Issues

• 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
  

• GRPO baseline groups rows by prompt text instead of by group id (single controller)

  accuracybug
  NVIDIA-NeMo/RL#4371 · 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-request
  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
  

• DTensor v2 fails to load Nemotron-H checkpoints (backbone.embedding vs backbone.embeddings)

  bug
  NVIDIA-NeMo/RL#4211 · Filter by author yuki-97 opened on Sep 19, 2026
  3

  

• Evaluation-path processors render a duplicated default system block (two chat-template calls)

  bugcommunity-request
  NVIDIA-NeMo/RL#4185 · Filter by author Alvorecer721 opened on Sep 17, 2026
  

• Support generation shard restarts with Mooncake data-plane checkpointing

  Feature
  NVIDIA-NeMo/RL#4178 · Filter by author zyzhou5 opened on Sep 17, 2026

• [Feature] Provide a validated Megatron GRPO recipe for the original Nemotron 3.5 Super VL checkpoint

  Feature
  NVIDIA-NeMo/RL#4175 · Filter by author xinyu-dev opened on Sep 17, 2026
  

• [Bug] select_teacher can pick different teachers after changing only the microbatch size

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4173 · Filter by author gss10282023 opened on Sep 17, 2026
  

• [Bug] Dynamic CE–KD scaling changes when the same training batch is split into microbatches

  accuracybugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4172 · Filter by author gss10282023 opened on Sep 17, 2026
  

• [Feature] Support R3 router replay with FSDP2/AutoModel and NeMo Gym

  DocumentationFeature
  NVIDIA-NeMo/RL#4170 · Filter by author xinyu-dev opened on Sep 17, 2026
  

• [Feature] Add tiled output projection and log-probability backward for FSDP2/AutoModel GRPO

  FeatureMemory
  NVIDIA-NeMo/RL#4169 · Filter by author xinyu-dev opened on Sep 17, 2026

• [Enhancement] Validate 32K Nemotron FSDP2 dynamic batching beyond the first update

  FeatureMemory
  NVIDIA-NeMo/RL#4168 · Filter by author xinyu-dev opened on Sep 17, 2026

• [Bug] Packed Nemotron sequences attend across trajectory boundaries in FSDP2/AutoModel

  bug
  NVIDIA-NeMo/RL#4167 · Filter by author xinyu-dev opened on Sep 17, 2026
  1

• [Feature] Support activation CPU offload with FSDP2/AutoModel

  FeatureMemory
  NVIDIA-NeMo/RL#4166 · Filter by author xinyu-dev opened on Sep 17, 2026