MAD Max Beyond Single-Node: Enabling Large Machine Learning Model Acceleration on Distributed Systems