<p>Achieving perfect channel state information at the transmitter (CSIT) in vehicle-to-everything (V2X) communication is often impractical due to the dynamic nature of vehicular environments and the inherent feedback and processing delays. To address this challenge, we propose a novel roadway-geometry-aware predictive beamforming design for rate-splitting multiple access (RSMA)-enabled V2X. Our approach designs an RSMA-based large multi-modal model (LMM) to enhance the fairness and robustness in complex V2X-enabled traffic systems. This innovative methodology employs spatial-temporal traffic data to significantly enhance the quality and performance of beamforming in V2X communications. Specifically, we first represent the complicated road-geometry data by a connected lane graph, and then extract the spatial features via a graph convolutional network (GCN) module. Meanwhile, we extract the temporal dependencies between the historical CSIT and the desired beamformer by leveraging a Transformer encoder module. Moreover, we incorporate road semantic information as an additional input to the decoder module, enabling the generation of a more context-aware beamforming design. Simulation results demonstrate that our proposed LMM outperforms the conventional deep learning approach and optimization approach. It significantly improves the effectiveness and robustness of V2X communication in diverse traffic conditions.</p>