All pull requests

Pull requests

• fix(grpo): group advantages by prompt identity, not prompt tokens

  #4495·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• feat(nemo-gym): optionally mask unusable rollouts instead of failing the batch

  Documentation
  #4494·linj-glitch opened Oct 5, 2026·4 tasks done·

  ·
  1

• chore(async): log the stream error on each NeMo-Gym batch retry

  #4493·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• perf(nemo-gym): add decode_prompt_strs knob to skip prompt_str decoding

  #4492·linj-glitch opened Oct 5, 2026·4 tasks done·

  ·
  1

• fix(nemo-gym): detect context exhaustion in multi-turn hit_max_tokens

  #4491·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(grpo): exclude masked rows from the GRPO baseline/std

  #4490·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(nemo-gym): honor top-level mask_sample on rollout results

  #4489·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(ray.sub): end the job when the head outlives the driver

  Documentation
  #4488·linj-glitch opened Oct 5, 2026·4 tasks done·

  ·
  1

• fix(reward-shaping): measure overlong length over all assistant turns

  #4487·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(vllm): cancel abandoned chat completion requests on client disconnect

  #4486·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(venvs): retry uv sync on transient failures

  #4485·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• feat(packing): give sequences longer than bin capacity their own bin

  #4484·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(utils): TimeoutChecker budgets for the slowest step and upcoming validation

  #4483·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• feat(nemo-gym): opt-in straggler cutoff for synchronous Gym rollouts

  Documentation
  #4482·linj-glitch opened Oct 5, 2026·4 tasks done·

  ·
  1

• fix(megatron): make the NCCL collective timeout configurable

  #4481·linj-glitch opened Oct 5, 2026·3 of 4 tasks·

  ·
  1

• fix(recipe): stabilize Qwen3 235B MXFP8 rollouts

  cherry-pick-to-main
  #4480·humairafirdowse18 opened Oct 5, 2026Contributor·4 tasks done·

  ·
  1

• fix(vllm): stabilize Qwen3 30B and 32B MXFP8 rollouts

  cherry-pick-to-main
  #4479·humairafirdowse18 opened Oct 5, 2026Contributor·3 tasks done·

  ·
  1

• feat(distributed): print compact actor placement layouts

  CI:L0Documentation
  #4477·terrykong opened Oct 4, 2026Collaborator·2/3·

  ·
  3

• build(vlm): pin the validated video R3 dependency stack

  CI:L2
  #4476·ehosseiniasl opened Oct 4, 2026Contributor·

  ·
  5

  

• test(vlm): validate the SA-V Gym integration

  CI:L1
  #4475·ehosseiniasl opened Oct 4, 2026Contributor·

  ·
  5

  

• fix(grpo): stabilize Gym routing for video RL

  CI:L1
  #4474·ehosseiniasl opened Oct 4, 2026Contributor·

  ·
  1

  

• feat(vlm): materialize video frames for RL and MOPD

  CI:L1
  #4473·ehosseiniasl opened Oct 4, 2026Contributor·

  ·
  1

  

• feat(vlm): support Super video RL data and model loading

  CI:L1
  #4472·ehosseiniasl opened Oct 4, 2026Contributor·

  ·
  1

• fix(megatron): stream generate_async without .options() so Ray 2.58 accepts it

  CI:L1
  #4468·terrykong opened Oct 4, 2026Collaborator·

  ·
  2

• cp: refactor(config): centralize the outdated-config checks in one module 4271

  cherry-pickCI:L1Documentation
  #4467·svcnvidia-nemo-ci opened Oct 4, 2026Contributor·

  ·
  2