VeRL-Omni v0.3.0 adds async multimodal RL training and makes V1 the default
AIVeRL-Omni v0.3.0 extends asynchronous execution to more multimodal RL workloads and makes the V1 trainer the default. The release adds recipes for MiniMax H3 audio-video post-training and MiniCPM-o 4.5 GSPO, with some legacy paths requiring explicit configuration.
AIWhy it matters.
The post gives per-prompt timing data for sync versus separate-async placement at several response lengths, useful for deciding which execution mode fits a workload.
Source details
vLLM Blog · VeRL-Omni Team · vllm.aiPublished · added here
Original title: VeRL-Omni v0.3.0: Async Training and Broader Multimodal Post-Training