vLLM-Omni: Fully Disaggregated Serving for Any-to-Any Multimodal Models
How do you serve a model that works across text, images, video and audio without creating bottlenecks between its components?
Any-to-any multimodal models combine components with different compute requirements, including autoregressive language models and diffusion transformers. Coordinating them efficiently is a challenge for serving systems designed around a single type of model.
Join the Nebius Science Paper Club with Han Gao, a co-founder of vLLM-Omni and co-author of the paper, to explore how vLLM-Omni breaks these complex pipelines into stages that can run independently. He’ll discuss the system’s design and share broader insights into optimizing multimodal LLM inference.
The talk will be followed by live Q&A and open discussion, moderated by Nebius R&D team.
What we’ll cover
- The challenges of multimodal serving: Why combining different model components creates coordination and performance bottlenecks.
- Inside vLLM-Omni: How a graph of interconnected stages supports any-to-any model architectures.
- Optimizing execution across stages: How independent batching, flexible GPU allocation and inter-stage data transfer help improve resource use and throughput.
Read the paper before the session →
Who should attend
ML researchers, ML engineers and inference infrastructure engineers working on multimodal models and serving systems. Students and anyone interested in multimodal AI are also welcome
About Nebius Science Paper Club
Nebius Science Paper Club is a webinar series led by Nebius researchers. Each session brings together paper authors and practitioners to discuss new ideas, research, and discoveries in AI.
The session is open to researchers, engineers, students, and anyone interested in knowledge distillation, model compression, and computer vision.