Introduce OVQuantizationConfig for nncf.quantize() parameters (huggingface#638)

* Introduce OVQuantizationConfig for nncf.quantize() parameters

* Ignored scope tweaks

* Added **kwargs to quantization call. Added config serialization test.

* Ignored scope changes. Tests pass.

* Added documentation

* Linters

* Linters

* Tweak ignored scope serialization

* Added deprecation errors, tweak docs

* Addressed minor comments

* Make quantization config contain only serializable properties.

* Small tweaks

* Address comments

* Fix ruff

* Fix ruff 2

Author: nikita-savelyevv

optimum/intel/__init__.py

@@ -124,6 +124,7 @@
             "OVModelForVision2Seq",
             "OVModelForSequenceClassification",
             "OVModelForTokenClassification",
+            "OVQuantizationConfig",
             "OVWeightQuantizationConfig",
             "OVConfig",
         ]
@@ -243,6 +244,7 @@
             OVModelForSpeechSeq2Seq,
             OVModelForTokenClassification,
             OVModelForVision2Seq,
+            OVQuantizationConfig,
             OVWeightQuantizationConfig,
         )
 

optimum/intel/openvino/__init__.py

@@ -43,7 +43,7 @@
         from .trainer import OVTrainer
 
 
-from .configuration import OVConfig, OVWeightQuantizationConfig
+from .configuration import OVConfig, OVQuantizationConfig, OVWeightQuantizationConfig
 from .modeling import (
     OVModelForAudioClassification,
     OVModelForAudioFrameClassification,

optimum/intel/openvino/configuration.py

@@ -100,13 +100,21 @@ def __init__(
             self._openvino_config = OVConfig(quantization_config=quantization_config)
 
     @staticmethod
-    def load_model(file_name: Union[str, Path], quantization_config: Union[OVWeightQuantizationConfig, Dict] = None):
+    def load_model(
+        file_name: Union[str, Path],
+        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
+        calibration_dataset: Optional = None,
+    ):
         """
         Loads the model.
 
         Arguments:
             file_name (`str` or `Path`):
                 The path of the model ONNX or XML file.
+            quantization_config (`OVWeightQuantizationConfig` or `Dict`, *optional*):
+                Quantization config to apply after model is loaded.
+            calibration_dataset (`nncf.Dataset`, *optional*):
+                Optional nncf.Dataset to feed to model weight compression when quantization config is provided.
         """
 
         def fix_op_names_duplicates(model: openvino.runtime.Model):
@@ -135,7 +143,7 @@ def fix_op_names_duplicates(model: openvino.runtime.Model):
 
             from optimum.intel.openvino.quantization import _weight_only_quantization
 
-            model = _weight_only_quantization(model, quantization_config)
+            model = _weight_only_quantization(model, quantization_config, calibration_dataset=calibration_dataset)
 
         return model
 

optimum/intel/openvino/modeling_base.py

@@ -572,7 +572,7 @@ def _from_pretrained(
         from_onnx: bool = False,
         local_files_only: bool = False,
         load_in_8bit: bool = False,
-        quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
+        quantization_config: Optional[Union[OVWeightQuantizationConfig, Dict]] = None,
         **kwargs,
     ):
         model_path = Path(model_id)
@@ -596,7 +596,12 @@ def _from_pretrained(
         quantization_config = cls._prepare_weight_quantization_config(quantization_config, load_in_8bit)
 
         load_in_4bit = quantization_config.bits == 4 if quantization_config else False
-        model = cls.load_model(model_cache_path, quantization_config=None if load_in_4bit else quantization_config)
+        calibration_dataset = kwargs.get("calibration_dataset", None)
+        model = cls.load_model(
+            model_cache_path,
+            quantization_config=None if load_in_4bit else quantization_config,
+            calibration_dataset=calibration_dataset,
+        )
 
         model_type = config.model_type.replace("_", "-")
         if model_type == "bloom":
@@ -632,7 +637,7 @@ def _from_pretrained(
                     f"For the given model, we recommend the following `quantization_config` : {default_config}"
                 )
 
-            if isinstance(quantization_config.dataset, str):
+            if calibration_dataset is None and isinstance(quantization_config.dataset, str):
                 tokenizer = quantization_config.tokenizer or AutoTokenizer.from_pretrained(model_id)
 
                 from optimum.gptq.data import get_dataset, prepare_dataset
@@ -644,9 +649,9 @@ def _from_pretrained(
                 dataset = get_dataset(quantization_config.dataset, tokenizer, seqlen=32, nsamples=nsamples)
                 dataset = prepare_dataset(dataset)
                 quantization_config = copy.deepcopy(quantization_config)
-                quantization_config.dataset = nncf.Dataset(dataset, lambda x: causal_model.prepare_inputs(**x))
+                calibration_dataset = nncf.Dataset(dataset, lambda x: causal_model.prepare_inputs(**x))
 
-            _weight_only_quantization(model, quantization_config)
+            _weight_only_quantization(model, quantization_config, calibration_dataset)
 
         return causal_model