You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi @DavidGOrtega it is possible if you implement the fp16 calculations in C++ and use the C++ interface.
We probably going to add python fp16 support in the near future.
Im about to tackle it. Did you have something in mind?
My first design would be abstract all the math ops and be able to set the fp in the Algorithm constructor so thats propagated to the math operations and of course the load and save.
I have the impression that the library is storing data in fp32, at least doing the math.
Im working with fp16 models mostly.
Its that possible?
The text was updated successfully, but these errors were encountered: