Why are you horrified?

In designing software, there's often a trade off between (i) generality / configurability, and (ii) performance.

llama.cpp is built for inference, not for training or model architecture research. It seems reasonable to optimize for performance, which is what ~100% of llama.cpp users care about.