Evaluating randomness in a model performance

Question

I'm evaluating the variability in performance (AUC) in the test set of a machine learning model with an intrinsic random component (xgboost). How many sources of variation should I use?

• Just replicate the analysis harnessing the model intrinsic randomness?
• Bootstrap the train set?
• Bootstrap the test set?
• Bootstrap both train and test?

Thanks