diff --git a/python/paddle/distributed/auto_parallel/static/engine.py b/python/paddle/distributed/auto_parallel/static/engine.py
index 3d7096441dac5c7dd169e21ec3e8c19b33ef5467..0ef5102a5499099e8ed3014955a019d46b415f27 100644
--- a/python/paddle/distributed/auto_parallel/static/engine.py
+++ b/python/paddle/distributed/auto_parallel/static/engine.py
@@ -833,14 +833,6 @@ class Engine:
                 dist_main_program, self._place, dist_context
             )
 
-        # NOTE(zhaoyinglia): Skip startup program when use new ir temporarily.
-        use_new_ir = False
-        if auto_utils.use_new_ir():
-            use_new_ir = True
-            paddle.framework.set_flags(
-                {"FLAGS_enable_new_ir_in_executor": False}
-            )
-
         if self._executor is None:
             self._executor = paddle.static.Executor(self._place)
             uninitialized = []
@@ -868,11 +860,6 @@ class Engine:
             ]
             self._executor.run(dist_startup_prog)
 
-        if use_new_ir:
-            paddle.framework.set_flags(
-                {"FLAGS_enable_new_ir_in_executor": True}
-            )
-
     def fit(
         self,
         train_data,
diff --git a/python/paddle/distributed/auto_parallel/static/utils.py b/python/paddle/distributed/auto_parallel/static/utils.py
index 79597fd693218acaf201e9b2db0367d80d040799..fa12cfd68e3b2e0d80d8900a3ba72feb9fe88f99 100644
--- a/python/paddle/distributed/auto_parallel/static/utils.py
+++ b/python/paddle/distributed/auto_parallel/static/utils.py
@@ -2423,19 +2423,6 @@ def use_new_executor():
     ]
 
 
-def use_new_ir():
-    enable_new_ir_in_executor = os.environ.get(
-        'FLAGS_enable_new_ir_in_executor', None
-    )
-    return enable_new_ir_in_executor in [
-        1,
-        '1',
-        True,
-        'True',
-        'true',
-    ]
-
-
 def get_pp_stage(dist_context, rank):
     pp_idx = None
     for idx, process_mesh in enumerate(dist_context.process_meshes):