How to perform int8 quantisation (not uint8) using ONNX?

[paul-ang]

Hi team, I am having issue quantizing the network consisting of Conv and Linear layers using int8 weights and activations in ONNX. I have tried setting it using op_type_dict, however it doesn't work. The activation is still using uint8. I am using version 2.3.1 neural compressor.

[mengniwang95]

Hi @paul-ang , we only support U8S8 by default because on x86-64 machines with AVX2 and AVX512 extensions, ONNX Runtime uses the VPMADDUBSW instruction for U8S8 for performance. I am so sorry you need to update the code by yourself to use S8S8. Please add 'int8' in activations' dtype list: https://github.com/intel/neural-compressor/blob/master/neural_compressor/adaptor/onnxrt.yaml. We will enhance it in our 3.0 API.