[Feature request] Integrate Python-only API to ONNX optimizer proposal

[thiagocrepaldi]

Hello @daquexian, I am with Microsoft developing the PyTorch ONNX Exporter project.

Currently we are working on a PyTorch Dynamo-based ONNX exporter (torch.onnx.dynamo_export) which will replace the existing Torchscript-based ONNX exporter (torch.onnx.export) once it is production-ready.

Due to how Dynamo works, the ONNX graphs produced doesn't really perform well as the ones acquired through the previous exporter. As a response, we are internally working on an ONNX optimizer API which is Python only. As the workstream approaches stability, we are looking into how to open source the library.

One of the possibilities would be to create a new library, but we are also interested in maybe merging the new APIs in this repository. The goals of our library and ONNX optimizer's align well and the new implementation might help growing the project.

Is this something ONNX org would be interested at? We can discuss more about what it would encompasses the new API, etc

Thanks

[daquexian]

Hello, sorry for the late reply. A Python-only ONNX optimizer is very attractive. But I'm afraid I do not have enough bandwidth to overhaul this project. Did you notice the ONNX optimizer-sig? It is a great place to discuss further.

[thiagocrepaldi]

Thank you for the reply. It is ok not having bandwidth. We have a team dedicated to this at Microsoft. We could migrate from the private repo to this one and keep driving development. Maybe that could give some momentum to have an official Release within github instead of Tags? Would sharing ownership of this repo something you would consider?

I am not too familiar with optimizer-sig, but doesn't them actually discusses only Intel's quantization optimization as opposed to a general-purpose optimization for onnx models?