<p>Vehicular communication systems are responsible for providing robust and timely intervehicle communication to ensure driving safety. However, high mobility, variable topology, and high density represent great challenges for medium access control (MAC) protocols to deal with communication collisions, which are the origin of network disruption, high delay, and packet loss. The majority of proposed solutions are based on vehicle localization information and random time-slot assignment, often incurring high overhead and energy costs. To tackle these problems, we present a preventive time-slot allocation framework for hybrid RSU-assisted vehicular networks, named "PTA-MAC". The core novelty of our work is the integration of a machine learning (ML)-based sojourn time prediction model with a collision-aware time-slot assignment protocol. First, we propose a sojourn time prediction model for vehicles entering a road segment. Then, based on this prediction, our time-slot assignment protocol proactively allocates resources to decrease the probability of future collisions and resolves them if they occur. The real-time prediction and allocation for thousands of vehicles in dense networks necessitate significant computational power, highlighting the need for High-Performance Computing (HPC) at the network edge (i.e., RSUs). Simulation results, generated through large-scale parallelized scenarios, demonstrate the superior performance of the proposed framework in terms of collision rate and overhead compared to state-of-the-art protocols for VANETs.</p>