* solving bug where for small epochs and large gradient_accumulation_steps we never train * black formatting * no need to change these files