Our work introduces a GNN-based method for learning SPAI preconditioners, enabling efficient and robust training across various datasets and outperforming traditional and learning-based approaches in accelerating Conjugate Gradient solvers on GPUs.