Abstract: The finite-difference method with vertically variable grid enhances the computational efficiency of reverse time migration (RTM) by employing adaptive difference grids in the depth direction, which leverages the general trend of increasing seismic velocity with depth. This paper addresses two key aspects of this method: fundamental principles and stability conditions, and presents a strategy for its implementation on GPU. We then turn to the technical issues related to finite-difference RTM using vertically variable grid and validate our approach through tests on synthetic data and field data with a high-precision velocity model from a marine area in China. The results demonstrate that our approach significantly reduces GPU memory usage for RTM and enhances computational efficiency without compromising accuracy. It yields superior imaging quality for deep complex buried-hill structures compared to the ray-based prestack depth migration method.