[phi] Transfer merged_momentum yaml to phi (#45359)

* add legacy_api.yaml

* set merged_momentum inplace only

* support inplace optional<vector<tensor>>

* add dygraph_mode api

* optimize TensorToConstDenseTensorPtr

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

@@ -1007,7 +1007,11 @@ class DygraphForwardFunctionGenerator(DygraphFunctionGeneratorBase):
             else:
                 assert IsVectorTensorType(ttype)
                 if is_optional:
-                    arg_str = f"const paddle::optional<std::vector<paddle::experimental::Tensor>>& {name}"
+                    if self.is_forward_only and is_inplaced and forward_inplace_map and name in forward_inplace_map.keys(
+                    ):
+                        arg_str = f"paddle::optional<std::vector<paddle::experimental::Tensor>>& {name}"
+                    else:
+                        arg_str = f"const paddle::optional<std::vector<paddle::experimental::Tensor>>& {name}"
                     amp_tensors_vector_optional_list.append(
                         f"if ({name}) amp_tensors_vector.push_back( *{name} );\n"
                     )
@@ -1015,7 +1019,11 @@ class DygraphForwardFunctionGenerator(DygraphFunctionGeneratorBase):
                         f"auto NEW_{name} = egr::EagerAmpAutoCasts(\"{name}\", {name}, amp_dst_dtype, op_name);\n"
                     )
                 else:
-                    arg_str = f"const std::vector<paddle::experimental::Tensor>& {name}"
+                    if is_inplaced and forward_inplace_map and name in forward_inplace_map.keys(
+                    ):
+                        arg_str = f"std::vector<paddle::experimental::Tensor>& {name}"
+                    else:
+                        arg_str = f"const std::vector<paddle::experimental::Tensor>& {name}"
                     amp_tensors_vector_list.append(f"{name}")
                     amp_autocast_list.append(
                         f"auto NEW_{name} = egr::EagerAmpAutoCasts(\"{name}\", {name}, amp_dst_dtype, op_name);\n"

paddle/phi/api/lib/api_gen_utils.cc

@@ -44,6 +44,34 @@ std::unique_ptr<std::vector<phi::DenseTensor*>> TensorToDenseTensor(
   return pt_tensors;
 }
 
+std::vector<const phi::DenseTensor*> TensorToConstDenseTensorPtr(
+    const std::vector<Tensor>& tensors) {
+  std::vector<const phi::DenseTensor*> pt_tensors(tensors.size());
+
+  for (size_t i = 0; i < tensors.size(); ++i) {
+    pt_tensors[i] = static_cast<phi::DenseTensor*>(tensors[i].impl().get());
+  }
+
+  return pt_tensors;
+}
+
+paddle::optional<std::vector<const phi::DenseTensor*>>
+TensorToConstDenseTensorPtr(
+    const paddle::optional<std::vector<Tensor>>& tensors) {
+  paddle::optional<std::vector<const phi::DenseTensor*>> pt_tensors;
+
+  if (tensors) {
+    pt_tensors =
+        paddle::optional<std::vector<const phi::DenseTensor*>>(tensors->size());
+    for (size_t i = 0; i < tensors->size(); ++i) {
+      pt_tensors->at(i) =
+          static_cast<phi::DenseTensor*>(tensors->at(i).impl().get());
+    }
+  }
+
+  return pt_tensors;
+}
+
 std::shared_ptr<phi::SelectedRows> TensorToSelectedRows(const Tensor& tensor) {
   return std::static_pointer_cast<phi::SelectedRows>(tensor.impl());
 }
@@ -140,6 +168,29 @@ std::vector<phi::DenseTensor*> SetKernelOutput(size_t out_size,
   return results;
 }
 
+std::vector<phi::DenseTensor*> SetInplaceVectorKernelOutput(
+    size_t out_size, Backend backend, std::vector<Tensor>* out) {
+  std::vector<phi::DenseTensor*> results(out->size(), nullptr);
+  for (size_t i = 0; i < out->size(); ++i) {
+    results[i] = static_cast<phi::DenseTensor*>(out->at(i).impl().get());
+  }
+  return results;
+}
+
+std::vector<phi::DenseTensor*> SetInplaceOptionalVectorKernelOutput(
+    size_t out_size,
+    Backend backend,
+    const paddle::optional<std::vector<Tensor>>& out) {
+  std::vector<phi::DenseTensor*> results;
+  if (out) {
+    results = std::vector<phi::DenseTensor*>(out->size(), nullptr);
+    for (size_t i = 0; i < out->size(); ++i) {
+      results[i] = static_cast<phi::DenseTensor*>(out->at(i).impl().get());
+    }
+  }
+  return results;
+}
+
 std::vector<phi::DenseTensor*> SetKernelOutput(std::vector<Tensor*>* out) {
   std::vector<phi::DenseTensor*> results(out->size(), nullptr);
   for (size_t i = 0; i < out->size(); ++i) {

paddle/phi/api/lib/api_gen_utils.h

@@ -38,6 +38,13 @@ paddle::optional<phi::DenseTensor> TensorToDenseTensor(
 std::unique_ptr<std::vector<phi::DenseTensor*>> TensorToDenseTensor(
     const std::vector<Tensor>& tensors);
 
+std::vector<const phi::DenseTensor*> TensorToConstDenseTensorPtr(
+    const std::vector<Tensor>& tensors);
+
+paddle::optional<std::vector<const phi::DenseTensor*>>
+TensorToConstDenseTensorPtr(
+    const paddle::optional<std::vector<Tensor>>& tensors);
+
 std::shared_ptr<phi::SelectedRows> TensorToSelectedRows(const Tensor& tensor);
 
 paddle::optional<phi::SelectedRows> TensorToSelectedRows(
@@ -72,6 +79,14 @@ std::vector<phi::DenseTensor*> SetKernelOutput(size_t out_size,
                                                Backend backend,
                                                std::vector<Tensor>* out);
 
+std::vector<phi::DenseTensor*> SetInplaceVectorKernelOutput(
+    size_t out_size, Backend backend, std::vector<Tensor>* out);
+
+std::vector<phi::DenseTensor*> SetInplaceOptionalVectorKernelOutput(
+    size_t out_size,
+    Backend backend,
+    const paddle::optional<std::vector<Tensor>>& out);
+
 // For backward api
 std::vector<phi::DenseTensor*> SetKernelOutput(std::vector<Tensor*>* out);
 

paddle/phi/api/yaml/generator/api_base.py

@@ -58,9 +58,14 @@ class BaseAPI(object):
     def get_input_tensor_args(self, inplace_flag=False):
         input_args = []
         inplace_type_map = {
-            "const Tensor&": "Tensor&",
-            "const paddle::optional<Tensor>&": "paddle::optional<Tensor>&",
-            "const std::vector<Tensor>&": "std::vector<Tensor>&"
+            "const Tensor&":
+            "Tensor&",
+            "const paddle::optional<Tensor>&":
+            "paddle::optional<Tensor>&",
+            "const std::vector<Tensor>&":
+            "std::vector<Tensor>&",
+            "const paddle::optional<std::vector<Tensor>>&":
+            "paddle::optional<std::vector<Tensor>>&"
         }
         for name in self.inputs['names']:
             name = name.split('@')[0]
@@ -595,9 +600,16 @@ PADDLE_API {self.get_return_type(inplace_flag=True)} {api_func_name}({self.get_d
                     if input_name in self.optional_vars:
                         if self.inputs['input_info'][
                                 input_name] == "const paddle::optional<std::vector<Tensor>>&":
-                            input_name_tensor_map[input_name].append(
-                                (f"{PREFIX_TENSOR_NAME}{input_name}_vec", True))
-                            input_tensor_code = input_tensor_code + f"""
+                            if input_name in self.inplace_map.values():
+                                input_name_tensor_map[input_name].append(
+                                    (f"{PREFIX_TENSOR_NAME}{input_name}", True))
+                                input_tensor_code = input_tensor_code + f"""
+{code_indent}  paddle::optional<std::vector<const phi::DenseTensor*>> {PREFIX_TENSOR_NAME}{input_name} = TensorToConstDenseTensorPtr({input_name});"""
+                            else:
+                                input_name_tensor_map[input_name].append(
+                                    (f"{PREFIX_TENSOR_NAME}{input_name}_vec",
+                                     True))
+                                input_tensor_code = input_tensor_code + f"""
 {code_indent}  auto {PREFIX_TENSOR_NAME}{input_name}_vec = PrepareData({input_name}, kernel.InputAt({kernel_param.index(input_name)}), {trans_flag});
 {code_indent}  paddle::optional<std::vector<const phi::DenseTensor*>> {PREFIX_TENSOR_NAME}{input_name};
 {code_indent}  if ({PREFIX_TENSOR_NAME}{input_name}_vec){{
@@ -622,9 +634,16 @@ PADDLE_API {self.get_return_type(inplace_flag=True)} {api_func_name}({self.get_d
 
                         elif self.inputs['input_info'][
                                 input_name] == "const std::vector<Tensor>&":
-                            input_name_tensor_map[input_name].append(
-                                (f"{PREFIX_TENSOR_NAME}{input_name}_vec", True))
-                            input_tensor_code = input_tensor_code + f"""
+                            if input_name in self.inplace_map.values():
+                                input_name_tensor_map[input_name].append(
+                                    (f"{PREFIX_TENSOR_NAME}{input_name}", True))
+                                input_tensor_code = input_tensor_code + f"""
+{code_indent}  std::vector<const phi::DenseTensor*> {PREFIX_TENSOR_NAME}{input_name} = TensorToConstDenseTensorPtr({input_name});"""
+                            else:
+                                input_name_tensor_map[input_name].append(
+                                    (f"{PREFIX_TENSOR_NAME}{input_name}_vec",
+                                     True))
+                                input_tensor_code = input_tensor_code + f"""
 {code_indent}  auto {PREFIX_TENSOR_NAME}{input_name}_vec = PrepareData({input_name}, kernel.InputAt({kernel_param.index(input_name)}), {trans_flag});
 {code_indent}  std::vector<const phi::DenseTensor*> {PREFIX_TENSOR_NAME}{input_name}({PREFIX_TENSOR_NAME}{input_name}_vec->size());
 {code_indent}  for (size_t i = 0; i < {PREFIX_TENSOR_NAME}{input_name}.size(); ++i) {{
@@ -699,19 +718,23 @@ PADDLE_API {self.get_return_type(inplace_flag=True)} {api_func_name}({self.get_d
 {code_indent}     ddims_vec.clear();"""
             for input_tensor, is_vector in input_name_tensor_map[input_name]:
                 if is_vector:
+                    input_tensor_truncate = input_tensor[:-4]
+                    if input_name in self.inplace_map.values():
+                        input_tensor_truncate = input_tensor
+
                     if input_name in self.optional_vars:
                         input_tensor_code = input_tensor_code + f"""
-{code_indent}     if ({input_tensor[:-4]}){{
-{code_indent}       ddims_vec.reserve({input_tensor[:-4]}->size());
-{code_indent}       for (size_t i = 0; i < {input_tensor[:-4]}->size(); ++i) {{
-{code_indent}         ddims_vec.emplace_back((*{input_tensor[:-4]}->at(i)).dims());
+{code_indent}     if ({input_tensor_truncate}){{
+{code_indent}       ddims_vec.reserve({input_tensor_truncate}->size());
+{code_indent}       for (size_t i = 0; i < {input_tensor_truncate}->size(); ++i) {{
+{code_indent}         ddims_vec.emplace_back((*{input_tensor_truncate}->at(i)).dims());
 {code_indent}       }}
 {code_indent}     }}"""
                     else:
                         input_tensor_code = input_tensor_code + f"""
-{code_indent}     ddims_vec.reserve({input_tensor[:-4]}.size());
-{code_indent}     for (size_t i = 0; i < {input_tensor[:-4]}.size(); ++i) {{
-{code_indent}       ddims_vec.emplace_back((*{input_tensor[:-4]}[i]).dims());
+{code_indent}     ddims_vec.reserve({input_tensor_truncate}.size());
+{code_indent}     for (size_t i = 0; i < {input_tensor_truncate}.size(); ++i) {{
+{code_indent}       ddims_vec.emplace_back((*{input_tensor_truncate}[i]).dims());
 {code_indent}     }}"""
                 else:
                     input_tensor_code = input_tensor_code + f"""  

paddle/phi/api/yaml/generator/api_gen.py

@@ -179,8 +179,7 @@ class ForwardAPI(BaseAPI):
 
                 for out_name in self.outputs['names']:
                     if out_name in self.inplace_map:
-                        output_create = output_create + self.inplace_map[
-                            out_name] + ', '
+                        output_create += self.inplace_map[out_name] + ', '
                     else:
                         output_create += 'Tensor(), '
                 output_create = output_create[:-2] + '};'
@@ -200,6 +199,13 @@ class ForwardAPI(BaseAPI):
                 if out_dtype_list[i] == 'std::vector<Tensor>':
                     assert self.outputs['out_size_expr'][i] is not None, \
                         f"{self.api}: The out size expr : '{{expr}}' should be set when output has Tensor[]. You can refer 'split' api."
+                    # Special case for inplace vector and inplace optional<vector>
+                    if self.outputs['names'][i] in self.inplace_map:
+                        set_out_func = "SetInplaceVectorKernelOutput"
+                        if self.inplace_map[self.outputs['names']
+                                            [i]] in self.optional_vars:
+                            set_out_func = "SetInplaceOptionalVectorKernelOutput"
+                            get_out_code = f"std::get<{i}>(api_output)"
                     output_create = output_create + f"""
 {code_indent}  auto kernel_out_{i} = {set_out_func}({self.outputs['out_size_expr'][i]}, kernel_backend, {get_out_code});"""
 

paddle/phi/api/yaml/legacy_api.yaml

@@ -1780,6 +1780,17 @@
     func : mean_all
   backward : mean_all_grad
 
+- api : merged_momentum_
+  args : (Tensor[] param, Tensor[] grad, Tensor[] velocity, Tensor[] learning_rate, Tensor[] master_param, float mu, bool use_nesterov = false, str[] regularization_method = {}, float[] regularization_coeff = {}, bool multi_precision = false, float rescale_grad = 1.0f)
+  output : Tensor[](param_out){param.size()}, Tensor[](velocity_out){param.size()}, Tensor[](master_param_out){param.size()}
+  infer_meta :
+    func : MergedMomentumInferMeta
+  optional: master_param
+  kernel :
+    func : merged_momentum
+    data_type : param
+  inplace : (param -> param_out), (velocity -> velocity_out), (master_param -> master_param_out)
+
 - api : meshgrid
   args : (Tensor[] inputs)
   output : Tensor[]{inputs.size()}

python/paddle/optimizer/momentum.py

@@ -473,17 +473,28 @@ class Momentum(Optimizer):
                 find_master = self._multi_precision and key == 'FP16_LODTensor'
 
                 if framework._non_static_mode():
-                    _, _, _ = _C_ops.merged_momentum(
-                        self._param_dict[key], grad_dict[key],
-                        self._velocity_dict[key], lr_dict[key],
-                        self._master_weight_dict[key], self._param_dict[key],
-                        self._velocity_dict[key], self._master_weight_dict[key],
-                        'mu', self._momentum, 'use_nesterov',
-                        self._use_nesterov, 'regularization_method',
-                        self._regularization_method_dict[key],
-                        'regularization_coeff',
-                        self._regularization_coeff_dict[key], 'multi_precision',
-                        find_master)
+                    if in_dygraph_mode():
+                        _, _, _ = _C_ops.final_state_merged_momentum_(
+                            self._param_dict[key], grad_dict[key],
+                            self._velocity_dict[key], lr_dict[key],
+                            self._master_weight_dict[key], self._momentum,
+                            self._use_nesterov,
+                            self._regularization_method_dict[key],
+                            self._regularization_coeff_dict[key], find_master,
+                            self._rescale_grad)
+                    else:
+                        _, _, _ = _C_ops.merged_momentum(
+                            self._param_dict[key], grad_dict[key],
+                            self._velocity_dict[key], lr_dict[key],
+                            self._master_weight_dict[key],
+                            self._param_dict[key], self._velocity_dict[key],
+                            self._master_weight_dict[key], 'mu', self._momentum,
+                            'use_nesterov', self._use_nesterov,
+                            'regularization_method',
+                            self._regularization_method_dict[key],
+                            'regularization_coeff',
+                            self._regularization_coeff_dict[key],
+                            'multi_precision', find_master)
                 else:
                     inputs = {
                         "Param": self._param_dict[key],