AttributeError when sampling with PyMC model using conditional probabilities

[ricardoV94]

Discussed in #7618

^{Originally posted by yk4978 December 18, 2024}

I am encountering an error when trying to sample from a PyMC model. In the model, I am defining a set of Bernoulli variables with conditional probabilities based on another parent Bernoulli variable A. The model definition itself works fine, but when I attempt to sample and calculate the posterior, I get the following error:
AttributeError: 'Scratchpad' object has no attribute 'ufunc'

Here is the code I am using:

import pymc as pm

n = 40
with pm.Model() as model:
    A = pm.Bernoulli('A', p=0.001)
    for i in range(n):
        if i < n // 2:
            pm.Bernoulli(f"A{i}", p=pm.math.switch(A, 0.6, 0.01), observed=1)
        else:
            pm.Bernoulli(f"A{i}", p=pm.math.switch(A, 0.6, 0.01))
    trace = pm.sample(1000, return_inferencedata=False, cores=1)

probs = sum(trace['A']) / len(trace['A'])
print(f"probs: {probs}")

Can anyone help me identify what might be causing this error and suggest a possible solution?

[ricardoV94]

@yk4978 I've seen that when PyTensor fails to fuse long graphs. Did you see any warnings prior to the failure?

How did you install pymc? And can you try on the latest version?