Whisper: Support ORT 1.17.1

[jambayk]

Describe your changes

ORT 1.17.1 made some changes to the beam search ops. For version 1.17.1, the beam search op used will be WhisperBeamSearch.

Fixes:
#1014

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.
• Is this PR including examples changes? If yes, please remember to update example documentation in a follow-up PR.

(Optional) Issue link

[xiaoyu-work]

I'm not sure if this is the root cause, cause I'm able to run test_transcription.py with ORT 1.17.1 successfully without --predict_timestamps. Do you know the reason?

[jambayk]

I'm not sure if this is the root cause, cause I'm able to run test_transcription.py with ORT 1.17.1 successfully without --predict_timestamps. Do you know the reason?

@xiaoyu-work you can see in this PR microsoft/onnxruntime#19509 that some token ids are now part of the model attributes and not inferred using offsets. This affects the logits processor which does the time stamp prediction. I think the lack of these values in the beam search node's attributes causes the issue.

[xiaoyu-work]

I'm not sure if this is the root cause, cause I'm able to run test_transcription.py with ORT 1.17.1 successfully without --predict_timestamps. Do you know the reason?

@xiaoyu-work you can see in this PR microsoft/onnxruntime#19509 that some token ids are now part of the model attributes and not inferred using offsets. This affects the logits processor which does the time stamp prediction. I think the lack of these values in the beam search node's attributes causes the issue.

Can you test if this change works?

[jambayk]

I'm not sure if this is the root cause, cause I'm able to run test_transcription.py with ORT 1.17.1 successfully without --predict_timestamps. Do you know the reason?

@xiaoyu-work you can see in this PR microsoft/onnxruntime#19509 that some token ids are now part of the model attributes and not inferred using offsets. This affects the logits processor which does the time stamp prediction. I think the lack of these values in the beam search node's attributes causes the issue.

Can you test if this change works?

It's already tested.