Unexpected Symbolic tensor in Tensorflow Probability tensor_coercible object (mixture layer)

Complex interpretation of Tensorflow Probability tensor_coercible object
I have a tensorflow model (keras sequential) that ends with a Tensorflow Probability (TFP) mixture layer. My goal is to fit this network with a custom loss function. The unexpected behaviour is the following:

When I pass a custom loss like this:

def wass_loss(y_true, y_pred):
    print(type(y_pred))
    print(type(y_true))
    # Do multiple 'tf' operations, one of them, sample the _TensorCoercible
    return ...

It prints:

<class 'tensorflow_probability.python.layers.internal.distribution_tensor_coercible._TensorCoercible'>
<class 'tensorflow.python.framework.ops.SymbolicTensor'>

Intepreting each argument correctly and results are ok /coherent.

Now, when I compile this same network with another loss func. and use the wass_loss above as metric.

def another_loss(y_true, y_pred):
    return -y_pred.log_prob(y_true)

my_model.compile(optimizer = kr.optimizers.Adam(1e-3),
                 loss = another_loss,
                 metrics = [wass_loss,])
my_model.fit(...)

I get:

<class 'tensorflow.python.framework.ops.SymbolicTensor'>
<class 'tensorflow.python.framework.ops.SymbolicTensor'>

Which is unexpected, I would expect the Tensor Coercible from TFP to still be a Tensor Coercible. Yet, looks like that after the new loss (which might be computed first), it is already a Symbolic Tensor Why is it happening? What am I missing?

Thanks a lot!

Start by reproducing the behavior with the Keras Sequential model, TensorFlow Probability mixture layer, and the two compile/fit configurations shown in the issue. Compare the custom loss and metric inputs, then document the cause and expected behavior; no source file or test is named in the report.