Leverage graph mode computation for pure c++ nn.module

Hi everyone, I might have missed that answer… I have a nn.module defined in c++ and I want to leverage graph mode computation to both improve computation time and hide its implementation.
AFAIK there’s no way to use jit::trace in c++ to export it to TorchScript.

There is thus little options left. Is embedding TorchScript code into c++ source code a viable solution? I didn’t find any example to get started.
Or is there other options ?