Issues

• refactor(reward-penalties): share one malformed-think-tag predicate between the echo and capture paths

  enhancement
  NVIDIA-NeMo/RL#5008 · Filter by author bg51717 opened on Oct 31, 2026
  

• data-plane: keep CUDA work on each rank's own GPU (follow-up to #4966)

  bug
  NVIDIA-NeMo/RL#5004 · Filter by author zyzhou5 opened on Oct 30, 2026

• feat(generation): add CoreWeave Dedicated Inference backend with checkpoint hot-load

  community-requestFeaturewaiting-on-maintainers
  NVIDIA-NeMo/RL#4964 · Filter by author anson627 opened on Oct 28, 2026

• Generation-node restart: replacement refit fails and router ignores new endpoints

  bug
  NVIDIA-NeMo/RL#4937 · Filter by author zyzhou5 opened on Oct 27, 2026

• bbox_giou_reward is unbounded: reversed-corner or NaN boxes score far above 1 (RefCOCO env)

  bugcommunity-request
  NVIDIA-NeMo/RL#4871 · Filter by author jimmy-holdout opened on Oct 23, 2026

• feat: generalize run_rollout_only.py into a single-controller simulator

  Feature
  NVIDIA-NeMo/RL#4860 · Filter by author yuki-97 opened on Oct 22, 2026
  

• Consider native async TQ APIs for async data-plane callers

  Speed
  NVIDIA-NeMo/RL#4848 · Filter by author zyzhou5 opened on Oct 21, 2026

• Megatron policy can't start from a builder-format Megatron Bridge checkpoint (pretrained_checkpoint.format: megatron_bridge)

  bug
  NVIDIA-NeMo/RL#4841 · Filter by author kamran-nvidia opened on Oct 21, 2026

• fix(xtoken): ordinary chat mode mis-supervises a final assistant turn with prose and tool_calls

  bug
  NVIDIA-NeMo/RL#4831 · Filter by author yuki-97 opened on Oct 21, 2026
  

• Use Gym’s exported server-type keys

  bug
  NVIDIA-NeMo/RL#4800 · Filter by author zyzhou5 opened on Oct 20, 2026

• Track Nano 3.5 PIECEWISE logprob-error tail spikes in r0.8.0 validation

  accuracybug
  NVIDIA-NeMo/RL#4777 · Filter by author fujial-code opened on Oct 18, 2026

• PPO and xtoken off-policy distillation ignore checkpointing.pretrained_checkpoint

  bug
  NVIDIA-NeMo/RL#4768 · Filter by author terrykong opened on Oct 17, 2026
  1

• refactor: consolidate Rollout Routing Replay (R3) code under nemo_rl/router_replay/

  NVIDIA-NeMo/RL#4767 · Filter by author yuki-97 opened on Oct 17, 2026

• [Bug] tilelang==0.1.11 pin in the automodel extra is x86-only; on aarch64 (GB200) the Automodel venv gets tilelang 0.1.12, tile_kernels cannot import, and DeepSeek-V4.1 training fails at the first forward

  bugcommunity-request
  NVIDIA-NeMo/RL#4754 · Filter by author linmuchuiyang opened on Oct 16, 2026

• Carry terminal truncation metadata through Gym token-capture receipts

  bug
  NVIDIA-NeMo/RL#4737 · Filter by author linj-glitch opened on Oct 15, 2026
  1

• [RFC] Smaller bit-exact refit payloads for cross-cluster sparse-delta refit (old-value-conditioned coding)

  community-requestFeatureSpeedwaiting-on-maintainers
  NVIDIA-NeMo/RL#4724 · Filter by author Shafwansafi06 opened on Oct 14, 2026

• Async replay-buffer restore doesn't re-age trajectories against resume step; restored buffer replays already-consumed trajectories

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA-NeMo/RL#4708 · Filter by author rkoivunen-sw opened on Oct 13, 2026

• Checkpoint finalize-thread errors surface one step late — training commits past a failed save

  bugcommunity-request
  NVIDIA-NeMo/RL#4707 · Filter by author rkoivunen-sw opened on Oct 13, 2026

• Unify draft embedding/lm_head ownership across Megatron and Automodel (and within Megatron)

  Feature
  NVIDIA-NeMo/RL#4691 · Filter by author yuekaizhang opened on Oct 13, 2026

• Clarify the role of NeMo RL’s built-in evaluator alongside NeMo Gym

  Feature
  NVIDIA-NeMo/RL#4688 · Filter by author cwing-nvidia opened on Oct 13, 2026

• Telemetry: per-process host memory (Ray actor/task breakdown) alongside rl.host.memory.*

  FeatureMemory
  NVIDIA-NeMo/RL#4681 · Filter by author terrykong opened on Oct 12, 2026
  1

• Centralize temperature scaling in loss_input.py instead of the per-backend train paths

  NVIDIA-NeMo/RL#4671 · Filter by author yuki-97 opened on Oct 12, 2026

• Track unmerged r0.8.0 changes (not yet merged back to main)

  NVIDIA-NeMo/RL#4666 · Filter by author yuki-97 opened on Oct 12, 2026

• Megatron token capture: the TQ media write blocks the engine step

  PerformanceSpeed
  NVIDIA-NeMo/RL#4663 · Filter by author terrykong opened on Oct 12, 2026
  1

  

• Support discard_weights_on_sleep (vLLM sleep level 2) in grpo_train_sync, single controller and async GRPO

  Feature
  NVIDIA-NeMo/RL#4659 · Filter by author yuki-97 opened on Oct 11, 2026