🐛 Describe the bug

After #74280 (cc @mruberry), the __rpow__ OpInfo has a sample input where other is a Tensor. This cannot happen during normal execution: to get to Tensor.__rpow__ a user does the following:

# self = some_tensor
# other = not_a_tensor
not_a_tensor ** some_tensor

If instead not_a_tensor is a Tensor, this ends up calling __pow__ in Python which will then handle the case.

Are there any legitimate cases where we do want this to happen?

Context

This caused some functorch tests to fail because we don't support the route where both self and other are Tensors. pytorch/pytorch also has some cryptic warning in that route:

but it's not clear to me if we want to support this or not.

Versions

pytorch main branch