All pull requests

Pull requests

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

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

  ·
  1

• 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·

  ·
  1

• 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

• [DO NOT MERGE] test: inherit mcore deps from Megatron-LM #8462 (variant B)

  CI:L0
  #5112·kajalj22 opened Nov 6, 2026Contributor·

  ·
  4

• [DO NOT MERGE] test: inherit mcore deps from Megatron-LM #8462 (variant A)

  CI:L0
  #5111·kajalj22 opened Nov 6, 2026Contributor·

  ·
  4

• build: make Apptainer an opt-in image install

  CI:L1CIDocumentation
  #5110·kajalj22 opened Nov 6, 2026Contributor·2 of 5 tasks·

  ·
  2

• chore: upgrade Python to 3.13.18 and uv to 0.13.7

  CI:L1
  #5109·kajalj22 opened Nov 6, 2026Contributor·1 of 3 tasks·

  ·
  3

• fix(data_plane): close a failed spill's fd before publishing the failure

  #5108·ZhiyuLi-Nvidia opened Nov 6, 2026Contributor·

  ·
  1

• feat(sc): add elastic scheduler design, scheduling policy and replay tool

  community-requestDocumentation
  #5107·LoganVegnaSHOP opened Nov 6, 2026·

  ·
  1

• feat(sc): add optimizer_steps_per_refit for multi-step bursts

  community-requestDocumentation
  #5106·LoganVegnaSHOP opened Nov 6, 2026·

  ·
  1

• feat(sc): score NeMo-Gym rollouts that reach the rollout deadline

  community-requestDocumentation
  #5105·LoganVegnaSHOP opened Nov 6, 2026·

  ·
  1

• feat(sc): add per-rollout loss normalization for forest training

  community-requestDocumentation
  #5104·LoganVegnaSHOP opened Nov 6, 2026·

  ·
  1

• feat(sc): cap forest training rows per rollout

  community-requestDocumentation
  #5103·LoganVegnaSHOP opened Nov 6, 2026·

  ·
  1

• fix(generation): name the rejected entry in score centering's top-k check

  CI:Lfast
  #5102·yfw opened Nov 6, 2026Contributor·4 tasks done·

  ·
  3

• test(data): cover OpenCode's tool-image layout in captured media

  CI:Lfast
  #5101·yfw opened Nov 6, 2026Contributor·4 tasks done·

  ·
  2

• feat(single-controller): support the TRT-LLM generation backend

  #5100·shuyixiong opened Nov 6, 2026Contributor·4 tasks·

  ·
  1

• feat(gym): do not capture tokens for validation rollouts

  #5098·zyzhou5 opened Nov 5, 2026Contributor·

  ·
  1

• feat(megatron): serve NeMo-Gym from the engines behind the Dynamo frontend

  CI:LfastDocumentation
  #5097·lauradang opened Nov 5, 2026Contributor·3 of 4 tasks·

  ·
  2

• feat(research): JustGRPO with padded block attention

  CI:L1
  #5096·sajadn opened Nov 5, 2026Contributor·

  ·
  3

• feat(sc): mask NeMo-Gym rollouts by response-field rules

  CI:Lfast
  #5094·bg51717 opened Nov 5, 2026Member·

  ·
  1

• fix(nemo_gym): give flat rows a Gym episode identity on the synchronous path

  CI:Lfast
  #5093·yfw opened Nov 5, 2026Contributor·3 of 4 tasks·

  ·
  3

• feat(grpo): add full-group std, per-sample vLLM output caps and placeholder-aware prompt validation

  CI:LfastDocumentation
  #5092·vmendelev opened Nov 5, 2026·4 tasks done·

  ·
  2