Towards Superior Quantization Accuracy: A Layer-sensitive Approach