推論系統
PhyAI Unifies Inference for VLAs and World-Action Models, Delivering 1.40× to 4.65× Speedups over Official Implementations
PhyAI uses model adapters to preserve each architecture’s conditioning, solver, and action semantics while sharing graph execution, kernels, caching, and multi-GPU serving. Public benchmarks cover π0, π0.5, GR00T N1.7, Cosmos3, and MiniCPM-Robot, although some comparisons do not use exactly the same numerical precision.
Embodied models often use different inference code for cloud-based reinforcement learning rollouts, offline evaluation, and on-device robot control. Even with the same checkpoint, differences in preprocessing, iterative action solvers, or output transformations can cause deployment behavior to diverge from training. Teams from Beijing University of Posts and Telecommunications, Peking University, Tsinghua University, MThreads, ModelBest, and other organizations have proposed PhyAI, aiming to cover both vision-language-action models (VLAs) and WAMs that jointly predict world states and actions through a single model execution path.
PhyAI does not force these models into the prefill/decode interface used by LLMs. Model adapters preserve conditioning inputs, iterative solver state, reusable tensors, scheduling, and action transformations. The underlying runtime shares CUDA Graphs, optimized kernels, memory and cache management, quantization, communication, and parallel execution. π0.5 can reuse a single-GPU runner for data parallelism, while Cosmos3 can combine tensor parallelism with classifier-free guidance parallelism. The open-source implementation currently supports deployments on Jetson, RTX 5090, A40, and H100 hardware, as well as multi-GPU configurations using H20s, A100s, and other accelerators.
Across 11 comparable model-and-device configurations, the paper reports speedups of 1.40× to 4.65× over official implementations. For Cosmos3-Nano-Policy-DROID running on eight H20 GPUs with CFG=2 and TP=4, latency fell from 2.46 seconds to 1.18 seconds. In a simulated reinforcement learning rollout using eight A100 GPUs, per-step time decreased by 26.5%, corresponding to an estimated training-throughput improvement of approximately 1.36×. Performance profiling also showed that π0.5’s action expert accounted for only 8.8% of FLOPs at batch size 1 but 57.2% of latency, illustrating how compute volume alone can misidentify bottlenecks in embodied models.
The team also introduces a control-time Roofline model: once inference is faster than the robot’s execution of its current action, further latency reductions may not increase control frequency and may instead provide only additional timing headroom or allow a single GPU to serve more policies. This is a more suitable metric for control systems than tokens/s alone. However, the authors acknowledge that specialized runtimes remain faster in some configurations and that certain official baselines were not evaluated at exactly the same precision. The RLinf integration also remains a simulation result. Future evaluation should examine end-to-end latency on physical robots, whether action quality matches official implementations, and whether quantization and multi-tenant scheduling can maintain control stability.