使用tf.GradientTape() wrt输入的渐变为None (TensorFlow2.4)

import tensorflow as tf

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(64, 3, activation='relu', input_shape=(32, 32, 3)),
    tf.keras.layers.MaxPooling2D(),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(10, activation='softmax')
])

inputs = tf.random.normal((1, 32, 32, 3))
targets = tf.constant([0])

with tf.GradientTape(watch_accessed_variables=False) as tape:
    tape.watch(model.trainable_variables)  # 只监视模型参数的变化
    logits = model(inputs)
    loss_value = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)(targets, logits)

grads = tape.gradient(loss_value, model.trainable_variables)