Infer if the model needs to be exported (#825)

* Infer if the model needs to be exported

* add test

* fix test revision

* remove infer export in pipeline

* apply comments

* fix offline mode

Author: echarlaix

docs/source/openvino/reference.mdx

@@ -19,7 +19,7 @@ limitations under the License.
 ## Generic model classes
 
 [[autodoc]] openvino.modeling_base.OVBaseModel
-    - _from_pretrained
+    - from_pretrained
     - reshape
 
 ## Natural Language Processing

docs/source/openvino/tutorials/diffusers.mdx

@@ -50,18 +50,14 @@ To further speed up inference, the model can be statically reshaped :
 
 ```python
 # Define the shapes related to the inputs and desired outputs
-batch_size = 1
-num_images_per_prompt = 1
-height = 512
-width = 512
-
+batch_size, num_images, height, width = 1, 1, 512, 512
 # Statically reshape the model
-pipeline.reshape(batch_size=batch_size, height=height, width=width, num_images_per_prompt=num_images_per_prompt)
+pipeline.reshape(batch_size=batch_size, height=height, width=width, num_images_per_prompt=num_images)
 # Compile the model before the first inference
 pipeline.compile()
 
 # Run inference
-images = pipeline(prompt, height=height, width=width, num_images_per_prompt=num_images_per_prompt).images
+images = pipeline(prompt, height=height, width=width, num_images_per_prompt=num_images).images
 ```
 
 In case you want to change any parameters such as the outputs height or width, you'll need to statically reshape your model once again.

optimum/intel/openvino/modeling.py

@@ -14,7 +14,6 @@
 
 import logging
 import os
-import warnings
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import Dict, Optional, Union
@@ -417,7 +416,6 @@ def _from_transformers(
         cls,
         model_id: str,
         config: PretrainedConfig,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -430,15 +428,6 @@ def _from_transformers(
         quantization_config: Union[OVWeightQuantizationConfig, Dict] = None,
         **kwargs,
     ):
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         save_dir = TemporaryDirectory()
         save_dir_path = Path(save_dir.name)
         # This attribute is needed to keep one reference on the temporary directory, since garbage collecting
@@ -591,7 +580,6 @@ def from_pretrained(
         model_id: Union[str, Path],
         export: bool = False,
         config: Optional["PretrainedConfig"] = None,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -602,15 +590,6 @@ def from_pretrained(
         trust_remote_code: bool = False,
         **kwargs,
     ):
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         # Fix the mismatch between timm_config and huggingface_config
         local_timm_model = _is_timm_ov_dir(model_id)
         if local_timm_model or (not os.path.isdir(model_id) and model_info(model_id).library_name == "timm"):

optimum/intel/openvino/modeling_base.py

@@ -28,12 +28,14 @@
 from transformers import GenerationConfig, PretrainedConfig
 from transformers.file_utils import add_start_docstrings
 from transformers.generation import GenerationMixin
+from transformers.utils import is_offline_mode
 
 from optimum.exporters.onnx import OnnxConfig
-from optimum.modeling_base import OptimizedModel
+from optimum.modeling_base import FROM_PRETRAINED_START_DOCSTRING, OptimizedModel
 
 from ...exporters.openvino import export, main_export
 from ..utils.import_utils import is_nncf_available
+from ..utils.modeling_utils import _find_files_matching_pattern
 from .configuration import OVConfig, OVDynamicQuantizationConfig, OVWeightQuantizationConfig
 from .utils import ONNX_WEIGHTS_NAME, OV_TO_PT_TYPE, OV_XML_FILE_NAME, _print_compiled_model_properties
 
@@ -220,7 +222,6 @@ def _from_pretrained(
         cls,
         model_id: Union[str, Path],
         config: PretrainedConfig,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -242,8 +243,6 @@ def _from_pretrained(
                 Can be either:
                     - The model id of a pretrained model hosted inside a model repo on huggingface.co.
                     - The path to a directory containing the model weights.
-            use_auth_token (Optional[Union[bool, str]], defaults to `None`):
-                Deprecated. Please use `token` instead.
             token (Optional[Union[bool, str]], defaults to `None`):
                 The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
                 when running `huggingface-cli login` (stored in `~/.huggingface`).
@@ -263,15 +262,6 @@ def _from_pretrained(
             load_in_8bit (`bool`, *optional*, defaults to `False`):
                 Whether or not to apply 8-bit weight quantization.
         """
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         model_path = Path(model_id)
         default_file_name = ONNX_WEIGHTS_NAME if from_onnx else OV_XML_FILE_NAME
         file_name = file_name or default_file_name
@@ -312,6 +302,87 @@ def _from_pretrained(
             **kwargs,
         )
 
+    @classmethod
+    @add_start_docstrings(FROM_PRETRAINED_START_DOCSTRING)
+    def from_pretrained(
+        cls,
+        model_id: Union[str, Path],
+        export: bool = False,
+        force_download: bool = False,
+        use_auth_token: Optional[Union[bool, str]] = None,
+        token: Optional[Union[bool, str]] = None,
+        cache_dir: str = HUGGINGFACE_HUB_CACHE,
+        subfolder: str = "",
+        config: Optional[PretrainedConfig] = None,
+        local_files_only: bool = False,
+        trust_remote_code: bool = False,
+        revision: Optional[str] = None,
+        **kwargs,
+    ):
+        if use_auth_token is not None:
+            warnings.warn(
+                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
+                FutureWarning,
+            )
+            if token is not None:
+                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
+            token = use_auth_token
+
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
+        _export = export
+        try:
+            if local_files_only:
+                object_id = model_id.replace("/", "--")
+                cached_model_dir = os.path.join(cache_dir, f"models--{object_id}")
+                refs_file = os.path.join(os.path.join(cached_model_dir, "refs"), revision or "main")
+                with open(refs_file) as f:
+                    revision = f.read()
+                model_dir = os.path.join(cached_model_dir, "snapshots", revision)
+            else:
+                model_dir = model_id
+
+            ov_files = _find_files_matching_pattern(
+                model_dir,
+                pattern=r"(.*)?openvino(.*)?\_model.xml",
+                subfolder=subfolder,
+                use_auth_token=token,
+                revision=revision,
+            )
+            _export = len(ov_files) == 0
+            if _export ^ export:
+                if export:
+                    logger.warning(
+                        f"The model {model_id} was already converted to the OpenVINO IR but got `export=True`, the model will be converted to OpenVINO once again. "
+                        "Don't forget to save the resulting model with `.save_pretrained()`"
+                    )
+                    _export = True
+                else:
+                    logger.warning(
+                        f"No OpenVINO files were found for {model_id}, setting `export=True` to convert the model to the OpenVINO IR. "
+                        "Don't forget to save the resulting model with `.save_pretrained()`"
+                    )
+        except Exception as exception:
+            logger.warning(
+                f"Could not infer whether the model was already converted or not to the OpenVINO IR, keeping `export={export}`.\n{exception}"
+            )
+
+        return super().from_pretrained(
+            model_id,
+            export=_export,
+            force_download=force_download,
+            token=token,
+            cache_dir=cache_dir,
+            subfolder=subfolder,
+            config=config,
+            local_files_only=local_files_only,
+            trust_remote_code=trust_remote_code,
+            revision=revision,
+            **kwargs,
+        )
+
     @staticmethod
     def _prepare_weight_quantization_config(
         quantization_config: Optional[Union[OVWeightQuantizationConfig, Dict]] = None, load_in_8bit: bool = False
@@ -337,7 +408,6 @@ def _set_ov_config_parameters(self):
     @staticmethod
     def _cached_file(
         model_path: Union[Path, str],
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -346,15 +416,6 @@ def _cached_file(
         subfolder: str = "",
         local_files_only: bool = False,
     ):
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         # locates a file in a local folder and repo, downloads and cache it if necessary.
         model_path = Path(model_path)
         if model_path.is_dir():
@@ -385,7 +446,6 @@ def _from_transformers(
         cls,
         model_id: str,
         config: PretrainedConfig,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -409,8 +469,6 @@ def _from_transformers(
                     - The path to a directory containing the model weights.            save_dir (`str` or `Path`):
                 The directory where the exported ONNX model should be saved, default to
                 `transformers.file_utils.default_cache_path`, which is the cache directory for transformers.
-            use_auth_token (`Optional[str]`, defaults to `None`):
-                Deprecated. Please use `token` instead.
             token (Optional[Union[bool, str]], defaults to `None`):
                 The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
                 when running `huggingface-cli login` (stored in `~/.huggingface`).
@@ -419,15 +477,6 @@ def _from_transformers(
             kwargs (`Dict`, *optional*):
                 kwargs will be passed to the model during initialization
         """
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         save_dir = TemporaryDirectory()
         save_dir_path = Path(save_dir.name)
         # This attribute is needed to keep one reference on the temporary directory, since garbage collecting
@@ -469,7 +518,6 @@ def _to_load(
         model,
         config: PretrainedConfig,
         onnx_config: OnnxConfig,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -478,15 +526,6 @@ def _to_load(
         stateful: bool = False,
         **kwargs,
     ):
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         save_dir = TemporaryDirectory()
         save_dir_path = Path(save_dir.name)
 

optimum/intel/openvino/modeling_base_seq2seq.py

@@ -14,7 +14,6 @@
 
 import logging
 import os
-import warnings
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import Dict, Optional, Union
@@ -120,7 +119,6 @@ def _from_pretrained(
         cls,
         model_id: Union[str, Path],
         config: PretrainedConfig,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -144,8 +142,6 @@ def _from_pretrained(
                 Can be either:
                     - The model id of a pretrained model hosted inside a model repo on huggingface.co.
                     - The path to a directory containing the model weights.
-            use_auth_token (Optional[Union[bool, str]], defaults to `None`):
-                Deprecated. Please use `token` instead.
             token (Optional[Union[bool, str]], defaults to `None`):
                 The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
                 when running `huggingface-cli login` (stored in `~/.huggingface`).
@@ -169,15 +165,6 @@ def _from_pretrained(
             local_files_only(`bool`, *optional*, defaults to `False`):
                 Whether or not to only look at local files (i.e., do not try to download the model).
         """
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         default_encoder_file_name = ONNX_ENCODER_NAME if from_onnx else OV_ENCODER_NAME
         default_decoder_file_name = ONNX_DECODER_NAME if from_onnx else OV_DECODER_NAME
         default_decoder_with_past_file_name = ONNX_DECODER_WITH_PAST_NAME if from_onnx else OV_DECODER_WITH_PAST_NAME
@@ -256,7 +243,6 @@ def _from_transformers(
         cls,
         model_id: str,
         config: PretrainedConfig,
-        use_auth_token: Optional[Union[bool, str]] = None,
         token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -282,8 +268,6 @@ def _from_transformers(
             save_dir (`str` or `Path`):
                 The directory where the exported ONNX model should be saved, defaults to
                 `transformers.file_utils.default_cache_path`, which is the cache directory for transformers.
-            use_auth_token (`Optional[str]`, defaults to `None`):
-                Deprecated. Please use `token` instead.
             token (Optional[Union[bool, str]], defaults to `None`):
                 The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
                 when running `huggingface-cli login` (stored in `~/.huggingface`).
@@ -292,15 +276,6 @@ def _from_transformers(
             kwargs (`Dict`, *optional*):
                 kwargs will be passed to the model during initialization
         """
-        if use_auth_token is not None:
-            warnings.warn(
-                "The `use_auth_token` argument is deprecated and will be removed soon. Please use the `token` argument instead.",
-                FutureWarning,
-            )
-            if token is not None:
-                raise ValueError("You cannot use both `use_auth_token` and `token` arguments at the same time.")
-            token = use_auth_token
-
         save_dir = TemporaryDirectory()
         save_dir_path = Path(save_dir.name)