Yesterday I showed you how to 2x your local LLM speed for free with Ngram-mod in llama.cpp. Today, we look under the hood.Spoiler: it’s not deep learning. It’s just brutal, O(1) computer science.Here is the exact hashing black magic llama.cpp uses to skip the matrix math and cheat its way to 100+ t/s. Most people think making an LLM faster requires pruning layers, quantization, or distilling models. But if you’re generating code, JSON, or editing documents, a massive percentage of your output al