[AMD] Update QuarkQuantization pass for Quark 0.12

[thpereir]

Describe your changes

Updates QuarkQuantization to work with AMD Quark 0.12, which introduced two breaking API changes:

ONNX path — calibration reader fix:
Quark 0.12 rejects calibration inputs that are not model inputs. Olive was forwarding the full dataset batch (including label columns like class) to the reader, causing INVALID_ARGUMENT: Invalid input name during calibration — the workflow would complete with exit code 0 but produce no output model. Fixed by passing model_path/io_config to create_calibration_dataloader() so non-input columns are filtered out, matching the pattern already used by OnnxQuantization.

Torch path — API update:

• revert_model_patching was removed in Quark 0.12 (patching is now reverted internally during export). Removed the import and call.
• prepare_for_moe_quant is deprecated and raises on transformers>=5. Replaced with preprocess_for_quantization.

Version gating:
Both ONNX and Torch paths now gate on amd-quark>=0.12.0 with a clear ValueError so users get an actionable error message instead of a cryptic ImportError.

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.

  The QuarkQuantization pass now requires amd-quark>=0.12.0. Users on 0.11.x will receive a clear ValueError with an upgrade prompt. The ONNX calibration path and Torch path are updated to work correctly with the 0.12 API.

(Optional) Issue link

[Copilot]

Pull request overview

Updates Olive’s QuarkQuantization pass to align with AMD Quark 0.12 behavior and APIs across both ONNX and Torch backends.

Changes:

• Updated Torch quantization integration to use Quark 0.12 APIs (preprocess_for_quantization) and removed now-obsolete patch reversion.
• Updated ONNX calibration dataloader construction to pass model_path/io_config, filtering non-model-input columns (to avoid Quark 0.12 invalid input errors).
• Raised the ONNX-side minimum Quark version check to >=0.12.0 and updated log/docs to reference 0.12.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
olive/passes/quark_quantizer/torch/quark_torch_quantization.py	Migrates Torch quantization flow to Quark 0.12 APIs and adjusts the quantize/freeze/export sequence accordingly.
olive/passes/quark_quantizer/quark_quantization.py	Updates Quark 0.12 messaging/version gating and fixes ONNX calibration dataloader creation to filter non-input columns.

[xiaoyu-work]

We can also add a version pin here: https://github.com/microsoft/Olive/blob/main/olive/olive_config.json#L601-L619

[thpereir]

We can also add a version pin here: https://github.com/microsoft/Olive/blob/main/olive/olive_config.json#L601-L619

Good call, updated