[cherry pick]disable lora

llm/merge_lora_params.py

@@ -127,8 +127,8 @@ def merge():
             config=config,
             low_cpu_mem_usage=True,
         )
-        lora_config.merge_weights = True
         model = LoRAModel.from_pretrained(model=model, lora_path=args.lora_path, lora_config=lora_config)
+        model.merge()
         model.eval()
         model_state_dict = model.model.state_dict()
         for key in list(model_state_dict):

llm/predictor.py

@@ -49,8 +49,8 @@
     AutoConfig,
     AutoModelForCausalLM,
     AutoTokenizer,
-    ChatGLMv2Tokenizer,
     ChatGLMTokenizer,
+    ChatGLMv2Tokenizer,
     LlamaTokenizer,
     PretrainedModel,
     PretrainedTokenizer,
@@ -242,7 +242,8 @@ def _preprocess(self, source):
             padding=True,
             # when use chat_template, it should not add special tokens
             # chatglm2 prefix-tokens can not be tokenized into ids
-            add_special_tokens=self.tokenizer.chat_template is None or isinstance(self.tokenizer, (ChatGLMv2Tokenizer, ChatGLMTokenizer)),
+            add_special_tokens=self.tokenizer.chat_template is None
+            or isinstance(self.tokenizer, (ChatGLMv2Tokenizer, ChatGLMTokenizer)),
         )
         return tokenized_source
 
@@ -272,7 +273,6 @@ def __init__(
         if config.lora_path is not None:
             lora_config = LoRAConfig.from_pretrained(config.lora_path)
             dtype = lora_config.dtype
-            lora_config.merge_weights = True
         elif config.prefix_path is not None:
             prefix_config = PrefixConfig.from_pretrained(config.prefix_path)
             dtype = prefix_config.dtype
@@ -289,6 +289,7 @@ def __init__(
                 tensor_parallel_degree=self.tensor_parallel_degree,
                 tensor_parallel_rank=self.tensor_parallel_rank,
             )
+            self.model.merge()
 
         if config.lora_path is not None:
             self.model = LoRAModel.from_pretrained(

paddlenlp/peft/lora/__init__.py

@@ -13,11 +13,5 @@
 # limitations under the License.
 
 from .lora_config import LoRAConfig
-from .lora_layers import (
-    ColumnParallelLoRALinear,
-    ColumnParallelLoRAMergedLinear,
-    LoRALinear,
-    LoRAMergedLinear,
-    RowParallelLoRALinear,
-)
+from .lora_layers import ColumnParallelLoRALinear, LoRALinear, RowParallelLoRALinear
 from .lora_model import LoRAModel

paddlenlp/peft/lora/lora_config.py

@@ -94,6 +94,11 @@ def __post_init__(self):
                 "We will automatically set `use_quick_lora` to `False` to avoid potential inconsistencies."
             )
             self.use_quick_lora = False
+        if self.merge_weights:
+            logger.error(
+                "'merge_weights' is deprecated and will be removed in a future version. "
+                "Please apply model.merge() or model.unmerge() to merge/unmerge LoRA weight to base model."
+            )
 
     @property
     def scaling(self):