Offline reinforcement learning aims to get a stable policy from fixed dataset. Methods in offline reinforcement learning are usually reformed from those in online environment with actor-critic architecture, suffering from the issue of distribution shift. Thus, planning-based offline reinforcement learning established on transformer architecture is brought in to solve this problem. Despite of working well in low dimensional environments, planning brings complex computation and leads to higher decision latency with action space scaled up. In this work, we introduce an architecture combining vector-quantised transformer with generative model. Utilizing generative model to produce low dimensional latent codes that represent high capacity trajectories and producing long horizon trajectories with the decoder of vector-quantised transformer makes it possible to balance performance and decision latency. The inference process is composed of two components-generating and decoding. Specifically, given a starting state, the generative model could search for actions in latent space and produce trajectories representing by latent codes. Thus, the decoder of transformer then restores latent codes to high dimensional actions. Experiments show that decision latency could be lowered strikingly in this way.