Why does Trainer.predict() overwrite drop_last of the dataloader?

[martinwimpff]

I observed the strange behaviour that the predict function of Trainer overwrites the drop_last argument of my DataLoader to False.
I know that this is useful for the majority of applications, as one typically wants to have the predictions for all datapoints. Yet, I am not sure whether Lightning should override the original datamodule. For me it is not really intuitive and kind of an odd behaviour. The fact that the test function does not have this behaviour is even stranger. In my opinion, such a "feature" should either be incorporated in both functions or neither one of them.