[DoubleGrad] Enabled double grad test cases in eager_mode for test_im…

…perative_double_grad (#41451)

* [DoubleGrad] Enabled double grad test cases in eager_mode for test_imperative_double_grad

* Fixed elementwise issue

* Addressed CI failures

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

@@ -23,7 +23,7 @@
 ########################
 ops_to_fill_zero_for_empty_grads = set([
     "split_grad", "rnn_grad", "matmul_double_grad", "matmul_triple_grad",
-    "sigmoid_triple_grad"
+    "sigmoid_triple_grad, add_double_grad"
 ])
 
 # For API dispatch used at python-level

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

@@ -205,6 +205,7 @@ class {} : public egr::GradNodeBase {{
 #endif
     }}
     // Forward API Call
+    VLOG(3) << \"Final State Running: \" << \"{}\"; 
 {}
     // Get Outputs
 {}
@@ -505,15 +506,11 @@ def ForwardsValidationCheck(self):
 
         for i in range(len(forward_attrs_list)):
             orig_attr_type = orig_forward_attrs_list[i][1]
-            orig_attr_default = orig_forward_attrs_list[i][2]
             orig_attr_pos = orig_forward_attrs_list[i][3]
             forward_attr_type = forward_attrs_list[i][1]
-            forward_attr_default = forward_attrs_list[i][2]
             forward_attr_pos = forward_attrs_list[i][3]
             assert orig_attr_type == forward_attr_type, AssertMessage(
                 orig_attr_type, forward_attr_type)
-            assert orig_attr_default == forward_attr_default, AssertMessage(
-                orig_attr_default, forward_attr_default)
             assert orig_attr_pos == forward_attr_pos, AssertMessage(
                 orig_attr_pos, forward_attr_pos)
 
@@ -753,6 +750,15 @@ def GenerateNodeCreationCodes(self):
         set_grad_out_meta_list = []
         set_edges_list = []
         for name, (_, pos) in forward_inputs_position_map.items():
+            # Has corresponding grad output
+            has_corresponding_grad_output = False
+            for _, (_, corresponding_pos,
+                    _) in backward_grad_outputs_map.items():
+                if pos == corresponding_pos:
+                    has_corresponding_grad_output = True
+            if not has_corresponding_grad_output:
+                continue
+
             input_autograd_meta_name = GetAutoGradMetaName(name)
             is_optional = (name in self.optional_inputs)
             if is_optional:
@@ -1063,9 +1069,10 @@ def GenerateForwardDefinition(self, is_inplaced):
         self.forward_definition_str += FORWARD_FUNCTION_TEMPLATE.format(
             returns_type_str, forward_function_name, inputs_args_definition_str,
             dygraph_event_str, amp_logic_str, inputs_autograd_meta_str,
-            forward_call_str, get_outputs_str, outputs_autograd_meta_str,
-            compute_require_grad_args_str, check_inplace_str,
-            bump_inplace_version_str, node_creation_str, returns_str)
+            forward_function_name, forward_call_str, get_outputs_str,
+            outputs_autograd_meta_str, compute_require_grad_args_str,
+            check_inplace_str, bump_inplace_version_str, node_creation_str,
+            returns_str)
         self.forward_declaration_str += f"{returns_type_str} {forward_function_name}({inputs_args_declaration_str});\n"
 
         logging.info(
@@ -1439,28 +1446,18 @@ def GenerateNodeDefinition(self, grad_node_creation_str):
             compute_require_grad_str += f"{indent}bool require_any_grad = egr::EagerUtils::ComputeRequireGrad({compute_require_grad_args_str});"
 
         # Construct grad_api returns
-        num_bwd_outputs = len(backward_grad_outputs_map.keys())
         slot_num_bwd_outputs = len(self.forward_inputs_position_map.keys())
         returns_str = f"{indent}std::vector<std::vector<paddle::experimental::Tensor>> returns({slot_num_bwd_outputs});\n"
         for name, (ttype, fwd_position,
                    grad_api_position) in backward_grad_outputs_map.items():
             transformed_tensor_name = self.TransformToNextGradName(name)
 
-            # Infer Grad API Return Type
-            if num_bwd_outputs == 1:
-                # Single tensor output, return as is
-                if IsPlainTensorType(ttype):
-                    returns_str += f"{indent}returns[0] = {{ {transformed_tensor_name} }};\n"
-                else:
-                    assert IsVectorTensorType(ttype)
-                    returns_str += f"{indent}returns[0] = {transformed_tensor_name};\n"
+            # Rearrange output order accordingly
+            if IsPlainTensorType(ttype):
+                returns_str += f"{indent}returns[{fwd_position}] = {{ {transformed_tensor_name} }};\n"
             else:
-                # Rearrange output order accordingly
-                if IsPlainTensorType(ttype):
-                    returns_str += f"{indent}returns[{fwd_position}] = {{ {transformed_tensor_name} }};\n"
-                else:
-                    assert IsVectorTensorType(ttype)
-                    returns_str += f"{indent}returns[{fwd_position}] = {transformed_tensor_name};\n"
+                assert IsVectorTensorType(ttype)
+                returns_str += f"{indent}returns[{fwd_position}] = {transformed_tensor_name};\n"
 
         returns_str += f"{indent}if(NeedComplexToRealConversion()) HandleComplexGradToRealGrad(&returns);\n"
         returns_str += f"{indent}return returns;\n"

paddle/fluid/eager/backward.cc

@@ -485,6 +485,7 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
       }
     }
   }
+
   return node_in_degree_map;
 }
 
@@ -526,6 +527,7 @@ std::vector<paddle::experimental::Tensor> RunBackward(
     bool allow_unused = false,
     const std::vector<paddle::experimental::Tensor>& no_grad_vars = {}) {
   VLOG(6) << "Start Backward";
+
   // *Gradient Hook should happen at node-level
   // *Inplace version check should perform at node-level
   // *Cross-batch accumulation happens at forward pass
@@ -729,6 +731,16 @@ std::vector<paddle::experimental::Tensor> RunBackward(
           continue;
         }
 
+        auto* next_node = next_node_shared.get();
+        if (!node_input_buffers_dict.count(next_node)) {
+          const auto& input_meta = next_node->InputMeta();
+          auto grad_tensor_holder =
+              std::make_unique<GradTensorHolder>(input_meta);
+          VLOG(6) << "Construct GradTensorHolder for grad node: "
+                  << next_node->name();
+          node_input_buffers_dict[next_node] = std::move(grad_tensor_holder);
+        }
+
         PADDLE_ENFORCE_LT(
             j, grad_output_tensors[i].size(),
             paddle::platform::errors::Fatal(
@@ -748,15 +760,6 @@ std::vector<paddle::experimental::Tensor> RunBackward(
                 << ", rank: " << j
                 << " 's name is: " << grad_output_tensor.name();
 
-        auto* next_node = next_node_shared.get();
-        if (!node_input_buffers_dict.count(next_node)) {
-          const auto& input_meta = next_node->InputMeta();
-          auto grad_tensor_holder =
-              std::make_unique<GradTensorHolder>(input_meta);
-          VLOG(6) << "Construct GradTensorHolder for grad node: "
-                  << next_node->name();
-          node_input_buffers_dict[next_node] = std::move(grad_tensor_holder);
-        }
         VLOG(6) << "Sum grad inputs for edge slot: " << edge_rank.first
                 << ", rank: " << edge_rank.second;
         node_input_buffers_dict[next_node]->add(

paddle/phi/kernels/cpu/elementwise_grad_kernel.cc

@@ -63,9 +63,9 @@ void AddGradKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void AddDoubleGradKernel(const Context& dev_ctx,
                          const DenseTensor& y,
+                         const DenseTensor& dout,
                          paddle::optional<const DenseTensor&> ddx,
                          paddle::optional<const DenseTensor&> ddy,
-                         const DenseTensor& dout,
                          int axis,
                          DenseTensor* ddout) {
   phi::AddDoubleGradImpl<T>(dev_ctx, y, ddx, ddy, dout, axis, ddout);

paddle/phi/kernels/elementwise_grad_kernel.h

@@ -31,9 +31,9 @@ void AddGradKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void AddDoubleGradKernel(const Context& dev_ctx,
                          const DenseTensor& y,
+                         const DenseTensor& dout,
                          paddle::optional<const DenseTensor&> ddx,
                          paddle::optional<const DenseTensor&> ddy,
-                         const DenseTensor& dout,
                          int axis,
                          DenseTensor* ddout);
 

paddle/phi/kernels/gpu/elementwise_grad_kernel.cu

@@ -56,9 +56,9 @@ void AddGradKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void AddDoubleGradKernel(const Context& dev_ctx,
                          const DenseTensor& y,
+                         const DenseTensor& dout,
                          paddle::optional<const DenseTensor&> ddx,
                          paddle::optional<const DenseTensor&> ddy,
-                         const DenseTensor& dout,
                          int axis,
                          DenseTensor* ddout) {
   phi::AddDoubleGradImpl<T>(dev_ctx, y, ddx, ddy, dout, axis, ddout);

paddle/phi/ops/compat/elementwise_sig.cc

@@ -115,7 +115,7 @@ KernelSignature ElementwiseAddGradOpArgumentMapping(
 KernelSignature ElementwiseAddDoubleGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   return KernelSignature(
-      "add_double_grad", {"Y", "DDX", "DDY", "DOut"}, {"axis"}, {"DDOut"});
+      "add_double_grad", {"Y", "DOut", "DDX", "DDY"}, {"axis"}, {"DDOut"});
 }
 
 KernelSignature ElementwiseAddTripleGradOpArgumentMapping(

python/paddle/fluid/dygraph/math_op_patch.py

@@ -15,7 +15,7 @@
 from __future__ import print_function
 
 from .. import core
-from ..framework import Variable, convert_np_dtype_to_dtype_, _varbase_creator
+from ..framework import Variable, convert_np_dtype_to_dtype_, _varbase_creator, _in_legacy_dygraph, in_dygraph_mode
 from ..layers.layer_function_generator import OpProtoHolder
 from . import no_grad
 from .. import framework
@@ -62,6 +62,15 @@
 _already_patch_varbase = False
 _already_patch_eager_tensor = False
 
+# Dispatch to final state Python-C functions
+_final_state_op_type_mapping = {
+    "elementwise_add": "final_state_add",
+    "elementwise_sub": "final_state_subtract",
+    "elementwise_div": "final_state_divide",
+    "elementwise_mul": "final_state_multiply",
+    "matmul_v2": "final_state_matmul",
+}
+
 
 def monkey_patch_math_varbase():
     """
@@ -105,10 +114,15 @@ def astype(self, dtype):
         """
         if not isinstance(dtype, core.VarDesc.VarType):
             dtype = convert_np_dtype_to_dtype_(dtype)
-        return _C_ops.cast(self, 'in_dtype', self.dtype, 'out_dtype', dtype)
+
+        if _in_legacy_dygraph():
+            return _C_ops.cast(self, 'in_dtype', self.dtype, 'out_dtype', dtype)
+        return _C_ops.final_state_cast(self, dtype)
 
     def _scalar_elementwise_op_(var, scale, bias):
-        return _C_ops.scale(var, 'scale', scale, 'bias', bias)
+        if _in_legacy_dygraph():
+            return _C_ops.scale(var, 'scale', scale, 'bias', bias)
+        return _C_ops.final_state_scale(var, float(scale), bias, True)
 
     def _neg_(var):
         return _scalar_elementwise_op_(var, -1.0, 0.0)
@@ -164,7 +178,10 @@ def _T_(var):
         perm = []
         for i in range(len(var.shape)):
             perm.insert(0, i)
-        out, _ = _C_ops.transpose2(var, 'axis', perm)
+        if _in_legacy_dygraph():
+            out, _ = _C_ops.transpose2(var, 'axis', perm)
+        else:
+            out = _C_ops.final_state_transpose(var, perm)
         return out
 
     def _scalar_add_(var, value):
@@ -270,11 +287,13 @@ def __impl__(self, other_var):
 
             # 4. calculation
             axis = -1
-            if framework._in_eager_mode_ and op_type == 'elementwise_add':
-                math_op = getattr(_C_ops, 'final_state_add')
+            if in_dygraph_mode(
+            ) and op_type in _final_state_op_type_mapping.keys():
+                math_op = getattr(_C_ops, _final_state_op_type_mapping[op_type])
+                return math_op(self, other_var)
             else:
                 math_op = getattr(_C_ops, op_type)
-            return math_op(self, other_var, 'axis', axis)
+                return math_op(self, other_var, 'axis', axis)
 
         comment = OpProtoHolder.instance().get_op_proto(op_type).comment
 

python/paddle/fluid/layers/nn.py

@@ -9036,7 +9036,10 @@ def relu(x, name=None):
                 # [[0.  0. ]
                 #  [1.  2.6]]
 """
-    if _non_static_mode():
+
+    if in_dygraph_mode():
+        return _C_ops.final_state_relu(x)
+    if _in_legacy_dygraph():
         return _C_ops.relu(x)
 
     check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'relu')

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

@@ -385,26 +385,23 @@ def func_example_with_gradient_accumulation_and_create_graph(self):
                        (x_np > 0) * 2).astype('float32')
         self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
 
-        if not _in_legacy_dygraph():
-            pass
-        else:
-            loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
-            loss.backward(retain_graph=True)
+        loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
+        loss.backward(retain_graph=True)
 
+        x_grad_actual = x.gradient()
+        x_grad_expected = (2.0 / float(numel) *
+                           (x_np + dx_expected *
+                            (x_np > 0) * 2 / float(numel))).astype('float32')
+        self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
+
+        for i in range(5):
+            loss.backward(retain_graph=True)
             x_grad_actual = x.gradient()
-            x_grad_expected = (2.0 / float(numel) * (
+            x_grad_expected = (i + 2) * (2.0 / float(numel) * (
                 x_np + dx_expected *
                 (x_np > 0) * 2 / float(numel))).astype('float32')
             self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
 
-            for i in range(5):
-                loss.backward(retain_graph=True)
-                x_grad_actual = x.gradient()
-                x_grad_expected = (i + 2) * (2.0 / float(numel) * (
-                    x_np + dx_expected *
-                    (x_np > 0) * 2 / float(numel))).astype('float32')
-                self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
-
     def test_example_with_gradient_accumulation_and_create_graph(self):
         with _test_eager_guard():
             self.func_example_with_gradient_accumulation_and_create_graph()
@@ -426,7 +423,10 @@ def func_example_with_gradient_accumulation_and_no_grad_vars(self):
         del y1, z, w
 
         dx_actual, = self.grad(
-            [w_mean], [x], create_graph=True, no_grad_vars=[y2])
+            [w_mean], [x],
+            retain_graph=True,
+            create_graph=True,
+            no_grad_vars=[y2])
 
         self.assertFalse(y2.stop_gradient)
         self.assertFalse(dx_actual.stop_gradient)
@@ -435,17 +435,14 @@ def func_example_with_gradient_accumulation_and_no_grad_vars(self):
                        (x_np > 0) * 2).astype('float32')
         self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
 
-        if not _in_legacy_dygraph():
-            pass
-        else:
-            loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
-            loss.backward()
+        loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
+        loss.backward()
 
-            x_grad_actual = x.gradient()
-            x_grad_expected = (2.0 / float(numel) * (
-                x_np + dx_expected *
-                (x_np > 0) * 4 / float(numel))).astype('float32')
-            self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
+        x_grad_actual = x.gradient()
+        x_grad_expected = (2.0 / float(numel) *
+                           (x_np + dx_expected *
+                            (x_np > 0) * 4 / float(numel))).astype('float32')
+        self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
 
     def test_example_with_gradient_accumulation_and_no_grad_vars(self):
         with _test_eager_guard():
@@ -476,15 +473,12 @@ def func_example_with_gradient_accumulation_and_not_create_graph(self):
 
         self.assertTrue(np.allclose(dx_actual.numpy(), dx_expected))
 
-        if not _in_legacy_dygraph():
-            pass
-        else:
-            loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
-            loss.backward()
+        loss = fluid.layers.reduce_mean(dx_actual * dx_actual + x * x)
+        loss.backward()
 
-            x_grad_actual = x.gradient()
-            x_grad_expected = (2.0 * x_np / float(numel)).astype('float32')
-            self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
+        x_grad_actual = x.gradient()
+        x_grad_expected = (2.0 * x_np / float(numel)).astype('float32')
+        self.assertTrue(np.allclose(x_grad_actual, x_grad_expected))
 
     def test_example_with_gradient_accumulation_and_not_create_graph(self):
         with _test_eager_guard():