Deep Dive: Integer Quantization Math from INT8 Foundations to 4-bit 70B Models

A technical series covering integer quantization for transformers from first principles, tracing the field's evolution from early INT8 struggles on 7B models to routinely fitting 70B models in 4-bit on a single GPU. The first post covers fixed-point math, quantization error sources, and memory reduction tradeoffs, with subsequent posts planned on specific techniques.