Issues

• [RFC] Unified Diffusion Replay Architecture: Batched Steps, Chunked Backward, and Per-Sample Replay

  enhancement
  Tencent-Hunyuan/UniRL#615 · Filter by author CjhHa1 opened on Oct 17, 2026

• [RFC] One rollout sample-observation path: shared records, W&B + local sinks

  enhancement
  Tencent-Hunyuan/UniRL#591 · Filter by author CjhHa1 opened on Oct 12, 2026

• [Benchmark] MiniMax-H3 T2VA: UniRL vs veRL-Omni on 8xH20 (native-configuration throughput)

  Tencent-Hunyuan/UniRL#588 · Filter by author leviking98z-rgb opened on Oct 11, 2026

• [RFC] Validate full-parameter RL training for MiniMax-H3 T2VA

  enhancement
  Tencent-Hunyuan/UniRL#587 · Filter by author CjhHa1 opened on Oct 11, 2026
  2

  

• FSDP2 collective mismatch when only some DP ranks skip a 0-token AR micro-batch

  bug
  Tencent-Hunyuan/UniRL#581 · Filter by author MinhaoLi0318 opened on Oct 11, 2026

• [RFC] Single-owner loss normalization: algorithms declare units, the train stack owns the denominator

  enhancement
  Tencent-Hunyuan/UniRL#580 · Filter by author CjhHa1 opened on Oct 11, 2026
  2

  

• [RFC] Which multimodal rollout/training extension should UniRL prioritize?

  Tencent-Hunyuan/UniRL#570 · Filter by author ruiling-smartbear opened on Oct 5, 2026

• cosmos3: H20 cuBLAS workarounds are gated to torch 2.10/cu128 and never run on the pinned torch 2.13/cu130 stack

  Tencent-Hunyuan/UniRL#567 · Filter by author CjhHa1 opened on Oct 5, 2026

• Monitor full-vocabulary AR policy entropy without regularization

  Tencent-Hunyuan/UniRL#564 · Filter by author ruiling-smartbear opened on Oct 5, 2026
  

• Add bounded dynamic sampling for synchronous AR GRPO

  Tencent-Hunyuan/UniRL#561 · Filter by author ruiling-smartbear opened on Oct 5, 2026
  

• [RFC] Opt-in token-level rollout correction for replay-anchored AR GRPO

  Tencent-Hunyuan/UniRL#556 · Filter by author ruiling-smartbear opened on Oct 5, 2026

• [Feature] TempFlow-GRPO: per-step branching with noise-aware weights

  enhancement
  Tencent-Hunyuan/UniRL#551 · Filter by author ruiling-smartbear opened on Oct 4, 2026
  

• [RFC] Mixed-resolution rollouts, long-tail scheduling, sibling batching, and train/inference consistency checks

  enhancement
  Tencent-Hunyuan/UniRL#537 · Filter by author nussejzz opened on Sep 30, 2026

• [RFC] Sibling group-affinity routing for AR rollout (#40 item 1 v2, follow-up to #262)

  Tencent-Hunyuan/UniRL#529 · Filter by author BeatSeat opened on Sep 27, 2026
  1

• train(unified): UnifiedModelTrainStack skips the eval-mode anchor freeze and prepare_part that TrainStack runs

  Tencent-Hunyuan/UniRL#523 · Filter by author CjhHa1 opened on Sep 27, 2026

• train/stack/planner: underscore helpers are imported and re-exported across modules

  enhancement
  Tencent-Hunyuan/UniRL#522 · Filter by author CjhHa1 opened on Sep 27, 2026

• fix(hi3): batched chat-template tokenization encodes only the first sample

  bug
  Tencent-Hunyuan/UniRL#520 · Filter by author CjhHa1 opened on Sep 27, 2026
  2

  

• [Feature] MixGRPO-Flash: higher-order ODE solver with step compression after the SDE window

  enhancement
  Tencent-Hunyuan/UniRL#516 · Filter by author CjhHa1 opened on Sep 26, 2026

• Frozen teacher resolution can fail on one rank and leave the others hanging in the next collective

  enhancement
  Tencent-Hunyuan/UniRL#515 · Filter by author CjhHa1 opened on Sep 26, 2026
  3

  

• [vLLM-Omni] AR rollout ignores ARSamplingParams.seed; multi-replica AR engines share one RNG stream

  enhancement
  Tencent-Hunyuan/UniRL#507 · Filter by author CjhHa1 opened on Sep 24, 2026

• [RFC] Complete decay schedulers and explore TempFlow-GRPO / Flash-GRPO

  Tencent-Hunyuan/UniRL#488 · Filter by author ruiling-smartbear opened on Sep 18, 2026
  4

  

• [RFC] Optional AISimulate integration for trace-driven rollout replay and validation

  Tencent-Hunyuan/UniRL#484 · Filter by author 0z5a opened on Sep 18, 2026
  1

• [RFC] Trainside optimization roadmap: VAE decode on UP, SM-free all-gather, micro-batch scheduling, long-tail batching, embed cache, quantitative offload

  enhancement
  Tencent-Hunyuan/UniRL#483 · Filter by author nussejzz opened on Sep 18, 2026
  3

• RFC: rollout memory and context budget for agentic RL

  Tencent-Hunyuan/UniRL#481 · Filter by author Dayuxiaoshui opened on Sep 17, 2026
  1

  

• [Tracking] Audit and retire vLLM/vLLM-Omni patches for 0.28/0.29

  Tencent-Hunyuan/UniRL#473 · Filter by author leviking98z-rgb opened on Sep 16, 2026
  4