"How to quantize the bias of a convolution in QAT (Quantization Aware Training) mode?

[seamoonlight-YBY]

My question is similar to the discussions

How pytorch simulates bias during quantization aware training

It seems that pytorch qat doesn’t simulate bias quantization error during qat. And I found that qat.Conv2d only fake-quantize weight and activation. So pytorch’s quantization strategy does not quantize the bias, right?

Is bias quantized while doing pytorch static quantization?

Why is bias not quantized upon pytorch static quantization? Or is it not required for deployment?

From these two discussions, I roughly understand the reason why quantizing bias is not possible in eager mode. I also gained some insights into quantizing bias in FX. However, I would like to confirm: In the latest version of eager mode QAT quantization, is it still not possible to quantize bias?

[supriyar]

Hi, we haven’t made changes to how we quantize bias in eager model. In the graph mode flows fx and pt2 export based - you can configure this yourself by annotating the graph correctly if your backend supports quantized bias operators.