Opset Coverage - Binary Size Tradeoff

[adityagoel4512]

We express models and feature engineering inference pipelines our clients have developed using ONNX. As it stands, the policy to only include operators (and specifically, support all argument types supported by the standard for a given operator) as and when needed out of concern for binary size can significantly slow down iteration speed and make user experience significantly worse since you only find out ORT doesn't support something you need (and cannot work around) when you need it most!

Given custom build options exist for reduced operator builds, users already have a straightforward way to build onnxruntime with a smaller binary footprint should they need this.

I would make the argument that aiming to support the entire standard and giving users optionality to build ORT with a smaller binary footprint strikes the right balance. We currently cannot specify reduced type builds where we only include certain types supported by an operator so this might be some further tooling that could be supported.

I would be interested to hear any thoughts the community has.