An alternative AD approach that seems to be faster than autograd

I am looking for people interested in benchmark testing, or helping with he development, of the following alternative approach to AD in libtorch: