Tensorflow: gradients suddenly became NaNs during training

Question

I am training simple variational autoencoder with negative binomial likelihood for decoder. I used python 3.7.1 and tensorflow 2.0.0.

The model was trained well witho