QuIP: 2-Bit Quantization of Large Language Models With Guarantees