vLLM 官方博客T1·Open Source
62 pts
VeRL-Omni released v0.3.0, extending asynchronous execution to more multimodal RL workloads and making the V1 trainer the default. The release unifies generation, reward evaluation, policy updates and weight synchronization across image, video, audio and multimodal understanding post-training, with end-to-end recipes for MiniMax H3 and MiniCPM-o 4.5.