Reproduce “PerLLM: Personalized Inference Scheduling with Edge-Cloud Collaboration for Diverse LLM Services” 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
Zheming Yang, Yuanhao Yang, Chang Zhao, Qi Guo, Wenkai He, Wen Ji
PerLLM introduces a personalized inference scheduling framework that leverages edge-cloud collaboration to efficiently serve diverse Large Language Model (LLM) requests. It achieved over 97% success in meeting personalized processing time requirements, delivered up to 2.2x higher throughput, and reduced energy costs by over 50% compared to baseline methods.
Key insights
- Modeling personalized LLM inference scheduling in dynamic edge-cloud environments as a Combinatorial Multi-Armed Bandit (CMAB) problem can lead to adaptive and efficient solutions.
- Integrating a constraint satisfaction mechanism within the UCB framework allows for real-time decision-making that explicitly ensures personalized processing time requirements and resource limits are met.
- Optimizing for both processing time requirements and energy costs simultaneously within a heterogeneous edge-cloud setup significantly enhances overall system efficiency and sustainability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.