disable lora (#8674)

* disable lora

* lora disable & merge

* fix

* fix unitest

* fix the finetune and predict in llm

---------

Co-authored-by: wawltor <fangzeyang0904@hotmail.com>

Author: lugimzzz

docs/llm/docs/peft.md

@@ -42,7 +42,6 @@ Prefix-tuning[论文](https://arxiv.org/abs/2101.00190)
         target_modules=target_modules,
         r=lora_rank,
         lora_alpha=2 * lora_rank,
-        merge_weights=True
     )
     model = LoRAModel(model, lora_config)
     model.mark_only_lora_as_trainable()
@@ -92,7 +91,7 @@ Parameters:
                         默认为 0.0，dropout的比例设置，float 类型
 
     --merge_weights
-                        默认为 False，模型推理时，是否进行base model 权重和 LoRA 权重的合参操作，bool 类型
+                        默认为 False，接口将被废弃。请使用model.merge()或model.unmerge()替代。
 
     --trainable_bias
                         指定可训练的 bias, 可选项 ['lora', 'all']

llm/predict/predictor.py

@@ -270,7 +270,6 @@ 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
@@ -292,6 +291,7 @@ def __init__(
             self.model = LoRAModel.from_pretrained(
                 model=self.model, lora_path=config.lora_path, lora_config=lora_config
             )
+            self.model.merge()
         if config.prefix_path is not None:
             prefix_tuning_params = get_prefix_tuning_params(self.model)
             self.model = PrefixModelForCausalLM.from_pretrained(

llm/tools/merge_lora_params.py

@@ -113,13 +113,17 @@ def lora_process(name, lora_config, state_dict, device, lora_state_dict=None):
 
 
 def merge_old_lora(lora_config, args):
-    lora_config.merge_weight = True
+    lora_config.merge_weights = True
     model = AutoModelForCausalLM.from_pretrained(
         args.model_name_or_path,
         dtype=lora_config.dtype,
     )
     model = LoRAModel.from_pretrained(model, args.lora_path)
-    model.eval()
+    try:
+        model.merge()
+        model.eval()
+    except:
+        model.eval()
     model_state_dict = model.model.state_dict()
     for key in list(model_state_dict):
         if "lora" in key:

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):