Same binary for CPU and GPU

[pkese]

Is there any chance of having a single binary built for both CPU and GPU:

  <ItemGroup>
    <PackageReference Include="TorchSharp-cpu" Version="0.96.7" />
    <PackageReference Include="TorchSharp-cuda-linux" Version="0.96.7" />
  </ItemGroup>

... and then choose what to use in runtime?

Currently the above breaks complaining that CPU and CUDA dlls don't like each other:
Two TorchSharp runtime packages have been referenced (both libtorch-cpu and libtorch-cuda)

My naive workaround is to have 3 projects:

1. actual ML code
2. a cpu-training project which refrences code and includes TorchSharp-cpu
3. a gpu-training project which refrences code and includes TorchSharp-cuda-linux

but it somehow feels wrong having to switch directories all the time.
And I also have to build and distribute two versions of inference binaries as well.

[lostmsu]

Can't you use cuda-linux binary for both purposes?

[pkese]

Oh ... 🤦‍♂️ yes I can.