How do you build an Attention Model with more outputs then inputs

It makes sences to build an Attention Model with more inputs as outputs but is it possible to build one with more ouptuts than inputs?

You can build and train attention models with more outputs than inputs.
Let me know the point of your confusion.