Remove datasets as required dependency (#1466)

* Remove datasets as required dependency

* move datasets import

* add datasets to inc extra

---------

Co-authored-by: Ilyas Moutawwakil <[email protected]>

Author: echarlaix

.github/workflows/build_documentation.yml

@@ -51,7 +51,7 @@ jobs:
         run: |
           pip install --upgrade pip uv
           uv pip install git+https://github.com/huggingface/doc-builder
-          uv pip install .[quality] nncf openvino neural-compressor[pt]>3.4 diffusers accelerate
+          uv pip install .[quality] nncf openvino neural-compressor[pt]>3.4 diffusers accelerate datasets
 
       - name: Make documentation
         shell: bash

.github/workflows/build_pr_documentation.yml

@@ -38,7 +38,7 @@ jobs:
         run: |
           pip install --upgrade pip uv
           uv pip install git+https://github.com/huggingface/doc-builder
-          uv pip install .[quality] nncf openvino neural-compressor[pt]>3.4 diffusers accelerate
+          uv pip install .[quality] nncf openvino neural-compressor[pt]>3.4 diffusers accelerate datasets
 
       - name: Make documentation
         shell: bash

optimum/intel/openvino/quantization.py

@@ -23,7 +23,6 @@
 from pathlib import Path
 from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple, Union
 
-import datasets
 import nncf
 import numpy as np
 import openvino
@@ -360,7 +359,15 @@ def build_from_quantization_config(self, config: OVQuantizationConfigBase) -> OV
                     streaming=dataset_metadata["streaming"],
                 )
             elif isinstance(config.dataset, list) and all(isinstance(it, str) for it in config.dataset):
-                dataset = datasets.Dataset.from_list([{"text": it} for it in config.dataset])
+                if not is_datasets_available():
+                    raise ValueError(
+                        DATASETS_IMPORT_ERROR.format("OVCalibrationDatasetBuilder.build_from_quantization_config")
+                    )
+
+                from datasets import Dataset
+
+                dataset = Dataset.from_list([{"text": it} for it in config.dataset])
+
             else:
                 raise ValueError(
                     "Please provide dataset as one of the accepted dataset labels or as a list of strings."

setup.py

@@ -30,9 +30,7 @@
     "torch>=2.1",
     "optimum-onnx==0.0.*",
     "transformers>=4.45,<4.56",
-    "datasets>=1.4.0",
     "setuptools",
-    "scipy",
 ]
 
 TESTS_REQUIRE = [
@@ -68,7 +66,7 @@
 EXTRAS_REQUIRE = {
     "nncf": ["nncf>=2.18.0"],
     "openvino": ["nncf>=2.18.0", "openvino>=2025.1.0", "openvino-tokenizers>=2025.1.0"],
-    "neural-compressor": ["neural-compressor[pt]>=3.4.1", "accelerate", "transformers<4.46"],
+    "neural-compressor": ["neural-compressor[pt]>=3.4.1", "accelerate", "transformers<4.46", "datasets"],
     "ipex": ["intel-extension-for-pytorch>=2.8", "transformers>4.54,<4.56", "accelerate"],
     "diffusers": ["diffusers"],
     "quality": QUALITY_REQUIRE,