inference efficiency

A hands-on guide to making local LLMs feel fast: why inference is memory/bandwidth bound, and 3–5 practical tweaks—pick the right quantised GGUF model, compile with OpenBLAS, pin threads and respect NUMA, right-size context, tune THP/swap, and set a performance governor—complete with apt/dnf/zypper installs, llama.cpp builds, monitoring tips, and real-world gains of 1.5–2× tokens/sec.