Deal with serve separately

Author: Imss27

vllm/utils.py

@@ -1201,13 +1201,16 @@ def _pull_args_from_config(args: List[str]) -> List[str]:
         config_args = FlexibleArgumentParser._load_config_file(file_path)
 
         # 0th index is for {serve,chat,complete}
-        # followed by model_tag
+        # followed by model_tag (only for serve)
         # followed by config args
         # followed by rest of cli args.
         # maintaining this order will enforce the precedence
         # of cli > config > defaults
-        args = [args[0]] + [args[1]] + config_args + \
-            args[2:index] + args[index+2:]
+        if args[0] == "serve":
+            args = [args[0]] + [args[1]
+                                ] + config_args + args[2:index] + args[index + 2:]
+        else:
+            args = [args[0]] + config_args + args[1:index] + args[index + 2:]
 
         return args