I currently use libtorch with torchscript to do gpu inference from c++ and wanted to understand what is the forthcoming replacement for this path. It seems tathat torch_.export.aoti_compile is likely the basis for generating libraries to call from native c++.
I am somewhat concerned however that this path is not generating source code that can be fed into my build system - it’s generating a built shared library. It seems that AOTInductor manages the toolchain. This is effectively a non-starter for me. I can’t control the arguments to the host compiler or nvcc to support different gpu architectures. I can’t cross compile. I can’t even set the host optimization flags. It definitely does not mesh well with usage of an existing c++/nvcc build system. What’s my future path forward here? Am I fundamentally misunderstanding something?