JARVIS: Disjoint Large Language Models on Radio VLANs for Intelligent Services

Large Language Models (LLMs) have changed the way we access and interpret information, communicate with each other and even operate computer systems through autonomous code generation. Typically, these billion-parameter models rely on cloud storage and execution due to their computational demands. In this paper, we challenge this status quo by proposing JARVIS, a distributed LLM framework that splits model layers across edge devices with limited compute resources, trading off computation for increased peer-level communication. JARVIS is robust to individual node failures, including recovery methods for lost layers via peer-level duplication. We evaluate JARVIS using Google’s open-source Gemma LLM (2B parameters) deployed over 18 software-defined radios in the NSF Colosseum RF emulator. Our evaluation explores LLM performance degradation from node losses, providing insights into node prioritization in tactical environments. The JARVIS software code is released for community exploration and adoption.