fix pretrained_config save dtype (#2587)

Author: cheng221

paddleformers/transformers/configuration_utils.py

@@ -667,6 +667,8 @@ def __init__(self, **kwargs):
                 "Transformers. Using `model.gradient_checkpointing_enable()` instead, or if you are using the "
                 "`Trainer` API, pass `gradient_checkpointing=True` in your `TrainingArguments`."
             )
+        self._save_to_hf = kwargs.pop("save_to_hf", False)
+        self._unsavable_keys.add("_save_to_hf")
 
         # Additional attributes without default values
         for key, value in kwargs.items():
@@ -759,6 +761,8 @@ def save_pretrained(self, save_directory: Union[str, os.PathLike], **kwargs):
 
         os.makedirs(save_directory, exist_ok=True)
 
+        self._save_to_hf = kwargs.pop("save_to_hf", False)
+
         # If we have a custom config, we copy the file defining it in the folder and set the attributes so it can be
         # loaded from the Hub.
         if self._auto_class is not None:
@@ -1068,7 +1072,7 @@ def to_dict(self, saving_file=False) -> Dict[str, Any]:
             del output["_auto_class"]
         if "moe_group" in output:
             del output["moe_group"]
-        if "dtype" in output:
+        if self._save_to_hf and "dtype" in output:
             output["torch_dtype"] = str(output["dtype"])
             del output["dtype"]