diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc
index fe5555d5cb100ad791774e504f1f95364007463c..46eca6df552c6fd7705a2c1e8a70d75a28c6d8e7 100644
--- a/paddle/fluid/inference/api/paddle_pass_builder.cc
+++ b/paddle/fluid/inference/api/paddle_pass_builder.cc
@@ -162,6 +162,7 @@ const std::vector<std::string> kLiteSubgraphPasses({
 // support fp16/bf16 precision, temporarily use low precision pass to prevent
 // running errors. After fusion operator supports low precision, delete this.
 const std::vector<std::string> kGpuLowerPrecisionPasses{
+    "identity_scale_op_clean_pass",
     "simplify_with_basic_ops_pass",
     "delete_quant_dequant_linear_op_pass",
     "delete_weight_dequant_linear_op_pass",
@@ -208,8 +209,8 @@ const std::vector<std::string> kCINNCompilerPasses{
 
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
-    //   "identity_scale_op_clean_pass",             //
-    "is_test_pass",                                                     //
+    "identity_scale_op_clean_pass",                                     //
+        "is_test_pass",                                                 //
         "simplify_with_basic_ops_pass",                                 //
         "delete_quant_dequant_linear_op_pass",                          //
         "delete_weight_dequant_linear_op_pass",                         //