Fix Smart Search when using OpenVINO

[agrawalsourav98]

This PR fixes smart search when using OpenVINO. See discussion #6869

The issue is caused due to the following bug in upstream, see here and fixed by this.

As this is the part of the v1.17.0 release and the onnxruntime_openvino package has not been updated in a while, this PR fixes the issue by changing the dir to where the external_data is loaded before the onnxruntime session is created.

[mertalev]

Looks good! There's a possible issue in that changing the cwd might be unsound if multiple models are loaded at the same time. Fortunately, we only allow one model to load at a time so this isn't an issue in practice.

[agrawalsourav98]

Looks good! There's a possible issue in that changing the cwd might be unsound if multiple models are loaded at the same time. Fortunately, we only allow one model to load at a time so this isn't an issue in practice.

Yes, We can always remove this in the future once the package is updated to include the fix.

[mertalev]

It looks like the tests need to be updated.

Also, I don't have an Intel device handy. Can you check if it works without the static_path stuff? It'd be nice if we could remove that if it isn't needed.

[agrawalsourav98]

It looks like the tests need to be updated.

Also, I don't have an Intel device handy. Can you check if it works without the static_path stuff? It'd be nice if we could remove that if it isn't needed.

Just checked the onnxruntime changelog and openvino supports dynamic shapes since v1.15, so the static shapes might not be required, also verified loading the weights locally. I think we can also get rid of the external_data stuff entirely if none of the models exceed 2GiBs in a single file (protobuf limit).

[mertalev]

Sweet, then let's just remove the code for static inputs.

I think we can also get rid of the external_data stuff entirely if none of the models exceed 2GiBs in a single file (protobuf limit).

Some models do exceed this, so not really an option.