[Feature Request] using SetSeed on C++ inference

[k5-inoue]

Describe the feature request

Will a future update be able to fix the random number seed during inference as well?

Describe scenario use case

Hi, I am making a C++ program for inference with a neural network model.
The neural network model includes a module that generates random numbers (RandomNormalLike).
Therefore, the output values are slightly different even for the same input.

I want to implement a function that turns this randomness on and off.
Example: We would like to store and use the seed of an excellent result.
Looking at the API reference, it seems that the SetSeed() function can be used to fix the random number seed during training (onnxruntime_training_c_api.h), but it is not available during inference (onnxruntime_cxx_api.h).
https://onnxruntime.ai/docs/api/c/namespace_ort.html

[Craigacp]

We've been discussing RNG behaviour in ONNX models here - onnx/onnx#6408, and I presented the gist linked at the bottom to the ONNX Operators SIG meeting in December. I think that would be a better solution to the seed issue, but does require ONNX spec changes.

[k5-inoue]

Thank you for your response.
I understand that related discussions are ongoing and that implementing this will require updates to the ONNX model specification.
I look forward to seeing these changes reflected in the future.