Does ONNX Runtime support building new Execution Provider outside of the ONNX Runtime code base?

[zhao-han]

Suppose I want to create a new Execution Provider, could I build this Execution Provider outside of the ONNX Runtime code base as a plugin?

For example, I can first install standard ONNX Runtime:

pip install onnxruntime

Then install my Execution Provider:

pip install onnxruntime-my-provider

Then I can run ONNX on my Provider.
Do we support this way?

[xadupre]

The current C API does not allow this scenario but it allows the implementation of custom kernels for existing providers: https://onnxruntime.ai/docs/reference/operators/add-custom-op.html. Would that work in your case?

[tomeuv]

@xadupre Are there any plans to allow for external delegates such as those that TensorFlow Lite support?

https://www.tensorflow.org/lite/performance/implementing_delegate#option_2_leverage_external_delegate

[xadupre]

• @yuslepukhin

[pranavsharma]

@jywu-msft and his team are working on this as we speak.

[tomeuv]

Thanks. Is this being developed within a SIG or working group? Or, what would be the best way to track this development?

[tomeuv]

Can anybody give some insight on what is going on within Microsoft regarding this?