There are in principle four distinct ways of using parallelisation when doing machine learning. Any combination of these can speed up the whole pipeline significantly.
1) Using spark distributed processing in feature engineering
2) When the data set that you want to train your model is large and can not be fit into a single machine, you need to use libraries which can natively distribute the training. Spark ML, or Horovod are examples of such libraries
3) You can train many versions of a same model on different datasets all at once using Pandas UDF. Like training a model for many different stores, marketing campagne, sensors and so on
4) You train different models on a same data set by using parallelisation on the hyperparameter search.