Add SDPA implementation to more models

[avishaiElmakies]

Feature request

many of the newer models have an sdpa implementation. It could be good that some of the older models get up to speed.

Motivation

this could be good to unify the API more.
and improve training\inference speed for some of the older models like OPT.

Your contribution

Might be able to help with a few. this one seems bigger than the other feature request i opened, so will probably need more help

[LysandreJik]

Hey @avishaiElmakies, this seems like a duplicate of #26350

We'd very much welcome some SDPA additions 🤗

[avishaiElmakies]

@LysandreJik, it says flash attention. So I thought it was a different thing 😅. You can close this if you consider them the same issue. I will try to work on some models in my free time.

[vasqu]

It's #28005 ;)

[avishaiElmakies]

Thanks

[LysandreJik]

Thanks @vasqu!