Add transformers v4.45 support

optimum/exporters/openvino/convert.py

@@ -39,6 +39,7 @@
     _torch_version,
     _transformers_version,
     compare_versions,
+    is_transformers_version,
 )
 from optimum.utils import DEFAULT_DUMMY_SHAPES, is_diffusers_available
 from optimum.utils.save_utils import maybe_save_preprocessors
@@ -615,6 +616,18 @@ def export_from_model(
     logging.disable(logging.NOTSET)
 
     if library_name != "diffusers":
+        if is_transformers_version(">=", "4.44.99"):
+            misplaced_generation_parameters = model.config._get_non_default_generation_parameters()
+            if model.can_generate() and len(misplaced_generation_parameters) > 0:
+                logger.warning(
+                    "Moving the following attributes in the config to the generation config: "
+                    f"{misplaced_generation_parameters}. You are seeing this warning because you've set "
+                    "generation parameters in the model config, as opposed to in the generation config.",
+                )
+                for param_name, param_value in misplaced_generation_parameters.items():
+                    setattr(model.generation_config, param_name, param_value)
+                    setattr(model.config, param_name, None)
+
         # Saving the model config and preprocessor as this is needed sometimes.
         model.config.save_pretrained(output)
         generation_config = getattr(model, "generation_config", None)

optimum/intel/openvino/modeling_base.py

@@ -34,7 +34,7 @@
 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.import_utils import is_nncf_available, is_transformers_version
 from ..utils.modeling_utils import _find_files_matching_pattern
 from .configuration import OVConfig, OVDynamicQuantizationConfig, OVWeightQuantizationConfig
 from .utils import (
@@ -125,11 +125,23 @@ def __init__(
 
         self.output_names = output_names
         self.output_dtypes = output_dtypes
-
         self.model = model
         self.request = None if not self._compile_only else self.model
         if self.can_generate():
             self.generation_config = kwargs.get("generation_config", GenerationConfig.from_model_config(config))
+
+            if is_transformers_version(">=", "4.44.99"):
+                misplaced_generation_parameters = self.config._get_non_default_generation_parameters()
+                if len(misplaced_generation_parameters) > 0:
+                    logger.warning(
+                        "Moving the following attributes in the config to the generation config: "
+                        f"{misplaced_generation_parameters}. You are seeing this warning because you've set "
+                        "generation parameters in the model config, as opposed to in the generation config.",
+                    )
+                    for param_name, param_value in misplaced_generation_parameters.items():
+                        setattr(self.generation_config, param_name, param_value)
+                        setattr(self.config, param_name, None)
+
         else:
             self.generation_config = None
 
@@ -351,19 +363,6 @@ def _from_pretrained(
                 model_save_dir=model_cache_path.parent,
             )
 
-        try:
-            generation_config = GenerationConfig.from_pretrained(
-                model_id,
-                token=token,
-                revision=revision,
-                subfolder=subfolder,
-                force_download=force_download,
-                cache_dir=cache_dir,
-            )
-            kwargs["generation_config"] = generation_config
-        except Exception:
-            pass
-
         return cls(
             model,
             config=config,

optimum/intel/openvino/modeling_base_seq2seq.py

@@ -26,6 +26,7 @@
 from transformers.file_utils import add_start_docstrings
 
 from ...exporters.openvino import main_export
+from ..utils.import_utils import is_transformers_version
 from .configuration import OVConfig, OVWeightQuantizationConfig
 from .modeling_base import OVBaseModel
 from .utils import (
@@ -78,10 +79,21 @@ def __init__(
         self.encoder_model = encoder
         self.decoder_model = decoder
         self.decoder_with_past_model = decoder_with_past
-        if self.can_generate():
-            self.generation_config = kwargs.get("generation_config", GenerationConfig.from_model_config(config))
-        else:
-            self.generation_config = None
+
+        self.generation_config = kwargs.get("generation_config", GenerationConfig.from_model_config(config))
+
+        if is_transformers_version(">=", "4.44.99"):
+            misplaced_generation_parameters = self.config._get_non_default_generation_parameters()
+            if len(misplaced_generation_parameters) > 0:
+                logger.warning(
+                    "Moving the following attributes in the config to the generation config: "
+                    f"{misplaced_generation_parameters}. You are seeing this warning because you've set "
+                    "generation parameters in the model config, as opposed to in the generation config.",
+                )
+                for param_name, param_value in misplaced_generation_parameters.items():
+                    setattr(self.generation_config, param_name, param_value)
+                    setattr(self.config, param_name, None)
+
         self._openvino_config = None
         if quantization_config:
             self._openvino_config = OVConfig(quantization_config=quantization_config)
@@ -166,6 +178,9 @@ 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).
         """
+        generation_config = kwargs.pop("generation_config", None)
+        subfolder = kwargs.pop("subfolder", "")
+
         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
@@ -229,6 +244,7 @@ def _from_pretrained(
                     cache_dir=cache_dir,
                     force_download=force_download,
                     local_files_only=local_files_only,
+                    subfolder=subfolder,
                 )
                 file_names[name] = model_cache_path
 
@@ -252,18 +268,24 @@ def _from_pretrained(
                         kwargs.get("ov_config"),
                         model_save_dir,
                     )
-        try:
-            generation_config = GenerationConfig.from_pretrained(
-                model_id,
-                token=token,
-                revision=revision,
-                cache_dir=cache_dir,
-                force_download=force_download,
-                local_files_only=local_files_only,
-            )
-            kwargs["generation_config"] = generation_config
-        except Exception:
-            pass
+
+        if generation_config is None:
+            try:
+                generation_config = GenerationConfig.from_pretrained(
+                    model_id,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                )
+                if getattr(generation_config, "cache_implementation", None) is not None:
+                    generation_config.cache_implementation = None
+            except OSError:
+                logger.info(
+                    "Generation config file not found, using a generation config created from the model config."
+                )
 
         return cls(
             encoder=encoder,
@@ -272,6 +294,7 @@ def _from_pretrained(
             config=config,
             model_save_dir=model_save_dir,
             quantization_config=quantization_config,
+            generation_config=generation_config,
             **kwargs,
         )
 

optimum/intel/openvino/modeling_decoder.py

@@ -143,9 +143,8 @@ def __init__(
         self.num_pkv = 2
         self.key_value_input_names = [key for key in self.input_names if "key_values" in key]
         self.key_value_output_names = [key for key in self.output_names if "present" in key]
-        self._original_model = (
-            self.model.clone() if not compile_only else None
-        )  # keep original model for serialization
+        # Keeping the original model for serialization
+        self._original_model = self.model.clone() if not compile_only else None
         self._pkv_precision = Type.f32
         self.next_beam_idx = None
         self._past_length = 0
@@ -787,6 +786,7 @@ def _from_pretrained(
         quantization_config: Optional[Union[OVWeightQuantizationConfig, Dict]] = None,
         **kwargs,
     ):
+        generation_config = kwargs.pop("generation_config", None)
         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
@@ -827,20 +827,23 @@ def _from_pretrained(
 
         enable_compilation = kwargs.pop("compile", True) and not quantization_config
 
-        try:
-            generation_config = GenerationConfig.from_pretrained(
-                model_id,
-                token=token,
-                revision=revision,
-                cache_dir=cache_dir,
-                force_download=force_download,
-                local_files_only=local_files_only,
-            )
-            if getattr(generation_config, "cache_implementation", None) is not None:
-                generation_config.cache_implementation = None
-            kwargs["generation_config"] = generation_config
-        except Exception:
-            pass
+        if generation_config is None:
+            try:
+                generation_config = GenerationConfig.from_pretrained(
+                    model_id,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                )
+                if getattr(generation_config, "cache_implementation", None) is not None:
+                    generation_config.cache_implementation = None
+            except OSError:
+                logger.info(
+                    "Generation config file not found, using a generation config created from the model config."
+                )
 
         causal_model = init_cls(
             model=model,
@@ -849,6 +852,7 @@ def _from_pretrained(
             compile=enable_compilation,
             compile_only=compile_only,
             quantization_config=quantization_config,
+            generation_config=generation_config,
             **kwargs,
         )
 

setup.py

@@ -28,8 +28,10 @@
 
 INSTALL_REQUIRE = [
     "torch>=1.11",
-    "transformers>=4.36,<4.45",
-    "optimum~=1.22",
+    "transformers @ git+https://github.com/huggingface/transformers.git",
+    "optimum @ git+https://github.com/huggingface/optimum.git@trfs-4.45",
+    # "transformers>=4.36,<4.46",
+    # "optimum~=1.22",
     "datasets>=1.4.0",
     "sentencepiece",
     "setuptools",