Beam training is regarded as an efficient beamforming strategy for Terahertz (THz) communications, where the line-of-sight propagation component is dominant in THz signal propagation. Nevertheless, the near-field effect boosts the codebook size for the beam training since each near-field beam merely concentrates the radiation in a small spatial area, rather than along a direction regardless of distance. In this paper, we propose to utilize grating lobes to accelerate the beam training process. Specifically, an explicit model is proposed for characterizing the beam patterns of the main lobe and grating lobes. Based on this model, we propose a codebook design and a fast beam training scheme for THz near-field communications, where multi-beams are generated simultaneously with only one RF chain. Numerical results show the effectiveness and superiority of using grating lobes compared to the benchmarks, with less beam training overhead while achieving almost the same rate performance as that by the perfect CSI case.