diff --git a/paddle/fluid/operators/arg_max_op_mlu.cc b/paddle/fluid/operators/arg_max_op_mlu.cc
index ca8966a795fbff41eb48b2b4c64e17994370532a..44f74f016c065155affeb2d79aef4fc35de57363 100644
--- a/paddle/fluid/operators/arg_max_op_mlu.cc
+++ b/paddle/fluid/operators/arg_max_op_mlu.cc
@@ -72,7 +72,7 @@ class ArgMaxMLUKernel : public framework::OpKernel<T> {
     MLUCnnlTensorDesc input_desc(
         flatten_x, CNNL_LAYOUT_ARRAY, ToCnnlDataType(flatten_x.dtype()));
     MLUCnnlReduceDesc reduction_desc(reduce_dims,
-                                     CNNL_REDUCE_MAX_LAST_INDEX,
+                                     CNNL_REDUCE_MAX,
                                      ToCnnlDataType<T>(),
                                      CNNL_NOT_PROPAGATE_NAN,
                                      CNNL_REDUCE_ONLY_INDICES,
diff --git a/python/paddle/fluid/tests/unittests/mlu/test_arg_max_op_mlu.py b/python/paddle/fluid/tests/unittests/mlu/test_arg_max_op_mlu.py
index bd943e05b2d33efd6137608f130ac02b39d51588..45a79b5ece51bac75b8beddfdd665bd4f28e43d0 100644
--- a/python/paddle/fluid/tests/unittests/mlu/test_arg_max_op_mlu.py
+++ b/python/paddle/fluid/tests/unittests/mlu/test_arg_max_op_mlu.py
@@ -52,6 +52,38 @@ class BaseTestCase(OpTest):
         self.check_output_with_place(self.place)
 
 
+class TestArgMaxSameValue1(BaseTestCase):
+
+    def initTestCase(self):
+        self.op_type = 'arg_max'
+        self.dtype = 'float32'
+        self.axis = 0
+
+    def setUp(self):
+        self.set_mlu()
+        self.initTestCase()
+        self.x = np.array([1, 2, 3, 5, 4, 5]).astype(self.dtype)
+        self.inputs = {'X': self.x}
+        self.attrs = {'axis': self.axis}
+        self.outputs = {'Out': np.argmax(self.x, axis=self.axis)}
+
+
+class TestArgMaxSameValue2(BaseTestCase):
+
+    def initTestCase(self):
+        self.op_type = 'arg_max'
+        self.dtype = 'float16'
+        self.axis = 0
+
+    def setUp(self):
+        self.set_mlu()
+        self.initTestCase()
+        self.x = np.array([[2, 3, 5, 5], [3, 2, 5, 5]]).astype(self.dtype)
+        self.inputs = {'X': self.x}
+        self.attrs = {'axis': self.axis}
+        self.outputs = {'Out': np.argmax(self.x, axis=self.axis)}
+
+
 # test argmax, dtype: float16
 class TestArgMaxFloat16Case1(BaseTestCase):