This study investigates the role of grain-fluid interactions, and especially those related to viscous dissipation, in the generation process of landslide tsunamis. For this purpose, original laboratory experiments on the collapse of a dry granular mass into a liquid are presented, in which only the grain size and liquid viscosity are varied to modify grain-scale viscous dissipation. By following the temporal evolution of the granular and free-surface height profiles, our results first reveal a surprising wealth in dynamics of granular collapses entering the liquid. Three distinct regimes are identified, namely dense-inertial, dilute-inertial and dense-viscous regimes. These regimes are described in terms of the flow Reynolds number Re and the Stokes number St that prescribe the relative importance of flow and particle inertia, respectively, to viscous dissipation of the liquid. Whereas the two inertial regimes are characterized by relatively fast and far propagation of, either dense or dilute, granular flows once immersed, the dense-viscous regime consists of the sudden arrest of the granular material on the incline, possibly accompanied by lift-off from the front. Although the collapse dynamics and deposition are very different for each of these regimes, they barely affect the first and largest wave, which suggests that (i) the generation of the leading wave is mainly governed by the impact process depending on subaerial flow conditions, and (ii) large-amplitude waves are not necessarily associated with long depositional runouts. In fact, granular flow regimes contribute mainly to the total energy transferred to the free surface and the resulting wave train, rather than only the leading wave. Accordingly, we show that the leading wave to total wave train energy ratio follows a master curve depending only on St, for dense granular flow regimes. Altogether, these results help to understand better physical mechanisms involved in landslide-tsunami generation, for which St can varied significantly at large Re, while offering an original and relevant benchmark for numerical models.