Introduce Intel® OpenVINO Weight Compression Pass (microsoft#2180)

## Describe your changes
Introduce OpenVINO Weight Compression Pass. This pass performs Weight
Compression on Hugginface models to produce OpenVINO or ONNX models, as
well as on ONNX models to produce ONNX models using Intel® NNCF
compress_weights() functionality.

## Checklist before requesting a review
- [x] Add unit tests for this change.
- [x] Make sure all tests can pass.
- [x] Update documents if necessary.
- [x] Lint and apply fixes to your code by running `lintrunner -a`
- [x] Is this a user-facing change? If yes, give a description of this
change to be included in the release notes.

## (Optional) Issue link

---------

## Describe your changes

## 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.

## (Optional) Issue link

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Author: Anirudh-Swaminathan

docs/source/features/ihv-integration/openvino.md

@@ -12,7 +12,7 @@ For Generative AI models, install Optimum Intel® from [Optimum Intel® Installa
 
 ## Prerequisites
 
-Note: OpenVINO version in Olive: 2025.1.0
+Note: OpenVINO version in Olive >= 2025.3.0
 
 ### Option 1: install Olive with OpenVINO extras
 
@@ -23,15 +23,15 @@ pip install olive-ai[openvino]
 ### Option 2: Install OpenVINO Runtime and OpenVINO Development Tools from Pypi
 
 ```bash
-pip install openvino==2025.1.0
-pip install nncf==2.16.0
+pip install openvino>=2025.3.0
+pip install nncf>=2.18.0
 pip install onnxruntime-openvino
 ```
 
 ### Install Optimum Intel® for Generative AI Workloads
 
 ```bash
-pip install optimum[openvino]
+pip install optimum[openvino]<=1.24.0
 ```
 
 More detailed instructions are available at [Optimum Intel® Installation Instructions](https://huggingface.co/docs/optimum/main/en/intel/installation)
@@ -96,6 +96,27 @@ Please refer to [OpenVINOQuantizationWithAccuracy](https://microsoft.github.io/O
 }
 ```
 
+## Weight Compression
+
+`OpenVINOWeightCompression` pass runs [Weight Compression](https://docs.openvino.ai/2025/openvino-workflow/model-optimization-guide/weight-compression.html) to compress Huggingface to OpenVINO model and Huggingface to ONNX model, as well as ONNX to ONNX model using Intel® NNCF.
+
+Please refer to [OpenVINOWeightCompression](https://microsoft.github.io/Olive/reference/pass.html#openvinoweightcompression) for more details about the `OpenVINOWeightCompression` pass and its config parameters.
+
+### Example Weight Compression Configuration
+
+```json
+{
+    "type": "OpenVINOWeightCompression",
+    "data_config": "compress_data_config",
+    "transform_fn": "custom_transform_func",
+    "extra_args": { "tokenizer": true },
+    "compress_config": {
+        "mode": "INT4_SYM",
+        "ratio": 0.8
+    }
+}
+```
+
 ## Model Encapsulation
 
 `OpenVINOEncapsulation` pass is used to generate an onnx model that encapsulates a OpenVINO IR model. It supports `OpenVINOModelHandler` for now.

docs/source/reference/options.md

@@ -320,8 +320,9 @@ Please also find the detailed options from following table for each pass:
 | [LoftQ](pass.rst#loftq) | Run LoftQ fine-tuning on a Hugging Face PyTorch model. |
 | [OpenVINOConversion](pass.rst#openvinoconversion) | Converts PyTorch, ONNX or TensorFlow Model to OpenVINO Model. |
 | [OpenVINOIoUpdate](pass.rst#openvinoioupdate) | Converts dynamic OpenVINO Model to static OpenVINO Model and updates IO names. |
-| [OpenVINOQuantization](pass.rst#openvinoquantization) | Post-training quantization for OpenVINO models and ONNX models |
-| [OpenVINOQuantizationWithAccuracy](pass.rst#openvinoquantizationwithaccuracy) | Post-training quantization with accuracy for OpenVINO models and ONNX models |
+| [OpenVINOQuantization](pass.rst#openvinoquantization) | Post-training quantization for OpenVINO models and ONNX models using Intel® NNCF |
+| [OpenVINOQuantizationWithAccuracy](pass.rst#openvinoquantizationwithaccuracy) | Post-training quantization with accuracy for OpenVINO models and ONNX models using Intel® NNCF |
+| [OpenVINOWeightCompression](pass.rst#openvinoweightcompression) | Weight Compression to compress Huggingface to OpenVINO model and Huggingface to ONNX model as well as ONNX to ONNX model using Intel® NNCF |
 | [OpenVINOEncapsulation](pass.rst#openvinoencapsulation) | Generates an ONNX model that encapsulates an OpenVINO IR model. |
 | [OpenVINOOptimumConversion](pass.rst#openvinooptimumconversion) | Run [optimum-cli export openvino](https://huggingface.co/docs/optimum/main/en/intel/openvino/export) command using Optimum Intel® to convert Huggingface Model to OpenVINO Model and optionally perform weight compression or quantization. |
 | [QNNConversion](pass.rst#qnnconversion) | Convert ONNX, TensorFlow, or PyTorch model to QNN C++ model. Quantize the model if –input_list is provided as extra_args. Uses qnn-[framework]-converter tool from the QNN SDK. |

docs/source/reference/pass.rst

@@ -333,6 +333,12 @@ OpenVINOQuantizationWithAccuracy
 --------------------------------
 .. autoconfigclass:: olive.passes.OpenVINOQuantizationWithAccuracy
 
+.. _openvino_weight_compression:
+
+OpenVINOWeightCompression
+--------------------------
+.. autoconfigclass:: olive.passes.OpenVINOWeightCompression
+
 .. _openvino_encapsulation:
 
 OpenVINOEncapsulation

olive/olive_config.json

@@ -403,6 +403,15 @@
             "supported_quantization_encodings": [  ],
             "extra_dependencies": [ "openvino" ]
         },
+        "OpenVINOWeightCompression": {
+            "module_path": "olive.passes.openvino.compression.OpenVINOWeightCompression",
+            "supported_providers": [ "*" ],
+            "supported_accelerators": [ "*" ],
+            "supported_precisions": [ "*" ],
+            "supported_algorithms": [  ],
+            "supported_quantization_encodings": [  ],
+            "extra_dependencies": [ "openvino" ]
+        },
         "OptimumConversion": {
             "module_path": "olive.passes.onnx.optimum_conversion.OptimumConversion",
             "supported_providers": [ "*" ],
@@ -618,8 +627,8 @@
         "lora": [ "accelerate>=0.30.0", "peft", "scipy" ],
         "nvmo": [ "nvidia-modelopt[onnx]" ],
         "openvino": [
-            "openvino>=2025.1.0",
-            "nncf>=2.16.0",
+            "openvino>=2025.3.0",
+            "nncf>=2.18.0",
             "numpy<2.0",
             "optimum[openvino]<=1.24",
             "onnxruntime-openvino"