Considerable advances have been made recently in applying reinforcement learning (RL) to packing problems. However, most of these methods lack scalability and cannot be applied in dynamic environments. To address this research gap, we propose a hybrid algorithm called GraphPack to solve the strip packing problem. Two graph neural networks are designed to fully incorporate the problem’s structure and enhance generalization performance. SkylineNet encodes the geometry of free space as the context feature, while PackNet, supporting the symmetry of rectangles, chooses the next rectangle to pack from the remaining rectangles at each timestep. We conduct fixed-scale, variable rectangle number and variable strip width experiments to test our method. The experimental results show that our method outperforms classical heuristic methods as well as previous RL methods. Notably, our method exhibits strong generalization ability and produces stable results even when the number of rectangles or strip width differs from that during training.