Fix embedding name matching for nested paths and ensure weight tying …

…in inject_adapter

src/peft/tuners/trainable_tokens/model.py

@@ -71,17 +71,29 @@ def inject_adapter(
                 if matched_keys:
                     parent, target, target_name = _get_submodules(model, name)
 
-                    peft_config = self.peft_config[adapter_name].to_dict()
-                    peft_config["tied_adapter"] = self.model.get_input_embeddings()
-
-                    self._create_and_replace_dict(
-                        peft_config,
-                        adapter_name,
-                        target,
-                        target_name,
-                        parent,
-                        matched_keys[0],
-                    )
+                    # If the module is already a TrainableTokensLayer, we need to replace it with a tied version
+                    # instead of just updating it. This handles the case where the user explicitly targeted
+                    # both the embedding and tied layers in target_modules.
+                    if isinstance(target, TrainableTokensLayer):
+                        # Replace the existing layer with a new one that's tied to the embedding
+                        peft_config = self.peft_config[adapter_name].to_dict()
+                        peft_config["tied_adapter"] = self.model.get_input_embeddings()
+
+                        new_module = self._create_new_module(peft_config, adapter_name, target.base_layer, **peft_config)
+                        self._replace_module(parent, target_name, new_module, target.base_layer)
+                    else:
+                        # Module hasn't been wrapped yet, create and replace normally
+                        peft_config = self.peft_config[adapter_name].to_dict()
+                        peft_config["tied_adapter"] = self.model.get_input_embeddings()
+
+                        self._create_and_replace_dict(
+                            peft_config,
+                            adapter_name,
+                            target,
+                            target_name,
+                            parent,
+                            matched_keys[0],
+                        )
 
     def _get_tied_target_modules(self, *args, **kwargs):
         # Normally this method would return the layers that target tied layers.

src/peft/utils/other.py

@@ -1488,8 +1488,8 @@ def set_additional_trainable_modules(model, peft_config, model_config, adapter_n
 
             # Find which target layers are in the tied weights (including the embedding source)
             for target_layer_name in target_layers:
-                # Check if this is the embedding layer
-                if target_layer_name == embedding_name:
+                # Check if this is the embedding layer (check both exact match and endswith for nested structures)
+                if target_layer_name == embedding_name or target_layer_name.endswith(embedding_name):
                     tied_layer_keys.append(target_layer_name)
                     continue
                 # Check if this target layer matches any tied module (considering nested structures)