This is the final result of my experimentation with the simplest solution I have found. One more step could be done by adding a clever script to pre-configure llama.cpp with the required settings (GPU, AVX2, platform...) and even full automatic build.