How to boost other gpus' util in distributed training with tf.keras?

Question

python3.6, tf2.3, win10, cuda10.1

When I test a-layer ConvLSTM model for distributed training, there is an issue about GPU-Util.

Here is my code.