LLM quantization severely damages model quality and perplexitygithub.com/ggerganov2 pointsbehnamoh3 years ago