Motion Prediction is a crucial part of Autonomous technique for autonomous driving. In order to predict entire scene agents, some occupancy flow prediction methods are developed. However, these methods exist the shortcomings of incomplete information representation or low computational efficiency. Therefore, this paper proposes a lightning hierarchical convolution network for occupancy flow prediction. The multiple inputs (map, flow and occupancy) are encoded by multi-stage hrnet module and swin-transformer module separately which can learn high-resolution information. And the hierarchical convolution decoder captures multi-scale features efficiently. The proposed method is comprehensively validated on the Waymo Open Motion Dataset. Our model achieves a Flow-Grounded Occupancy AUC of 0.8184. And compared with other comparable performance models, our lightning model inference time is as low as 0.1065 s.