update eager code gen (#40924)

* update

* remove useless code

* remove label smooth test

* polish code

* polish code

* polish code

* remove _in_eager_mode error;

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

@@ -690,8 +690,13 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
         for name, (_, is_fwd_input,
                    grad_api_position), in backward_forward_inputs_map.items():
             tensor_wrapper_name = GetSavedName(name)
-            grad_api_args[
-                grad_api_position] = f"egr::EagerUtils::RecoverTensorWrapper(&this->{tensor_wrapper_name}, nullptr)"
+            is_optional = (name in self.optional_inputs)
+            if is_optional:
+                grad_api_args[
+                    grad_api_position] = f"egr::EagerUtils::RecoverOptionalTensorWrapper(&this->{tensor_wrapper_name}, nullptr)"
+            else:
+                grad_api_args[
+                    grad_api_position] = f"egr::EagerUtils::RecoverTensorWrapper(&this->{tensor_wrapper_name}, nullptr)"
 
         for _, (ttype, fwd_position,
                 grad_api_position) in backward_grad_inputs_map.items():
@@ -710,7 +715,8 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
 
         # Construct grad_api returns
         num_bwd_outputs = len(backward_grad_outputs_map.keys())
-        returns_str = f"std::vector<std::vector<paddle::experimental::Tensor>> returns({num_bwd_outputs});\n"
+        slot_num_bwd_outputs = len(self.forward_inputs_position_map.keys())
+        returns_str = f"std::vector<std::vector<paddle::experimental::Tensor>> returns({slot_num_bwd_outputs});\n"
         for _, (ttype, fwd_position,
                 grad_api_position) in backward_grad_outputs_map.items():
             # Infer Grad API Return Type
@@ -768,7 +774,7 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
             is_optional = (name in optional_inputs)
             if IsPlainTensorType(ttype):
                 if is_optional:
-                    arg_str = f"const paddle::optional<paddle::experimental::Tensor>& {name}"
+                    arg_str = f"const paddle::optional<const paddle::experimental::Tensor&> {name}"
                 else:
                     if inplace_map and name in inplace_map.keys():
                         arg_str = f"paddle::experimental::Tensor& {name}"
@@ -917,9 +923,9 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
             bump_inplace_version_str += BUMP_INPLACE_VERSION_TEMPLATE.format(
                 inplace_name, inplace_name)
 
-        # Node Construction
-        num_backward_inputs = len(backward_grad_inputs_map.keys())
-        num_backward_outputs = len(backward_grad_outputs_map.keys())
+        # Node Construction        
+        num_backward_inputs = len(forward_outputs_position_map.keys())
+        num_backward_outputs = len(forward_inputs_position_map.keys())
         grad_node_name = GetGradNodeName(forward_api_name)
 
         node_construction_str = f"            auto grad_node = std::make_shared<{grad_node_name}>({num_backward_inputs}, {num_backward_outputs});"
@@ -946,7 +952,7 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
 
             if is_fwd_input:
                 if is_optional:
-                    set_tensor_wrappers = f"        if({name}.is_initialized()) grad_node->SetTensorWrapper{name}({name}, true);"
+                    set_tensor_wrappers = f" if({name}.get_ptr() != nullptr) grad_node->SetTensorWrapper{name}(*({name}.get_ptr()), true);"
                 else:
                     set_tensor_wrappers = f"        grad_node->SetTensorWrapper{name}({name}, true);"
             else:
@@ -960,7 +966,7 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
                     tw_name = f"api_result"
 
                 if is_optional:
-                    set_tensor_wrappers = f"        if({tw_name}.is_initialized()) grad_node->SetTensorWrapper{name}({tw_name}, false);"
+                    set_tensor_wrappers = f" if({tw_name}.get_ptr() != nullptr) grad_node->SetTensorWrapper{name}(*({tw_name}.get_ptr()), false);"
                 else:
                     set_tensor_wrappers = f"        grad_node->SetTensorWrapper{name}({tw_name}, false);"
             set_tensor_wrappers_list.append(set_tensor_wrappers)
@@ -971,8 +977,13 @@ class DygraphSingleFunctionGenerator(FunctionGeneratorBase):
         set_edges_list = []
         for name, (_, pos) in forward_inputs_position_map.items():
             input_autograd_meta_name = GetAutoGradMetaName(name)
-            set_grad_out_meta = f"        grad_node->SetGradOutMeta({name}, {pos});"
-            set_edges = f"        grad_node->AddEdges({input_autograd_meta_name}, {pos});"
+            is_optional = (name in self.optional_inputs)
+            if is_optional:
+                set_grad_out_meta = f"            if({name}.get_ptr() != nullptr) grad_node->SetGradOutMeta(*({name}.get_ptr()), {pos});"
+                set_edges = f"            if({name}.get_ptr() != nullptr)  grad_node->AddEdges({input_autograd_meta_name}, {pos});"
+            else:
+                set_grad_out_meta = f"            grad_node->SetGradOutMeta({name}, {pos});"
+                set_edges = f"            grad_node->AddEdges({input_autograd_meta_name}, {pos});"
             set_grad_out_meta_list.append(set_grad_out_meta)
             set_edges_list.append(set_edges)
         set_grad_out_meta_str = "\n".join(set_grad_out_meta_list)

paddle/fluid/eager/utils.cc

@@ -72,6 +72,14 @@ AutogradMeta* EagerUtils::nullable_autograd_meta(
   return static_cast<AutogradMeta*>(p_autograd_meta);
 }
 
+AutogradMeta* EagerUtils::nullable_autograd_meta(
+    paddle::optional<const paddle::experimental::Tensor&> target) {
+  if (target.get_ptr() != nullptr) {
+    return EagerUtils::nullable_autograd_meta(*(target.get_ptr()));
+  }
+  return nullptr;
+}
+
 std::vector<AutogradMeta*> EagerUtils::nullable_autograd_meta(
     const std::vector<paddle::experimental::Tensor>& targets) {
   std::vector<AutogradMeta*> metas;
@@ -327,6 +335,22 @@ paddle::experimental::Tensor EagerUtils::RecoverTensorWrapper(
   return tw->recover(grad_node);
 }
 
+paddle::optional<const paddle::experimental::Tensor&>
+EagerUtils::RecoverOptionalTensorWrapper(
+    TensorWrapper* tw, const std::shared_ptr<GradNodeBase>& grad_node) {
+  PADDLE_ENFORCE_NOT_NULL(
+      tw, phi::errors::InvalidArgument("TensorWrapper in "
+                                       "RecoverOptionalTensorWrapper function "
+                                       "should not be null"));
+  auto tmp = tw->recover(grad_node);
+
+  paddle::optional<const paddle::experimental::Tensor&> res{paddle::none};
+  if (tmp.initialized()) {
+    res = tmp;
+  }
+  return res;
+}
+
 std::vector<paddle::experimental::Tensor> EagerUtils::RecoverTensorWrapper(
     std::vector<TensorWrapper>* tw,
     const std::shared_ptr<GradNodeBase>& grad_node) {

paddle/fluid/eager/utils.h

@@ -121,6 +121,8 @@ class EagerUtils {
   // This method will return an AutogradMeta pointer unsafely.
   static AutogradMeta* nullable_autograd_meta(
       const paddle::experimental::Tensor& target);
+  static AutogradMeta* nullable_autograd_meta(
+      paddle::optional<const paddle::experimental::Tensor&> target);
   static std::vector<AutogradMeta*> nullable_autograd_meta(
       const std::vector<paddle::experimental::Tensor>& targets);
   static AutogradMeta* unsafe_autograd_meta(
@@ -164,6 +166,9 @@ class EagerUtils {
   static std::vector<paddle::experimental::Tensor> RecoverTensorWrapper(
       std::vector<TensorWrapper>* tw,
       const std::shared_ptr<GradNodeBase>& grad_node);
+  static paddle::optional<const paddle::experimental::Tensor&>
+  RecoverOptionalTensorWrapper(TensorWrapper* tw,
+                               const std::shared_ptr<GradNodeBase>& grad_node);
 
   // Intermidate needed remove this once we don't need legacy
   // Inner Method

paddle/fluid/pybind/eager_utils.cc

@@ -606,7 +606,7 @@ PyObject* ToPyObject(
 
 // For Final State Dygraph,
 // We directly use paddle::optional(Tensor) as dispensable Tensor
-paddle::optional<paddle::experimental::Tensor> GetOptionalTensorFromArgs(
+paddle::optional<const paddle::experimental::Tensor&> GetOptionalTensorFromArgs(
     const std::string& op_type, const std::string& arg_name, PyObject* args,
     ssize_t arg_idx, bool dispensable) {
   PyObject* obj = PyTuple_GET_ITEM(args, arg_idx);
@@ -621,10 +621,10 @@ paddle::optional<paddle::experimental::Tensor> GetOptionalTensorFromArgs(
           "%s(): argument '%s' (position %d) must be Tensor, but got None",
           op_type, arg_name, arg_idx));
     }
-    return {};
+    return paddle::none;
   }
 
-  return paddle::make_optional<paddle::experimental::Tensor>(
+  return paddle::make_optional<const paddle::experimental::Tensor&>(
       reinterpret_cast<TensorObject*>(obj)->tensor);
 }
 

paddle/fluid/pybind/eager_utils.h

@@ -162,7 +162,7 @@ paddle::experimental::DataType CastPyArg2DataType(PyObject* obj,
                                                   const std::string& op_type,
                                                   ssize_t arg_pos);
 
-paddle::optional<paddle::experimental::Tensor> GetOptionalTensorFromArgs(
+paddle::optional<const paddle::experimental::Tensor&> GetOptionalTensorFromArgs(
     const std::string& op_type, const std::string& arg_name, PyObject* args,
     ssize_t arg_idx, bool dispensable = false);
 

paddle/phi/api/lib/data_transform.cc

@@ -191,7 +191,6 @@ std::shared_ptr<phi::DenseTensor> PrepareData(
            dense_tensor.layout(), target_args_def.layout, transform_flag))) {
     return std::static_pointer_cast<phi::DenseTensor>(tensor_in);
   }
-
   phi::DenseTensor out =
       TransformData(dense_tensor, target_args_def, transform_flag);
   return std::make_shared<phi::DenseTensor>(std::move(out));
@@ -207,6 +206,17 @@ std::shared_ptr<phi::DenseTensor> PrepareData(
   return {nullptr};
 }
 
+std::shared_ptr<phi::DenseTensor> PrepareData(
+    const paddle::optional<const Tensor&> input,
+    const phi::TensorArgDef& target_args_def,
+    const TransformFlag& transform_flag) {
+  if (input.get_ptr() != nullptr) {
+    return PrepareData(*(input.get_ptr()), target_args_def, transform_flag);
+  }
+
+  return {nullptr};
+}
+
 std::unique_ptr<std::vector<phi::DenseTensor>> PrepareData(
     const std::vector<Tensor>& inputs,
     const phi::TensorArgDef& target_args_def,

paddle/phi/api/lib/data_transform.h

@@ -76,5 +76,10 @@ std::unique_ptr<std::vector<phi::DenseTensor>> PrepareData(
     const phi::TensorArgDef& target_args_def,
     const TransformFlag& transform_flag);
 
+std::shared_ptr<phi::DenseTensor> PrepareData(
+    const paddle::optional<const Tensor&> input,
+    const phi::TensorArgDef& target_args_def,
+    const TransformFlag& transform_flag);
+
 }  // namespace experimental
 }  // namespace paddle

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

@@ -17,11 +17,13 @@ from __future__ import print_function
 import unittest
 import numpy as np
 from op_test import OpTest
+import paddle
 
 
 class TestLabelSmoothOp(OpTest):
     def config(self):
         self.op_type = "label_smooth"
+        self.python_api = paddle.nn.functional.label_smooth
         self.epsilon = 0.1
         batch_size, self.label_dim = 10, 12
         self.label = np.zeros((batch_size, self.label_dim)).astype("float64")
@@ -37,10 +39,10 @@ class TestLabelSmoothOp(OpTest):
         self.outputs = {'Out': smoothed_label}
 
     def test_check_output(self):
-        self.check_output()
+        self.check_output(check_eager=False)
 
     def test_check_grad(self):
-        self.check_grad(["X"], "Out")
+        self.check_grad(["X"], "Out", check_eager=False)
 
 
 class TestLabelSmoothOpWithPriorDist(TestLabelSmoothOp):
@@ -72,4 +74,5 @@ class TestLabelSmoothOpWithPriorDist3D(TestLabelSmoothOpWithPriorDist):
 
 
 if __name__ == '__main__':
+    paddle.enable_static()
     unittest.main()

python/paddle/utils/code_gen/api.yaml

@@ -544,6 +544,7 @@
 
 
 
+
 #  =================================== sep0
 
 

python/paddle/utils/code_gen/api_base.py

@@ -109,7 +109,7 @@ class BaseAPI(object):
             'int[]': 'const std::vector<int>&'
         }
         optional_types_trans = {
-            'Tensor': 'const paddle::optional<Tensor>&',
+            'Tensor': 'paddle::optional<const Tensor&>',
             'Tensor[]': 'const paddle::optional<std::vector<Tensor>>&',
             'int': 'paddle::optional<int>',
             'int32_t': 'paddle::optional<int32_t>',
@@ -502,7 +502,9 @@ PADDLE_API {self.gene_return_type_code()} {self.get_api_func_name() + '_'}({self
             'const Tensor&': 'const phi::DenseTensor&',
             'const std::vector<Tensor>&':
             'const std::vector<const phi::DenseTensor*>&',
-            'const paddle::optional<Tensor>&':
+            'const paddle::optional<Tensor&>':
+            'paddle::optional<const phi::DenseTensor&>',
+            'paddle::optional<const Tensor&>':
             'paddle::optional<const phi::DenseTensor&>',
             'const paddle::optional<std::vector<Tensor>>&':
             'paddle::optional<const std::vector<phi::DenseTensor>&>'

python/paddle/utils/code_gen/backward.yaml

@@ -307,6 +307,7 @@
   kernel :
     func : mv_grad
 
+    
 - backward_api : cast_grad
   forward : cast (Tensor x, DataType out_dtype) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)

python/paddle/utils/code_gen/wrapped_infermeta_gen.py

@@ -44,13 +44,17 @@ PD_REGISTER_INFER_META_FN({api.kernel['func'][0]}, phi::{api.infer_meta['func']}
                 'Tensor': 'MetaTensor*',
                 'std::vector<Tensor>': 'std::vector<MetaTensor>*',
                 'const paddle::optional<Tensor&>':
-                'const paddle::optional<MetaTensor&>'
+                'const paddle::optional<MetaTensor&>',
+                'paddle::optional<const Tensor&>':
+                'paddle::optional<const MetaTensor&>'
             }
 
             wrapped_infermeta_name = get_wrapped_infermeta_name(api.api)
             args = []
+            print("@@@", api.api)
             for input_name in api.inputs['names']:
                 if input_name in kernel_params:
+                    print("type", api.inputs['input_info'])
                     args.append(tensor_type_map[api.inputs['input_info'][
                         input_name]] + ' ' + input_name)
             for attr_name in api.attrs['names']: