fix

src/pet/pet_model.py

@@ -46,10 +46,10 @@ class PETModel(torch.nn.Module):
         num_transformer_submodules = 0
         transformer_backbone = None
         for name, module in self.base_model.named_children():
+            for param in module.parameters():
+                param.requires_grad = False
             if isinstance(module, PreTrainedModel):
                 # Make sure to freeze Tranformers model
-                for param in module.parameters():
-                    param.requires_grad = False
                 if transformer_backbone is None:
                     transformer_backbone = module
                     self.transformer_backbone_name = name

src/pet/utils/other.py

@@ -44,3 +44,15 @@ def _set_trainable(model):
                 param.requires_grad = True
             else:
                 param.requires_grad = False
+
+
+# def fsdp_auto_wrap_policy():
+#     from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy, lambda_auto_wrap_policy, _or_policy
+
+#     def lambda_policy(module):
+#         if len(module.named_children()) != 0 and
+#             return True
+#         return False
+
+
+#     pass