Difference returning dtype of tfp.layers.IndependentNormal.params_size on Windows and Google Colab (Linux )

I found the following alike example from the document of TensorFlow probability.

model = tf.keras.Sequential([
    tfp.layers.DenseFlipout(800, activation=tf.nn.tanh),
    tfp.layers.DenseFlipout(tfp.layers.IndependentNormal.params_size(1)), # This line raise InvalidArgumentError (Original exception throw from tf.concat function.)
    # tfp.layers.DenseFlipout(np.int32(tfp.layers.IndependentNormal.params_size(1))), # This line is working both environments.
    tfp.layers.IndependentNormal(1)
])

The example is working on Windows but it is NOT working on Google Colab (Linux) by the eager execution.
The problem function is tfp.layers.IndependentNormal.params_size.
This return dtype is int32 on Windows but it is int64 on Google Colab (Linux),

Is this problem well known or unknown?
Or my usage is wrong?
I seem correct behavior is the Windows environment because of the "int" type will interpret "int32" on the numpy.

I made a small example, please verify the usage and the problem.
https://gist.github.com/elda27/54989bb20383ae29dc981bf36bb29a73

Start with the linked gist and the tfp.layers.IndependentNormal.params_size entry point, then reproduce the dtype difference in the Windows and Google Colab environments described. Done means the platform-dependent result is resolved or clearly documented, with the provided model example behaving consistently.