fix quantlora bug (#10562)

Author: lugimzzz

paddlenlp/peft/lora/lora_model.py

@@ -607,8 +607,12 @@ def _find_and_replace_module(self, model, module_name, lora_config):
             lora_module = QuantizationLoRALinear(module, lora_config)
         elif isinstance(module, ColumnParallelQuantizationLinear):
             lora_module = ColumnParallelQuantizationLoRALinear(module, lora_config)
+            # Lora column parallel will spilt lora B matrix
+            self.add_lora_split_mapping(module_name + ".lora_B", is_column=True)
         elif isinstance(module, RowParallelQuantizationLinear):
             lora_module = RowParallelQuantizationLoRALinear(module, lora_config)
+            # Lora row parallel will spilt lora A matrix
+            self.add_lora_split_mapping(module_name + ".lora_A", is_column=False)
         if lora_module is None:
             raise ValueError(
                 f"LoRA strategy only supports paddle.nn.Linear or paddle.distributed.fleet.meta_parallel.ColumnParallelLinear or paddlenlp.transformers.sequence_utils. {module}({module_name} {type(module).__name__}) is not supported。"