Tag
ReRound is a post-training quantization method that uses a conditional diffusion model to guide rounding of near-midpoint weights, improving small LLM accuracy at 3-bit and 4-bit quantization without inference overhead.