Have you ever had to deal with training machine learning models where the data is very large? If the data does not fit in main memory, then how can you use GPUs if their memories are even smaller? Many of these cases require strategies for handling data sets. In this presentation, we will introduce DASF, a framework that brings together lazy data loading techniques using Dask, acceleration techniques using RAPIDS AI, and other techniques that facilitate the use of large data in ML pipelines locally or in HPC environments. We will also present a show case carried out with a company in the oil and gas sector.