diff --git a/python/paddle/fluid/tests/unittests/ir/inference/auto_scan_test.py b/python/paddle/fluid/tests/unittests/ir/inference/auto_scan_test.py
index fa09ef19977b077c336b5ab4074729aef915667a..08c634d58cafb5f609531ef095bccefad8fc7cfb 100755
--- a/python/paddle/fluid/tests/unittests/ir/inference/auto_scan_test.py
+++ b/python/paddle/fluid/tests/unittests/ir/inference/auto_scan_test.py
@@ -330,8 +330,7 @@ class PassAutoScanTest(AutoScanTest):
                        reproduce=None,
                        min_success_num=25,
                        max_duration=180,
-                       passes=None,
-                       use_gpu_run_baseline=False):
+                       passes=None):
         if os.getenv('HYPOTHESIS_TEST_PROFILE', 'ci') == "dev":
             max_examples *= 10
             min_success_num *= 10
@@ -356,10 +355,7 @@ class PassAutoScanTest(AutoScanTest):
             return self.sample_program_config(draw)
 
         def run_test(prog_config):
-            return self.run_test(
-                quant=quant,
-                prog_configs=[prog_config],
-                use_gpu_run_baseline=use_gpu_run_baseline)
+            return self.run_test(quant=quant, prog_configs=[prog_config])
 
         generator = st.composite(program_generator)
         loop_func = given(generator())(run_test)
@@ -396,10 +392,7 @@ class PassAutoScanTest(AutoScanTest):
                 format(max_duration))
             assert False
 
-    def run_test(self,
-                 quant=False,
-                 prog_configs=None,
-                 use_gpu_run_baseline=False):
+    def run_test(self, quant=False, prog_configs=None):
         status = True
 
         for prog_config in prog_configs:
@@ -418,22 +411,13 @@ class PassAutoScanTest(AutoScanTest):
                     'data': tensor_config.data,
                     'lod': tensor_config.lod
                 }
-            results: List[Dict[str, np.ndarray]] = []
-
-            # baseline: cpu no ir_optim run
 
-            base_config = self.create_inference_config(
-                ir_optim=False, use_gpu=use_gpu_run_baseline)
             logging.info('RUN program_config: ' + str(prog_config))
-            results.append(
-                self.run_test_config(model, params, prog_config, base_config,
-                                     feed_data))
-            self.success_log('RUN_CPU_BASELINE done')
-
             self.num_predictor_kinds = 0
             for pred_config, op_list, (
                     atol, rtol) in self.sample_predictor_configs(prog_config):
                 self.num_predictor_kinds += 1
+
                 # skip info
                 ignore_flag = False
                 for ignore_info in self.ignore_cases:
@@ -454,12 +438,26 @@ class PassAutoScanTest(AutoScanTest):
                 if not os.path.exists(self.cache_dir):
                     os.mkdir(self.cache_dir)
 
+                # baseline: no ir_optim run
+                base_config = self.create_inference_config(
+                    ir_optim=False,
+                    use_gpu=pred_config.use_gpu(),
+                    use_mkldnn=pred_config.mkldnn_enabled(), )
                 try:
-                    results.append(
-                        self.run_test_config(model, params, prog_config,
-                                             pred_config, feed_data))
-                    self.assert_tensors_near(atol, rtol, results[-1],
-                                             results[0])
+                    # baseline
+                    base_result = self.run_test_config(
+                        model, params, prog_config, base_config, feed_data)
+                    self.success_log('RUN_BASELINE ' +
+                                     self.inference_config_str(
+                                         base_config) + ' done')
+
+                    if os.path.exists(self.cache_dir):
+                        shutil.rmtree(self.cache_dir)
+
+                    pred_result = self.run_test_config(
+                        model, params, prog_config, pred_config, feed_data)
+                    self.assert_tensors_near(atol, rtol, pred_result,
+                                             base_result)
                     if not ignore_flag:
                         self.assert_op_list(op_list)
 
diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_fc_elementwise_layernorm_fuse_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_fc_elementwise_layernorm_fuse_pass.py
index 2ccb9de5d5470f977956142bbbc46555544bd7bb..26f91092d2af826edcec6dfa4e8de2ade24798f8 100644
--- a/python/paddle/fluid/tests/unittests/ir/inference/test_fc_elementwise_layernorm_fuse_pass.py
+++ b/python/paddle/fluid/tests/unittests/ir/inference/test_fc_elementwise_layernorm_fuse_pass.py
@@ -127,8 +127,7 @@ class TestFCElementwiseLayerNormFusePass(PassAutoScanTest):
         self.run_and_statis(
             quant=False,
             max_examples=300,
-            passes=["fc_elementwise_layernorm_fuse_pass"],
-            use_gpu_run_baseline=True)
+            passes=["fc_elementwise_layernorm_fuse_pass"])
 
 
 if __name__ == "__main__":
diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_simplify_with_basic_ops_pass_autoscan.py b/python/paddle/fluid/tests/unittests/ir/inference/test_simplify_with_basic_ops_pass_autoscan.py
index 03e9feb418a82b890f2d986c1d1b95caa6807ff6..cb55dc64445ec324f912959400baf1c1b19ef3c3 100644
--- a/python/paddle/fluid/tests/unittests/ir/inference/test_simplify_with_basic_ops_pass_autoscan.py
+++ b/python/paddle/fluid/tests/unittests/ir/inference/test_simplify_with_basic_ops_pass_autoscan.py
@@ -46,7 +46,8 @@ class TestSimplifyWithBasicOpsPassUpscale(PassAutoScanTest):
         dropout_op = OpConfig(
             "dropout",
             inputs={"X": ["input_data"]},
-            outputs={"Out": ["dropout_output"]},
+            outputs={"Out": ["dropout_output"],
+                     "Mask": ["mask"]},
             fix_seed=fix_seed,
             dropout_implementation=dropout_implementation,
             dropout_prob=dropout_prob,
@@ -107,7 +108,8 @@ class TestSimplifyWithBasicOpsPassDowngrade(PassAutoScanTest):
         dropout_op = OpConfig(
             "dropout",
             inputs={"X": ["input_data"]},
-            outputs={"Out": ["dropout_output"]},
+            outputs={"Out": ["dropout_output"],
+                     "Mask": ["mask"]},
             fix_seed=fix_seed,
             dropout_implementation=dropout_implementation,
             dropout_prob=dropout_prob,