ImportError dependent on import order

[itrharrison]

Hi,

Working on an M3 macbook pro in macOS 15.6.

Using miniforge3.

Creating a conda env from scratch.

Running

conda install python pip ipython
pip install tensorflow tensorflow_probablity tensorflow-keras

to install python 3.13 and

tensorflow==2.20.0
tensorflow-probability==0.24.0
tensorflow_keras==0.1

then in a python terminal

In [1]: import tensorflow_probability
In [2]: import tensorflow

succeeds, and I can use tf and tf keras and import and use my own modules which use them . But

In [1]: import tensorflow
In [2]: import tensorflow_probability

returns the (familiar) error

File ~/miniforge3/envs/tensorflow/lib/python3.13/site-packages/tensorflow_probability/python/__init__.py:79, in _validate_tf_environment(package)
     77 if required_tensorflow_version[0] == '2':
     78   try:
---> 79     import tf_keras  # pylint: disable=unused-import
     80   except (ImportError, ModuleNotFoundError):
     81     # Print more informative error message, then reraise.
     82     print('\n\nFailed to import TF-Keras. Please note that TF-Keras is not '
     83           'installed by default when you install TensorFlow Probability. '
     84           'This is so that JAX-only users do not have to install TensorFlow '
   (...)     87           'This can be be done through installing the '
     88           'tensorflow-probability[tf] extra.\n\n')

ModuleNotFoundError: No module named 'tf_keras'

When importing tf and tfp at the top of my own modules, this error also appears when I import mymodule no matter what order the import of tf and tfp happens in the module file.

I am note even sure how to debug this! I guess something to do with namespaces getting messed up, but I am very much a user not a dev...

[csuter]

Please confirm that you've followed the recommendations in the release notes for keras compatibility, and let us know if that doesn't fix it.