Master weights meaning

In the context of mixed precision training, I’ve found the following:

Using mixed precision training requires three steps:
1/Converting the model to use the float16 data type where possible.
2/Keeping float32 master weights to accumulate per-iteration weight updates.
3/Using loss scaling to preserve small gradient values.

What does master weights mean exactly?


I am not aware of half-precision training, but it looks like the original/float32-type weights is what being referred to as master weights here.