Supporting Pytorch for Custom Compiler Backend

jit
1

Hi Pytorch Team,

We have Custom Deep Learning hardware and a Compiler for the Hardware. For now, we have only Caffe framework support.

Having said that, we want to work on adding support for the Pytorch framework.

Is there any reference for the same?

Thanks,
Darshan C G

2

You might look into:

  1. Convert from PyTorch to ONNX using torch.onnx.export. Compile from ONNX format to your machine code. This is the approach taken by Intel OpenVINO and a few other hardware vendors, and it will let you easily add support for TensorFlow as well, since there is also a TensorFlow->ONNX converter.

  2. I think the Glow compiler can load TorchScript, so you could look into that.

What’s the name of your project / company?