EfficientQAT: LLM Quantization, gets a 2-bit llama2-70B outperform regular 13B (old.reddit.com)