VeRL-Omni is a reinforcement learning training framework purpose-built for multimodal generative models, spun out of the multimodal RL work inside the popular verl project so it can evolve on its own release cadence. It targets three model families: diffusion generators for image/video/audio (Qwen-Image, Wan2.2, LTX2.3), unified understanding-plus-generation models (BAGEL, HunyuanImage-3.0), and omni-modality models that jointly handle text, image, audio, and video (Qwen3-Omni). Training runs on a vLLM-Omni rollout backend with rollout routing, batching, and embedding-cache optimizations for faster multimodal generation, alongside asynchronous multi-reward serving (HPSv3, GenRM-OCR, UnifiedReward, or a custom HTTP scorer) so reward computation overlaps the rollout phase instead of blocking it. Modular VeOmni and FSDP2 training backends support combinable USP/TP/DP parallelism, and rollout correction skips redundant logP recomputation for more stable, reproducible diffusion RL runs. On its reference Qwen-Image FlowGRPO setup the project reports roughly 25% higher end-to-end throughput than the diffusers-based flow_grpo baseline.