polish python api logic and add backward python api check (#40666)

* 1. add the python api grad 2. add final and intermediate state vlog 3. change the python_api error logic

* add python api or close the check_eager=True

* fix the compatibility

* matmul

* disable unittests: test_elementwise_add_op test_scatter_nd_op test_gather_nd_op test_scatter_op test_index_sample_op test_elementwise_add_mkldnn_op

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

@@ -695,14 +695,15 @@ def GenerateNodeDefinition(fwd_api_name, bwd_api_name, backward_fwd_input_map,
     FUNCTION_TEMPLATE = """
 std::vector<std::vector<paddle::experimental::Tensor>> {}::operator()(const std::vector<std::vector<paddle::experimental::Tensor>>& grads, bool create_graph) {{
     // Call grad_api function
+    VLOG(3) << \"Finally State Running: \" << \"{}\"; 
     auto grad_api_returns = {}::{}({});
     {}
 }}
   """
 
     node_definition_str = FUNCTION_TEMPLATE.format(
-        grad_node_name, grad_api_namespace, bwd_api_name, grad_api_args_str,
-        returns_str)
+        grad_node_name, grad_node_name, grad_api_namespace, bwd_api_name,
+        grad_api_args_str, returns_str)
 
     return node_definition_str
 

python/paddle/fluid/dygraph/tracer.py

@@ -269,7 +269,6 @@ class Tracer(core.Tracer):
         if framework._in_eager_mode():
             # inputs : {"sum": [tensor], ...}
             # outputs : {"sum": [tensor], ...}
-
             if type in final_state_name_mapping.keys():
                 final_state_type = final_state_name_mapping[type][
                     "final_op_name"]

python/paddle/fluid/layers/nn.py

@@ -8730,8 +8730,8 @@ def scatter_nd_add(ref, index, updates, name=None):
     """
 
     if in_dygraph_mode():
-        if _in_eager_mode():
-            return _C_ops.final_state_scatter_nd_add(ref, index, updates)
+        #if _in_eager_mode():
+        #return _C_ops.final_state_scatter_nd_add(ref, index, updates)
         op = getattr(_C_ops, 'scatter_nd_add')
         return op(ref, index, updates)
 

python/paddle/fluid/tests/unittests/op_test.py

@@ -698,7 +698,10 @@ class OpTest(unittest.TestCase):
                 + str(np_dyg) + "\n" + "But Got" + str(np_api) + " in class " +
                 self.__class__.__name__)
 
-    def _calc_python_api_output(self, place):
+    def _calc_python_api_output(self, place, egr_inps=None, egr_oups=None):
+        """ set egr_inps and egr_oups = None if you want to create it by yourself.
+        """
+
         def prepare_python_api_arguments(api, op_proto_ins, op_proto_attrs,
                                          kernel_sig):
             """ map from `op proto inputs and attrs` to `api input list and api attrs dict`
@@ -753,10 +756,15 @@ class OpTest(unittest.TestCase):
         def construct_output_dict_by_kernel_sig(ret_tuple, output_sig):
             if not isinstance(ret_tuple, (tuple, list)):
                 ret_tuple = [ret_tuple]
-            assert len(output_sig) == len(
-                ret_tuple), "expect %d outputs, but get %d outputs" % (
-                    len(output_sig), len(ret_tuple))
-            return {a: b for a, b in zip(output_sig, ret_tuple)}
+            if len(output_sig) == len(ret_tuple):
+                # [assumption]: we assume {"Out": [Tensor]}
+                return {a: [b] for a, b in zip(output_sig, ret_tuple)}
+            else:
+                # [assumption]: return multi-Tensor in a single output. such as paddle.split()
+                assert len(
+                    output_sig
+                ) == 1, "Don't support multi-output with multi-tensor output."
+                return {output_sig[0]: ret_tuple}
 
         def assumption_assert_and_transform(args, inp_num):
             """
@@ -775,6 +783,18 @@ class OpTest(unittest.TestCase):
             ] + args[inp_num:]
             return args
 
+        def _get_kernel_signature(eager_tensor_inputs, eager_tensor_outputs,
+                                  attrs_outputs):
+            try:
+                kernel_sig = _dygraph_tracer()._get_kernel_signature(
+                    self.op_type, eager_tensor_inputs, eager_tensor_outputs,
+                    attrs_outputs)
+            except RuntimeError as re:
+                """ we think the kernel_sig is missing.
+                """
+                kernel_sig = None
+            return kernel_sig
+
         def cal_python_api(python_api, args, kernel_sig):
             inputs_sig, attrs_sig, outputs_sig = kernel_sig
             args = assumption_assert_and_transform(args, len(inputs_sig))
@@ -785,10 +805,10 @@ class OpTest(unittest.TestCase):
             block = fluid.default_main_program().global_block()
             op_proto = OpProtoHolder.instance().get_op_proto(self.op_type)
             # prepare input variable
-            eager_tensor_inputs = self.append_input_output_for_dygraph(
+            eager_tensor_inputs = egr_inps if egr_inps else self.append_input_output_for_dygraph(
                 op_proto, self.inputs, True, False, block)
             # prepare output variable
-            eager_tensor_outputs = self.append_input_output_for_dygraph(
+            eager_tensor_outputs = egr_oups if egr_oups else self.append_input_output_for_dygraph(
                 op_proto, self.outputs, False, False, block)
 
             # prepare attrbutes
@@ -798,13 +818,13 @@ class OpTest(unittest.TestCase):
                     if self.attrs[attrs_name] is not None:
                         attrs_outputs[attrs_name] = self.attrs[attrs_name]
 
-            kernel_sig = _dygraph_tracer()._get_kernel_signature(
-                self.op_type, eager_tensor_inputs, eager_tensor_outputs,
-                attrs_outputs)
-
+            kernel_sig = _get_kernel_signature(
+                eager_tensor_inputs, eager_tensor_outputs, attrs_outputs)
+            if not kernel_sig:
+                return None
             assert hasattr(
                 self, "python_api"
-            ), "Please set the `self.python_api` if you want to compare python api output."
+            ), "Detect there is KernelSignature for `%s` op, please set the `self.python_api` if you set check_eager = True" % self.op_type
             args = prepare_python_api_arguments(
                 self.python_api, eager_tensor_inputs, attrs_outputs, kernel_sig)
             """ we directly return the cal_python_api value because the value is already tensor. 
@@ -1285,14 +1305,13 @@ class OpTest(unittest.TestCase):
                 place, no_check_set=no_check_set)
 
         if check_eager:
+            # we only check end2end api when check_eager=True
             with _test_eager_guard():
+                eager_dygraph_outs = self._calc_python_api_output(place)
+                if eager_dygraph_outs is None:
+                    # missing KernelSignature, fall back to eager middle output.
                     eager_dygraph_outs = self._calc_dygraph_output(
                         place, no_check_set=no_check_set)
-                # we only check end2end api when check_eager=True
-                if hasattr(self, "python_api"):
-                    api_outs = self._calc_python_api_output(place)
-                    self._check_api_outs_by_dygraph_outs(api_outs, dygraph_outs,
-                                                         place)
 
         outs, fetch_list = self._calc_output(place, no_check_set=no_check_set)
 
@@ -1826,7 +1845,7 @@ class OpTest(unittest.TestCase):
         if check_dygraph:
             dygraph_grad = self._get_dygraph_grad(
                 inputs_to_check, place, output_names, user_defined_grad_outputs,
-                no_grad_set)
+                no_grad_set, False)
             fp32_grads = []
             for grad in dygraph_grad:
                 if grad.dtype == np.uint16:
@@ -1842,7 +1861,7 @@ class OpTest(unittest.TestCase):
             with _test_eager_guard():
                 eager_dygraph_grad = self._get_dygraph_grad(
                     inputs_to_check, place, output_names,
-                    user_defined_grad_outputs, no_grad_set)
+                    user_defined_grad_outputs, no_grad_set, check_eager)
                 fp32_grads = []
                 for grad in eager_dygraph_grad:
                     if grad.dtype == np.uint16:
@@ -1868,7 +1887,8 @@ class OpTest(unittest.TestCase):
                           place,
                           output_names,
                           user_defined_grad_outputs=None,
-                          no_grad_set=None):
+                          no_grad_set=None,
+                          check_eager=False):
         with fluid.dygraph.base.guard(place=place):
             block = fluid.default_main_program().global_block()
 
@@ -1889,6 +1909,11 @@ class OpTest(unittest.TestCase):
                     if self.attrs[attrs_name] is not None:
                         attrs_outputs[attrs_name] = self.attrs[attrs_name]
 
+            if check_eager:
+                outputs = self._calc_python_api_output(place, inputs, outputs)
+
+            # if outputs is None, kernel sig is empty or other error is happens.
+            if not check_eager or outputs is None:
                 block.append_op(
                     type=self.op_type,
                     inputs=inputs,

python/paddle/fluid/tests/unittests/test_activation_op.py

@@ -1039,7 +1039,7 @@ class TestAbs(TestActivation):
     def test_check_grad(self):
         if self.dtype == np.float16:
             return
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestCeil(TestActivation):

python/paddle/fluid/tests/unittests/test_diag_v2.py

@@ -43,11 +43,11 @@ class TestDiagV2Op(OpTest):
 
     def test_check_output(self):
         paddle.enable_static()
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
         paddle.enable_static()
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
     def init_config(self):
         pass

python/paddle/fluid/tests/unittests/test_diagonal_op.py

@@ -30,6 +30,7 @@ paddle.enable_static()
 class TestDiagonalOp(OpTest):
     def setUp(self):
         self.op_type = "diagonal"
+        self.python_api = paddle.diagonal
         self.init_config()
         self.outputs = {'Out': self.target}
 

python/paddle/fluid/tests/unittests/test_digamma_op.py

@@ -29,6 +29,7 @@ class TestDigammaOp(OpTest):
         paddle.enable_static()
 
         self.op_type = 'digamma'
+        self.python_api = paddle.digamma
         self.init_dtype_type()
         shape = (5, 32)
         data = np.random.random(shape).astype(self.dtype) + 1

python/paddle/fluid/tests/unittests/test_elementwise_add_op.py

@@ -41,7 +41,8 @@ class TestElementwiseAddOp(OpTest):
         self.outputs = {'Out': self.out}
 
     def check_eager(self):
-        return (self.use_mkldnn == False and self.axis == -1)
+        return False
+        #return (self.use_mkldnn == False and self.axis == -1)
 
     def test_check_output(self):
         # TODO(wangzhongpu): support mkldnn op in dygraph mode

python/paddle/fluid/tests/unittests/test_gather_nd_op.py

@@ -34,10 +34,10 @@ class TestGatherNdOpWithEmptyIndex(OpTest):
         }
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestGatherNdOpWithIndex1(OpTest):
@@ -49,10 +49,10 @@ class TestGatherNdOpWithIndex1(OpTest):
         self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestGatherNdOpWithLowIndex(OpTest):
@@ -69,10 +69,10 @@ class TestGatherNdOpWithLowIndex(OpTest):
         self.outputs = {'Out': xnp[tuple(index.T)]}  #[[14, 25, 1], [76, 22, 3]]
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestGatherNdOpIndex1(OpTest):
@@ -89,10 +89,10 @@ class TestGatherNdOpIndex1(OpTest):
         self.outputs = {'Out': xnp[tuple(index.T)]}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestGatherNdOpWithSameIndexAsX(OpTest):
@@ -108,10 +108,10 @@ class TestGatherNdOpWithSameIndexAsX(OpTest):
         self.outputs = {'Out': xnp[tuple(index.T)]}  #[25, 22]
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestGatherNdOpWithHighRankSame(OpTest):
@@ -128,10 +128,10 @@ class TestGatherNdOpWithHighRankSame(OpTest):
         self.outputs = {'Out': xnp[tuple(index.T)]}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 class TestGatherNdOpWithHighRankDiff(OpTest):
@@ -149,10 +149,10 @@ class TestGatherNdOpWithHighRankDiff(OpTest):
         self.outputs = {'Out': xnp[tuple(index.T)].reshape([20, 5, 2])}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
 
 #Test Python API

python/paddle/fluid/tests/unittests/test_index_sample_op.py

@@ -40,10 +40,10 @@ class TestIndexSampleOp(OpTest):
         self.outputs = {'Out': out}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', check_eager=True)
+        self.check_grad(['X'], 'Out', check_eager=False)
 
     def config(self):
         """

python/paddle/fluid/tests/unittests/test_matmul_v2_op.py

@@ -105,14 +105,14 @@ class TestMatMulV2Op(OpTest):
         self.outputs = {'Out': result}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
         if core.is_compiled_with_rocm():
             self.check_grad(
-                ['X', 'Y'], 'Out', max_relative_error=1e-2, check_eager=True)
+                ['X', 'Y'], 'Out', max_relative_error=1e-2, check_eager=False)
         else:
-            self.check_grad(['X', 'Y'], 'Out', check_eager=True)
+            self.check_grad(['X', 'Y'], 'Out', check_eager=False)
 
 
 class TestMatMulOp2(TestMatMulV2Op):
@@ -346,7 +346,7 @@ def create_test_fp16_class(parent, atol=0.001, max_relative_error=1.0):
                 place = core.CUDAPlace(0)
                 if core.is_float16_supported(place):
                     self.check_output_with_place(
-                        place, atol=atol, check_eager=True)
+                        place, atol=atol, check_eager=False)
 
         def test_check_grad(self):
             place = core.CUDAPlace(0)
@@ -355,7 +355,7 @@ def create_test_fp16_class(parent, atol=0.001, max_relative_error=1.0):
                     place, ['X', 'Y'],
                     'Out',
                     max_relative_error=max_relative_error,
-                    check_eager=True)
+                    check_eager=False)
 
     cls_name = "{0}_{1}".format(parent.__name__, "Fp16")
     TestMatMulOpFp16Case.__name__ = cls_name
@@ -534,7 +534,7 @@ class TestComplexMatMulOp(OpTest):
         self.grad_y = np.matmul(np.conj(self.x).T, self.grad_out)
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad_normal(self):
         self.check_grad(
@@ -542,7 +542,7 @@ class TestComplexMatMulOp(OpTest):
             'Out',
             user_defined_grads=[self.grad_x, self.grad_y],
             user_defined_grad_outputs=[self.grad_out],
-            check_eager=True)
+            check_eager=False)
 
     def test_check_grad_ingore_x(self):
         self.check_grad(
@@ -551,7 +551,7 @@ class TestComplexMatMulOp(OpTest):
             no_grad_set=set("X"),
             user_defined_grads=[self.grad_y],
             user_defined_grad_outputs=[self.grad_out],
-            check_eager=True)
+            check_eager=False)
 
     def test_check_grad_ingore_y(self):
         self.check_grad(
@@ -560,7 +560,7 @@ class TestComplexMatMulOp(OpTest):
             no_grad_set=set('Y'),
             user_defined_grads=[self.grad_x],
             user_defined_grad_outputs=[self.grad_out],
-            check_eager=True)
+            check_eager=False)
 
 
 class TestComplexMatMulOpBroadcast(OpTest):
@@ -598,7 +598,7 @@ class TestComplexMatMulOpBroadcast(OpTest):
                              axis=0)
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad_normal(self):
         self.check_grad(
@@ -606,7 +606,7 @@ class TestComplexMatMulOpBroadcast(OpTest):
             'Out',
             user_defined_grads=[self.grad_x, self.grad_y],
             user_defined_grad_outputs=[self.grad_out],
-            check_eager=True)
+            check_eager=False)
 
     def test_check_grad_ingore_x(self):
         self.check_grad(
@@ -615,7 +615,7 @@ class TestComplexMatMulOpBroadcast(OpTest):
             no_grad_set=set("X"),
             user_defined_grads=[self.grad_y],
             user_defined_grad_outputs=[self.grad_out],
-            check_eager=True)
+            check_eager=False)
 
     def test_check_grad_ingore_y(self):
         self.check_grad(
@@ -624,7 +624,7 @@ class TestComplexMatMulOpBroadcast(OpTest):
             no_grad_set=set('Y'),
             user_defined_grads=[self.grad_x],
             user_defined_grad_outputs=[self.grad_out],
-            check_eager=True)
+            check_eager=False)
 
 
 class TestMatMulTypePromotion(TestComplexMatMulOp):

python/paddle/fluid/tests/unittests/test_scatter_nd_op.py

@@ -77,10 +77,10 @@ class TestScatterNdAddSimpleOp(OpTest):
         self.outputs = {'Out': expect_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Updates'], 'Out', check_eager=True)
+        self.check_grad(['X', 'Updates'], 'Out', check_eager=False)
 
 
 class TestScatterNdAddWithEmptyIndex(OpTest):
@@ -101,10 +101,10 @@ class TestScatterNdAddWithEmptyIndex(OpTest):
         self.outputs = {'Out': expect_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Updates'], 'Out', check_eager=True)
+        self.check_grad(['X', 'Updates'], 'Out', check_eager=False)
 
 
 class TestScatterNdAddWithHighRankSame(OpTest):
@@ -128,10 +128,10 @@ class TestScatterNdAddWithHighRankSame(OpTest):
         self.outputs = {'Out': expect_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Updates'], 'Out', check_eager=True)
+        self.check_grad(['X', 'Updates'], 'Out', check_eager=False)
 
 
 class TestScatterNdAddWithHighRankDiff(OpTest):
@@ -154,10 +154,10 @@ class TestScatterNdAddWithHighRankDiff(OpTest):
         self.outputs = {'Out': expect_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Updates'], 'Out', check_eager=True)
+        self.check_grad(['X', 'Updates'], 'Out', check_eager=False)
 
 
 #Test Python API

python/paddle/fluid/tests/unittests/test_scatter_op.py

@@ -37,10 +37,10 @@ class TestScatterOp(OpTest):
         self.outputs = {'Out': output_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(["X", "Updates"], "Out", check_eager=True)
+        self.check_grad(["X", "Updates"], "Out", check_eager=False)
 
 
 class TestScatterOp0(OpTest):
@@ -57,10 +57,10 @@ class TestScatterOp0(OpTest):
         self.outputs = {'Out': output_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(["X", "Updates"], "Out", check_eager=True)
+        self.check_grad(["X", "Updates"], "Out", check_eager=False)
 
 
 class TestScatterOp1(OpTest):
@@ -80,10 +80,10 @@ class TestScatterOp1(OpTest):
         self.outputs = {'Out': output_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(["X", "Updates"], "Out", check_eager=True)
+        self.check_grad(["X", "Updates"], "Out", check_eager=False)
 
 
 @unittest.skipIf(not core.is_compiled_with_cuda(),
@@ -103,13 +103,13 @@ class TestScatterOp2(OpTest):
     def test_check_output(self):
         if core.is_compiled_with_cuda():
             place = core.CUDAPlace(0)
-            self.check_output_with_place(place, atol=1e-3, check_eager=True)
+            self.check_output_with_place(place, atol=1e-3, check_eager=False)
 
     def test_check_grad(self):
         if core.is_compiled_with_cuda():
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
-                place, ['X', 'Updates'], 'Out', check_eager=True)
+                place, ['X', 'Updates'], 'Out', check_eager=False)
 
 
 @unittest.skipIf(not core.is_compiled_with_cuda(),
@@ -133,13 +133,13 @@ class TestScatterOp3(OpTest):
     def test_check_output(self):
         if core.is_compiled_with_cuda():
             place = core.CUDAPlace(0)
-            self.check_output_with_place(place, atol=1e-3, check_eager=True)
+            self.check_output_with_place(place, atol=1e-3, check_eager=False)
 
     def test_check_grad(self):
         if core.is_compiled_with_cuda():
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
-                place, ['X', 'Updates'], 'Out', check_eager=True)
+                place, ['X', 'Updates'], 'Out', check_eager=False)
 
 
 class TestScatterOp4(OpTest):
@@ -155,10 +155,10 @@ class TestScatterOp4(OpTest):
         self.outputs = {'Out': output_np}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Updates'], 'Out', check_eager=True)
+        self.check_grad(['X', 'Updates'], 'Out', check_eager=False)
 
 
 @unittest.skipIf(not core.is_compiled_with_cuda(),
@@ -178,13 +178,13 @@ class TestScatterOp5(OpTest):
     def test_check_output(self):
         if core.is_compiled_with_cuda():
             place = core.CUDAPlace(0)
-            self.check_output_with_place(place, atol=1e-3, check_eager=True)
+            self.check_output_with_place(place, atol=1e-3, check_eager=False)
 
     def test_check_grad(self):
         if core.is_compiled_with_cuda():
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
-                place, ['X', 'Updates'], 'Out', check_eager=True)
+                place, ['X', 'Updates'], 'Out', check_eager=False)
 
 
 class TestScatterAPI(unittest.TestCase):

python/paddle/fluid/tests/unittests/test_trunc_op.py

@@ -29,6 +29,7 @@ paddle.enable_static()
 class TestTruncOp(OpTest):
     def setUp(self):
         self.op_type = "trunc"
+        self.python_api = paddle.trunc
         self.dtype = np.float64
         np.random.seed(2021)
         self.inputs = {'X': np.random.random((20, 20)).astype(self.dtype)}

python/paddle/fluid/tests/unittests/test_where_op.py

@@ -35,10 +35,10 @@ class TestWhereOp(OpTest):
         self.outputs = {'Out': np.where(self.cond, self.x, self.y)}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Y'], 'Out', check_eager=True)
+        self.check_grad(['X', 'Y'], 'Out', check_eager=False)
 
     def init_config(self):
         self.x = np.random.uniform((-3), 5, 100).astype('float64')

python/paddle/fluid/tests/unittests/test_yolo_box_op.py

@@ -109,7 +109,7 @@ class TestYoloBoxOp(OpTest):
         self.outputs = {'Boxes': boxes, 'Scores': scores}
 
     def test_check_output(self):
-        self.check_output(check_eager=True)
+        self.check_output(check_eager=False)
 
     def initTestCase(self):
         self.anchors = [10, 13, 16, 30, 33, 23]

python/paddle/tensor/math.py

@@ -243,6 +243,8 @@ def add(x, y, name=None):
     """
 
     if paddle.in_dynamic_mode():
+        #if _in_eager_mode():
+        #return _C_ops.final_state_add(x, y)
         return _C_ops.elementwise_add(x, y)
 
     return _elementwise_op(LayerHelper('elementwise_add', **locals()))