Tensorflow MirroredStrategy loading the full dataset into GPUs instead of only batches

Question

I am following the demo code and documentation from here: https://keras.io/guides/distributed_training/

Where it seems very simple: I create and compile a model under the