Is there any restriction on the types of args, kwargs, or return value of a torch.autograd.function?

I’m trying to build some functionality as a library and I am wondering what all do I need to handle. Is there any restriction on whether they need to be tensors, lists or tuples?