Reproduce “BrownoutServe: SLO-Aware Inference Serving under Bursty Workloads for MoE-based LLMs” 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
Jianmin Hu, Minxian Xu, Kejiang Ye, Chengzhong Xu
BrownoutServe adapts mixture-of-experts serving by distilling groups of original experts into united experts and using latency feedback to route more tokens to them during bursts. On Qwen1.5-MoE-A2.7B-Chat across four A100 GPUs, it reduced measured SLO-violation rates and improved throughput over non-fused vLLM, including 2.07×, while gains were smaller against native fused-MoE vLLM and accuracy losses were around 5%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.