How to set symmetric zero points for TRT deployment

[rui-shen-afk]

I read in the docs that "Quantization zero points are set to be asymmetric for activations and symmetric for weights". I trained a yolov5l quantized model using recipe and found it to be true. I also found out that there is a "QuantizationModifier" class that has a "tensorrt" param that alters the zero points settings.

I want to deploy my model with TensorRT which requires symmetric zero points. Is incorporating QuantizationModifier class into the current training procedure the suggested way? How should I go about it?

[bfineran]

Yes, currently the best supported pathway is to set tensorrt: True for any QuantizationModifier in your sparsification recipe