In this session, we covered the latest updates from the vLLM ecosystem, followed by a deep dive into vLLM-Omni, a new effort to make omni-modal serving easy, fast, and accessible for everyone. vLLM Core Committer kicked things off with a project update, highlighting recent developments across the community and core system improvements that were released in vLLM v0.17.0, v0.18.0, and v0.19.0. We then heard from Gao Han, maintainer of vLLM-Omni, who introduced a unified approach to multimodal comprehension and generation within the vLLM project. vLLM-Omni extends vLLM’s high-performance inference engine to support text, image, video, and audio workloads, covering both autoregressive and diffusion-based architectures. The session explored how the project works today, how to get started, and what’s ahead for omni-modality serving in the vLLM ecosystem. Timestamps: 00:00 – Introduction and welcome 03:24 – vLLM community update 13:03 – What's new in vLLM v0.17.0 16:27 – What's new in vLLM v0.18.0 18:49 – What's new in vLLM v0.19.0 22:05 – Intro to vLLM-Omni: What it is, how it works, how to get started, and more 1:10:25 – Q&A and discussion Session slides: https://docs.google.com/presentation/d/1UcbvMcMM2SEr82OJLyALa43mwQy_uQmGooQEZRIEhmc/ Explore and join our bi-weekly vLLM office hours: https://red.ht/office-hours

 544      17