All pull requests

Pull requests

• cp: fix: make dpo-nanov3-30B3AB-1n8g-automodel-ep8.v2 deterministic 4619

  cherry-pickCI:LfastDocumentation
  #4740·svcnvidia-nemo-ci opened Oct 15, 2026Contributor·

  ·
  2

• test(ppo): add Qwen3-30B-A3B PPO + router replay single-controller nightly

  CI:Lfast
  #4739·bg51717 opened Oct 15, 2026Member·3 of 4 tasks·

  ·
  2

• chore: Bump dependencies to address CVES for r0.8.0

  cherry-pickCI:L1
  #4736·chtruong814 opened Oct 15, 2026Contributor·4 tasks·

  ·
  2

• feat(telemetry): add executor and event-loop diagnostics

  CI:L0Documentation
  #4732·macandro96 opened Oct 14, 2026Contributor·4 tasks done·

  ·
  2

• perf: add a bounded queued vLLM HTTP bridge

  Documentation
  #4731·arnavkomaragiri opened Oct 14, 2026·3 of 5 tasks·

  ·
  1

• feat(telemetry): export rollout checkpoint metrics to Lens

  CI:L0Documentation
  #4730·macandro96 opened Oct 14, 2026Contributor·4 tasks done·

  ·
  3

• fix(checkpoint): stop training at the next step after a failed background save

  community-request
  #4729·Thanatos9404 opened Oct 14, 2026·2 of 4 tasks·

  ·
  1

  1

• feat(xtoken): add opt-in overlength rejection and full-turn CE to collator

  #4728·avenkateshha opened Oct 14, 2026Contributor·

  ·
  1

• perf(dataplane): add opt-in native async TQ path for A/B tests

  Documentation
  #4727·zyzhou5 opened Oct 14, 2026Contributor·4 tasks·

  ·
  1

• feat(dynamo): refit, FP8 and speculative decoding via the vLLM worker extension

  CI:L1CIDocumentation
  #4725·jthomson04 opened Oct 14, 2026Contributor·4 tasks done·

  ·
  19

• feat(grpo): add deterministic async rollout scheduling

  CI:L2Documentation
  #4723·wdykas opened Oct 14, 2026Contributor·4 tasks·

  ·
  9

• feat(xtoken): share same-tokenizer KD helper; top-K from KD positions only

  #4721·avenkateshha opened Oct 14, 2026Contributor·3 of 4 tasks·

  ·
  1

• fix: fence packed refit buffer reuse across all CUDA streams

  CI:Lfast
  #4720·yuekaizhang opened Oct 14, 2026Contributor·

  ·
  8

• feat(dynamo): port to ai-dynamo 1.5.1 and run under the SingleController

  CI:L1Documentation
  #4719·jthomson04 opened Oct 14, 2026Contributor·4 tasks done·

  ·
  31

• feat: train full rollout trees with GRPO

  Documentation
  #4718·arnavkomaragiri opened Oct 14, 2026·4 of 5 tasks·

  ·
  4

• perf: add direct vLLM token capture for Gym rollouts

  Documentation
  #4717·arnavkomaragiri opened Oct 14, 2026·5 of 6 tasks·

  ·
  4

• (perf): Dblanaru/persistent compile caches

  Documentation
  #4715·DanBlanaru opened Oct 13, 2026·3 of 4 tasks·

  ·
  1

• perf(megatron): parallel reads for torch_dist checkpoint loads

  Documentation
  #4714·DanBlanaru opened Oct 13, 2026·4 tasks·

  ·
  1

• perf: EXTERNAL_VLLM_LB_CONTAINER runs the load balancers from a small image

  #4713·DanBlanaru opened Oct 13, 2026·4 tasks·

  ·
  1

• fix(grpo): port Gym log dir fix (#3956), warn on overridden output dirs

  #4712·pjin-nvidia opened Oct 13, 2026Contributor·

  ·
  1

• fix(gym rollouts): mark rollouts cut by max_output_tokens as truncated

  #4711·arendu opened Oct 13, 2026Contributor·

  ·
  3

• fix(rollout): redispatch checkpoint-refused siblings independently

  CI:L0
  #4710·macandro96 opened Oct 13, 2026Contributor·

  ·
  2

• fix(docker): drop all of Ray's vendored runtime-env agent packages

  CI:L1
  #4709·jthomson04 opened Oct 13, 2026Contributor·

  ·
  2

• fix(flow-grpo): install cv2 system libs, cut OCR nightly to 240 min

  cherry-pick-to-main
  #4696·yuki-97 opened Oct 13, 2026Contributor·

  ·
  1

• NCCL-EP: one gpu per rank

  community-requestwaiting-on-maintainers
  #4690·rapatel opened Oct 13, 2026·4 tasks done·

  ·
  2