Reproduce “Kafila: Serving Large Language Models on a Trusted Set of Heterogeneous Commodity Machines” 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
Murtaza Rangwala, Richard O. Sinnott, Rajkumar Buyya
Kafila lets an invited group run a model across machines that lack sufficient memory individually by forming a ring through NATs, measuring device capabilities, and assigning contiguous model blocks and the model head accordingly. On a four-device LAN, Qwen3-8b reached 21.5 ± 0.2 tokens/s, versus 13.7 ± 1.3 for uniform allocation. Compute balance does not remove network bottlenecks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.