Add OnnxDiscrepancyCheck speedup metric with default timing updates

[xadupre]

Describe your changes

Added speedup measurement for OnnxDiscrepancyCheck and updated behavior based on review feedback:

• Changed timing_iterations default from 10 to 5.
• If timing_iterations is set to 0, speedup measurement is skipped.
• Added unit tests to validate the new default and the skip behavior.
• Fixed test mocks to properly configure device attribute for compare_generation tests.

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.

[Copilot]

Pull request overview

This PR enhances the OnnxDiscrepancyCheck pass by adding an inference speedup measurement (ONNX vs PyTorch) and introducing configurable warmup/timing iteration settings, with updated defaults and tests to validate the new behavior.

Changes:

• Added warmup_iterations and timing_iterations config parameters (defaulting timing_iterations to 5) and implemented speedup measurement with an option to skip when timing_iterations=0.
• Updated session/device setup to target the configured accelerator (with CPU fallback) and run PyTorch on the matched torch device.
• Added unit tests to validate the new default and the skip behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
olive/passes/onnx/discrepancy_check.py	Adds speedup timing, new config params/defaults, and device-aware session/model handling.
test/passes/onnx/test_discrepancy_check.py	Adds tests for the new timing_iterations default and the “skip when 0” behavior.

[Copilot]

This is already addressed in commit 595bffc. input_ids is moved to ref_model.device (line 389) and the output is moved back to CPU via .cpu() before .tolist() (line 398).

[xadupre]

@copilot test/passes/onnx/test_discrepancy_check.py:105: in test_compare_generation_returns_common_prefix_length
result = pass_instance.compare_generation(config, mock_ref_model)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
olive/passes/onnx/discrepancy_check.py:389: in compare_generation
input_ids = input_ids.to(ref_model.device)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E TypeError: to() received an invalid combination of arguments - got (MagicMock), but expected one of:
E * (torch.device device = None, torch.dtype dtype = None, bool non_blocking = False, bool copy = False, *, torch.memory_format memory_format = None)
E * (torch.dtype dtype, bool non_blocking = False, bool copy = False, *, torch.memory_format memory_format = None)
E * (Tensor tensor, bool non_blocking = False, bool copy = False, *, torch.memory_format memory_format = None)

[Copilot]

@copilot test/passes/onnx/test_discrepancy_check.py:105: in test_compare_generation_returns_common_prefix_length
result = pass_instance.compare_generation(config, mock_ref_model)
^^...

Fixed in 25c6ab4. Added mock_ref_model.device = torch.device("cpu") to both compare_generation test cases so that the mock returns a proper torch.device object instead of a MagicMock.