diff --git a/python/paddle/fluid/tests/unittests/seresnext_net.py b/python/paddle/fluid/tests/unittests/seresnext_net.py
index c31471efbde84da1ea1b3141f344bbefda8134c2..252ffe2c22ad2e9d0b48e89d9dc0a7d695e1c0a4 100644
--- a/python/paddle/fluid/tests/unittests/seresnext_net.py
+++ b/python/paddle/fluid/tests/unittests/seresnext_net.py
@@ -172,8 +172,11 @@ def optimizer(learning_rate=0.01):
 model = SE_ResNeXt50Small
 
 
-def batch_size():
-    return 8
+def batch_size(use_cuda):
+    if use_cuda:
+        # Paddle uses 8GB P4 GPU for unittest so we decreased the batch size.
+        return 8
+    return 12
 
 
 def iter(use_cuda):
@@ -183,9 +186,9 @@ def iter(use_cuda):
 
 
 gpu_img, gpu_label = init_data(
-    batch_size=batch_size(), img_shape=img_shape, label_range=999)
+    batch_size=batch_size(use_cuda=True), img_shape=img_shape, label_range=999)
 cpu_img, cpu_label = init_data(
-    batch_size=batch_size(), img_shape=img_shape, label_range=999)
+    batch_size=batch_size(use_cuda=False), img_shape=img_shape, label_range=999)
 feed_dict_gpu = {"image": gpu_img, "label": gpu_label}
 feed_dict_cpu = {"image": cpu_img, "label": cpu_label}
 
diff --git a/python/paddle/fluid/tests/unittests/seresnext_test_base.py b/python/paddle/fluid/tests/unittests/seresnext_test_base.py
index 65879d39d91145b2403ac1b0c29e51df1960c8d1..9f055191b11a5c3708d19f984996149db37e798a 100644
--- a/python/paddle/fluid/tests/unittests/seresnext_test_base.py
+++ b/python/paddle/fluid/tests/unittests/seresnext_test_base.py
@@ -32,7 +32,7 @@ class TestResnetBase(TestParallelExecutorBase):
             seresnext_net.model,
             feed_dict=seresnext_net.feed_dict(use_cuda),
             iter=seresnext_net.iter(use_cuda),
-            batch_size=seresnext_net.batch_size(),
+            batch_size=seresnext_net.batch_size(use_cuda),
             use_cuda=use_cuda,
             use_reduce=False,
             optimizer=seresnext_net.optimizer)
@@ -41,7 +41,7 @@ class TestResnetBase(TestParallelExecutorBase):
             seresnext_net.model,
             feed_dict=seresnext_net.feed_dict(use_cuda),
             iter=seresnext_net.iter(use_cuda),
-            batch_size=seresnext_net.batch_size(),
+            batch_size=seresnext_net.batch_size(use_cuda),
             use_cuda=use_cuda)
 
         if compare_seperately:
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext_with_reduce_cpu.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext_with_reduce_cpu.py
index 74c5999c4fd3e4be82e9a5b2484efe69a0271baf..62eb7e1155e6f2ffa6e4e7b32ff1cfcf8eeb6ea3 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext_with_reduce_cpu.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext_with_reduce_cpu.py
@@ -28,7 +28,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
             seresnext_net.model,
             feed_dict=seresnext_net.feed_dict(use_cuda),
             iter=seresnext_net.iter(use_cuda),
-            batch_size=seresnext_net.batch_size(),
+            batch_size=seresnext_net.batch_size(use_cuda),
             use_cuda=use_cuda,
             use_reduce=False,
             optimizer=seresnext_net.optimizer)
@@ -36,7 +36,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
             seresnext_net.model,
             feed_dict=seresnext_net.feed_dict(use_cuda),
             iter=seresnext_net.iter(use_cuda),
-            batch_size=seresnext_net.batch_size(),
+            batch_size=seresnext_net.batch_size(use_cuda),
             use_cuda=use_cuda,
             use_reduce=True,
             optimizer=seresnext_net.optimizer)
@@ -53,7 +53,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
             seresnext_net.model,
             feed_dict=seresnext_net.feed_dict(use_cuda),
             iter=seresnext_net.iter(use_cuda),
-            batch_size=seresnext_net.batch_size(),
+            batch_size=seresnext_net.batch_size(use_cuda),
             use_cuda=use_cuda,
             use_reduce=False,
             optimizer=seresnext_net.optimizer,
@@ -63,7 +63,7 @@ class TestResnetWithReduceBase(TestParallelExecutorBase):
             seresnext_net.model,
             feed_dict=seresnext_net.feed_dict(use_cuda),
             iter=seresnext_net.iter(use_cuda),
-            batch_size=seresnext_net.batch_size(),
+            batch_size=seresnext_net.batch_size(use_cuda),
             use_cuda=use_cuda,
             use_reduce=True,
             optimizer=seresnext_net.optimizer,