The aim of this work is to propose an hybrid implementation of a semi-Lagrangian particle method on a multi-CPU and multi-GPU architecture. The applications we have in view deal with the transport of a passive scalar in a turbulent flow. For high Schmidt numbers (ratio of flow viscosity to scalar diffusivity), these problems exhibit two different scales: one related to the flow and the other -a smaller scale - to the scalar fluctuations. This scale separation motivates the use of hybrid methods where scalar and flow dynamics can be solved with different algorithms and at different resolutions. The coupling between these scales is done through the velocity field.