PATCHEDSERVE: A Patch Management Framework for SLO-Optimized Hybrid Resolution Diffusion Serving

The Text-to-Image (T2I) diffusion model is one of the most popular models in the world. However, serving diffusion models at the entire image level faces several problems, especially when there are multiple candidate resolutions. First, image based serving system prevents requests with different resolutions from batching together. On the other hand, requests with hybrid resolutions also indicate diverse locality features, which makes it hard to apply the same cache policy to all of them. To this end, we propose PATCHEDSERVE, A Patch Management Framework for SLO-Optimized Hybrid Resolution Diffusion Serving that provides a patch-level management strategy to gather hybrid resolution requests into batches. Specifically, PATCHEDSERVE incorporates a novel patch-based processing workflow, significantly enhancing throughput for hybrid resolution inputs. Furthermore, PATCHEDSERVE designs a patch-level cache reuse policy to fully exploit the redundancy in diffusion. In addition, PATCHEDSERVE features an SLO-aware scheduling algorithm with lightweight online latency prediction, achieving higher SLO satisfaction rates. We show that PATCHEDSERVE can achieve 30.1 % higher SLO satisfaction compared to SOTA diffusion serving system while not hurt the image quality.