diff --git a/python/paddle/fluid/tests/unittests/test_diff_op.py b/python/paddle/fluid/tests/unittests/test_diff_op.py
index b89f7759a73a6164f9dd2f5704d509ee5442cb53..1c1a8639ee2b67483861c98e9c7cc55c392117cc 100644
--- a/python/paddle/fluid/tests/unittests/test_diff_op.py
+++ b/python/paddle/fluid/tests/unittests/test_diff_op.py
@@ -228,6 +228,36 @@ class TestDiffOpPreAppendAxis(TestDiffOp):
         self.append = np.array([[2, 3, 4, 7], [1, 3, 5, 6]]).astype('float32')
 
 
+class TestDiffOpFp16(TestDiffOp):
+    def test_fp16_with_gpu(self):
+        paddle.enable_static()
+        if paddle.fluid.core.is_compiled_with_cuda():
+            place = paddle.CUDAPlace(0)
+            with paddle.static.program_guard(
+                paddle.static.Program(), paddle.static.Program()
+            ):
+                input = np.random.random([4, 4]).astype("float16")
+                x = paddle.static.data(
+                    name="input", shape=[4, 4], dtype="float16"
+                )
+                exe = paddle.static.Executor(place)
+                out = paddle.diff(
+                    x,
+                    n=self.n,
+                    axis=self.axis,
+                    prepend=self.prepend,
+                    append=self.append,
+                )
+                fetches = exe.run(
+                    paddle.static.default_main_program(),
+                    feed={
+                        "input": input,
+                    },
+                    fetch_list=[out],
+                )
+        paddle.disable_static()
+
+
 if __name__ == '__main__':
     paddle.enable_static()
     unittest.main()
diff --git a/python/paddle/tensor/math.py b/python/paddle/tensor/math.py
index 3fbc72c01f6bdfa61ec69a889a07c05a438232f1..d040ba1673ff45e19b34ec7262fd6c7c9ce1da26 100644
--- a/python/paddle/tensor/math.py
+++ b/python/paddle/tensor/math.py
@@ -4611,7 +4611,7 @@ def diff(x, n=1, axis=-1, prepend=None, append=None, name=None):
     Only n=1 is currently supported.
 
     Args:
-        x (Tensor): The input tensor to compute the forward difference on
+        x (Tensor): The input tensor to compute the forward difference on, the data type is float16(GPU), float32, float64, bool, int32, int64.
         n (int, optional): The number of times to recursively compute the difference.
                           Only support n=1. Default:1
         axis (int, optional): The axis to compute the difference along. Default:-1
@@ -4706,7 +4706,10 @@ def diff(x, n=1, axis=-1, prepend=None, append=None, name=None):
             return _C_ops.subtract(input_back, input_front)
     else:
         check_variable_and_dtype(
-            x, 'x', ['float32', 'float64', 'bool', 'int32', 'int64'], 'diff'
+            x,
+            'x',
+            ['float16', 'float32', 'float64', 'bool', 'int32', 'int64'],
+            'diff',
         )
         check_type(axis, 'axis', (int), 'diff')
         helper = LayerHelper('diff', **locals())