Add auto completion module for auto parallel

paddle/fluid/framework/op_desc.cc

@@ -353,6 +353,14 @@ void OpDesc::CopyFrom(const OpDesc &op_desc) {
   outputs_ = op_desc.outputs_;
   attrs_ = op_desc.attrs_;
   need_update_ = true;
+  // When creating graph from program, the creation of op node will create a new
+  // OpDesc instead of
+  // referring to the original one. To find the original OpDesc of the op node,
+  // the id have to be
+  // copied to the new OpDesc. The var node has the same situation, but the
+  // default copy constructor
+  // can copy the id automatically.
+  id_ = op_desc.id_;
 }
 
 OpDesc::OpDesc(const proto::OpDesc &desc, BlockDesc *block)

paddle/fluid/framework/op_desc.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <atomic>
 #include <string>
 #include <unordered_map>
 #include <utility>
@@ -151,6 +152,18 @@ class OpDesc {
 
   const BlockDesc *Block() const { return this->block_; }
 
+  // This thread-safe implementation seems to be redudent since the neural
+  // networks
+  // are usually constructed in a single thread
+  static uint64_t GenerateId() {
+    static std::atomic<std::uint64_t> id{0};
+    return ++id;
+  }
+
+  // Note: the identity only used as a key for referring to its
+  // distributed attribute now.
+  uint64_t Id() { return id_; }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {
@@ -173,6 +186,8 @@ class OpDesc {
   // need_update_ indicate there some local changes not be synchronized. If
   // local changes should be synchronized, need_update_ should be set to true.
   bool need_update_{false};
+
+  uint64_t id_ = GenerateId();
 };
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/var_desc.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
+#include <atomic>
 #include <string>
 #include <vector>
 
@@ -150,6 +151,17 @@ class VarDesc {
 
   Attribute GetAttr(const std::string &name) const;
 
+  // This thread-safe implementation seems to be redudent since the neural
+  // networks are usually constructed in a single thread.
+  static uint64_t GenerateId() {
+    static std::atomic<std::uint64_t> uid{0};
+    return ++uid;
+  }
+
+  // Note: the identity only used as a key for referring to its
+  // distributed attribute now.
+  uint64_t Id() { return id_; }
+
  private:
   const proto::VarType::TensorDesc &tensor_desc() const;
   std::vector<proto::VarType::TensorDesc> tensor_descs() const;
@@ -158,6 +170,7 @@ class VarDesc {
 
   proto::VarDesc desc_;
   AttributeMap attrs_;
+  uint64_t id_ = GenerateId();
 };
 
 bool operator==(const VarDesc &left, const VarDesc &right);

paddle/fluid/pybind/protobuf.cc

@@ -24,7 +24,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/framework/version.h"
-
 #include "paddle/fluid/pybind/pybind_boost_headers.h"
 
 namespace paddle {
@@ -201,6 +200,7 @@ void BindVarDsec(pybind11::module *m) {
       .def("attr_names", &pd::VarDesc::AttrNames)
       .def("_set_attr", &pd::VarDesc::SetAttr)
       .def("remove_attr", &pd::VarDesc::RemoveAttr)
+      .def("id", &pd::VarDesc::Id)
       .def("attr", &pd::VarDesc::GetAttr);
 
   pybind11::enum_<pd::proto::VarType::Type> vartype(var_desc, "VarType", "");
@@ -293,6 +293,7 @@ void BindOpDesc(pybind11::module *m) {
       .def("serialize_to_string", SerializeMessage<pd::OpDesc>)
       .def("block", [](pd::OpDesc &self) { return self.Block(); },
            pybind11::return_value_policy::reference)
+      .def("id", &pd::OpDesc::Id)
       .def("inputs", &pd::OpDesc::Inputs)
       .def("outputs", &pd::OpDesc::Outputs);
 }

python/paddle/distributed/__init__.py

@@ -53,7 +53,8 @@
 from . import cloud_utils  # noqa: F401
 from . import utils  # noqa: F401
 
-__all__ = [     #noqa
+
+__all__ = [  # noqa
       "spawn",
       "scatter",
       "broadcast",

python/paddle/distributed/auto_parallel/__init__.py

@@ -18,5 +18,6 @@
 from .interface import set_offload_device  # noqa: F401
 from .interface import set_pipeline_stage  # noqa: F401
 from .interface import ProcessMesh  # noqa: F401
+from .completion import complete_annotation  # noqa: F401
 
 __all__ = []