fix the fsdp peft autowrap policy (#1694)

* fix the fsdp peft autowrap policy

* address comment wrt backwards compatibility

src/peft/utils/other.py

@@ -393,6 +393,11 @@ def fsdp_auto_wrap_policy(model):
     import os
 
     from accelerate import FullyShardedDataParallelPlugin
+
+    if hasattr(FullyShardedDataParallelPlugin, "get_module_class_from_name"):
+        get_module_class_from_name = FullyShardedDataParallelPlugin.get_module_class_from_name
+    else:
+        from accelerate.utils.dataclasses import get_module_class_from_name
     from torch.distributed.fsdp.wrap import _or_policy, lambda_auto_wrap_policy, transformer_auto_wrap_policy
 
     from ..tuners import PrefixEncoder, PromptEmbedding, PromptEncoder
@@ -405,7 +410,7 @@ def fsdp_auto_wrap_policy(model):
     ).split(",")
     transformer_cls_to_wrap = {PrefixEncoder, PromptEncoder, PromptEmbedding}
     for layer_class in transformer_cls_names_to_wrap:
-        transformer_cls = FullyShardedDataParallelPlugin.get_module_class_from_name(model, layer_class)
+        transformer_cls = get_module_class_from_name(model, layer_class)
         if transformer_cls is None:
             raise Exception("Could not find the transformer layer class to wrap in the model.")
         else: