We are seeing some errors in scikit-learn CI using numpy dev recently see scikit-learn/scikit-learn#17707 (comment) for some context.

After debugging the problem it comes from the fact that np.dtype('float64').__class__ is not picklable in numpy development version. I haven't done a git bisect but I can do one if that would help.

For completeness the reason we care about this lies within joblib:
https://github.com/joblib/joblib/blob/ac0f1528ffec9cd53cabbafaca887ac880697862/joblib/hashing.py#L222-L235

Reproducing code example:

import pickle

import numpy as np


print('numpy version:', np.__version__)
dt = np.dtype('float64')
print('class: ', dt.__class__)
try:
    pickle.dumps(dt.__class__)
    print('Can pickle dt.__class__')
except Exception as exc:
    print('Exception while pickling dt.__class__')
    print(exc)

Output numpy development version:

numpy version: 1.20.0.dev0+be8ab91
class:  <class 'numpy.dtype[float64]'>
Exception while pickling dt.__class__
Can't pickle <class 'numpy.dtype[float64]'>: attribute lookup dtype[float64] on numpy failed

Ouptut numpy 1.19.0:

numpy version: 1.19.0
class:  <class 'numpy.dtype'>
Can pickle dt.__class__