Implementation of ICLR 2018 paper "Loss-aware Weight Quantization of Deep Networks"

TMG-Q: Tanh-Nonlinear Mixed-Precision Genetic Quantization. A post-training quantization framework for LLMs using evolved tanh-based formulas to achieve 2x compression with only +2.5% perplexity loss. Features dynamic outlier saliency and activation-aware fitness for high-fidelity weight reconstruction.