Fix table transformers cuda export

[mht-sharma]

As per title!

Issue: #1774

[mht-sharma]

Little confused by comment, does the export of gpt_neo fails with DISABLE?

[fxmarty]

According to the comment at least

[mht-sharma]

I tried with EleutherAI/gpt-neo-1.3B and now the gpt-neo export seems to work with the DISABLE_ALL.

[fxmarty]

@mht-sharma I'm not sure about this - actually we may be better off sticking with GraphOptimizationLevel::ORT_ENABLE_ALL during validation as this is the default of ORT, better to catch issues ahead?

But then again someone may use the ONNX export and then not use ORT.

[mht-sharma]

@mht-sharma I'm not sure about this - actually we may be better off sticking with GraphOptimizationLevel::ORT_ENABLE_ALL during validation as this is the default of ORT, better to catch issues ahead?
But then again someone may use the ONNX export and then not use ORT.

I agree with your points. Perhaps such a test would be more suitable for ORTModel than export? Given the current error message, users might assume that the model isn't supported, when in fact they simply need to disable the optimizations?

Alternative, is obviously to have conditional statements to make it pass.