Predictive Model Switching for Generative AI: Anticipating Future Model Demand to Coordinate Proactive Loading, Residency, and Unloading

The increasing deployment of generative artificial intelligence (AI) systems has resulted in inference infrastructures supporting multiple models with heterogeneous capabilities, computational requirements, memory footprints, and request distributions. As the number of available large language models (LLMs) and specialized generative models increases, maintaining every model permanently resident in GPU memory becomes increasingly expensive or technically infeasible. Conventional model-serving systems commonly activate, load, swap, or unload models in response to current requests and observed resource utilization. Although dynamic model management can improve GPU memory utilization, reactive switching may introduce model-loading delays, cold-start latency, memory-transfer overhead, and unnecessary model replacement during rapidly changing workloads. Furthermore, a model that becomes necessary during an upcoming demand surge may require substantial initialization and loading time if preparation begins only after its first request arrives. This paper introduces Predictive Model Switching for Generative AI (PMS-GAI), a conceptual framework that investigates whether future model demand can be predicted and translated into proactive model lifecycle decisions before corresponding requests materialize. Instead of determining model residency exclusively from current requests, PMS-GAI estimates future model-specific demand distributions and anticipates which models should be loaded, retained, partially prepared, migrated, or unloaded within a given prediction horizon. Its architecture integrates predictive model-demand forecasting, workload-to-model mapping, model-loading time estimation, GPU memory availability prediction, proactive model activation, residency management, cost-aware unloading, and adaptive model switching. A central distinction is established between request-level model routing and predictive model switching: routing determines which model should process an incoming request, whereas predictive switching determines which models should already be available when future requests arrive. The framework develops a mathematical formulation for selecting future model-residency configurations under GPU memory, loading bandwidth, switching cost, inference latency, and service-level constraints. It introduces an expected switching utility function that considers predicted request volume, model activation costs, memory occupation, unloading costs, prediction uncertainty, and the expected latency penalties associated with unavailable models. Furthermore, a multi-stage residency architecture distinguishes GPU-resident models, host-memory-resident models, storage-resident models, and models requiring remote retrieval. This enables the framework to investigate not only whether a model should be loaded or unloaded, but also where it should be maintained before anticipated demand materializes. The proposed methodology establishes experimental comparisons between reactive model switching, frequency-based residency management, threshold-based model loading, conventional predictive preloading, and prediction-driven joint loading and unloading. Evaluation metrics include model-demand prediction accuracy, model loading latency, time-to-first-token, GPU memory utilization, switching frequency, model eviction rate, total computational cost, energy consumption, and service-level compliance. Particular attention is given to the risks of prediction errors, excessive switching, premature unloading, memory contention, and interference with active inference sessions. PMS-GAI extends predictive computing from preparing anticipated computational workloads toward anticipating and managing the availability of the models required to execute those workloads. It does not assume that predictive switching is universally beneficial; rather, it investigates the conditions under which future-aware model lifecycle management can reduce model activation overhead and improve resource utilization without compromising inference quality or service reliability. Its practical effectiveness remains subject to empirical validation.