is there any reason you are using frozenset for the keys?

Looking at the PR in more details, I have the following questions:

1. If we have 5 adapter steps for HF checkpoints, and 1 step that only needs to run on HF checkpoints on CPU, we would register 2 adapters, one for (source: "hf",) with 5 steps and one for (source: "hf", device: "cpu") with 5+1 steps?
2. Then, we would pick on CPU the most specific "hf"+"cpu" adapter, while on GPU we would pick the "hf" one only because there is nothing more specific?
3. Are we sure we want to use itertools.combinations? How are we ensuring we are not creating far too many combinations to the point our program slows down to a crawl? We can maybe take a look at how pytorch handles its dispatcher keys, as I think we are trying to solve the same problem here and they have done it very efficiently