Training with fp16?

[longyangqi]

Great work!
However, training on my own data consumes too much GPU memory.
As fp16 is supported in Memory-efficient inference, can we also train the model with fp16?

Thanks!