Backends get installed in bin, confusing llama.cpp? #3420
Description
whisper.cpp, at least when built by the whisper-cpp package in Nixpkgs, seems to end up installing the various libggml-*.so backend files into the bin directory, rather than lib like llama.cpp (package) does. As a result, if I install both programs into the same profile, llama.cpp ends up trying to load the backends from whisper.cpp's version of libggml in addition to its own - and if I update llama.cpp to a version that's too far away from whisper.cpp, it ends up segfaulting.
I'm not sure whether this is an upstream issue here or a packaging issue in Nixpkgs. I decided to report it here first because a) I figured the devs here were more likely to know what the expected/correct layout is for libggml, and b) I don't see anything in either of the Nix derivations that would affect which directory the backends were installed into. I'm fine with reopening it on the Nixpkgs side if you think that's more appropriate, though.
(Currently trying this on macOS 15.5 Sequoia on an Intel-based Mac, if that matters. Trying to run the llama-cpp-vulkan and whisper-cpp-vulkan variants of the packages.)