

  ·

• feat(dataset): add offline validation CLI and custom-dataset walkthrough

  #1852·wesssI opened Oct 30, 2026·6 of 15 tasks·

  ·
  1

  1

• fix(engine): fix Qwen-VL FSDP on transformers 5.2 and 5.16+

  #1851·SjeYinTeoIntel opened Oct 30, 2026Contributor·5 of 9 tasks·

  ·
  1

  2

• feat(tools): add offline Autopilot analysis

  #1849·bianyh opened Oct 30, 2026Contributor·8 of 9 tasks·

  ·
  1

• feat(infra): balance v1 rollout sessions across inference backends

  safe-to-test
  #1847·dingzhiqiang opened Oct 30, 2026Collaborator·7 of 9 tasks·

  ·
  2

• feat(trainer): add optional trajectory capture and offline replay

  #1842·sitabulaixizawaluduo opened Oct 29, 2026Collaborator·6 of 15 tasks·

  ·

• fix(trainer): make on-policy distillation follow the teacher

  #1838·kkli08 opened Oct 28, 2026·4 of 15 tasks·

  ·
  1

  3

• fix(engine): automatically protect FlashInfer GDN decode padding

  safe-to-test
  #1837·yulangz opened Oct 28, 2026Collaborator·4 of 6 tasks·

  ·
  3

• fix(engine): make trainer-rollout colocation work end to end

  #1835·xiazhahe opened Oct 28, 2026Contributor·3 of 15 tasks·

  ·

• feat(engine): train Qwen3.5 PointerHead decision models on Megatron

  #1810·parhamyassini opened Oct 21, 2026·6 of 15 tasks·

  ·

• perf: bound R3 routing storage and transport memory

  safe-to-test
  #1808·yulangz opened Oct 19, 2026Collaborator·5 of 15 tasks·

  ·
  8

• feat: add Score Centering off-policy drift correction

  #1797·Aression opened Oct 16, 2026·6 of 15 tasks·

  ·
  3

• feat: distinguish consumed rollout and retained training metrics

  safe-to-test
  #1796·fishcrap opened Oct 16, 2026Collaborator·8 of 9 tasks·

  ·

• feat(engine): support native Flash-Next MTP-only SFT

  stale
  #1792·dingzhiqiang opened Oct 15, 2026Collaborator·

  ·
  1

• feat(trainer): add composable actor loss objectives

  #1785·fishcrap opened Oct 14, 2026Collaborator·8 of 9 tasks·

  ·
  1

• feat: support recovery across v2 trainers

  stale
  #1783·sitabulaixizawaluduo opened Oct 14, 2026Collaborator·1 task done·

  ·
  4