AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration