diff --git a/python/paddle/fluid/tests/unittests/test_assign_op.py b/python/paddle/fluid/tests/unittests/test_assign_op.py
index 5efae3fdd5d6cb4c5d68eb6e6a55ac355e0702db..c7fc518986d3513659268d66292ba6aaa3a7b924 100644
--- a/python/paddle/fluid/tests/unittests/test_assign_op.py
+++ b/python/paddle/fluid/tests/unittests/test_assign_op.py
@@ -18,6 +18,7 @@ import eager_op_test
 import gradient_checker
 import numpy as np
 from decorator_helper import prog_scope
+from eager_op_test import convert_float_to_uint16, convert_uint16_to_float
 
 import paddle
 from paddle import fluid
@@ -47,6 +48,9 @@ class TestAssignOp(eager_op_test.OpTest):
         paddle.disable_static()
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda(), "FP16 test runs only on GPU"
+)
 class TestAssignFP16Op(eager_op_test.OpTest):
     def setUp(self):
         self.python_api = paddle.assign
@@ -69,6 +73,32 @@ class TestAssignFP16Op(eager_op_test.OpTest):
         paddle.disable_static()
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda(), "BFP16 test runs only on GPU"
+)
+class TestAssignBFP16Op(eager_op_test.OpTest):
+    def setUp(self):
+        self.python_api = paddle.assign
+        self.public_python_api = paddle.assign
+        self.op_type = "assign"
+        self.prim_op_type = "prim"
+        self.enable_cinn = False
+        x = np.random.uniform(0, 1, [100, 10]).astype(np.float32)
+        x = convert_float_to_uint16(x)
+        self.inputs = {'X': x}
+        self.outputs = {'Out': x}
+
+    def test_forward(self):
+        paddle.enable_static()
+        self.check_output()
+        paddle.disable_static()
+
+    def test_backward(self):
+        paddle.enable_static()
+        self.check_grad(['X'], 'Out', check_prim=True)
+        paddle.disable_static()
+
+
 class TestAssignOpWithLoDTensorArray(unittest.TestCase):
     def test_assign_LoDTensorArray(self):
         paddle.enable_static()
@@ -160,32 +190,12 @@ class TestAssignOApi(unittest.TestCase):
         paddle.disable_static()
 
     def test_assign_NumpyArray(self):
-        with fluid.dygraph.guard():
-            array = np.random.random(size=(100, 10)).astype(np.bool_)
-            result1 = paddle.zeros(shape=[3, 3], dtype='float32')
-            paddle.assign(array, result1)
-        np.testing.assert_allclose(result1.numpy(), array, rtol=1e-05)
-
-    def test_assign_NumpyArray1(self):
-        with fluid.dygraph.guard():
-            array = np.random.random(size=(100, 10)).astype(np.float32)
-            result1 = paddle.zeros(shape=[3, 3], dtype='float32')
-            paddle.assign(array, result1)
-        np.testing.assert_allclose(result1.numpy(), array, rtol=1e-05)
-
-    def test_assign_NumpyArray2(self):
-        with fluid.dygraph.guard():
-            array = np.random.random(size=(100, 10)).astype(np.int32)
-            result1 = paddle.zeros(shape=[3, 3], dtype='float32')
-            paddle.assign(array, result1)
-        np.testing.assert_allclose(result1.numpy(), array, rtol=1e-05)
-
-    def test_assign_NumpyArray3(self):
-        with fluid.dygraph.guard():
-            array = np.random.random(size=(100, 10)).astype(np.int64)
-            result1 = paddle.zeros(shape=[3, 3], dtype='float32')
-            paddle.assign(array, result1)
-        np.testing.assert_allclose(result1.numpy(), array, rtol=1e-05)
+        for dtype in [np.bool_, np.float32, np.int32, np.int64]:
+            with fluid.dygraph.guard():
+                array = np.random.random(size=(100, 10)).astype(dtype)
+                result1 = paddle.zeros(shape=[3, 3], dtype='float32')
+                paddle.assign(array, result1)
+            np.testing.assert_allclose(result1.numpy(), array, rtol=1e-05)
 
     def test_assign_List(self):
         l = [1, 2, 3]
@@ -232,6 +242,31 @@ class TestAssignOApi(unittest.TestCase):
         paddle.disable_static()
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda(), "FP16 test runs only on GPU"
+)
+class TestAssignOApiFP16(unittest.TestCase):
+    def test_assign_fp16(self):
+        x = np.random.uniform(0, 10, [3, 3]).astype(np.float16)
+        x = paddle.to_tensor(x)
+        result = paddle.zeros(shape=[3, 3], dtype='float16')
+        paddle.assign(x, result)
+        np.testing.assert_equal(result.numpy(), x.numpy())
+
+    def test_assign_bfp16(self):
+        x_f = np.random.uniform(0, 10, [3, 3]).astype(np.float32)
+        x = convert_float_to_uint16(x_f)
+        x = paddle.to_tensor(x)
+        result = paddle.zeros(shape=[3, 3], dtype='bfloat16')
+        paddle.assign(x, result)
+        np.testing.assert_allclose(
+            convert_uint16_to_float(result.numpy()), x_f, rtol=1e-02
+        )
+        np.testing.assert_equal(
+            convert_uint16_to_float(result.numpy()), convert_uint16_to_float(x)
+        )
+
+
 class TestAssignOpErrorApi(unittest.TestCase):
     def test_errors(self):
         paddle.enable_static()
@@ -284,6 +319,7 @@ class TestAssignDoubleGradCheck(unittest.TestCase):
             places.append(fluid.CUDAPlace(0))
         for p in places:
             self.func(p)
+        paddle.disable_static()
 
 
 class TestAssignTripleGradCheck(unittest.TestCase):
@@ -315,6 +351,7 @@ class TestAssignTripleGradCheck(unittest.TestCase):
             places.append(fluid.CUDAPlace(0))
         for p in places:
             self.func(p)
+        paddle.disable_static()
 
 
 if __name__ == '__main__':