Predictive Model Loading for Generative AI: Anticipating Future Model Invocation and Proactively Preparing Model Residency

The increasing deployment of generative artificial intelligence (AI) systems has introduced increasingly heterogeneous model-serving environments in which multiple large language models (LLMs), image generation models, video generation models, speech models, embedding models, and specialized expert models may share limited computational infrastructure. Although dynamic model loading and GPU memory management enable inference systems to support multiple models without permanently allocating dedicated GPU resources to every model, conventional on-demand loading remains reactive: model parameters and associated runtime resources are prepared primarily after a request requiring the corresponding model has arrived. This behavior can introduce substantial cold-start latency, model initialization overhead, GPU memory contention, and additional data-transfer expenditure, particularly when workloads exhibit sudden changes in model popularity or require coordinated execution across multiple models. This paper introduces Predictive Model Loading for Generative AI (PML-GAI), a conceptual framework that investigates whether future model invocation requirements can be anticipated and corresponding models can be selectively prepared before actual requests materialize. Rather than predicting only aggregate GPU utilization or future inference demand, PML-GAI estimates which specific models, model families, model versions, adapters, and computational configurations are likely to be required within a future time horizon. Its architecture integrates model invocation forecasting, model selection prediction, model-loading demand estimation, hierarchical model residency management, predictive weight prefetching, GPU memory allocation, model switching optimization, and adaptive model eviction. Potential prediction signals include historical model invocation sequences, application behavior, temporal workload patterns, scheduled tasks, emerging information trends, agent execution plans, and multimodal request distributions. A central component of the framework is a model-level optimization formulation that evaluates the expected benefit of preloading a model against its loading cost, memory occupancy, transfer expenditure, initialization overhead, eviction consequences, and the probability that the model will actually be invoked. The framework further distinguishes between complete model loading, partial model prefetching, CPU-memory residency, GPU-memory residency, and runtime-state preparation. These mechanisms are treated as different preparation levels rather than interchangeable operations. Particular attention is given to the coordination between predictive model loading and predictive computational orchestration, including the possibility of performing model transfers during periods of low GPU utilization or favorable resource availability. The paper also positions predictive model loading in relation to existing multi-model serving and GPU prewarming research, recognizing that workload-aware model prewarming has already been investigated. Its intended contribution is a broader model-invocation prediction framework that considers heterogeneous generative AI model catalogs, multimodal and agent-driven invocation dependencies, hierarchical residency decisions, and the joint optimization of loading probability, loading lead time, memory capacity, and economic cost. A mathematical formulation and experimental methodology are developed to evaluate whether predictive loading can reduce model cold-start latency and total serving expenditure without compromising inference throughput, memory safety, or service-level requirements. The proposed framework does not assume that predictive loading is universally beneficial; its effectiveness depends on the predictability of model invocation, model loading characteristics, available memory capacity, and the economic value of avoiding future cold starts. Its practical effectiveness remains subject to empirical validation.