Predictive Batch Composition for Generative AI: Forecast-Aware Grouping and Joint Execution of Anticipated Precomputation Tasks

The increasing computational demands of generative artificial intelligence (AI) systems have created significant opportunities for improving GPU utilization through batching and parallel execution. Conventional batching mechanisms primarily organize computational tasks according to requests that have already arrived, while emerging predictive computing approaches investigate the anticipation and advance preparation of future inference workloads. However, predicting future computational demand and preparing individual tasks do not necessarily ensure that these tasks are organized into computationally efficient execution groups. Multiple anticipated tasks may have different computational requirements, execution characteristics, memory footprints, deadlines, and expected reuse probabilities. Executing them independently may introduce avoidable scheduling overhead and underutilize available GPU capacity, whereas combining incompatible tasks may increase memory pressure, execution imbalance, and completion latency. This paper introduces Predictive Batch Composition for Generative AI (PBCC-GAI), a conceptual framework that investigates how multiple predicted future computational tasks can be intelligently grouped into batches before their corresponding user requests materialize. The proposed framework extends predictive computational preparation by introducing forecast-aware task grouping, computational compatibility estimation, batch utility prediction, resource-constrained batch formation, and adaptive batch execution. Rather than merely predicting when individual requests will arrive, PBCC-GAI estimates which anticipated computational tasks are likely to occur within compatible temporal intervals and determines which combinations can provide favorable execution efficiency. Its architecture integrates future workload prediction, candidate task generation, computational feature extraction, compatibility-aware batch composition, multi-objective batch optimization, predictive execution scheduling, and post-execution adaptation. A unified mathematical formulation is developed to maximize expected computational benefits while accounting for GPU utilization, batch execution overhead, memory capacity, task deadlines, expected reuse, execution variability, preparation expenditure, and prediction uncertainty. The framework further distinguishes between predictive precomputation batches and conventional online inference batches, recognizing that speculative workloads and actual user requests may require different scheduling priorities and reliability constraints. An experimental methodology is established to evaluate whether predictive batch composition can improve GPU utilization, reduce total computational expenditure, increase effective batch efficiency, and preserve response quality compared with individual predictive execution and conventional batching approaches. The proposed framework does not assume that larger batches are universally more efficient; rather, it investigates how the composition, timing, compatibility, and expected future reuse of computational tasks jointly determine the effectiveness of predictive batch execution. Its practical effectiveness remains subject to empirical validation.