【Eager】Allow return none when stop_gradient=False (#51740)

* allow return none when stop_gradient=True

* remove useless code

* refine code

* refine code

* fix test cast

* change more test

* add more tests

paddle/fluid/eager/pylayer/py_layer_node.cc

@@ -155,9 +155,21 @@ GradNodePyLayer::operator()(
         grad_out.push_back({});
       } else {
         if (ctx->forward_input_tensor_is_duplicable[i]) {
-          grad_out.push_back(paddle::pybind::GetTensorListFromPyObject(obj));
+          grad_out.push_back(
+              paddle::pybind::GetTensorListFromPyObject(obj, true));
         } else {
-          grad_out.push_back({paddle::pybind::GetTensorFromPyObject(obj)});
+          if (paddle::pybind::PyCheckTensor(obj)) {
+            grad_out.push_back(
+                {paddle::pybind::UnSafeGetTensorFromPyObject(obj)});
+          } else if (obj == Py_None) {
+            VLOG(4) << "Got None for Tensor with pos: " << i;
+            grad_out.push_back({});
+          } else {
+            PADDLE_THROW(phi::errors::InvalidArgument(
+                "argument must be "
+                "Tensor or None, but got %s",
+                reinterpret_cast<PyTypeObject*>(obj->ob_type)->tp_name));
+          }
         }
       }
     } else {

paddle/fluid/pybind/eager_utils.cc

@@ -1219,7 +1219,8 @@ std::vector<paddle::Tensor*> GetTensorPtrListFromPyObject(PyObject* obj) {
   return result;
 }
 
-std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj) {
+std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj,
+                                                      bool allow_none) {
   std::vector<paddle::Tensor> result;
   if (PyList_Check(obj)) {
     Py_ssize_t len = PyList_Size(obj);
@@ -1229,6 +1230,9 @@ std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj) {
       if (PyObject_IsInstance(item,
                               reinterpret_cast<PyObject*>(p_tensor_type))) {
         result.emplace_back(reinterpret_cast<TensorObject*>(item)->tensor);
+      } else if (allow_none && (item == Py_None)) {
+        VLOG(4) << "Got None in Tensor list: " << i;
+        result.emplace_back();
       } else {
         PADDLE_THROW(platform::errors::InvalidArgument(
             "argument must be "
@@ -1245,6 +1249,9 @@ std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj) {
       if (PyObject_IsInstance(item,
                               reinterpret_cast<PyObject*>(p_tensor_type))) {
         result.emplace_back(reinterpret_cast<TensorObject*>(item)->tensor);
+      } else if (allow_none && (item == Py_None)) {
+        VLOG(4) << "Got None in Tensor list: " << i;
+        result.emplace_back();
       } else {
         PADDLE_THROW(platform::errors::InvalidArgument(
             "argument must be "
@@ -1262,16 +1269,9 @@ std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj) {
   return result;
 }
 
-paddle::Tensor& GetTensorFromPyObject(PyObject* obj) {
-  if (!PyCheckTensor(obj)) {
-    PADDLE_THROW(platform::errors::InvalidArgument(
-        "argument must be "
-        "Tensor, but got %s",
-        reinterpret_cast<PyTypeObject*>(obj->ob_type)->tp_name));
-  }
+paddle::Tensor& UnSafeGetTensorFromPyObject(PyObject* obj) {
   return reinterpret_cast<TensorObject*>(obj)->tensor;
 }
-
 paddle::experimental::Scalar CastNumpy2Scalar(PyObject* obj,
                                               const std::string& op_type,
                                               ssize_t arg_pos) {

paddle/fluid/pybind/eager_utils.h

@@ -327,9 +327,9 @@ std::vector<paddle::Tensor*> GetTensorPtrListFromArgs(
 
 std::vector<paddle::Tensor*> GetTensorPtrListFromPyObject(PyObject* obj);
 
-std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj);
-
-paddle::Tensor& GetTensorFromPyObject(PyObject* obj);
+std::vector<paddle::Tensor> GetTensorListFromPyObject(PyObject* obj,
+                                                      bool allow_none = false);
+paddle::Tensor& UnSafeGetTensorFromPyObject(PyObject* obj);
 
 // end of Slice related methods
 

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

@@ -122,6 +122,36 @@ class TestPyLayer(unittest.TestCase):
             np.max(np.abs((input1.grad.numpy() - input2.grad.numpy()))) < 1e-10
         )
 
+    def test_simple_pylayer_multi_output(self):
+        class tanh(PyLayer):
+            @staticmethod
+            def forward(ctx, x1, func1, func2=paddle.split):
+                ctx.func = func2
+                y1 = func1(x1)
+                ctx.save_for_backward(y1)
+                return y1
+
+            @staticmethod
+            def backward(ctx, dy1):
+                (y1,) = ctx.saved_tensor()
+                re1 = ctx.func(dy1, 3)
+                return re1
+
+        input1 = paddle.randn([2, 3]).astype("float64")
+        input2 = paddle.randn([2, 3]).astype("float64")
+        input3 = paddle.randn([2, 3]).astype("float64")
+        input1.stop_gradient = False
+        input2.stop_gradient = False
+        input3.stop_gradient = False
+        z = tanh.apply(x1=[input1, input2, input3], func1=paddle.concat)
+        z.mean().backward()
+        z2 = paddle.concat([input1, input2, input3])
+        z2.mean().backward()
+
+        self.assertTrue(
+            np.max(np.abs((input1.grad.numpy() - input2.grad.numpy()))) < 1e-10
+        )
+
     def test_pylayer_num_output_match(self):
         class tanh(PyLayer):
             @staticmethod
@@ -269,8 +299,8 @@ class TestPyLayer(unittest.TestCase):
         input2.stop_gradient = False
         z = Layer_bk_none1.apply(input2)
 
-        with self.assertRaises(ValueError):
-            z.sum().backward()
+        z.sum().backward()
+        self.assertEqual(input2.grad, None)
 
         class Layer_bk_none2(PyLayer):
             @staticmethod
@@ -285,8 +315,8 @@ class TestPyLayer(unittest.TestCase):
         input1.stop_gradient = False
         z = Layer_bk_none2.apply(input1, input1)
 
-        with self.assertRaises(ValueError):
-            z.mean().backward()
+        z.mean().backward()
+        self.assertIsNone(z.grad)
 
         class Layer_bk_one1(PyLayer):
             @staticmethod