Issues

• [RFC] Ascend A5 support in verl-omni

  verl-project/verl-omni#773 · Filter by author mengchengTang opened on Oct 18, 2026

• [reward] whisper_cer scorer cannot run on non-CUDA accelerators: --device auto resolves to CPU and trips the float16 guard

  verl-project/verl-omni#768 · Filter by author li-lizhe opened on Oct 16, 2026

• [RFC] Call for contributions: DiffusionOPSD & AWM for diffusion post-training

  enhancementhelp wanted
  verl-project/verl-omni#763 · Filter by author NancyFyong opened on Oct 16, 2026
  2

• [Bug] Flaky ci-core EADDRINUSE in test_diffusion_agent_loop: ephemeral rendezvous port race in standalone rollout worker groups

  bug
  verl-project/verl-omni#759 · Filter by author zhtmike opened on Oct 15, 2026

• [RFC] Video Generation RSI: Long-Horizon Story-Driven Video Generation on MiniMax-H3

  verl-project/verl-omni#748 · Filter by author HaFred opened on Oct 13, 2026
  1

• [RFC] Add DiffusionART (aggregation-first reward tilting) for diffusion RL

  verl-project/verl-omni#747 · Filter by author KeWang0622 opened on Oct 13, 2026
  

• [RFC] Speed up AVQA Megatron RL by reusing actor log-probabilities

  verl-project/verl-omni#746 · Filter by author hbhflw2000 opened on Oct 13, 2026

• [RFC] Add single-turn AlphaGRPO for BAGEL

  verl-project/verl-omni#728 · Filter by author ruiling-smartbear opened on Oct 10, 2026
  6

  

• [RFC] A Layered Reward Scorer Interface for verl-omni

  verl-project/verl-omni#723 · Filter by author zhtmike opened on Oct 9, 2026
  12

  

• [RFC] Call for contributions: RL & distillation algorithms for diffusion & omni post-training

  enhancementhelp wanted
  verl-project/verl-omni#717 · Filter by author zhtmike opened on Oct 8, 2026
  10

• [RFC] SkipManager for every RL pipeline stage in verl-omni: generalize rollout skip to reward, log-prob, ref, actor update, and weight sync

  verl-project/verl-omni#716 · Filter by author knlnguyen1802 opened on Oct 8, 2026

• [RFC][tracking] Fully async optimization for verl-omni: keep-mode weight sync, denoise-step resume, streaming diffusion V1

  verl-project/verl-omni#715 · Filter by author knlnguyen1802 opened on Oct 8, 2026

• [RFC] Optimize video payload transfer in the diffusion V1 trainer: trajectory field groups, worker-side fetch, single-flight reward

  verl-project/verl-omni#714 · Filter by author knlnguyen1802 opened on Oct 8, 2026

• [RFC][tracking] TransferQueue V1 optimization across verl and verl-omni: puts, reads, red writes, poll latency

  verl-project/verl-omni#713 · Filter by author knlnguyen1802 opened on Oct 8, 2026

• [RFC] Improve fully async profiling in verl-omni

  verl-project/verl-omni#712 · Filter by author knlnguyen1802 opened on Oct 8, 2026

• [Bug] Offline MLLM DPO truncation drops multimodal placeholders but keeps their features

  verl-project/verl-omni#707 · Filter by author OnePunchMonk opened on Oct 7, 2026

• [bug] Offline MLLM DPO dataset rejects prompts that reuse the same media file

  verl-project/verl-omni#705 · Filter by author OnePunchMonk opened on Oct 7, 2026
  1

• [bug] Offline MLLM DPO collator returns an object array when one branch lacks a tensor key

  verl-project/verl-omni#704 · Filter by author OnePunchMonk opened on Oct 7, 2026
  1

• [RFC] VeRL-Omni Q4 Road Map - request collection & dicussion phase

  verl-project/verl-omni#700 · Filter by author SamitHuang opened on Oct 7, 2026
  12

  

• [RFC] LoRA refactor: one general adapter abstraction, no per-model patches

  verl-project/verl-omni#685 · Filter by author knlnguyen1802 opened on Sep 30, 2026
  1

  

• [Bug] Multimodal video payloads cross-contaminated between concurrent requests in multi-replica (2-node) deployments — vLLM-Omni 0.28.0rc1

  bug
  verl-project/verl-omni#660 · Filter by author zyxzyxzyx opened on Sep 23, 2026
  4

• choice_reward.compute_score scores empty/malformed responses as correct when ground_truth has no <answer> tag

  verl-project/verl-omni#643 · Filter by author OnePunchMonk opened on Sep 20, 2026
  1

• [Feature] Report per-sample progress inside a diffusion training step

  verl-project/verl-omni#638 · Filter by author ruiling-smartbear opened on Sep 20, 2026

• [RFC] 0-diff train-rollout log-probs for the Qwen3-Omni thinker

  enhancement
  verl-project/verl-omni#637 · Filter by author cr-gao opened on Sep 20, 2026
  5

  

• [RFC] Add GPU and NPU profiling support to the diffusion V1 trainer

  verl-project/verl-omni#632 · Filter by author SelfDLer opened on Sep 18, 2026