ONNXRT Quantization: Recalibrating PTQ scales using QAT conv scales during ONNXRT QDQ export

[Darshcg]

Hi Team,

I'm working on ONNXRT QDQ export. I have QAT-derived scales for convolution weights and activations, and I want to export a QDQ model that uses those QAT scales for conv layers (via an override config) while using PTQ scales for the other ops.

My question: the PTQ scales for the other ops were calibrated from float activations. Is there any option to recalibrate those PTQ scales so they are computed with respect to the QAT conv scales (i.e., take the QAT conv weight/activation scales into account)?

Any pointers or config options would be appreciated.

Thanks!