Experiences with `_keras_mask` in custom keras-based ModelV2?

Have anyone experiences with _keras_mask which is attached to the tensor output returned by the Masking or Embedding layer?
Whenever I try to use this _keras_mask tensor (of bool type) in a custom keras-based ModelV2, for example, like this way

padded_timesteps = tf.keras.layers.Masking()(inputTensor)
padded_timesteps_mask = padded_timesteps._keras_mask
expanded_padded_timesteps_mask = tf.expand_dims(tf.cast(padded_timesteps_mask, replicated_self_hoist.dtype), axis=-1)
concat = tf.keras.layers.concatenate([inputTensor, expanded_padded_timesteps_mask])

then I get an InvalidArgumentError: You must feed a value for placeholder tensor ‘policy/inputTensor’.
This error appears in both graph mode (framewok=tf) and eager mode (framework=tf2).

Does anyone know how to use this _keras_mask tensor with a keras-based model in RLlib?