How to calculate second derivative for mixed precision

[soonchangAI (soonchang)]

How to calculate second derivative gradients for mixed precision. I try calling backward() twice but it gave the same gradients as first derivative.

self.scaler.scale(loss).backward()
self.scaler.scale(loss).backward()

[ptrblck]

Calling backward multiple times will calculate the same gradients and accumulate them.
You would need to backward the derivatives again as described in this post.