Recently, with the advent of deepseek, time of training and inference, and generally time processing have been considered as a major constraint. For example, to limit exchange between CPU memory and GPU memory, some architecture could be developed. However, many optimizations could be developed in CUDA manipulations. In the following description, we focus on optimisation to limit communication between CPU and GPU.