From the autograd and amp docs, I gather that the @custom_fwd
and @custom_bwd
decorators let your custom autograd Function
compose nicely w/ mixed precision. In what scenarios, then, would you not wrap your Function
’s methods with these decorators?
I’m asking because I was a little surprised that custom Function
s and amp
don’t compose by default, but I’m sure there’s a good reason for it. Thanks!