Please separate orchestration and execution (Or: Packaging hell is dragging ML)
run.houseInteresting! Is it possible to mix and match GPUs from different providers e.g. AWS, GCP?
Yes! It's natively supported in the cluster object. You can specify rh.cluster(name="my-a10", instance_type="A10:1", provider="aws"), and gcp, azure, and lambda labs are also supported, and if you leave provider empty it will pick the cheapest one across whichever clouds you have set up locally (e.g. have a ~/.aws/credentials)