EfficientQAT: Efficient Quantization-Aware Training for Large Language Models