Merge branch 'main' into QNN_Qwen_Coder

qc-tbhardwa · web-flow · commit 56aa71860a3e · 2026-04-17T15:36:43.000+05:30
diff --git a/.aitk/requirements/Intel/Test_py3.12.9.txt b/.aitk/requirements/Intel/Test_py3.12.9.txt
@@ -47,7 +47,7 @@ optimum[openvino]==1.24.0
 optuna==4.2.1
 packaging==24.2
 pandas==2.2.3
-pillow==12.1.1
+pillow==12.2.0
 protobuf==5.29.6
 psutil==7.0.0
 pyarrow==19.0.1
diff --git a/.aitk/requirements/requirements-IntelNPU.txt b/.aitk/requirements/requirements-IntelNPU.txt
@@ -76,7 +76,7 @@ optimum[openvino]==1.26.1
 optuna==4.2.1
 packaging==24.2
 pandas==2.2.3
-pillow==12.1.1
+pillow==12.2.0
 propcache==0.3.2
 protobuf==5.29.6
 psutil==7.0.0
diff --git a/.aitk/requirements/requirements-NvidiaGPU-GptqModel.txt b/.aitk/requirements/requirements-NvidiaGPU-GptqModel.txt
@@ -10,7 +10,7 @@ logbar==0.0.4
 maturin==1.9.3
 numpy==2.2.6
 packaging==24.2
-pillow==12.1.1
+pillow==12.2.0
 protobuf==6.32.0
 random_word==1.0.13
 safetensors==0.6.2
diff --git a/.aitk/requirements/requirements-NvidiaGPU.txt b/.aitk/requirements/requirements-NvidiaGPU.txt
@@ -49,7 +49,7 @@ optimum==1.26.1
 optuna==4.3.0
 packaging==24.2
 pandas==2.2.3
-pillow==12.1.1
+pillow==12.2.0
 propcache==0.3.2
 protobuf==5.29.6
 pyarrow==19.0.1
diff --git a/.aitk/requirements/requirements-QNN.txt b/.aitk/requirements/requirements-QNN.txt
@@ -43,7 +43,7 @@ onnxscript==0.5.3
 optuna==4.2.1
 packaging==24.2
 pandas==2.2.3
-pillow==12.1.1
+pillow==12.2.0
 propcache==0.3.2
 protobuf==5.29.6
 pyarrow==19.0.1
diff --git a/.aitk/requirements/requirements-WCR-QAI.txt b/.aitk/requirements/requirements-WCR-QAI.txt
@@ -1,4 +1,4 @@
-gdown==5.2.0
+gdown==5.2.2
 gitpython==3.1.46
 librosa==0.11.0
 qai_hub==0.42.0
diff --git a/.aitk/requirements/requirements-WCR.txt b/.aitk/requirements/requirements-WCR.txt
@@ -71,7 +71,7 @@ optuna==4.2.1
 packaging==24.2
 pandas==2.2.3
 parso==0.8.5
-pillow==12.1.1
+pillow==12.2.0
 platformdirs==4.4.0
 prompt-toolkit==3.0.52
 propcache==0.3.2
diff --git a/.aitk/requirements/requirements-WCR_CUDA.txt b/.aitk/requirements/requirements-WCR_CUDA.txt
@@ -72,7 +72,7 @@ optuna==4.2.1
 packaging==24.2
 pandas==2.2.3
 parso==0.8.5
-pillow==12.1.1
+pillow==12.2.0
 platformdirs==4.4.0
 prompt-toolkit==3.0.52
 propcache==0.3.2
diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml
@@ -1,6 +1,9 @@
 name: main
 description: "CI workflow for Olive Recipes"
 
+permissions:
+  contents: read
+
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
diff --git a/README.md b/README.md
diff --git a/google-gemma-3-1b-it/OpenVINO/README.md b/google-gemma-3-1b-it/OpenVINO/README.md
@@ -0,0 +1,76 @@
+# Gemma-3-1b-it Compression
+
+This folder contains a sample use case of Olive to optimize the [google/gemma-3-1b-it](https://huggingface.co/google/gemma-3-1b-it) model using Intel® OpenVINO tools.
+
+- Intel® NPU: [Gemma 3 1b it Dynamic Shape model optimized for NPU](#gemma-3-1b-it-npu)
+- Intel® GPU: [Gemma 3 1b it Dynamic Shape model optimized for GPU](#gemma-3-1b-it-gpu)
+
+## Quantization Workflows
+
+This workflow performs quantization with Optimum Intel®. It performs the optimization pipeline:
+
+- *Huggingface Model -> Quantized OpenVINO model -> Quantized encapsulated ONNX OpenVINO IR model*
+
+### Gemma 3 1b it NPU
+
+The flow in the following config file executes the above workflow producing a dynamic shape model.
+
+1. [gemma_3_1b_it_context_ov_npu_config.json](gemma_3_1b_it_context_ov_npu_config.json)
+
+### Gemma 3 1b it GPU
+
+The flow in the following config file executes the above workflow producing a dynamic shape model.
+
+1. [gemma_3_1b_it_context_ov_gpu_config.json](gemma_3_1b_it_context_ov_gpu_config.json)
+
+## How to run
+
+### Setup
+
+Install the necessary python packages:
+
+```bash
+python -m pip install olive-ai[openvino]
+python -m pip install -r requirements.txt
+```
+
+### Run Olive config
+
+The optimization techniques to run are specified in the relevant config json file.
+
+Optimize the model using the following command:
+
+```bash
+olive run --config <config_file.json>
+```
+
+Example:
+
+```bash
+olive run --config gemma_3_1b_it_context_ov_npu_config.json
+```
+
+or run simply with python code:
+
+```python
+from olive import run
+workflow_output = run("<config_file.json>")
+```
+
+After running the above command, the model candidates and corresponding config will be saved in the output directory.
+
+### (Optional) Run Console-Based Chat Interface
+
+To run ONNX OpenVINO IR Encapsulated GenAI models, please setup latest ONNXRuntime GenAI with ONNXRuntime OpenVINO EP support.
+
+The sample chat app to run is found as [model-chat.py](https://github.com/microsoft/onnxruntime-genai/blob/main/examples/python/model-chat.py) in the [onnxruntime-genai](https://github.com/microsoft/onnxruntime-genai/) GitHub repository.
+
+```bash
+python model-chat.py -e follow_config -v -g -m models/<model_folder>/
+```
+
+Example:
+
+```bash
+python model-chat.py -e follow_config -v -g -m models/gemma_3_1b_it_context_ov_npu/
+```
diff --git a/google-gemma-3-1b-it/OpenVINO/gemma_3_1b_it_context_ov_gpu_config.json b/google-gemma-3-1b-it/OpenVINO/gemma_3_1b_it_context_ov_gpu_config.json
@@ -0,0 +1,57 @@
+{
+    "input_model": {
+        "type": "HfModel",
+        "model_path": "google/gemma-3-1b-it"
+    },
+    "systems": {
+        "local_system": {
+            "type": "LocalSystem",
+            "accelerators": [
+                {
+                    "device": "gpu",
+                    "execution_providers": [
+                        "OpenVINOExecutionProvider"
+                    ]
+                }
+            ]
+        }
+    },
+    "passes": {
+        "optimum_convert": {
+            "type": "OpenVINOOptimumConversion",
+            "extra_args": {
+                "device": "gpu"
+            },
+            "ov_quant_config": {
+                "weight_format": "int4",
+                "group_size": 128,
+                "ratio": 0.8,
+                "trust_remote_code": true
+            }
+        },
+        "encapsulation": {
+            "type": "OpenVINOEncapsulation",
+            "target_device": "gpu",
+            "keep_ov_dynamic_dims": true,
+            "ov_version": "2026.1",
+            "reuse_cache": true,
+            "genai_config_override": {
+                "model": {
+                    "context_length": 32768
+                },
+                "search": {
+                    "do_sample": true,
+                    "max_length": 32768,
+                    "repetition_penalty": 1.1,
+                    "temperature": 1.0,
+                    "top_k": 64,
+                    "top_p": 0.95
+                }
+            }
+        }
+    },
+    "search_strategy": false,
+    "target": "local_system",
+    "evaluate_input_model": false,
+    "output_dir": "models/gemma_3_1b_it_context_ov_gpu"
+}
diff --git a/google-gemma-3-1b-it/OpenVINO/gemma_3_1b_it_context_ov_npu_config.json b/google-gemma-3-1b-it/OpenVINO/gemma_3_1b_it_context_ov_npu_config.json
@@ -0,0 +1,73 @@
+{
+    "input_model": {
+        "type": "HfModel",
+        "model_path": "google/gemma-3-1b-it"
+    },
+    "systems": {
+        "local_system": {
+            "type": "LocalSystem",
+            "accelerators": [
+                {
+                    "device": "npu",
+                    "execution_providers": [
+                        "OpenVINOExecutionProvider"
+                    ]
+                }
+            ]
+        }
+    },
+    "passes": {
+        "optimum_convert": {
+            "type": "OpenVINOOptimumConversion",
+            "extra_args": {
+                "device": "npu"
+            },
+            "ov_quant_config": {
+                "weight_format": "int4",
+                "group_size": 128,
+                "ratio": 1,
+                "dataset": "wikitext2",
+                "sym": true,
+                "trust_remote_code": true,
+                "sensitivity_metric": "weight_quantization_error"
+            }
+        },
+        "encapsulation": {
+            "type": "OpenVINOEncapsulation",
+            "target_device": "npu",
+            "keep_ov_dynamic_dims": true,
+            "ov_version": "2026.1",
+            "reuse_cache": true,
+            "genai_config_override": {
+                "model": {
+                    "context_length": 32768,
+                    "decoder": {
+                        "session_options": {
+                            "provider_options": [
+                                {
+                                    "OpenVINO": {
+                                        "device_type": "NPU",
+                                        "enable_causallm": "True",
+                                        "load_config": "{\"NPU\":{\"MAX_PROMPT_LEN\":\"4096\",\"MIN_RESPONSE_LEN\":\"128\"}}"
+                                    }
+                                }
+                            ]
+                        }
+                    }
+                },
+                "search": {
+                    "do_sample": true,
+                    "max_length": 32768,
+                    "repetition_penalty": 1.1,
+                    "temperature": 1.0,
+                    "top_k": 64,
+                    "top_p": 0.95
+                }
+            }
+        }
+    },
+    "search_strategy": false,
+    "target": "local_system",
+    "evaluate_input_model": false,
+    "output_dir": "models/gemma_3_1b_it_context_ov_npu"
+}
diff --git a/google-gemma-3-1b-it/OpenVINO/info.yml b/google-gemma-3-1b-it/OpenVINO/info.yml
@@ -0,0 +1,13 @@
+keywords:
+    openvino
+    olive
+arch: gemma
+recipes:
+    - file: "gemma_3_1b_it_context_ov_npu_config.json"
+      devices:
+        - npu
+      ep: OpenVINOExecutionProvider
+    - file: "gemma_3_1b_it_context_ov_gpu_config.json"
+      devices:
+        - gpu
+      ep: OpenVINOExecutionProvider
diff --git a/google-gemma-3-1b-it/OpenVINO/requirements.txt b/google-gemma-3-1b-it/OpenVINO/requirements.txt
@@ -0,0 +1,2 @@
+optimum-intel
+transformers
diff --git a/openai-whisper-large-v3-turbo/OpenVINO/README.md b/openai-whisper-large-v3-turbo/OpenVINO/README.md
@@ -14,6 +14,10 @@ This example demonstrates how to export and optimize OpenAI's Whisper Large V3 T
 Install the required packages:
 
 ```bash
+# For recent pip versions (supports direct URL syntax)
+python -m pip install "olive-ai[openvino] @ git+https://github.com/microsoft/olive.git@main"
+
+# For older pip versions (fallback format)
 python -m pip install "git+https://github.com/microsoft/olive.git@main#egg=olive-ai[openvino]"
 ```
 
diff --git a/openai-whisper-large-v3-turbo/OpenVINO/convert_whisper_to_ovir.py b/openai-whisper-large-v3-turbo/OpenVINO/convert_whisper_to_ovir.py
@@ -125,11 +125,19 @@ def reshape_and_save_to_tmp(input_ov_path: str, reshape_dict: dict, tmp_ov_path:
     @param reshape_dict: Dictionary of input names to new shapes.
     @param tmp_ov_path: Path to save the reshaped OpenVINO model.
     """
-    from openvino.runtime import Core, serialize
-    core = Core()
-    ov_model = core.read_model(input_ov_path)
-    ov_model.reshape(reshape_dict)
-    serialize(ov_model, tmp_ov_path)
+    try:
+        from openvino.runtime import Core, serialize
+        core = Core()
+        ov_model = core.read_model(input_ov_path)
+        ov_model.reshape(reshape_dict)
+        serialize(ov_model, tmp_ov_path)
+    except (ImportError, AttributeError):
+        # fallback to the newer 2025+ version of openvino that doesn't have openvino.runtime
+        import openvino as ov
+        core = ov.Core()
+        ov_model = core.read_model(input_ov_path)
+        ov_model.reshape(reshape_dict)
+        ov.serialize(ov_model, tmp_ov_path)
 
 
 def update_genai_overrides(encapsulate_json, w_config: WhisperConfig, enable_npu_ws: bool, reshape: bool, output_directory: str):

Original file line number	Diff line number	Diff line change
`@@ -1,4 +1,4 @@`
`1`		`-gdown==5.2.0`
	`1`	`+gdown==5.2.2`
`2`	`2`	`gitpython==3.1.46`
`3`	`3`	`librosa==0.11.0`
`4`	`4`	`qai_hub==0.42.0`