Avoid adding unneccessary changes

Author: nikita-savelyevv

optimum/intel/openvino/modeling_base.py

@@ -791,11 +791,11 @@ def _export(
             )
             compile_only = False
 
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
+        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
             ov_config = None
+        else:
+            ov_config = OVConfig(dtype="fp32")
 
         variant = kwargs.pop("variant", None)
 

optimum/intel/openvino/modeling_decoder.py

@@ -321,11 +321,11 @@ def _export(
             if use_cache:
                 task = task + "-with-past"
 
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
-            ov_config = None
+        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
+            ov_export_config = None
+        else:
+            ov_export_config = OVConfig(dtype="auto")
 
         stateful = kwargs.pop("stateful", ensure_stateful_is_available(warn=False) and use_cache)
 
@@ -349,7 +349,7 @@ def _export(
             local_files_only=local_files_only,
             force_download=force_download,
             trust_remote_code=trust_remote_code,
-            ov_config=ov_config,
+            ov_config=ov_export_config,
             stateful=stateful,
             model_loading_kwargs=model_loading_kwargs,
             library_name=cls._library_name,

optimum/intel/openvino/modeling_diffusion.py

@@ -608,11 +608,12 @@ def _export(
             )
             compile_only = False
 
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
+        # If load_in_8bit and quantization_config not specified then ov_config is set
+        # to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
             ov_config = None
+        else:
+            ov_config = OVConfig(dtype="auto")
 
         torch_dtype = kwargs.pop("torch_dtype", None)
 

optimum/intel/openvino/modeling_open_clip.py

@@ -245,11 +245,11 @@ def _export(
         # would end-up removing the directory containing the underlying OpenVINO model
         cls._model_save_dir_tempdirectory_instance = save_dir
 
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
+        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
             ov_config = None
+        else:
+            ov_config = OVConfig(dtype="fp32")
 
         def fn_get_submodels(model):
             return {"model_text": model.text}
@@ -370,11 +370,11 @@ def _export(
         # would end-up removing the directory containing the underlying OpenVINO model
         cls._model_save_dir_tempdirectory_instance = save_dir
 
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
+        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
             ov_config = None
+        else:
+            ov_config = OVConfig(dtype="fp32")
 
         def fn_get_submodels(model):
             return {"model_vision": model.visual}

optimum/intel/openvino/modeling_seq2seq.py

@@ -593,11 +593,11 @@ def _export(
                 "Please provide openvino model obtained using optimum-cli or saved on disk using `save_pretrained`"
             )
             compile_only = False
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
+        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
             ov_config = None
+        else:
+            ov_config = OVConfig(dtype="fp32")
         stateful = kwargs.get("stateful", True)
         variant = kwargs.pop("variant", None)
 

optimum/intel/openvino/modeling_visual_language.py

@@ -653,11 +653,12 @@ def _export(
         if task is None:
             task = cls.export_feature
 
-        ov_config = OVConfig(dtype="auto")
-        if load_in_8bit is None and quantization_config is None:
-            # If load_in_8bit and quantization_config are not specified then ov_config is set to None, and
-            # models larger than 1B parameters will be quantized to int8
+        # If load_in_8bit and quantization_config not specified then ov_config is set to None and will be set by default in convert depending on the model size
+        if load_in_8bit is None and not quantization_config:
             ov_config = None
+        else:
+            # Export in fp32 if compression won't be applied later
+            ov_config = OVConfig(dtype="fp32" if load_in_8bit is False else "auto")
 
         stateful = kwargs.pop("stateful", ensure_stateful_is_available(warn=False) and use_cache)
         variant = kwargs.pop("variant", None)