S U M M A R YNumerical modelling is a useful tool to understand the role of different parameters (site geometry, impedance contrast, material properties, constitutive model, input motion) governing site effects. We focus here on the 2-D P-SV seismic wave propagation in a simple-shaped asymmetric model. We consider two ways of describing the sedimentary infilling properties: the basin is either composed of distinct homogeneous geological layers or is described through properties progressively changing as a function of depth (smooth model). In order to study the wave propagation in the two basins with similar soil properties, P and S velocities of the layered basin are deduced from the smooth basin ones, making both models compatible in a kinematic point of view. P and S velocities globally increase with depth. Following EPRI, non-linear properties are considered as constant in layers for the two basins. We assess the influence of the model choice on the wave propagation while taking into account linear and non-linear constitutive models. We test the influence of the input motion by propagating two different input motions, the first being a simple impulsive one and the second a real complex accelerogram. We find that transfer functions computed in each basin are quite similar irrespective of the soil constitutive behaviour and the input motion. Moreover, we show that the maximum shear strain spatial distribution depends on the input motion frequency content and maximum values depend on the input motion strength and complexity. Maximum shear strains are generally higher for superficial layers than for deeper ones. We show that the highest shear strain values are located above discontinuities, where velocities are lower and the soil is more non-linear than in the underlying layer, regardless of the input motion. Indeed, in a layered basin, the shear strain is higher above interfaces for all the soil constitutive models that were used. In the smooth visco-elastoplastic model, a kind of layering appears, absent in a viscoelastic model, due to the different non-linear properties in each layer. Moreover, no matter what the model structure and the input motion used, maximum shear strain levels are higher in the visco-elastoplastic soil than in the viscoelastic one. Such a numerical study helps to understand the non-linearity location and therefore the consequences of a model choice when performing a site-specific study.