Abstract
Multimodal Large Language Models (MLLMs) offer powerful capabilities for processing and generating text, image, and audio data, enabling real-time intelligence in diverse applications. Deploying MLLM services at the edge can reduce transmission latency and enhance responsiveness, but it also introduces significant challenges due to the high computational demands of these models and the heterogeneity of edge devices. In this paper, we propose DistMLLM, a profit-oriented framework that enables efficient MLLM service deployment in heterogeneous edge environments. DistMLLM disaggregates multimodal tasks into encoding and inference stages, assigning them to different devices based on capability. To optimize task allocation under uncertain device conditions and competing provider interests, it employs a multi-agent bandit algorithm that jointly learns and schedules encoder and inference tasks. Extensive simulations demonstrate that DistMLLM consistently achieves higher long-term profit and lower regret than strong baselines, offering a scalable and adaptive solution for edge-based MLLM services.