#quantization

LLM Inference: mmap Loading & Quantization Deep Dive

Efficient LLM inference hinges on mmap for lazy memory loading (e.g., <10s startup on llama.cpp) and quantization like GGUF K-Quants or AWQ/EXL2 to shrink 15GB models while preserving quality via salient weights and mixed precision.