Support for H100 "Transformer Engine"

[simusid]

I've tried to search the repo and I have not found obvious support for the NVidia Transformer Engine. I don't know if this is already on the developer roadmap but it sounds like it would improve performance, possibly significantly. The C/C++ API is here.

It doesn't look like a simple drop in replacement though. I have no idea the LOE necessary for something like this but I very selfishly would like to use this on our new DGX-H200.