Tensorflow: tape.gradient() returns None for GRU layer

Question

I build my model with the following code(tensorflow==1.14):

class Model(tf.keras.Model):
  def __init__(self):
    super(Model, self).__init__()

    self.embedding = tf.keras.layers.Embedding(10, 5)
    self.rnn = tf.keras.layers.GRU(100)  # neither GRU nor LSTM works
    self.final_layer = tf.keras.layers.Dense(10)
    self.loss_obj = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction='none')

  def call(self, inp):
    inp_em = self.embedding(inp)  # (batch_size, seq_len, embedding_size)
    inp_enc = self.rnn(inp_em)  # (batch_size, hidden_size)
    logits = self.final_layer(inp_enc)  # (batch_size, class_num)

    return logits

model = Model()

inp = np.random.randint(0, 10, [5, 50], dtype=np.int32)
out = np.random.randint(0, 10, [5], dtype=np.int32)

with tf.GradientTape() as tape:
  logits = model(inp)
  loss = model.loss_obj(out, logits)
  print(loss)

gradients = tape.gradient(tf.reduce_mean(loss), model.trainable_variables)

print('==========  Trainable Variables  ==========')
for v in model.trainable_variables:
  print(v)

print('==========  Gradients  ==========')
for g in gradients:
    print(g)

but when i print the gridents, the output is :

Tensor("categorical_crossentropy/weighted_loss/Mul:0", shape=(5,), dtype=float32)
==========  Trainable Variables  ==========
<tf.Variable 'model/embedding/embeddings:0' shape=(10, 5) dtype=float32>
<tf.Variable 'model/gru/kernel:0' shape=(5, 300) dtype=float32>
<tf.Variable 'model/gru/recurrent_kernel:0' shape=(100, 300) dtype=float32>
<tf.Variable 'model/gru/bias:0' shape=(300,) dtype=float32>
<tf.Variable 'model/dense/kernel:0' shape=(100, 10) dtype=float32>
<tf.Variable 'model/dense/bias:0' shape=(10,) dtype=float32>
==========  Gradients  ==========
None
None
None
None
Tensor("MatMul:0", shape=(100, 10), dtype=float32)
Tensor("BiasAddGrad:0", shape=(10,), dtype=float32)

the gridents for final layer works well, but None for GRU layers and so on.

I have tried both tf.keras.layers.LSTM and tf.keras.layers.GRU, the same problem exists.

Update

finally, I replace tf.GradientTape().gradient() with tf.graidents():

logits = model(inp)
loss = model.loss_obj(out, logits)

gradients = tf.gradients(tf.reduce_mean(loss), model.trainable_variables)

the gradients works. But I still don't know what's the difference between these two implements.