Add new _export method

optimum/intel/openvino/modeling_base.py

@@ -22,21 +22,16 @@
 from huggingface_hub import hf_hub_download
 from openvino import Core, convert_model
 from openvino._offline_transformations import apply_moc_transformations, compress_model_transformation
-from transformers import PretrainedConfig
+from transformers import PretrainedConfig, GenerationConfig
 from transformers.file_utils import add_start_docstrings
 
 from optimum.exporters.onnx import OnnxConfig
 from optimum.modeling_base import OptimizedModel
 
 from ...exporters.openvino import export, main_export
-from ..utils.import_utils import is_nncf_available, is_transformers_version
+from ..utils.import_utils import is_nncf_available
 from .utils import ONNX_WEIGHTS_NAME, OV_XML_FILE_NAME, _print_compiled_model_properties
-
-
-if is_transformers_version("<", "4.25.0"):
-    from transformers.generation_utils import GenerationMixin
-else:
-    from transformers.generation import GenerationMixin
+from transformers.generation import GenerationMixin
 
 core = Core()
 
@@ -92,12 +87,7 @@ def __init__(
         if enable_compilation:
             self.compile()
 
-        if is_transformers_version("<=", "4.25.1"):
-            self.generation_config = None
-        else:
-            from transformers import GenerationConfig
-
-            self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
+        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
 
     @staticmethod
     def load_model(file_name: Union[str, Path], load_in_8bit: bool = False):
@@ -247,7 +237,7 @@ def _cached_file(
         return model_cache_path
 
     @classmethod
-    def _from_transformers(
+    def _export(
         cls,
         model_id: str,
         config: PretrainedConfig,
@@ -304,6 +294,12 @@ def _from_transformers(
         config.save_pretrained(save_dir_path)
         return cls._from_pretrained(model_id=save_dir_path, config=config, load_in_8bit=False, **kwargs)
 
+    @classmethod
+    def _from_transformers(cls, *args, **kwargs):
+        # TODO : add warning when from_pretrained_method is set to cls._export instead of cls._from_transformers when export=True
+        # logger.warning("The method `_from_transformers` is deprecated, please use `_export` instead")
+        return cls._export(*args, **kwargs)
+
     @classmethod
     def _to_load(
         cls,

optimum/intel/openvino/modeling_base_seq2seq.py

@@ -21,11 +21,10 @@
 import openvino
 from huggingface_hub import hf_hub_download
 from openvino._offline_transformations import apply_moc_transformations, compress_model_transformation
-from transformers import PretrainedConfig
+from transformers import PretrainedConfig, GenerationConfig
 from transformers.file_utils import add_start_docstrings
 
 from ...exporters.openvino import main_export
-from ..utils.import_utils import is_transformers_version
 from .modeling_base import OVBaseModel
 from .utils import (
     ONNX_DECODER_NAME,
@@ -75,13 +74,7 @@ def __init__(
         self.encoder_model = encoder
         self.decoder_model = decoder
         self.decoder_with_past_model = decoder_with_past
-
-        if is_transformers_version("<=", "4.25.1"):
-            self.generation_config = None
-        else:
-            from transformers import GenerationConfig
-
-            self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
+        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
 
     def _save_pretrained(self, save_directory: Union[str, Path]):
         """

optimum/intel/openvino/modeling_decoder.py

@@ -30,16 +30,10 @@
 from optimum.utils import NormalizedConfigManager
 
 from ...exporters.openvino import main_export
-from ..utils.import_utils import is_transformers_version
 from ..utils.modeling_utils import MULTI_QUERY_ATTN_MODELS
 from .modeling import _TOKENIZER_FOR_DOC, INPUTS_DOCSTRING, MODEL_START_DOCSTRING, OVModel
 from .utils import ONNX_WEIGHTS_NAME, OV_XML_FILE_NAME, STR_TO_OV_TYPE
-
-
-if is_transformers_version("<", "4.25.0"):
-    from transformers.generation_utils import GenerationMixin
-else:
-    from transformers.generation import GenerationMixin
+from transformers.generation import GenerationMixin
 
 
 logger = logging.getLogger(__name__)

optimum/intel/openvino/modeling_seq2seq.py

@@ -36,16 +36,11 @@
 from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 from transformers.models.whisper.tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
 
-from ..utils.import_utils import is_transformers_version
 from .modeling_base_seq2seq import OVBaseModelForSeq2SeqLM
 from .utils import _print_compiled_model_properties
+from transformers.generation import GenerationMixin
 
 
-if is_transformers_version("<", "4.25.0"):
-    from transformers.generation_utils import GenerationMixin
-else:
-    from transformers.generation import GenerationMixin
-
 if TYPE_CHECKING:
     from transformers import PretrainedConfig
 

tests/openvino/test_modeling_basic.py

@@ -1,6 +1,6 @@
 """
 The goal of the test in this file is to test that basic functionality of optimum[openvino] works:
-- Load the model with `from_transformers=True`
+- Load the model with `export=True`
 - Do inference with appropriate pipeline
 - Save the model to disk
 
@@ -58,7 +58,7 @@ def test_pipeline(self, model_id):
         tokenizer = AutoTokenizer.from_pretrained(model_id)
         model_class_str = MODEL_NAMES[model_id]
         model_class = eval(model_class_str)
-        model = model_class.from_pretrained(model_id, from_transformers=True)
+        model = model_class.from_pretrained(model_id, export=True)
         model.save_pretrained(f"{model_id}_ov")
         model = model_class.from_pretrained(f"{model_id}_ov")
 
@@ -80,7 +80,7 @@ def test_openvino_methods(self):
         """
         model_id = "hf-internal-testing/tiny-random-distilbert"
         tokenizer = AutoTokenizer.from_pretrained(model_id)
-        model = OVModelForSequenceClassification.from_pretrained(model_id, from_transformers=True)
+        model = OVModelForSequenceClassification.from_pretrained(model_id, export=True)
         model.reshape(1, 16)
         model.half()
         model.to("cpu")