add default mapping to peft integration

src/transformers/integrations/peft.py

@@ -28,6 +28,7 @@ from ..utils import (
     is_torch_available,
     logging,
 )
+from ..modeling_utils import VLMS
 
 
 if is_torch_available():
@@ -151,6 +152,8 @@ class PeftAdapterMixin:
         # peft only supports low_cpu_mem_usage starting from v0.13.0
         peft_load_kwargs = {}
         key_mapping = adapter_kwargs.pop("key_mapping", None) if adapter_kwargs is not None else None
+        if key_mapping is None and any(allowed_name in self.__class__.__name__.lower() for allowed_name in VLMS):
+            key_mapping = self._checkpoint_conversion_mapping
         if low_cpu_mem_usage:
             min_version_lcmu = "0.13.0"
             if version.parse(importlib.metadata.version("peft")) >= version.parse(min_version_lcmu):

src/transformers/modeling_utils.py

@@ -4251,11 +4251,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, PushToHubMixin, PeftAdapterMi
         device_mesh = kwargs.pop("device_mesh", None)
         trust_remote_code = kwargs.pop("trust_remote_code", None)
 
-        # Load models with hardcoded key mapping on class for VLMs only,  to keep BC and standardize model
-        if any(allowed_name in cls.__name__.lower() for allowed_name in VLMS):
-            key_mapping = kwargs.pop("key_mapping", cls._checkpoint_conversion_mapping)
-        else:
-            key_mapping = kwargs.pop("key_mapping", None)
+        key_mapping = kwargs.pop("key_mapping", None)
+        # Load models with hardcoded key mapping on class for VLMs only, to keep BC and standardize model
+        if key_mapping is None and any(allowed_name in cls.__name__.lower() for allowed_name in VLMS):
+            key_mapping = cls._checkpoint_conversion_mapping
 
         # Not used anymore -- remove them from the kwargs
         _ = kwargs.pop("resume_download", None)