Reproduce “LegoDiffusion: Micro-Serving Text-to-Image Diffusion Workflows” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lingyun Yang, Suyi Li, Tianyu Feng, Xiaoxiao Jiang, Zhipeng Di, Weiyi Lu, Kan Liu, Yinghao Yu and 5 more
LegoDiffusion, developed by researchers at the Hong Kong University of Science and Technology and Alibaba Group, introduces a micro-serving architecture for Text-to-Image (T2I) diffusion workflows. This system decomposes complex T2I pipelines into independently managed microservices, achieving up to 3x higher request rates, meeting 6x more stringent SLOs, and utilizing 3x fewer GPUs compared to monolithic serving systems.
Key insights
- Decomposing T2I diffusion workflows into microservices, despite their complex iterative dependencies, can significantly improve efficiency if inter-GPU tensor communication is handled with zero-copy mechanisms.
- Fine-grained visibility into the workflow's computation graph enables advanced scheduling strategies, such as cross-workflow model sharing and adaptive parallelism, which are not possible with monolithic systems.
- A declarative programming interface coupled with a graph compiler allows for easy workflow definition and integration of system-level optimizations without altering core model logic.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.