Tensorflow gradient of loss with respect to model output gives None

Question

I'm trying to differentiate my loss function with respect to the model output in the training_step function of a tf.keras.Model. This is my attempt:

    def train_step(self, data):
        x, y = data
        with tf.GradientTape(persistent=True) as tape1:
            y_pred, dense_out = self(x, training=True)
            with tf.GradientTape() as tape2:
                tape2.watch(y_pred)
                loss = self.compiled_loss(y, y_pred)
        dy = tape2.gradient(loss, y_pred)

I would use tape1 for gradients I need later. First of all, dy gives None, how can I fix this? Secondly, is it allowed to return two output values when training=True in the call-method of my model and only one when training=False? Even if I don't do this dy is None?

Edit: If I do the following outside the train_step function, It does give a result different from None:

    with tf.GradientTape() as tape:
        y_pred, dense_out = model(x_train, training=True)
        loss = loss_fn(y_train, y_pred)
    print(tape.gradient(loss, y_pred))