XLM/src/trainer.py (2 lines):
	- line 468: pred_mask[0] = 0  # TODO: remove
	- line 577: if False:  # AMP checkpoint reloading is buggy, we cannot do that - TODO: fix - https://github.com/NVIDIA/apex/issues/250


XLM/src/model/__init__.py (1 line):
	- line 147: # TODO: only output when necessary - len(params.clm_steps + params.mlm_steps) > 0