DynamicScale on different platforms

[massquantity]

Hi, when browsing the imagenet example, I noticed mixed precision training is available.
The create model code shows we can use float16 on CPU/GPU and bfloat16 on TPU, but the training code suggests that DynamicScale is only used on GPU. Does this mean it is not applicable for bfloat16 on TPU and float16 on CPU?

Another question is why DynamicScale is not documented in API Reference ? Although I finally found its source code.