fix peft is_multimodal (#2462)

Author: Jintao-Huang

swift/tuners/peft.py

@@ -77,40 +77,22 @@ def from_pretrained(cls, pretrained_model_name_or_path: str, subfolder: Optional
         return self
 
 
-def _get_target(*args, **kwargs):
-    target = None
-    if 'target' in kwargs:
-        target = kwargs['target']
-    else:
-        for arg in args:
-            if isinstance(arg, torch.nn.Module):
-                target = arg
-                break
-    return target
-
-
-def _create_and_replace_hook(self, *args, **kwargs):
-    target = _get_target(*args, **kwargs)
-    if target and target.__class__.__name__ == 'NonDynamicallyQuantizableLinear':
-        return
-
-    return self._create_and_replace_origin(*args, **kwargs)
-
-
-def _create_and_replace_hook2(self, *args, **kwargs):
-    target = _get_target(*args, **kwargs)
-
+def _create_and_replace_hook(self, peft_config, adapter_name, target, *args, **kwargs):
     all_supported_names = ('linear', )
     all_supported_types = (torch.nn.Embedding, torch.nn.Conv2d, transformers.pytorch_utils.Conv1D)
+    target_modules = getattr(peft_config, 'target_modules', None)
+    if target is None:
+        return
 
-    is_multimodal = getattr(self.model, 'is_multimodal', False)
-
-    if is_multimodal and target is not None and (not any(
+    if isinstance(target_modules, str) and not any(
         [name in target.__class__.__name__.lower()
-         for name in all_supported_names]) and not any([isinstance(target, type) for type in all_supported_types])):
+         for name in all_supported_names]) and not any([isinstance(target, type_) for type_ in all_supported_types]):
         return
 
-    return _create_and_replace_hook(self, *args, **kwargs)
+    if target.__class__.__name__ == 'NonDynamicallyQuantizableLinear':
+        return
+
+    return self._create_and_replace_origin(peft_config, adapter_name, target, *args, **kwargs)
 
 
 def _convert_dtype(target: torch.nn.Module, adapter_name: str, lora_dtype: str):
@@ -296,28 +278,24 @@ def keep_device_forward(self, *args, **kwargs):
 
 def hot_patch_peft_module():
     from peft.tuners.lora import LoraLayer
+    if hasattr('LoraModel', '_create_and_replace_origin'):
+        return
 
     # Fix Lora does not support NonDynamicallyQuantizableLinear
     LoraModel._create_and_replace_origin = LoraModel._create_and_replace
     LoraModel._create_and_replace = _create_and_replace_hook
     VeraModel._create_and_replace_origin = VeraModel._create_and_replace
-    VeraModel._create_and_replace = _create_and_replace_hook2
+    VeraModel._create_and_replace = _create_and_replace_hook
     BOFTModel._create_and_replace_origin = BOFTModel._create_and_replace
-    BOFTModel._create_and_replace = _create_and_replace_hook2
+    BOFTModel._create_and_replace = _create_and_replace_hook
     IA3Model._create_and_replace_origin = IA3Model._create_and_replace
-    IA3Model._create_and_replace = _create_and_replace_hook2
+    IA3Model._create_and_replace = _create_and_replace_hook
     if FourierFTModel is not None:
         FourierFTModel._create_and_replace_origin = FourierFTModel._create_and_replace
-        FourierFTModel._create_and_replace = _create_and_replace_hook2
+        FourierFTModel._create_and_replace = _create_and_replace_hook
 
     # Support type conversion
     def init(self, model: torch.nn.Module, config: Dict[str, LoraConfig], adapter_name):
-        if isinstance(config, dict):
-            for _config in config.values():  # There is a target_modules as a string.
-                if isinstance(getattr(_config, 'target_modules', None), str):
-                    # Make sure the regex can find all linear in the module.
-                    LoraModel._create_and_replace = _create_and_replace_hook2
-                    break
 
         self.__init_origin__(model, config, adapter_name)
         if isinstance(self.active_adapter, list):