Continue LM pretraining with Huggingface - loss function clarification

Question

I'm trying to use Huggingface's tensorflow run_mlm.py script to continue pretraining a bert model, and didn't understand the following: in the above script, the model is loaded using from_pretrained and then compiled with a dummy_loss function before running model.fit(...). The dummy_loss function defined in the script ignores y_true and simply returns the mean of y_pred. Is this loss function overridden somehow? I can't understand how the script actually continues pretraining if this is the only loss function used.