All pull requests

Pull requests

• feat(telemetry): trace worker-side training and vision workload

  Documentation
  #5143·rrs45 opened Nov 9, 2026Collaborator·

  ·
  1

• perf(data_plane): stop the TQ controller rescanning nested shapes on clear

  Documentation
  #5141·ZhiyuLi-Nvidia opened Nov 9, 2026Contributor·

  ·
  3

• perf(checkpoint): defer rollout and trainer data-plane persistence

  #5139·macandro96 opened Nov 9, 2026Contributor·2/2·

  ·
  3

• fix(token-capture): preserve safe continuations after system updates

  Documentation
  #5138·aroshanghias-nvd opened Nov 9, 2026Contributor·

  ·
  1

• feat(data-plane): add two-phase SimpleStorage checkpoint capture

  CI:Lfast
  #5137·macandro96 opened Nov 9, 2026Contributor·1/2·

  ·
  3

• ci: re-run a test shard once on known transient errors

  CI:LfastCI
  #5136·kajalj22 opened Nov 9, 2026Contributor·4 of 5 tasks·

  ·
  13

• fix(cache): keep FlashInfer autotune results out of vLLM engine caches

  #5135·snivertynv opened Nov 9, 2026Contributor·

  ·
  1

  

• feat(telemetry): add an llm_span_breakdown_hook for Gym's model-call spans

  #5134·shuyixiong opened Nov 9, 2026Contributor·2 of 4 tasks·

  ·
  1

• docs(environments): explain where first-turn stop strings come from

  community-requestDocumentation
  #5133·sandheepp opened Nov 9, 2026·

  ·
  1

• feat(rewards): add SQuAD-normalized QA exact-match and token-F1 rewards

  community-request
  #5132·sandheepp opened Nov 9, 2026·

  ·
  1

• feat(sc): support in-loss sequence filtering for PPO

  CI:LfastDocumentation
  #5130·snivertynv opened Nov 9, 2026Contributor·

  ·
  2

  

• build(automodel): bump Automodel to 41f9262

  CI:L1Documentation
  #5129·jQizhang opened Nov 9, 2026Contributor·3 of 4 tasks·

  ·
  3

• feat(vllm): add captured Qwen3 TITO continuations

  Documentation
  #5128·fujial-code opened Nov 9, 2026Contributor·4 of 6 tasks·

  ·
  2

• feat(forest): mask rollouts whose later calls rewrite assistant history

  CI:LfastDocumentation
  #5127·bg51717 opened Nov 9, 2026Member·4 tasks done·

  ·
  11

• feat(forest): add continued_chains selection and a per-rollout row cap

  CI:LfastDocumentation
  #5126·bg51717 opened Nov 9, 2026Member·4 tasks done·

  ·
  19

• fix(router-replay): honor resolved routed_experts dtype end to end

  community-request
  #5125·saiyambharara opened Nov 9, 2026·

  ·
  1

  2

• feat(grpo): support detached MTP heads with in-loss sequence filtering

  CI:LfastDocumentation
  #5124·snivertynv opened Nov 8, 2026Contributor·

  ·
  3

  

• feat(gym): add Super VL single, multi and mixed visual coding recipes

  community-requestDocumentation
  #5122·FuxiaoLiu opened Nov 8, 2026·

  ·
  1

• feat(gym): pin matched recipes and reduce optimizer restore copies

  Documentation
  #5121·init-nikhil opened Nov 8, 2026Member·

  ·
  17

• fix(megatron): lift Ray's 1000-call cap on the Megatron policy worker

  #5120·wdykas opened Nov 7, 2026Contributor·3 of 4 tasks·

  ·
  4

• fix(data-plane): reject uint16/32/64 fields at put_samples

  #5118·ZhiyuLi-Nvidia opened Nov 6, 2026Contributor·3 of 4 tasks·

  ·
  1

• fix(gym): size every Gym actor's max_concurrency from the rollout fan-in

  #5117·jthomson04 opened Nov 6, 2026Contributor·

  ·
  2

• fix(megatron): colocated wake barrier before the engines resume

  #5116·wdykas opened Nov 6, 2026Contributor·3 of 4 tasks·

  ·
  1

• feat(dynamo)!: Dynamo 1.7 nightly, high session cap, frontends across nodes

  CI:L1Documentation
  #5115·jthomson04 opened Nov 6, 2026Contributor·

  ·
  1

• feat(examples): add Claude Code subagent training recipe

  Documentation
  #5113·aroshanghias-nvd opened Nov 6, 2026Contributor·

  ·
  1