Reproduce “LLM Serving Optimization with Variable Prefill and Decode Lengths” 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
Meixuan Wang, Yinyu Ye, Zijie Zhou
Sorted-F schedules LLM requests under a fixed KV-cache budget by selecting feasible groups that balance group size against total output-token work, then admitting requests only while live memory remains feasible. In a simulated 80/20 mixed workload, it achieved average latency of 154.1 ± 2.3 seconds for 2,000 requests, versus 321.5 ± 2.0 seconds for MC-SF. This is a simulator result, not a production measurement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.