Yuwen Zhou

hi @chensuyue, PR is ready for extension test

@chensuyue extension test: https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3784/artifact/report.html  performance regression is caused by switching performance dataset from dummy to real dataset.

https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3851/

https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3877/ Note: object detection models need new quantization recipe support from Strategy team and may not pass extension test now.

> https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3877/ Note: object detection models need new quantization recipe support from Strategy team and may not pass extension test now. NLP models failed due to some typos and code...

> > https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3877/ Note: object detection models need new quantization recipe support from Strategy team and may not pass extension test now. > > NLP models failed due to some...

> Retest: https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3890/ yolov3, yolov4 and tiny_yolov3 will not be enabled in this version because 'onnxrt.graph_optimization.level' is not supported now. 1. ssd-12, ssd-12_qdq, faster_rcnn, faster_rcnn_qdq, mask_rcnn, mask_rcnn_qdq will be re-enabled...

**passed:** bert_squad_model_zoo_dynamic, mobilebert_squad_mlperf_dynamic, mobilebert_squad_mlperf_qdq, duc, BiDAF_dynamic and huggingface question answering models **failed:** gpt2_lm_head_wikitext_model_zoo_dynamic and huggingface test classification models, retest: https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3913/

**passed:** bert_squad_model_zoo_dynamic, mobilebert_squad_mlperf_dynamic, mobilebert_squad_mlperf_qdq, duc, BiDAF_dynamic and huggingface question answering models: https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3908/artifact/report.html gpt2_lm_head_wikitext_model_zoo_dynamic and huggingface test classification models: https://inteltf-jenk.sh.intel.com/job/intel-lpot-validation-top-mr-extension/3919/artifact/report.html

Hi @clementpoiret , sorry for the late response. This is not an issue of export. The QAT PyTorch model you generated is invalid. Please refer to this [document](https://github.com/intel/neural-compressor/blob/master/docs/source/quantization.md#quantization-aware-training-1) for the...