Control per-device computation and make gradient averaging explicit in a multi-GPU training step.

Soon you can unlock how this was done.

Behind this: the tool used · the method · what actually resulted.