llama.cpp: speculative decoding with draft models for 2-3x speedup

llama.cpp is the open-source library that enables running AI models on modest hardware. With this addition, it implements a technique called speculative decoding: a small, fast model (called a draft) generates a text proposal, and the large model verifies and corrects it in bulk rather than generating word by word. The result is 2 to 3 times faster generation, even on CPU, without sacrificing response quality.