Please separate orchestration and execution (Or: Packaging hell is dragging ML)

Interesting! Is it possible to mix and match GPUs from different providers e.g. AWS, GCP?