The short answer is that Regression Learner uses the tic and toc commands around the model "predict" method call/s on the validation data.
Assume this results in a time of t seconds for prediction of n observations. This implies n/t observations are predicted per second. So, you can use this same method to measure the prediction time at the command line.
In practice, the reported elapsed time will vary depending on the time required to load the models and methods for prediction, other computation on the machine, etc.
