In this work, a method for inverse design of two-dimensional honeycomb pentamode metastructures (HPM) based on the Conditional Variational Auto-Encoder (CVAE) is proposed to achieve acoustic cloaking. The parameter distribution of the perfect acoustic cloak with two-dimensional cylindrical Kohn-Shen-Vogelius-Weinstein (KSVW) mapping is first derived. The CVAE model framework is then established along with its loss function in terms of the design parameters of the HPM. The inverse design performance of the deep generative model is evaluated using a large number of random test samples based on finite element simulations, showing that the equivalent mechanical parameters obtained from inverse design are highly consistent with the target parameters of the perfect acoustic cloak. For the HPM cloak design given by the trained deep generative model, the total scattering cross section (TSCS) is significantly reduced as compared to the case without a cloak, thereby demonstrating the effectiveness of the CVAE-based inverse design of acoustic cloak.