All pull requests

Pull requests

• feat(token-capture): read Gym manifests directly from local ledgers

  CI:L0Documentation
  #4247·pthombre opened Sep 23, 2026·

  ·
  5

• fix(sc): checkpoint monotonic NeMo Gym task IDs

  CI:L0
  #4245·mehraakash opened Sep 22, 2026·

  ·
  8

• feat(ppo): stream colocated policy updates in SingleController

  CI:L2Documentation
  #4244·youngeunkwon0405 opened Sep 22, 2026·

  ·
  5

  

• fix: routed-expert range check and default route in int64 for uint16 ids

  community-request
  #4240·zhiyuanma-agents opened Sep 22, 2026·

  ·
  1

• docs: add NeMo Gym environment guidance

  community-requestDocumentation
  #4239·judyheflin opened Sep 22, 2026·

  ·
  1

  1

• fix: Mask generated media toks and pipe style configs to MInf.

  CI:LfastDocumentationinferencetraining
  #4238·cspades opened Sep 21, 2026·

  ·
  1

  

• feat(dynamo): forward multimodal rollout data

  #4237·pstjohn opened Sep 21, 2026·

  ·
  1

• feat(recipes): add async SpatialClaw PivotRL

  CI:L1Documentation
  #4236·ehosseiniasl opened Sep 21, 2026·

  ·
  3

• feat(nemo-gym): support async SpatialClaw GRPO

  CI:L1Documentation
  #4235·ehosseiniasl opened Sep 21, 2026·

  ·
  3

• fix: expose grouped expert LoRA adapter sharing

  CI:Lfastcommunity-requestDocumentation
  #4234·romitjain-sarvam opened Sep 21, 2026·

  ·
  5

• Bump TensorRT-LLM pin from rc21 to feat/rl_rc

  #4230·hchings opened Sep 21, 2026·

  ·
  1

  

• feat(vllm): abort policy for generation log-prob validation

  community-requestDocumentation
  #4229·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• feat(model): add DeepSeek V4.1 Flash support

  Documentation
  #4228·zpqiu opened Sep 21, 2026·

  ·
  1

  2

• feat(dtensor): opt-in CUDA memory-history snapshot on worker OOM

  community-requestDocumentation
  #4227·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• feat(dtensor): configurable FSDP2 mixed-precision output dtype

  community-requestDocumentation
  #4226·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• feat(grpo): context-cost reward shaping for multi-turn rollouts

  community-requestDocumentation
  #4225·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• feat(async): FIFO target assignment for streamed prompt groups

  community-requestDocumentation
  #4224·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• feat(async): stream prompt groups from the native multi-turn collector

  community-requestDocumentation
  #4223·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• fix: chunk target logprobs without tensor parallelism (long-context OOM)

  community-requestDocumentation
  #4222·zhiyuanma-agents opened Sep 21, 2026·

  ·
  2

• fix(vllm): declare the LLM-class usage context for the async engine

  community-requestDocumentation
  #4221·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• feat(vllm): forward thinking_token_budget to vLLM SamplingParams

  community-requestDocumentation
  #4220·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• fix: report no OpenAI server as an empty list across generation backends

  community-request
  #4219·VishnuR23 opened Sep 21, 2026·

  ·
  1

  1

• fix(vllm): validate generation log-probs instead of zero-filling

  community-requestDocumentation
  #4218·zhiyuanma-agents opened Sep 21, 2026·

  ·
  1

• fix(automodel): plumb packing boundaries as top-level cu_seqlens for Nemotron

  community-requestwaiting-on-maintainers
  #4215·dundysm opened Sep 20, 2026·

  ·
  1

  2

• fix(ci): do not reuse a code snapshot that was never finished

  CI:L0
  #4214·NolenLiang opened Sep 20, 2026·

  ·
  3