Libtorch dll versioning

[spiegelball]

🚀 Feature

Build libraries like torch_cpu, c10, torch_cuda etc. with proper versioning, mode (cpu vs. gpu) and debug / release information in the name.

Motivation

Let's say I have two dll's, a.dll and b.dll. a.dll depends on libtorch 1.9 cpu version and b.dll depends on libtorch 1.10 cuda version. With the current naming scheme it's not possible to have an application load both a.dll and b.dll. After a.dll is loaded, e.g. c10.dll (cpu version, 1.9.) is already loaded so b.dll will be fed with a wrong version of c10.dll

Pitch

Libraries could be built in the format library_name_version_backend.dll. E.g., c10.dll in version 1.10 would become either c10_110_cpu.dll or c10_110_cuda.dll. This also implies, that torch_cpu_110_cpu.dll now depends on c10_110_cpu.dll.

Alternatives

Maybe side-by-side assemblies, however I haven't found a solution for the scenario described above yet.

Additional context

cc @ezyang @seemethere @malfet

[malfet]

Just adding extra suffixes to libraries would not solve side-by-side loading, as there would be numerous symbol clashes.

[spiegelball]

Would all libraries built by pytorch need correct manifests for Sxs loading then? What could be a solution?