BAAI introduces PhysisServe, a pipelined system for accelerating denoising inference

BAAI introduces PhysisServe, a system for autoregressive denoising inference that splits a model into stages, each covering one Transformer block at one denoising step.

BAAI says the system pipelines both denoising steps and Transformer blocks at once, so a single stream can use as many GPUs as the number of denoising steps multiplied by the number of blocks.

BAAI says that in steady state, output latency depends only on the slowest stage, and that adding denoising steps or model depth increases pipeline depth and GPU count without lowering output frequency. The report describes PhysisServe as the first of three system layers, and covers only single-stream acceleration.