diff --git a/vllm/entrypoints/cli/serve.py b/vllm/entrypoints/cli/serve.py
index a69363e3d98f..7dcba2cccdb5 100644
--- a/vllm/entrypoints/cli/serve.py
+++ b/vllm/entrypoints/cli/serve.py
@@ -140,11 +140,16 @@ def run_multi_api_server(args: argparse.Namespace):
     num_api_servers = args.api_server_count
     assert num_api_servers > 0
 
+    orig_disable_mm_preprocessor_cache = args.disable_mm_preprocessor_cache
+
     # set_process_title("ProcManager")
 
     if num_api_servers > 1:
         setup_multiprocess_prometheus()
 
+        # Not compatible with API server scale-out
+        args.disable_mm_preprocessor_cache = True
+
     listen_address, sock = setup_server(args)
 
     engine_args = vllm.AsyncEngineArgs.from_cli_args(args)
@@ -161,11 +166,9 @@ def run_multi_api_server(args: argparse.Namespace):
                              "with api_server_count > 1")
 
         if model_config.is_multimodal_model and not (
-                model_config.disable_mm_preprocessor_cache):
-            logger.warning(
-                "Multi-model preprocessor cache will be disabled for"
-                " api_server_count > 1")
-            model_config.disable_mm_preprocessor_cache = True
+                orig_disable_mm_preprocessor_cache):
+            logger.warning("Multi-model preprocessor cache will be disabled "
+                           "for api_server_count > 1")
 
     executor_class = Executor.get_class(vllm_config)
     log_stats = not engine_args.disable_log_stats