Reproduce “HETEROVLM: Edge Multi-Request Vision-Language Model Serving on a Heterogeneous Versal SoC with AI Engines and Programmable Logic” 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

Zhiheng Chen, Yu-An Chou, Yifan Zhang, Ye Qiao, Saptarshi Mitra, Sitao Huang

HETEROVLM serves concurrent vision-language requests by pausing work at short boundaries, saving progress in memory, and running urgent requests first. It shares decode steps so requests reuse model-weight reads. During prefill of a 4,088-token text prompt, an arriving urgent camera question has first-token time 1.05 times idle, versus 6.63 and 6.75 without preemption; this is one measured case, not a workload-wide guarantee.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.