Always Be Training

Working out some problems with pre-training a GPT-2 model with C4 dataset on a Google TPU v6e1.