Reproduce “Model Agnostic Hybrid Sharding For Heterogeneous Distributed Inference” 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
Claudio Angione, Yue Zhao, Harry Yang, Ahmad Farhan, Fielding Johnston, James Buban, Patrick Colangelo
Nesa proposes distributed inference that runs successive model blocks across heterogeneous machines. An orchestrator selects nodes using network and hardware characteristics, and the system can rebalance shards as conditions change. Mixtral 7x8B scores 0.78 ± 0.01 at 16-bit and 0.77 ± 0.01 at 8-bit on HellaSwag. These results do not establish that topology-aware routing improves throughput, because the paper reports no routing baseline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.