Accurate demand forecasting is crucial for effective supply chain planning. However, many existing models fail to account for the complex interdependencies among products, resulting in poor predictions and costly disruptions. Current methods often focus either on temporal patterns or structural relationships but not both, which limits their effectiveness. To address this issue, we present GraphBART, a hybrid model that combines Graph Neural Network (GNN) structural reasoning with Bidirectional and Auto-Regressive Transformer (BART) temporal modeling for multi-step demand forecasting. Evaluated on the real-world SupplyGraph dataset from a major fast-moving consumer goods company in Bangladesh, the dataset includes 41 products as nodes connected by 684 edges, with daily time-series data on production, sales orders, deliveries, and factory issues. GraphBART generates node embeddings using a GNN to capture cross-product dependencies, then uses a BART encoder-decoder to forecast future demand. Our model outperforms five state-of-the-art baselines, including Long Short-Term Memory (LSTM) and Spatio-Temporal Graph Convolutional Network (STGCN), achieving the lowest MAE (1.48) and MAPE (2.59%) and the highest R2(0.90). Notably, it excels in forecasting low-demand products, where data sparsity poses challenges. This work is the first to combine graph-aware learning with transformer-based forecasting in supply chains and includes a web-based decision support system for planners to interactively explore forecasts and make informed production decisions.