Internal computational problems using quantified model inference

[LZLwoaini]

Quantify the YOLO model to int8 using onnxruntime, and insert quantization and inverse quantization nodes when using QDQ quantization. As shown in the figure below, inverse quantization is performed before convolution calculation. So, should floating-point calculation still be maintained during convolution calculation? So the significance of quantification is only to make the model smaller? Moreover, it was observed that the convolution output tensor is of floating-point type. Why? Thank you!