[Performance] Accuracy dropped heavily using onnxruntime to inference a model quantized by QAT

[Performance] Accuracy dropped heavily using onnxruntime to inference a model quantized by QAT #4011