float with 0 decimal converted to float32 instead of float64

[ricardoV94]

Description

import pytensor.tensor as pt

pt.as_tensor(8.9).dtype, pt.as_tensor(9.0).dtype, pt.as_tensor(9.1).dtype
# ('float64', 'float32', 'float64')

Completely puzzling

[ayulockedin]

@ricardoV94 i will pick this up, lets see what can be done here

[ayulockedin]

@ricardoV94 I think I see what's happening here. In pytensor/scalar/basic.py , NumpyAutocaster is actively checking if the float's value fits perfectly into a float32 without losing precision (which 9.0 does, but 8.9 and 9.1 don't).

[ricardoV94]

Does the other cast mode not do this? See #1073 (comment)

If so, it may finally be time to pull the plug on this odd casting machinery

[ayulockedin]

@ricardoV94 You are right, I just tested it and the numpy+floatX policy natively fixes this issue, and it looks like it resolves the integer downcasting from #1073 as well.

I had a local fix going for the custom block, but pulling the plug on that old machinery makes much more sense.

Should I just open a PR changing the default cast_policy in configdefaults.py to see how the CI test suite handles it? Or would you prefer I try to rip out the old custom logic in scalar/basic.py while I am at it?

[ricardoV94]

Let's change the default and see what happens first