Current PR is based on 1.10 - should of course be rebased.

It might make more sense to name the libraries libfaiss_cuda and libfaiss_rocm than libfaiss_gpu - to enable a consumer to load both CUDA and ROCm libraries into the same process.

Sorry, I didn't quite get this part:

With the current build set-up, where the GPU objects are statically linked into libfaiss, a consumer might dlopen the non-CUDA libfaiss_c, and in turn the non-CUDA libfaiss. Then when loading the CUDA libfaiss_c, the dynamic linker will re-use the non-CUDA libfaiss and fail to resolve CUDA/GPU symbols in libfaiss:

Why can't the user just load the CUDA libfaiss_c if they are using GPU, i.e. skip the first step?

The reason the user can't just load the CUDA libfaiss_c is a combination of:

• It is preferable to provide access to Faiss from Julia with CUDA as add-on functionality.
• Julia is a dynamic environment, so there is not necessarily an ahead-of-time decision to not load the non-GPU libfaiss_c.

In addition, the preferred method for adding CUDA functionality is via Julia package extensions, where the Faiss (Julia) package may be extended, at run-time, with additional functionality, e.g., CUDA.

I.e., it should be enough to add the Faiss Julia package to the user environment, and then if the user also adds the CUDA (or CUDA_Runtime_jll) Julia package, the CUDA functionality will automatically be made available.

Hi @stemann sorry for the delay, was still discussing internally. But I see you have moved the PR to draft mode, so I will follow up after you publish again

No worries. Please do comment on the general idea. I only moved the PR to draft as it is a bit of a rough draft.

@stemann I am trying to see if making it just shared by default is workable and not having a static version, then we can just keep the 1 GPU flag instead of introducing a new one. If not, we can go with the approach in this PR.

Ah - that sounds good.

On a related note, I've also been toying around (also in JuliaPackaging/Yggdrasil#11057) with building the AVX2, and AVX512 versions, where it would actually be preferable to get a library with the same name, e.g. libfaiss.so, and just have the content be different depending on the targeted instruction set (e.g., plain x86_64, or avx2, avx512, ...). But no need to complicate this request/PR with that - that's a separate concern.