How to implement mixed precision for different layers

Any instructions about how to implement mixed precision in different layers? For example, I want the convolution in layer 1 to be int8 and the convolution in the layer 10 to be int 32.