Cannot sample from DistributionLayer

# Model
n_mvn = 5
output_dim = 2

input_layer = k.layers.Input(shape=(x.shape[1],))

learner = k.layers.Dense(tfp.layers.MixtureNormal.params_size(n_mvn, [output_dim]))(input_layer)
learner = tfp.layers.MixtureNormal(n_mvn, [output_dim], name='generator')(learner)

model = k.models.Model(input_layer, learner)
print(model.summary())

model.compile(
    loss=lambda y, stuff: -stuff.log_prob(y),
    optimizer=k.optimizers.Adam(0.01)
)
hist = model.fit(
    x, x, 
    batch_size=1024, epochs=500, 
    validation_split=0.1, 
    callbacks=[k.callbacks.EarlyStopping(patience=50)],
    verbose=True, 
)

Please notice I have a MixtureNormal layer in the keras model. I have search literally everywhere. Am I really supposed to believe that there is absolutely no way to get this layer to produce random numbers or call .sample() in some way?

Start with the MixtureNormal layer and DistributionLayer sampling API, then reproduce the provided Keras model to establish the current behavior. Determine whether a supported sampling path exists or what API change is needed, and add or update coverage so the requested random-number generation is demonstrably usable.