Possible issue / missing operator in Numba back end for betainc_grad

[DrEntropy]

I was working on a model involving a truncated negative binomial distribution using the nutpie sampler. I was getting warnings about the Numba backend falling back to using object-mode for betainc_grad. ("UserWarning: Numba will use object mode to run betainc_grad's perform method")

The betainc (incomplete beta function) comes up because pm.Truncate uses the logcdf which involves the incomplete beta function. Here is a simple example illustrating the warning in pytensor.

import pytensor
import pytensor.tensor as pt

x = pt.scalar("x")
a = pt scalar("a")
b = pt.scalar("b")

y = pt.betainc(a, b, x)
dy_da = pytensor.grad(y, a)

func = pytensor.function([a, b, x], dy_da, mode="NUMBA")
func(0.6, 2.0, 0.5)

I didn't post this as an issue on pytensor since I am not sure if it is a 'bug'. Nutpie can still sample (although slower), and further this is a bit of an edge case. For my particular use I was only truncating the zeros and this was best done with a pm.CustomDist and just subtracting the logp since it is just one (discrete) point.

Note on versions:

• PyTensor: 2.18.2
• Numba: 0.61.0

[ricardoV94]

This is already tracked in #1149 and #299. It's not a bug, as you said it just means numba will have to run python for that operator which can slow it down.

It's also correct that the use in the HurdleMixture is rather silly. It's just a hack to allow implementing Hurdle models with the PyMC Mixture machinery.