`pt.specify_broadcastable` does not work with negative axis values

[tomicapretto]

Description

If I use pt.specify_broadcastable with the second argument equal to -1 it does not do anything. It needs the value of the axis in order to work. See the example:

import pytensor.tensor as pt

tensor = pt.tensor(shape=(5, 1), broadcastable=(False, False))
print(tensor.broadcastable)

tensor2 = pt.specify_broadcastable(tensor, -1)
print(tensor2.broadcastable)

tensor3 = pt.specify_broadcastable(tensor, 1)
print(tensor3.broadcastable)

(False, False)
(False, False)
(False, True)

I'm using PyTensor version 2.20.0

[Dhruvanshu-Joshi]

What is the behaviour expected when axis is set to -1? I believe it should set the last dimension as broadcastable right? Or do we want to return an error?

[ricardoV94]

What is the behaviour expected when axis is set to -1? I believe it should set the last dimension as broadcastable right? Or do we want to return an error?

The first. There's a numpy helper used throughout the library to normalize axis, that converts negative ones to positive based on the input ndim