🐛 Describe the bug

We have simultaneously a Metal backend and an MPS backend in PyTorch. It's not clear to me we need both of them. In particular, if we want to deploy to iOS, if you can be iOS 9+ MPS should be usable. Well, I don't know if anyone has actually tried this in reality. And there might be deployment considerations (as binary size matters for the mobile target.) We are also still using Metal in prod here at Meta.

Any comments from folks about if we should get rid of Metal for MPS? Or if we don't get rid of it, should there be better interop between the two? E.g., not having a separate Tensor type for MPS versus Metal. The Metal support in PT is very limited, and a few years ago @xta0 was complaining about how views are not done correctly; MPS does do views correctly.

cc @malfet @seemethere @kulinseth @albanD @DenisVieriu97 @razarmehr @abhudev for perspectives on MPS

cc @dreiss @xta0 @SS-JIA for perspectives on Metal

Versions

master