diff --git a/paddle/fluid/pybind/eager_utils.cc b/paddle/fluid/pybind/eager_utils.cc
index 371ba65a46d156df44c02d5df393416435d81d34..b3ed1df95709f9a8b86ad0a6f4d056a9fe4e0742 100644
--- a/paddle/fluid/pybind/eager_utils.cc
+++ b/paddle/fluid/pybind/eager_utils.cc
@@ -1343,6 +1343,9 @@ paddle::experimental::Scalar CastNumpy2Scalar(PyObject* obj,
   } else if (type_name == "numpy.float32") {
     float value = CastPyArg2Float(obj, op_type, arg_pos);
     return paddle::experimental::Scalar(value);
+  } else if (type_name == "numpy.float16") {
+    float16 value = CastPyArg2Float16(obj, op_type, arg_pos);
+    return paddle::experimental::Scalar(value);
   } else if (type_name == "numpy.int64") {
     int64_t value = CastPyArg2Long(obj, op_type, arg_pos);
     return paddle::experimental::Scalar(value);
@@ -1352,7 +1355,7 @@ paddle::experimental::Scalar CastNumpy2Scalar(PyObject* obj,
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(
         "%s(): argument (position %d) must be "
-        "numpy.float32/float64, numpy.int32/int64, but got %s",
+        "numpy.float16/float32/float64, numpy.int32/int64, but got %s",
         op_type,
         arg_pos + 1,
         type_name));  // NOLINT
diff --git a/paddle/fluid/pybind/op_function_common.cc b/paddle/fluid/pybind/op_function_common.cc
index 5cdd9a0fa06688bec5bdede6cb4b7f01df7f1262..edab97c8b5e696fa72e9b89d732cb7cc121abc61 100644
--- a/paddle/fluid/pybind/op_function_common.cc
+++ b/paddle/fluid/pybind/op_function_common.cc
@@ -184,6 +184,12 @@ void CastPyArg2AttrLong(PyObject* obj,
   attrs[key] = CastPyArg2Long(obj, op_type, arg_pos);
 }
 
+float16 CastPyArg2Float16(PyObject* obj,
+                          const std::string& op_type,
+                          ssize_t arg_pos) {
+  return static_cast<float16>(CastPyArg2Double(obj, op_type, arg_pos));
+}
+
 float CastPyArg2Float(PyObject* obj,
                       const std::string& op_type,
                       ssize_t arg_pos) {
diff --git a/paddle/fluid/pybind/op_function_common.h b/paddle/fluid/pybind/op_function_common.h
index 686694631cc662e0c7ec42522ddec25835e8845d..57423bb3b74a31e3f0c9f8456fe56f740bbe2aa0 100644
--- a/paddle/fluid/pybind/op_function_common.h
+++ b/paddle/fluid/pybind/op_function_common.h
@@ -55,6 +55,9 @@ int CastPyArg2Int(PyObject* obj, const std::string& op_type, ssize_t arg_pos);
 int64_t CastPyArg2Long(PyObject* obj,
                        const std::string& op_type,
                        ssize_t arg_pos);
+float16 CastPyArg2Float16(PyObject* obj,
+                          const std::string& op_type,
+                          ssize_t arg_pos);
 float CastPyArg2Float(PyObject* obj,
                       const std::string& op_type,
                       ssize_t arg_pos);
diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py
index 1cb57e6d72fd574892e4a2601ca7c1a1ece7fbf4..3bf2b7cdcd703583e2cf0e00a9542b6130acef22 100644
--- a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py
+++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py
@@ -737,6 +737,24 @@ class TestElementwiseAddop1(unittest.TestCase):
         paddle.enable_static()
 
 
+class TestTensorAddNumpyScalar(unittest.TestCase):
+    def test_float32_add(self):
+        paddle.disable_static()
+        a = paddle.full([4, 5, 6], 1.5, dtype='float32')
+        b = np.array([1.5], dtype='float32')[0]
+        c = a + b
+        self.assertTrue(c.dtype == core.VarDesc.VarType.FP32)
+
+    def test_float16_add(self):
+        if not core.is_compiled_with_cuda():
+            return
+        paddle.disable_static()
+        a = paddle.full([4, 5, 6], 1.5, dtype='float16')
+        b = np.array([1.5], dtype='float16')[0]
+        c = a + b
+        self.assertTrue(c.dtype == core.VarDesc.VarType.FP16)
+
+
 if __name__ == '__main__':
     paddle.enable_static()
     unittest.main()
diff --git a/python/paddle/fluid/transpiler/collective.py b/python/paddle/fluid/transpiler/collective.py
index 870efa0968d72b04e87cb594a58936854b30de67..04bd68d2571635c19e55802305bc37149e6e7f08 100644
--- a/python/paddle/fluid/transpiler/collective.py
+++ b/python/paddle/fluid/transpiler/collective.py
@@ -516,12 +516,12 @@ class SingleProcessMultiThread(GradAllReduce):
     def _transpile_main_program(self):
         # not need loss scale and no dense param
         param_cnt = self._get_update_param_count()
-        if self.loss_scale is 0 and param_cnt is 0:
+        if self.loss_scale == 0 and param_cnt == 0:
             return
         # scale loss
         self._insert_scale_loss_grad_ops()
         # no param
-        if param_cnt is 0:
+        if param_cnt == 0:
             return
         # fuse allreduce
         if self.fuse_allreduce > 0: