GENSERVE: Efficient Co-Serving of Heterogeneous Diffusion Model Workloads

2026-04-05

AI Summary

This work addresses the challenge of co-serving heterogeneous diffusion models—such as text-to-image (T2I) and text-to-video (T2V)—on shared GPU clusters, where significant differences in computational demands, parallelization characteristics, and latency requirements often lead to severe violations of service-level objectives (SLOs). Leveraging the step-level predictability and preemptibility inherent in diffusion inference, the authors propose a novel co-optimization framework that integrates intelligent video preemption, elastic sequence parallelism, dynamic batching, and an SLO-aware scheduler to enable efficient resource allocation. This approach establishes a new paradigm for serving diverse generative workloads, achieving up to a 44% improvement in SLO compliance over the strongest baseline across various configurations while substantially enhancing both resource utilization and service quality.

Abstract

Diffusion models have emerged as the prevailing approach for text-to-image (T2I) and text-to-video (T2V) generation, yet production platforms must increasingly serve both modalities on shared GPU clusters while meeting stringent latency SLOs. Co-serving such heterogeneous workloads is challenging: T2I and T2V requests exhibit vastly different compute demands, parallelism characteristics, and latency requirements, leading to significant SLO violations in existing serving systems. We present GENSERVE, a co-serving system that leverages the inherent predictability of the diffusion process to optimize serving efficiency. A central insight is that diffusion inference proceeds in discrete, predictable steps and is naturally preemptible at step boundaries, opening a new design space for heterogeneity-aware resource management. GENSERVE introduces step-level resource adaptation through three coordinated mechanisms: intelligent video preemption, elastic sequence parallelism with dynamic batching, and an SLO-aware scheduler that jointly optimizes resource allocation across all concurrent requests. Experimental results show that GENSERVE improves the SLO attainment rate by up to 44% over the strongest baseline across diverse configurations.

Problem

• diffusion models
• heterogeneous workloads
• co-serving
• latency SLOs
• GPU clusters

Innovation

• diffusion models
• co-serving
• step-level preemption
• elastic sequence parallelism
• SLO-aware scheduling