Firstly, you get a prediction using model as a function, call this predition A; and then you use predict function of the model and get the same seemingly same prediction, call it B.

If you use CategoricalCrossentropy to calculate loss with same label, the loss are different between them.

class NLPModel(keras.Model):
  def __init__(self, **kwargs):
    super().__init__(**kwargs)
    self.hidden_layer1 = tf.keras.layers.GRU(400)
    self.outputs = tf.keras.layers.Dense(3, activation="sigmoid")
  def call(self, inputs):
    outputs = self.hidden_layer1(inputs)
    outputs = self.outputs(outputs)
    return outputs
model = NLPModel()
label = tf.constant([[1.,0.,0.]])
prediction = model(tf.constant([[[0.4, 0.5, 0.54,0.4, 0.5, 0.54,0.4, 0.5, 0.54,0.9],
                                  [0.4, 0.5, 0.54,0.4, 0.5, 0.54,0.4, 0.5, 0.54,0.9]]]))
cce = tf.keras.losses.CategoricalCrossentropy()
loss = cce(label, prediction)
def custom_cce(label, prediction):
  return tf.reduce_mean(tf.math.log(tf.matmul(label, tf.transpose(prediction)) / tf.reduce_sum(prediction)))
print("before numpy")
print(prediction)
print(loss)
print("before numpy")
print("after custom_cce")
print(custom_cce(label, prediction))
print("after custom_cce")
prediction = prediction.numpy()
loss = cce(label, prediction)
print("after numpy")
print(prediction)
print(loss)
print("after numpy")

before numpy
tf.Tensor([[0.5021686  0.48505145 0.5026401 ]], shape=(1, 3), dtype=float32)
tf.Tensor(1.0769439, shape=(), dtype=float32)
before numpy
after custom_cce
tf.Tensor(-1.0875016, shape=(), dtype=float32)
after custom_cce
after numpy
[[0.5021686  0.48505145 0.5026401 ]]
tf.Tensor(1.0875016, shape=(), dtype=float32)
after numpy