This PR integrates the Intel Extension for PyTorch (IPEX) into our build process so that PyTorch models quantized using the IPEX backend can be evaluated using the pytorch_inference processor. IPEX is currently only supported on Linux machines with x86_64 architecture.
Our research has shown that the main benefit of IPEX quantization comes from the use of the BFloat16 data type. Therefore, the pytorch inference processor has been extended to support this data type as well.
This PR integrates the Intel Extension for PyTorch (IPEX) into our build process so that PyTorch models quantized using the IPEX backend can be evaluated using the
pytorch_inference
processor. IPEX is currently only supported on Linux machines with x86_64 architecture.Our research has shown that the main benefit of IPEX quantization comes from the use of the BFloat16 data type. Therefore, the pytorch inference processor has been extended to support this data type as well.