Reduction in performance of quantized bert model

quantization
7

Hi,
I’m facing a similar issue when quantizing Efficientnet.
I opened a thread about it here, but i was wondering if you found any solutions for your problem