In encoder models, the weight dynamic range of attention layers is usually larger than that of FFN layers, and aggressive quantization easily distorts the attention distribution, while FFN layers are usually more robust, so mixed-precision strategies are often adopted in practice.