PhysisServe speeds up autoregressive denoising serving for world models

👋 Hello, X. We’re Physis. AI’s next frontier is the physical world. We’re building a general-purpose world foundation model—one that can accumulate and reuse experience across scenarios to understand, predict, and act. 🎯 Our goal is general physical intelligence, with applications across embodied AI, industrial simulation, and scientific prediction. This brings us to a foundational question: 🤔How can we build a foundation for physical intelligence that accumulates and reuses experience across scenarios—and understands, predicts, and influences the real world at the speed and accuracy each task requires? Over the coming months, we’ll share our thinking, models, and systems as we work toward this goal. Today, we begin with infrastructure. ✨ Together with vLLM, we’re introducing PhysisServe. Born at UC Berkeley’s Sky Computing Lab, vLLM now powers inference on an estimated 500,000 GPUs worldwide at any given time. Together, we’re advancing efficient, scalable serving for the next frontier: physical AI. Three core designs make this possible: • Stage-wise pipelining across denoising steps and Transformer blocks. • Time-sharing and batching to improve GPU utilization. • Virtual ranks and joint scheduling to share resources across streams. 📊 To explore how world action models (WAMs) and action-conditioned world models (ACWMs) can sustain higher output rates as they scale, we benchmark PhysisServe on five backbones, including MiniMax H3 33B. On 200 GPUs, this 33B backbone delivers a chunk every 4.083 ms—about 245 chunks/s.* The figure below shows how additional compute enables higher sustained output rates across model sizes. *DiT denoising only; excludes VAE decoding. Output interval is not end-to-end response latency. 👉 Read our first blog post on infra: https://physis-ai.com/blog/physisserve/en 👉 Explore the code: https://github.com/Physis-AI/PhysisServe