Large Model Inference Task Decomposition and Edge Collaborative Computing: A New Intelligent Scheduling Scheme in WiFi Offloading Networks

This article proposes a large model inference task decomposition and edge collaboration framework for resource-constrained wireless devices. Using an LLM planner to achieve subtask difficulty prediction and dynamic scheduling, it delivers significant results in WiFi network environments: 20% lower latency and an 80% increase in overall gain, providing key technical references for efficient deployment of large models in edge scenarios.

With the improvement of large language model capabilities, AI deployment on mobile terminals has become an industry trend. However, resource-constrained devices face bottlenecks in computing power and energy consumption when running inference directly. Traditional binary offloading strategies in edge computing struggle to adapt to the characteristics of large model inference, such as heterogeneous capabilities, semantic correlations, and uncertain output lengths. In WiFi environments, channel competition, multi-user scheduling, and task semantic relevance further increase the difficulty of offloading decisions. The core problem to solve is ensuring inference quality while minimizing end-to-end latency.

This study focuses on WiFi scenarios with multiple users and multiple edge nodes. Inference tasks can choose from three execution modes: local execution (low latency but high computing power requirements), full offloading (relying on stable wireless connections), and decomposed collaboration (local-edge collaboration). Decisions need to consider heterogeneous node computing power differences, dynamic changes in wireless links, subtask dependencies, and communication overhead. Additionally, the difficulty in predicting output length affects the accuracy of latency prediction.

The core of the framework is an LLM intelligent planner with dual prediction capabilities: subtask difficulty inference (estimating computational load by analyzing input complexity) and output length prediction (predicting token count based on semantic context). Based on the prediction results, a decomposition-aware joint scheduling strategy is designed to optimize subtask allocation, execution order, and result aggregation as a whole, satisfying constraints such as WiFi bandwidth competition, edge queue waiting, and node computing power availability.

Simulation experiments cover different network topologies, user distributions, and task loads, comparing against baselines of pure local execution and nearest edge offloading. The framework achieves a better latency-accuracy trade-off, with an average latency reduction of 20% and an overall gain increase of 80%. The lightweight planner transfers large model capabilities via knowledge distillation, maintaining performance while being suitable for edge deployment.

Insights: Large model inference offloading should leverage task decomposability for fine-grained optimization, and LLMs can act as decision participants (AI for AI). Future directions: Explore scalability for complex network topologies, reinforcement learning-based adaptive online scheduling, and application to more generative AI tasks. As edge computing power grows and communication technologies evolve, this scheme will play an increasingly important role.