Update VitisAIQuantization to use Quark

[vortex-captain]

Description

Example usage in Olive workflow json:

  "passes": {
    "conversion": {
      "device": "cpu",
      "type": "OnnxConversion",
      "target_opset": 17,
      "use_dynamo_exporter": false
    },
    "to_fixed_shape": {
      "type": "DynamicToFixedShape",
      "dim_param": ["batch_size", "sequence_length"],
      "dim_value": [1, 77]
    },
    "quantization": {
      "type": "QuarkQuantization",
      "data_config": "calib_data",
      "config_template": "XINT8",
      "enable_npu_transformer": true,
      "extra_options": {
        "OpTypesToExcludeOutputQuantization": ["MatMul", "Gemm"],
        "ActivationSymmetric": true
      },
      "debug_mode": true,
      "log_severity_level": 0,
      "ignore_warnings": false
    }
  }

Please refer to https://quark.docs.amd.com/latest/onnx/user_guide_config_description.html for the complete list of config_template options. All the other quantization options are listed in https://quark.docs.amd.com/latest/onnx/appendix_full_quant_config_features.html .

Examples

2 ResNet examples are added to examples/vai, which convert the models using Quark then evaluate on VitisAIExecutionProvider (run on NPU, RyzenAI 1.3.1, onnxruntime-vitisai 1.19).

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

[github-advanced-security]

lintrunner found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[ChaoLi-AMD]

Describe your changes

Example usage in Olive workflow json:

  "passes": {
    "conversion": {
      "device": "cpu",
      "type": "OnnxConversion",
      "target_opset": 17,
      "use_dynamo_exporter": false
    },
    "quantization": {
      "type": "VitisAIQuantization",
      "data_config": "calib_data",
      "config_template": "INT8_TRANSFORMER_ACCURATE",
      "extra_options": {
        "OpTypesToExcludeOutputQuantization": ["MatMul", "Gemm"],
        "ActivationSymmetric": true
      },
      "debug_mode": true,
      "log_severity_level": 0,
      "ignore_warnings": false
    }
  }

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

please improve the example to refer to Quark documentation: https://quark.docs.amd.com/latest/supported_accelerators/ryzenai/index.html

[xiaoyu-work]

/azp run

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[xiaoyu-work]

Can you explain more about how will you use this folder? The cache folder is designed to be pass-agnostic so i want to double confirm the use case here.

[vortex-captain]

The folder will be created at the beginning of the evaluation step, upon the creation of a VitisAIExecutionProvider inference session (used as model cache by EP). Is evaluation considered an Olive pass?

[xiaoyu-work]

No, if VitisAIEP will need to cache a model for evaluation, can we create a temporal folder for it? (and it will be deleted after all. I assume this model cache is not needed when the workflow finish.). We can create a temporary folder in cache.evaluations like temp_model_cache or something.

[shaahji]

Update the entry in olive_config.json to point to the correct location of the pass implementation in the module. Many of the tests are failing because of the wrong entry.

[vortex-captain]

Describe your changes

Example usage in Olive workflow json:

  "passes": {
    "conversion": {
      "device": "cpu",
      "type": "OnnxConversion",
      "target_opset": 17,
      "use_dynamo_exporter": false
    },
    "quantization": {
      "type": "VitisAIQuantization",
      "data_config": "calib_data",
      "config_template": "INT8_TRANSFORMER_ACCURATE",
      "extra_options": {
        "OpTypesToExcludeOutputQuantization": ["MatMul", "Gemm"],
        "ActivationSymmetric": true
      },
      "debug_mode": true,
      "log_severity_level": 0,
      "ignore_warnings": false
    }
  }

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

please improve the example to refer to Quark documentation: https://quark.docs.amd.com/latest/supported_accelerators/ryzenai/index.html

Added links of Quark documentation on quantization configurations

[ChaoLi-AMD]

Describe your changes

Example usage in Olive workflow json:

  "passes": {
    "conversion": {
      "device": "cpu",
      "type": "OnnxConversion",
      "target_opset": 17,
      "use_dynamo_exporter": false
    },
    "quantization": {
      "type": "VitisAIQuantization",
      "data_config": "calib_data",
      "config_template": "INT8_TRANSFORMER_ACCURATE",
      "extra_options": {
        "OpTypesToExcludeOutputQuantization": ["MatMul", "Gemm"],
        "ActivationSymmetric": true
      },
      "debug_mode": true,
      "log_severity_level": 0,
      "ignore_warnings": false
    }
  }

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

please improve the example to refer to Quark documentation: https://quark.docs.amd.com/latest/supported_accelerators/ryzenai/index.html

Added links of Quark documentation on quantization configurations

For a Ryzen AI example, please use XINT8 as the example instead of INT8_TRANSFORMER_ACCURATE. Just checking — does this example currently runnable on Olive?

[jambayk]

@vortex-captain please create a copy of your branch directly in this repo and open a new PR to be able to run the CI without the login issue.

[vortex-captain]

Describe your changes

Example usage in Olive workflow json:

  "passes": {
    "conversion": {
      "device": "cpu",
      "type": "OnnxConversion",
      "target_opset": 17,
      "use_dynamo_exporter": false
    },
    "quantization": {
      "type": "VitisAIQuantization",
      "data_config": "calib_data",
      "config_template": "INT8_TRANSFORMER_ACCURATE",
      "extra_options": {
        "OpTypesToExcludeOutputQuantization": ["MatMul", "Gemm"],
        "ActivationSymmetric": true
      },
      "debug_mode": true,
      "log_severity_level": 0,
      "ignore_warnings": false
    }
  }

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

please improve the example to refer to Quark documentation: https://quark.docs.amd.com/latest/supported_accelerators/ryzenai/index.html

Added links of Quark documentation on quantization configurations

For a Ryzen AI example, please use XINT8 as the example instead of INT8_TRANSFORMER_ACCURATE. Just checking — does this example currently runnable on Olive?

Updated example in description. And yes, such an example (BERT text model) is runnable on Olive, but in evaluation, the output model cannot run on NPU (all nodes assigned to CPU), unlike the ResNet examples. Any insights?

[VishalX]

This is adding dependency to a specific version. I don't think we should add this here.