fit() with multiple cores hangs with PyMC >= 5.27.1

[swademb]

Using JupyterLab with bambi 0.16.0

Fitting a generalised linear mixed effects model - the formula doesn't seem to matter much. Let's say that:

• height and group:height are fixed effects.
• (1|area) is a random effect.

E.g. we have the formula:

outcome ~ 0 + group + group:height + (1:area)

And we have:

import bambi as bmb

# df = ...

model = bmb.Model(
    'outcome ~ 0 + group + group:height+ (1|area)',
    data=df,
    family='bernoulli'
)
model.build()
model.fit(chains=4, cores=2)

• PyMC == 5.26.1 or 5.27.0: Can train with either cores >= 1.
• PyMC >= 5.27.1: Can train with cores == 1; cannot train with cores > 1

When using a version of PyMC that fails, the (python) processes briefly appear in top using CPU, then a second later they go to zero usage or disappear, without control returning to JupyterLab. No error or other output is encountered. It just hangs.