wrong inference results of qat int8 model with Conv2DTranspose.

Tensorflow == 2.4.0

Below is my image compression model which have two Conv2DTranspose layers.

model = keras.Sequential([
  keras.layers.InputLayer(input_shape=(256, 256, 3)),
  keras.layers.Conv2D(filters=64, kernel_size=(4, 4), strides=(2, 2), padding='same',activation=None),
  keras.layers.Conv2D(filters=96, kernel_size=(4, 4), strides=(2, 2), padding='same',activation=None),
  keras.layers.Conv2DTranspose(filters=128, kernel_size=(4, 4), strides=(2, 2), padding='same',activation=None),
  keras.layers.Conv2DTranspose(filters=3, kernel_size=(4, 4), strides=(2, 2), padding='same', activation=None)
])

After quant aware train the model, calculate MSE between inputs and outputs, inference result of tflite is totally wrong.
inference Keras with fake quant op: MSE is 1.6923981
inference tflite : MSE is 4376.958.

After debug, the error is caused by Conv2DTranspose. Just inference the last deconv layer，results as below.

>>> inference_tflite.reshape(-1)[:10]
array([-0.6456693 ,  0.        ,  0.        , -0.68503934,  0.        ,
        0.        , -0.70866144,  0.        ,  0.        , -0.70866144],
      dtype=float32)
>>> inference_keras.reshape(-1)[:10]
array([-0.6456693 , -0.5984252 , -0.53543305, -0.6850394 , -0.62204725,
       -0.53543305, -0.70866144, -0.62992126, -0.5511811 , -0.70866144],
      dtype=float32)

Compared results, only the first channel of tflite result is correct，the else are all zeros. Why?
tflite TestInputs