Do you think an additional backend layer here would be useful? For now if a model is exported to CoreML, we silently make the user download it. And the user may wish to run the XNNPack model on both platforms for whatever reason.

I'm fine with this structure to keep it simple for now, but I wonder if it's something we want to add in the future relases.

I think we should rather have models named LLAMA3_2_3B_XNNPACK etc. than adding another layer of nesting here.

Just like you said @chmjkb, this is simple for now, I can even remove the type classification (the main reason for this basic registry is that I need it right now to send proper model name to the database) and that might actually be the correct way to go as that would allow us to extend it later without any braking changes/backward compatibility worries.
As for the structure itself, we had a discussion today with Norbert and we didn't reach common ground on the structure so this is still open question and one we should discuss, I think tomorrow in the office is the best option.

Agreed.

can we jot down the conclusion from our discussion today somewhere? cc @mkopcins
I guess a good place would be #612.