fix docstring

vllm/config.py

@@ -2635,6 +2635,9 @@ class LoRAConfig:
     max_cpu_loras: Optional[int] = None
     """Maximum number of LoRAs to store in CPU memory. Must be >= than
     `max_loras`."""
+    merge_lora: bool = False
+    """Enable merge lora
+    """
     lora_target_modules: Optional[List[str]] = None
     """List of lora module name, If not specified, 
     modules will be chosen according to the model architecture.
@@ -2653,8 +2656,6 @@ class LoRAConfig:
     specified, only adapters trained with the base model scaling factor are
     allowed."""
     bias_enabled: bool = False
-    merge_lora: bool = False
-    lora_target_modules: Optional[List[str]] = None
     """Enable bias for LoRA adapters."""
 
     def compute_hash(self) -> str:
@@ -2677,6 +2678,7 @@ class LoRAConfig:
         factors.append(self.lora_extra_vocab_size)
         factors.append(self.long_lora_scaling_factors)
         factors.append(self.bias_enabled)
+        factors.append(self.merge_lora)
         hash_str = hashlib.md5(str(factors).encode(),
                                usedforsecurity=False).hexdigest()
         return hash_str

vllm/engine/arg_utils.py

@@ -280,7 +280,7 @@ class EngineArgs:
     max_lora_rank: int = LoRAConfig.max_lora_rank
     fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras
     max_cpu_loras: Optional[int] = LoRAConfig.max_cpu_loras
-    merge_lora: bool = False
+    merge_lora: bool = LoRAConfig.merge_lora
     lora_target_modules: Optional[List[str]] = LoRAConfig.lora_target_modules
     lora_dtype: Optional[Union[str, torch.dtype]] = LoRAConfig.lora_dtype
     lora_extra_vocab_size: int = LoRAConfig.lora_extra_vocab_size
@@ -706,8 +706,9 @@ class EngineArgs:
         lora_group.add_argument('--max-lora-rank',
                                 **lora_kwargs["max_lora_rank"])
         lora_group.add_argument('--merge-lora',
-                            action=argparse.BooleanOptionalAction,
-                            help='If set to True, the weights of the base layer will be merged with the weights of Lora.')
+                                **lora_kwargs["merge-lora"])
+                            # action=argparse.BooleanOptionalAction,
+                            # help='If set to True, the weights of the base layer will be merged with the weights of Lora.')
         lora_group.add_argument('--lora-target-modules',
                             **lora_kwargs["lora_target_modules"])
         lora_group.add_argument('--lora-extra-vocab-size',