Deprecate compression option

Author: echarlaix

README.md

@@ -126,7 +126,7 @@ from optimum.intel import OVQuantizer, OVModelForSequenceClassification
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 
 model_id = "distilbert-base-uncased-finetuned-sst-2-english"
-model = AutoModelForSequenceClassification.from_pretrained(model_id)
+model = OVModelForSequenceClassification.from_pretrained(model_id, export=True)
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 def preprocess_fn(examples, tokenizer):
     return tokenizer(

docs/source/optimization_ov.mdx

@@ -26,11 +26,11 @@ Here is how to apply static quantization on a fine-tuned DistilBERT:
 
 ```python
 from functools import partial
-from transformers import AutoModelForSequenceClassification, AutoTokenizer
-from optimum.intel import OVConfig, OVQuantizer
+from transformers import  AutoTokenizer
+from optimum.intel import OVConfig, OVQuantizer, OVModelForSequenceClassification,
 
 model_id = "distilbert-base-uncased-finetuned-sst-2-english"
-model = AutoModelForSequenceClassification.from_pretrained(model_id)
+model = OVModelForSequenceClassification.from_pretrained(model_id, export=True)
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 # The directory where the quantized model will be saved
 save_dir = "ptq_model"

optimum/exporters/openvino/__main__.py

@@ -21,7 +21,8 @@
 
 from optimum.exporters import TasksManager
 from optimum.exporters.onnx.base import OnnxConfig
-from optimum.intel.openvino.configuration import OVConfig
+
+
 from optimum.utils.save_utils import maybe_load_preprocessors
 
 from ...intel.utils.import_utils import (
@@ -152,18 +153,21 @@ def main_export(
     if compression_option is not None:
         logger.warning(
             "The `compression_option` argument is deprecated and will be removed in optimum-intel v1.17.0. "
-            "Please, pass a `OVWeightQuantizationConfig` object in `quantization_config` argument instead."
+            "Please, pass an `ov_config` argument instead `OVConfig(..., quantization_config=quantization_config)`."
         )
 
     if compression_ratio is not None:
         logger.warning(
             "The `compression_ratio` argument is deprecated and will be removed in optimum-intel v1.17.0. "
-            "Please, pass a `OVWeightQuantizationConfig(ratio=compression_ratio)` object in `quantization_config` argument instead."
+            "Please, pass an `ov_config` argument instead `OVConfig(quantization_config={ratio=compression_ratio})`."
         )
 
     # default_config = _check_default_4bit_configs(config)
 
     if ov_config is None and compression_option is not None:
+
+        from ...intel.openvino.configuration import OVConfig
+
         if compression_option == "fp16":
             ov_config = OVConfig(dtype="fp16")
         elif compression_option != "fp32":

optimum/exporters/openvino/convert.py

@@ -32,7 +32,7 @@
 from optimum.exporters.onnx.convert import check_dummy_inputs_are_allowed
 from optimum.exporters.onnx.convert import export_pytorch as export_pytorch_to_onnx
 from optimum.exporters.onnx.convert import export_tensorflow as export_tensorflow_onnx
-from optimum.intel.openvino.configuration import OVConfig
+
 from optimum.utils import DEFAULT_DUMMY_SHAPES, is_diffusers_available
 from optimum.utils.save_utils import maybe_save_preprocessors
 
@@ -48,7 +48,6 @@
     remove_none_from_dummy_inputs,
 )
 
-
 if is_optimum_version(">=", "1.16.99"):
     from optimum.exporters.onnx.utils import _get_submodels_and_onnx_configs
 
@@ -599,6 +598,9 @@ def export_from_model(
 
         if num_parameters >= _MAX_UNCOMPRESSED_SIZE:
             if is_nncf_available():
+
+                from ...intel.openvino.configuration import OVConfig
+
                 ov_config = OVConfig(quantization_config={"bits": 8})
 
                 logger.info("The model weights will be quantized to int8.")

optimum/intel/__init__.py

@@ -58,16 +58,12 @@
         raise OptionalDependencyNotAvailable()
 except OptionalDependencyNotAvailable:
     _import_structure["utils.dummy_openvino_and_nncf_objects"] = [
-        "OVConfig",
         "OVQuantizer",
         "OVTrainer",
         "OVTrainingArguments",
-        "OVWeightQuantizationConfig",
     ]
 else:
-    _import_structure["openvino"].extend(
-        ["OVConfig", "OVQuantizer", "OVTrainer", "OVTrainingArguments", "OVWeightQuantizationConfig"]
-    )
+    _import_structure["openvino"].extend(["OVQuantizer", "OVTrainer", "OVTrainingArguments"])
 
 try:
     if not (is_openvino_available() and is_diffusers_available()):
@@ -119,6 +115,8 @@
             "OVModelForSpeechSeq2Seq",
             "OVModelForSequenceClassification",
             "OVModelForTokenClassification",
+            "OVWeightQuantizationConfig",
+            "OVConfig",
         ]
     )
 
@@ -180,14 +178,12 @@
             raise OptionalDependencyNotAvailable()
     except OptionalDependencyNotAvailable:
         from .utils.dummy_openvino_and_nncf_objects import (
-            OVConfig,
             OVQuantizer,
             OVTrainer,
             OVTrainingArguments,
-            OVWeightQuantizationConfig,
         )
     else:
-        from .openvino import OVConfig, OVQuantizer, OVTrainer, OVTrainingArguments, OVWeightQuantizationConfig
+        from .openvino import OVQuantizer, OVTrainer, OVTrainingArguments
 
     try:
         if not (is_openvino_available() and is_diffusers_available()):
@@ -218,6 +214,7 @@
         from .utils.dummy_openvino_objects import *
     else:
         from .openvino import (
+            OVConfig,
             OVModelForAudioClassification,
             OVModelForAudioFrameClassification,
             OVModelForAudioXVector,
@@ -231,6 +228,7 @@
             OVModelForSequenceClassification,
             OVModelForSpeechSeq2Seq,
             OVModelForTokenClassification,
+            OVWeightQuantizationConfig,
         )
 
     try:

optimum/intel/openvino/__init__.py

@@ -36,11 +36,12 @@
 
     patch_torch_operators()
 
-    from .configuration import OVConfig, OVWeightQuantizationConfig
     from .quantization import OVQuantizer
     from .trainer import OVTrainer
     from .training_args import OVTrainingArguments
 
+
+from .configuration import OVConfig, OVWeightQuantizationConfig
 from .modeling import (
     OVModelForAudioClassification,
     OVModelForAudioFrameClassification,

optimum/intel/utils/dummy_openvino_and_nncf_objects.py

@@ -46,25 +46,3 @@ def __init__(self, *args, **kwargs):
     @classmethod
     def from_pretrained(cls, *args, **kwargs):
         requires_backends(cls, ["openvino", "nncf"])
-
-
-class OVConfig(metaclass=DummyObject):
-    _backends = ["openvino", "nncf"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["openvino", "nncf"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["openvino", "nncf"])
-
-
-class OVWeightQuantizationConfig(metaclass=DummyObject):
-    _backends = ["openvino", "nncf"]
-
-    def __init__(self, *args, **kwargs):
-        requires_backends(self, ["openvino", "nncf"])
-
-    @classmethod
-    def from_pretrained(cls, *args, **kwargs):
-        requires_backends(cls, ["openvino", "nncf"])

optimum/intel/utils/dummy_openvino_objects.py

@@ -167,3 +167,25 @@ def __init__(self, *args, **kwargs):
     @classmethod
     def from_pretrained(cls, *args, **kwargs):
         requires_backends(cls, ["openvino"])
+
+
+class OVConfig(metaclass=DummyObject):
+    _backends = ["openvino"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["openvino"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(self, ["openvino"])
+
+
+class OVWeightQuantizationConfig(metaclass=DummyObject):
+    _backends = ["openvino"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["openvino"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(self, ["openvino"])