When resuming from a saved checkpoint, the train_dataloader doesn't resume from the same saved step.

[cridin1]

In the run_lm.py of CodeCompletion-token, the "step, batch" loop doesn't take into account also the saved checkpoint step.