Paper introduces statistically-lossless quantization for LLMs

Addresses the trade-off between lossy compression methods like GPTQ and AWQ, which save memory but lose fidelity, and lossless techniques that preserve accuracy without accelerating inference.