diff --git a/python/paddle/fluid/dygraph/math_op_patch.py b/python/paddle/fluid/dygraph/math_op_patch.py
index 8ce56d5a92686da92d44798e2a7db3404950cc3f..8a19be640a7ff591158dffd34eaa0f594ec63a01 100644
--- a/python/paddle/fluid/dygraph/math_op_patch.py
+++ b/python/paddle/fluid/dygraph/math_op_patch.py
@@ -222,7 +222,9 @@ def monkey_patch_math_varbase():
                 # so the calculation result here and the calculation result of numpy are 
                 # different after 6 decimal point. If necessary, we can also use float64 here.
                 # torch's behavior here is consistent with ours
-                if op_type == 'elementwise_div' and self.dtype in _supported_int_dtype_:
+                if (op_type == "final_state_divide" or
+                        op_type == "elementwise_div"
+                    ) and self.dtype in _supported_int_dtype_:
                     self = astype(self, 'float32')
                 # here use `scale` replace `elementwise` to get better performance
                 # but only +, -, *, / can use this method
@@ -277,7 +279,8 @@ def monkey_patch_math_varbase():
                 self = other_var
                 other_var = tmp
 
-            if op_type == 'elementwise_div' and self.dtype in _supported_int_dtype_:
+            if (op_type == "final_state_divide" or op_type == "elementwise_div"
+                ) and self.dtype in _supported_int_dtype_:
                 self = astype(self, 'float32')
                 other_var = astype(other_var, 'float32')
 
diff --git a/python/paddle/fluid/tests/unittests/test_tensor_scalar_type_promotion_dynamic.py b/python/paddle/fluid/tests/unittests/test_tensor_scalar_type_promotion_dynamic.py
index c5e3cb29e0c208bd1bb89732e1371fcf55c34fb4..774d40a17c66d3a23b2b6eb59ccb7e94e0cdba15 100644
--- a/python/paddle/fluid/tests/unittests/test_tensor_scalar_type_promotion_dynamic.py
+++ b/python/paddle/fluid/tests/unittests/test_tensor_scalar_type_promotion_dynamic.py
@@ -18,8 +18,7 @@ import unittest
 import numpy as np
 
 import paddle
-from paddle.fluid.framework import _enable_legacy_dygraph
-_enable_legacy_dygraph()
+from paddle.fluid.framework import _test_eager_guard
 
 # Support types are ref from `paddle.tensor.math`
 # - Related paddle dtypes:
@@ -52,7 +51,7 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         self.assertEqual(c_rlt.dtype, c.dtype)
         self.assertTrue(np.array_equal(c_rlt.numpy(), c.numpy()))
 
-    def test_tensor_add_scalar(self):
+    def func_tensor_add_scalar(self):
         # tensor(int64) + scalar(int)
         a = paddle.ones([2, 2, 2], dtype='int64')
         b = 1
@@ -83,7 +82,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 2.5, dtype="float32")
         self.check_operation(a, b, c, '+')
 
-    def test_tensor_sub_scalar(self):
+    def test_tensor_add_scalar(self):
+        with _test_eager_guard():
+            self.func_tensor_add_scalar()
+        self.func_tensor_add_scalar()
+
+    def func_tensor_sub_scalar(self):
         # tensor(int64) - scalar(int)
         a = paddle.ones([2, 2, 2], dtype='int64')
         b = 1
@@ -114,7 +118,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 0.5, dtype="float32")
         self.check_operation(a, b, c, '-')
 
-    def test_scalar_sub_tensor(self):
+    def test_tensor_sub_scalar(self):
+        with _test_eager_guard():
+            self.func_tensor_sub_scalar()
+        self.func_tensor_sub_scalar()
+
+    def func_scalar_sub_tensor(self):
         # scalar(int) - tensor(int64)
         a = 1
         b = paddle.ones([2, 2, 2], dtype='int64')
@@ -145,7 +154,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], -0.5, dtype="float32")
         self.check_operation(a, b, c, '-')
 
-    def test_tensor_mul_tensor(self):
+    def test_scalar_sub_tensor(self):
+        with _test_eager_guard():
+            self.func_scalar_sub_tensor()
+        self.func_scalar_sub_tensor()
+
+    def func_tensor_mul_tensor(self):
         # tensor(int64) * scalar(int)
         a = paddle.ones([2, 2, 2], dtype='int64')
         b = 1
@@ -176,7 +190,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 1.5, dtype="float32")
         self.check_operation(a, b, c, '*')
 
-    def test_tensor_div_scalar(self):
+    def test_tensor_mul_tensor(self):
+        with _test_eager_guard():
+            self.func_tensor_mul_tensor()
+        self.func_tensor_mul_tensor()
+
+    def func_tensor_div_scalar(self):
         # tensor(int64) / scalar(int)
         a = paddle.ones([2, 2, 2], dtype='int64')
         b = 2
@@ -207,7 +226,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 2, dtype="float32")
         self.check_operation(a, b, c, '/')
 
-    def test_scalar_div_tensor(self):
+    def test_tensor_div_scalar(self):
+        with _test_eager_guard():
+            self.func_tensor_div_scalar()
+        self.func_tensor_div_scalar()
+
+    def func_scalar_div_tensor(self):
         # scalar(int) / tensor(int64)
         a = 1
         b = paddle.full([2, 2, 2], 2, dtype='int64')
@@ -232,7 +256,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 2, dtype="float32")
         self.check_operation(a, b, c, '/')
 
-    def test_tensor_pow_scalar(self):
+    def test_scalar_div_tensor(self):
+        with _test_eager_guard():
+            self.func_scalar_div_tensor()
+        self.func_scalar_div_tensor()
+
+    def func_tensor_pow_scalar(self):
         # tensor(int64) ** scalar(int)
         a = paddle.full([2, 2, 2], 2, dtype='int64')
         b = 3
@@ -257,7 +286,12 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 8, dtype="float32")
         self.check_operation(a, b, c, '**')
 
-    def test_scalar_pow_tensor(self):
+    def test_tensor_pow_scalar(self):
+        with _test_eager_guard():
+            self.func_tensor_pow_scalar()
+        self.func_tensor_pow_scalar()
+
+    def func_scalar_pow_tensor(self):
         # scalar(int) ** tensor(int64)
         a = 3
         b = paddle.full([2, 2, 2], 2, dtype='int64')
@@ -282,15 +316,25 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 9, dtype="float32")
         self.check_operation(a, b, c, '**')
 
+    def test_scalar_pow_tensor(self):
+        with _test_eager_guard():
+            self.func_scalar_pow_tensor()
+        self.func_scalar_pow_tensor()
+
     ## TODO: floordiv op kernel doesn't support float
-    def test_tensor_floordiv_scalar(self):
+    def func_tensor_floordiv_scalar(self):
         # tensor(int64) // scalar(int)
         a = paddle.full([2, 2, 2], 3, dtype='int64')
         b = 2
         c = paddle.full([2, 2, 2], 1, dtype="int64")
         self.check_operation(a, b, c, '//')
 
-    def test_tensor_mod_scalar(self):
+    def test_tensor_floordiv_scalar(self):
+        with _test_eager_guard():
+            self.func_tensor_floordiv_scalar()
+        self.func_tensor_floordiv_scalar()
+
+    def func_tensor_mod_scalar(self):
         # tensor(int64) % scalar(int)
         a = paddle.full([2, 2, 2], 3, dtype='int64')
         b = 2
@@ -315,6 +359,11 @@ class TestTensorScalarTypePromotionDynamic(unittest.TestCase):
         c = paddle.full([2, 2, 2], 1, dtype="float32")
         self.check_operation(a, b, c, '%')
 
+    def test_tensor_mod_scalar(self):
+        with _test_eager_guard():
+            self.func_tensor_mod_scalar()
+        self.func_tensor_mod_scalar()
+
 
 if __name__ == '__main__':
     unittest.main()