Issues

• invalidate_kv_cache() always returns True: worker methods return None, so all([]) makes the success guard dead code

  bugcommunity-request
  NVIDIA-NeMo/RL#4632 · Filter by author shivasathishs-rp opened on Oct 9, 2026

• r0.8.0: pyrefly rejects captured optional config in make_policy_factory, blocking CI

  bug
  NVIDIA-NeMo/RL#4599 · Filter by author fujial-code opened on Oct 8, 2026

• Quantized Nemotron-3 Super recipes fail on MTP quantization and vLLM refit

  bugLow Precisionsuper-v3t-mcoretestingvllm
  NVIDIA-NeMo/RL#4585 · Filter by author sharonyu-115 opened on Oct 8, 2026

• routed_experts: honor resolve_routed_experts_dtype end to end (several paths hard-code int16)

  Memory
  NVIDIA-NeMo/RL#4578 · Filter by author ZhiyuLi-Nvidia opened on Oct 7, 2026

• Sharded NeMo Gym spinup timeout is fixed at 1800s and can expire before external_service_readiness

  bug
  NVIDIA-NeMo/RL#4565 · Filter by author yfw opened on Oct 7, 2026

• Context compaction: support trainer and rollout checkpoint/resume

  Feature
  NVIDIA-NeMo/RL#4558 · Filter by author terrykong opened on Oct 7, 2026

• Context compaction: retry a failed prompt group instead of dropping it

  bug
  NVIDIA-NeMo/RL#4557 · Filter by author terrykong opened on Oct 7, 2026
  1

• Rename the remaining DTensor-named tests and wording to Automodel

  NVIDIA-NeMo/RL#4535 · Filter by author yuki-97 opened on Oct 6, 2026
  

• NCCL-Reshard refit refactoring

  enhancementFeature
  Task

  NVIDIA-NeMo/RL#4523 · Filter by author youngeunkwon0405 opened on Oct 6, 2026
  

• Megatron LoRA refit sends a TP-misordered Mamba in_proj delta to vLLM at TP>1 (Nemotron-H)

  accuracybugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4471 · Filter by author yoavf opened on Oct 4, 2026

• DTensor LoRA refit sends unmerged base weights to vLLM when activation_checkpointing is on

  accuracybugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4470 · Filter by author yoavf opened on Oct 4, 2026

• feat(automodel): expose frozen multimodal sharding configuration

  DocumentationFeature
  NVIDIA-NeMo/RL#4455 · Filter by author sharonyu-115 opened on Oct 3, 2026
  1

  

• Track AutoModel HybridEP OOM and metric regressions in excluded recipes

  Memory
  NVIDIA-NeMo/RL#4427 · Filter by author seonjinn opened on Oct 2, 2026

• Support shared-policy multi-agent RL training on the context-compaction capture path

  Feature
  NVIDIA-NeMo/RL#4400 · Filter by author aroshanghias-nvd opened on Oct 1, 2026

• mooncake_cpu: every CPU-only Ray actor with a TQ client pins a full 64+4 GiB segment (worst case per node)

  DocumentationFeatureMemory
  NVIDIA-NeMo/RL#4377 · Filter by author terrykong opened on Sep 30, 2026
  

• Advantage estimators should declare their own input columns, kwargs, and return shape (follow-up to #4326)

  Feature
  NVIDIA-NeMo/RL#4376 · Filter by author terrykong opened on Sep 30, 2026
  

• Router replay (vLLM): use -1 instead of arange(K) for the last token's route row

  accuracybug
  NVIDIA-NeMo/RL#4355 · Filter by author terrykong opened on Sep 29, 2026
  

• refactor(single-controller): move checkpoint and rollout-recovery code into a RolloutCheckpointManager

  NVIDIA-NeMo/RL#4325 · Filter by author terrykong opened on Sep 29, 2026

• Token capture: terminal fallback can select a discarded attempt when response attribution fails

  bug
  NVIDIA-NeMo/RL#4317 · Filter by author aroshanghias-nvd opened on Sep 28, 2026
  

• exact_answer_alphanumeric_reward returns 1.0 when both the answer and the ground truth normalize to an empty string

  accuracybugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4275 · Filter by author shaurya416 opened on Sep 25, 2026
  

• Leader broadcast corrupts noncontiguous tensor values

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4262 · Filter by author Alvorecer721 opened on Sep 25, 2026
  

• Support DeepSeek V4.1 Flash

  enhancementFeature
  NVIDIA-NeMo/RL#4246 · Filter by author zpqiu opened on Sep 23, 2026
  

• Nemotron-3-Nano-30B-A3B: train/generation logprob divergence persists on vLLM 0.20

  accuracybugDocumentation
  NVIDIA-NeMo/RL#4243 · Filter by author nvegesna-netizen opened on Sep 22, 2026

• Normalize the in-loss survivor denominator once instead of rescaling gradients by G/K

  accuracyenhancement
  NVIDIA-NeMo/RL#4242 · Filter by author yuki-97 opened on Sep 22, 2026
  

• Track OPD loss refactoring and full-vocabulary configuration documentation

  Documentation
  NVIDIA-NeMo/RL#4241 · Filter by author RayenTian opened on Sep 22, 2026