Multi-SKU demand forecasting poses considerable challenges for retail operations due to intricate interdependencies among products, seasonal fluctuations, and external market influences. Conventional forecasting approaches often fail to model these complexities effectively, resulting in poor inventory management and financial inefficiencies. This paper proposes a novel Cross-Modal Attention Transformer (CMAT) architecture for addressing multi-SKU demand prediction in retail contexts. The model employs a dual-attention mechanism to capture temporal patterns within individual SKU time series as well as cross-SKU relationships via an inter-product attention layer. It further integrates multimodal external variables such as promotional campaigns, weather data, and economic indicators using dedicated feature encoders. Evaluation on a large-scale retail dataset comprising 15,847 SKUs across 12 product categories shows a 23.4% reduction in Mean Absolute Percentage Error (MAPE) compared to standard transformer models and a 31.2% improvement over traditional statistical methods. The results indicate potential for improving supply chain efficiency and reducing inventory-related costs through more accurate multi-horizon demand predictions.