fix loraga merge (#9765)

* fix loraga merge

* change sign

Author: greycooker

paddlenlp/peft/lora/lora_model.py

@@ -327,12 +327,18 @@ def set_state_dict(self, state_dict):
         model_state_dict = self.model.state_dict()
         if self.lora_config.loraga:
 
-            def process_split_and_assign(name, concat_tensor, axis, init_dict, state_dict):
+            def process_split_and_assign(name, concat_tensor, init_dict, state_dict):
+                if "lora_A" in name:
+                    axis = 1
+                else:
+                    axis = 0
                 if isinstance(concat_tensor, np.ndarray):
                     final_lora, init_lora = np.split(concat_tensor, 2, axis=axis)
                     init_lora = paddle.to_tensor(init_lora)
                 else:
                     final_lora, init_lora = paddle.split(concat_tensor, 2, axis=axis)
+                if "lora_B" in name:
+                    init_lora *= -1
                 init_dict[name] = init_lora
                 state_dict[name] = final_lora
                 return init_lora
@@ -341,13 +347,13 @@ def process_split_and_assign(name, concat_tensor, axis, init_dict, state_dict):
                 if "lora_A" in name:
                     concat_lora_A = state_dict[name]
                     init_loraA = process_split_and_assign(
-                        name, concat_lora_A, axis=1, init_dict=self.loraga_init_dict, state_dict=state_dict
+                        name, concat_lora_A, init_dict=self.loraga_init_dict, state_dict=state_dict
                     )
 
                     loraB_name = name.replace("lora_A", "lora_B")
                     concat_lora_B = state_dict[loraB_name]
                     init_loraB = process_split_and_assign(
-                        loraB_name, concat_lora_B, axis=0, init_dict=self.loraga_init_dict, state_dict=state_dict
+                        loraB_name, concat_lora_B, init_dict=self.loraga_init_dict, state_dict=state_dict
                     )
 
                     base_name = name.replace("lora_A", "weight")
@@ -690,7 +696,7 @@ def get_trainable_state_dict(self, concat_init_lora=False):
                     if "lora_A" in name:
                         trainable_state_dict[name] = paddle.concat([weight, self.loraga_init_dict[name]], axis=1)
                     else:
-                        trainable_state_dict[name] = paddle.concat([weight, self.loraga_init_dict[name]], axis=0)
+                        trainable_state_dict[name] = paddle.concat([weight, -self.loraga_init_dict[name]], axis=0)
                 else:
                     trainable_state_dict[name] = weight