support eager switch system (#38170)

* support eager switch system

* polish code

paddle/fluid/eager/api/utils/CMakeLists.txt

 cc_library(tensor_utils SRCS tensor_utils.cc DEPS pten pten_api autograd_meta grad_node_info accumulation_node)
 cc_library(hook_utils SRCS hook_utils.cc DEPS pten tensor_utils autograd_meta grad_node_info utils accumulation_node)
-cc_library(global_utils SRCS global_utils.cc DEPS place)
+cc_library(global_utils SRCS global_utils.cc DEPS place tracer)

paddle/fluid/eager/api/utils/global_utils.h

@@ -17,7 +17,7 @@
 
 #include <atomic>
 #include <memory>
-#include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/imperative/tracer.h"
 
 namespace egr {
 
@@ -34,29 +34,49 @@ class UniqueNameGenerator {
 };
 
 // Global
+// TODO(jiabin): Now we are using imperative tracer, move it here when we
+// deprecate imperative.
+
 class Controller {
  public:
   static Controller& Instance() { return *controller_; }
-  const paddle::platform::Place& GetExpectedPlace() const {
-    return *expected_place_.get();
+  paddle::platform::Place GetExpectedPlace() const {
+    return tracer_->ExpectedPlace();
   }
   void SetExpectedPlace(const paddle::platform::Place& place) {
-    expected_place_ = std::make_shared<paddle::platform::Place>(place);
+    tracer_->SetExpectedPlace(place);
+  }
+  void SetAMPLevel(paddle::imperative::AmpLevel level) {
+    tracer_->SetAmpLevel(level);
   }
-  void SetAMPLevel(int level) { amp_level_ = level; }
-  int GetAMPLevel() const { return amp_level_; }
-  bool HasGrad() const { return has_grad_; }
+  paddle::imperative::AmpLevel GetAMPLevel() const {
+    return tracer_->GetAmpLevel();
+  }
+  bool HasGrad() const { return tracer_->HasGrad(); }
+  void SetHasGrad(bool has_grad) { tracer_->SetHasGrad(has_grad); }
   std::string GenerateUniqueName(std::string key = "eager_tmp") {
-    return generator_->Generate(key);
+    return tracer_->GenerateUniqueName(key);
+  }
+  const std::shared_ptr<paddle::imperative::Tracer>& GetCurrentTracer() {
+    return tracer_;
   }
+  void SetCurrentTracer(
+      const std::shared_ptr<paddle::imperative::Tracer>& tracer) {
+    tracer_ = tracer;
+    VLOG(6) << "Set current tracer: " << tracer_;
+  }
+
+  bool InEagerMode() const { return in_eager_mode_; }
+
+  void SetInEagerMode(bool in_eager_mode) { in_eager_mode_ = in_eager_mode; }
 
  private:
   Controller() = default;
   static Controller* controller_;
-  std::shared_ptr<paddle::platform::Place> expected_place_ = nullptr;
-  int amp_level_ = 0;
-  bool has_grad_ = true;
-  std::unique_ptr<UniqueNameGenerator> generator_{new UniqueNameGenerator()};
+  std::shared_ptr<paddle::imperative::Tracer> tracer_{
+      new paddle::imperative::Tracer()};
+  // TODO(jiabin): remove when we don't need imperative.
+  bool in_eager_mode_{false};
   DISABLE_COPY_AND_ASSIGN(Controller);
 };
 

paddle/fluid/eager/legacy/amp_auto_cast.cc

@@ -117,7 +117,7 @@ static inline std::shared_ptr<egr::EagerTensor> CastToType(
   NameTensorMap outs = {{"Out", {out}}};
 
   {
-    AutoCastGuard guard(0);
+    AutoCastGuard guard(paddle::imperative::AmpLevel::O0);
     paddle::framework::AttributeMap default_attrs;
     RunOp("cast", ins, outs, std::move(attrs), {}, &default_attrs, true);
   }

paddle/fluid/eager/legacy/amp_auto_cast.h

@@ -22,18 +22,11 @@
 #include "paddle/fluid/eager/api/utils/global_utils.h"
 #include "paddle/fluid/eager/eager_tensor.h"
 #include "paddle/fluid/eager/legacy/type_def.h"
+#include "paddle/fluid/imperative/amp_auto_cast.h"
 
 namespace egr {
 namespace legacy {
 
-// NOTE(zhiqiu): only O1 and O2 are valid now
-enum class AmpLevel {
-  O0 = 0,  // fp32
-  O1,      // amp, mixed fp32-fp16
-  O2,      // almost fp16
-  O3,      // fp16
-};
-
 class AmpOperators {
  public:
   ~AmpOperators();
@@ -69,7 +62,7 @@ std::ostream& operator<<(std::ostream& os, AmpOperators& ops);
 // NOTE(zhiqiu): AutoCastGuard is used for RAII.
 class AutoCastGuard {
  public:
-  explicit AutoCastGuard(int guard_level) {
+  explicit AutoCastGuard(paddle::imperative::AmpLevel guard_level) {
     pre_amp_level_ = Controller::Instance().GetAMPLevel();
 
     if (pre_amp_level_ != guard_level) {
@@ -84,7 +77,7 @@ class AutoCastGuard {
   AutoCastGuard& operator=(const AutoCastGuard& guard) = delete;
 
  private:
-  int pre_amp_level_;
+  paddle::imperative::AmpLevel pre_amp_level_;
 };
 
 NameTensorMap AutoCastInputs(const std::string& op_type,

paddle/fluid/eager/legacy/op_runner.cc

@@ -131,10 +131,10 @@ void RunOp(const std::string& type, const NameTensorMap& ins,
 
   auto amp_level = egr::Controller::Instance().GetAMPLevel();
   NameTensorMap new_ins = ins;
-  if (amp_level == 1) {
+  if (amp_level == paddle::imperative::AmpLevel::O1) {
     VLOG(5) << "Auto mixed precision run operator: " << type;
     new_ins = AutoCastInputs(type, ins);
-  } else if (amp_level == 2) {
+  } else if (amp_level == paddle::imperative::AmpLevel::O2) {
     VLOG(5) << "Pure fp16 run operator: " << type;
     new_ins = CastPureFp16Inputs(type, ins);
   }

paddle/fluid/pybind/CMakeLists.txt

@@ -2,7 +2,7 @@ set(PYBIND_DEPS pybind python proto_desc memory executor fleet_wrapper box_wrapp
   feed_fetch_method pass generate_pass pass_builder parallel_executor profiler layer tracer engine scope_pool
   analysis_predictor imperative_profiler imperative_flag save_load_util dlpack_tensor device_context
   gloo_wrapper infer_io_utils heter_wrapper generator op_version_registry ps_gpu_wrapper custom_operator
-  cost_model cuda_graph_with_memory_pool fleet_executor)
+  cost_model cuda_graph_with_memory_pool fleet_executor global_utils)
 
 if (WITH_PSCORE)
   set(PYBIND_DEPS ${PYBIND_DEPS} ps_service)

paddle/fluid/pybind/eager.cc

@@ -37,18 +37,17 @@ namespace py = ::pybind11;
 
 PyTypeObject* p_eager_tensor_type;
 
-PyObject* eagertensor_new(PyTypeObject* type, PyObject* args,
-                          PyObject* kwargs) {
+PyObject* EagerTensorNew(PyTypeObject* type, PyObject* args, PyObject* kwargs) {
   PyObject* obj = type->tp_alloc(type, 0);
   if (obj) {
     auto v = reinterpret_cast<EagerTensorObject*>(obj);
-    new (&(v->eagertensor)) egr::EagerTensor();
+    new (&(v->eager_tensor)) egr::EagerTensor();
   }
   return obj;
 }
 
 static void eagertensor_dealloc(EagerTensorObject* self) {
-  self->eagertensor.~EagerTensor();
+  self->eager_tensor.~EagerTensor();
   Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
 }
 
@@ -94,7 +93,7 @@ PyTypeObject eager_tensor_type = {
     0,                       /* tp_dictoffset */
     0,                       /* tp_init */
     0,                       /* tp_alloc */
-    eagertensor_new,         /* tp_new */
+    EagerTensorNew,          /* tp_new */
     0,                       /* tp_free */
     0,                       /* tp_is_gc */
     0,                       /* tp_bases */

paddle/fluid/pybind/eager_functions.cc

@@ -90,13 +90,20 @@ static PyObject* eager_api_set_expected_place(PyObject* self, PyObject* args,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
+static PyObject* eager_api_get_expected_place(PyObject* self, PyObject* args,
+                                              PyObject* kwargs) {
+  EAGER_TRY
+  return ToPyObject(egr::Controller::Instance().GetExpectedPlace());
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 static PyObject* eager_api_scale(PyObject* self, PyObject* args,
                                  PyObject* kwargs) {
   EAGER_TRY
   // TODO(jiabin): Sync Tensor and Variable here when we support
   egr::EagerTensor ret =
       egr::scale(reinterpret_cast<EagerTensorObject*>(PyTuple_GET_ITEM(args, 0))
-                     ->eagertensor,
+                     ->eager_tensor,
                  CastPyArg2AttrFloat(PyTuple_GET_ITEM(args, 1), 1),
                  CastPyArg2AttrFloat(PyTuple_GET_ITEM(args, 2), 2),
                  CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 3), 3),
@@ -128,10 +135,10 @@ static PyObject* eager_api_numpy_to_tensor(PyObject* numpy_data,
   PyObject* obj = p_eager_tensor_type->tp_alloc(p_eager_tensor_type, 0);
   if (obj) {
     auto v = reinterpret_cast<EagerTensorObject*>(obj);
-    new (&(v->eagertensor)) egr::EagerTensor();
-    v->eagertensor.set_impl(densetensor);
-    v->eagertensor.set_name(egr::Controller::Instance().GenerateUniqueName());
-    auto meta = egr::EagerUtils::autograd_meta(&(v->eagertensor));
+    new (&(v->eager_tensor)) egr::EagerTensor();
+    v->eager_tensor.set_impl(densetensor);
+    v->eager_tensor.set_name(egr::Controller::Instance().GenerateUniqueName());
+    auto meta = egr::EagerUtils::autograd_meta(&(v->eager_tensor));
     meta->SetStopGradient(stop_gradient);
 
     // Created tensor will be leaf tensor
@@ -204,6 +211,9 @@ PyMethodDef variable_functions[] = {
     {"_set_expected_place",
      (PyCFunction)(void (*)(void))eager_api_set_expected_place,
      METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_get_expected_place",
+     (PyCFunction)(void (*)(void))eager_api_get_expected_place,
+     METH_VARARGS | METH_KEYWORDS, NULL},
     {"retain_grad_for_tensor",
      (PyCFunction)(void (*)(void))eager_api_retain_grad_for_tensor,
      METH_VARARGS | METH_KEYWORDS, NULL},

paddle/fluid/pybind/eager_method.cc

@@ -36,15 +36,14 @@ extern PyTypeObject* pEagerTensorType;
 
 static PyObject* eager_tensor_method_numpy(EagerTensorObject* self,
                                            PyObject* args, PyObject* kwargs) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  if (!self->eagertensor.initialized()) {
+  EAGER_SYNC_TRY
+  if (!self->eager_tensor.initialized()) {
     Py_INCREF(Py_None);
     return Py_None;
   }
-  auto tensor_dims = self->eagertensor.shape();
-  auto numpy_dtype = TensorDtype2NumpyDtype(self->eagertensor.type());
-  auto sizeof_dtype = pten::DataTypeSize(self->eagertensor.type());
+  auto tensor_dims = self->eager_tensor.shape();
+  auto numpy_dtype = TensorDtype2NumpyDtype(self->eager_tensor.type());
+  auto sizeof_dtype = pten::DataTypeSize(self->eager_tensor.type());
   Py_intptr_t py_dims[paddle::framework::DDim::kMaxRank];
   Py_intptr_t py_strides[paddle::framework::DDim::kMaxRank];
   size_t numel = 1;
@@ -61,18 +60,18 @@ static PyObject* eager_tensor_method_numpy(EagerTensorObject* self,
           pybind11::detail::npy_api::NPY_ARRAY_WRITEABLE_,
       nullptr);
 
-  if (self->eagertensor.is_cpu()) {
+  if (self->eager_tensor.is_cpu()) {
     auto dense_tensor =
-        std::dynamic_pointer_cast<pten::DenseTensor>(self->eagertensor.impl());
+        std::dynamic_pointer_cast<pten::DenseTensor>(self->eager_tensor.impl());
     platform::CPUPlace place;
     // deep copy
     paddle::memory::Copy(place, reinterpret_cast<void*>(
                                     pybind11::detail::array_proxy(array)->data),
                          place, dense_tensor->data(), sizeof_dtype * numel);
 #if defined(PADDLE_WITH_CUDA)
-  } else if (self->eagertensor.is_cuda()) {
+  } else if (self->eager_tensor.is_cuda()) {
     auto dense_tensor =
-        std::dynamic_pointer_cast<pten::DenseTensor>(self->eagertensor.impl());
+        std::dynamic_pointer_cast<pten::DenseTensor>(self->eager_tensor.impl());
 
     paddle::platform::GpuMemcpySync(
         pybind11::detail::array_proxy(array)->data, dense_tensor->data(),
@@ -93,11 +92,8 @@ static PyObject* eager_tensor_method_numpy(EagerTensorObject* self,
 static PyObject* eager_tensor_method_is_initialized(EagerTensorObject* self,
                                                     PyObject* args,
                                                     PyObject* kwargs) {
-  EAGER_TRY
-  if (self->eagertensor.Var().IsInitialized()) {
-    self->eagertensor.SyncToTensor();
-  }
-  return ToPyObject(self->eagertensor.initialized());
+  EAGER_SYNC_TRY
+  return ToPyObject(self->eager_tensor.initialized());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 

paddle/fluid/pybind/eager_properties.cc

@@ -36,44 +36,39 @@ extern PyTypeObject* p_eager_tensor_type;
 
 PyObject* eager_tensor_properties_get_name(EagerTensorObject* self,
                                            void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  return ToPyObject(self->eagertensor.name());
+  EAGER_SYNC_TRY
+  return ToPyObject(self->eager_tensor.name());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
 int eager_tensor_properties_set_name(EagerTensorObject* self, PyObject* value,
                                      void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  self->eagertensor.set_name(CastPyArg2AttrString(value, 0));
+  EAGER_SYNC_TRY
+  self->eager_tensor.set_name(CastPyArg2AttrString(value, 0));
   return 0;
   EAGER_CATCH_AND_THROW_RETURN_ZERO
 }
 
 PyObject* eager_tensor_properties_get_stop_gradient(EagerTensorObject* self,
                                                     void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  auto meta = egr::EagerUtils::autograd_meta(&self->eagertensor);
+  EAGER_SYNC_TRY
+  auto meta = egr::EagerUtils::autograd_meta(&self->eager_tensor);
   return ToPyObject(meta->StopGradient());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
 PyObject* eager_tensor_properties_get_grad(EagerTensorObject* self,
                                            void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  auto meta = egr::EagerUtils::unsafe_autograd_meta(self->eagertensor);
+  EAGER_SYNC_TRY
+  auto meta = egr::EagerUtils::unsafe_autograd_meta(self->eager_tensor);
   return ToPyObject(meta->Grad());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
 int eager_tensor_properties_set_stop_gradient(EagerTensorObject* self,
                                               PyObject* value, void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  auto meta = egr::EagerUtils::autograd_meta(&self->eagertensor);
+  EAGER_SYNC_TRY
+  auto meta = egr::EagerUtils::autograd_meta(&self->eager_tensor);
   meta->SetStopGradient(CastPyArg2AttrBoolean(value, 0));
   return 0;
   EAGER_CATCH_AND_THROW_RETURN_ZERO
@@ -81,18 +76,16 @@ int eager_tensor_properties_set_stop_gradient(EagerTensorObject* self,
 
 PyObject* eager_tensor_properties_get_persistable(EagerTensorObject* self,
                                                   void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  auto meta = egr::EagerUtils::autograd_meta(&self->eagertensor);
+  EAGER_SYNC_TRY
+  auto meta = egr::EagerUtils::autograd_meta(&self->eager_tensor);
   return ToPyObject(meta->Persistable());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
 int eager_tensor_properties_set_persistable(EagerTensorObject* self,
                                             PyObject* value, void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  auto meta = egr::EagerUtils::autograd_meta(&self->eagertensor);
+  EAGER_SYNC_TRY
+  auto meta = egr::EagerUtils::autograd_meta(&self->eager_tensor);
   meta->SetPersistable(CastPyArg2AttrBoolean(value, 0));
   return 0;
   EAGER_CATCH_AND_THROW_RETURN_ZERO
@@ -100,9 +93,8 @@ int eager_tensor_properties_set_persistable(EagerTensorObject* self,
 
 PyObject* eager_tensor_properties_get_shape(EagerTensorObject* self,
                                             void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  auto ddim = self->eagertensor.shape();
+  EAGER_SYNC_TRY
+  auto ddim = self->eager_tensor.shape();
   std::vector<int64_t> value;
   size_t rank = static_cast<size_t>(ddim.size());
   value.resize(rank);
@@ -116,27 +108,24 @@ PyObject* eager_tensor_properties_get_shape(EagerTensorObject* self,
 
 PyObject* eager_tensor_properties_get_place(EagerTensorObject* self,
                                             void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  return ToPyObject(self->eagertensor.place());
+  EAGER_SYNC_TRY
+  return ToPyObject(self->eager_tensor.place());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
 PyObject* eager_tensor_properties_get_place_str(EagerTensorObject* self,
                                                 void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
+  EAGER_SYNC_TRY
   std::stringstream ostr;
-  ostr << self->eagertensor.place();
+  ostr << self->eager_tensor.place();
   return ToPyObject(ostr.str());
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
 PyObject* eager_tensor_properties_get_dtype(EagerTensorObject* self,
                                             void* closure) {
-  EAGER_TRY
-  self->eagertensor.SyncToTensor();
-  return ToPyObject(pten::DataType2String(self->eagertensor.type()));
+  EAGER_SYNC_TRY
+  return ToPyObject(pten::TransToProtoVarType(self->eager_tensor.type()));
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 

paddle/fluid/pybind/eager_utils.cc

@@ -33,6 +33,7 @@ namespace pybind {
 
 extern PyTypeObject* p_eager_tensor_type;
 
+extern PyTypeObject* g_vartype_pytype;
 extern PyTypeObject* g_place_pytype;
 extern PyTypeObject* g_cudaplace_pytype;
 extern PyTypeObject* g_cpuplace_pytype;
@@ -174,7 +175,7 @@ std::string CastPyArg2AttrString(PyObject* obj, ssize_t arg_pos) {
 egr::EagerTensor CastPyArg2EagerTensor(PyObject* obj, ssize_t arg_pos) {
   if (PyObject_IsInstance(obj,
                           reinterpret_cast<PyObject*>(p_eager_tensor_type))) {
-    return reinterpret_cast<EagerTensorObject*>(obj)->eagertensor;
+    return reinterpret_cast<EagerTensorObject*>(obj)->eager_tensor;
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(
         "argument (position %d) must be "
@@ -194,7 +195,7 @@ std::vector<egr::EagerTensor> CastPyArg2VectorOfEagerTensor(PyObject* obj,
       if (PyObject_IsInstance(
               item, reinterpret_cast<PyObject*>(p_eager_tensor_type))) {
         result.emplace_back(
-            reinterpret_cast<EagerTensorObject*>(item)->eagertensor);
+            reinterpret_cast<EagerTensorObject*>(item)->eager_tensor);
       } else {
         PADDLE_THROW(platform::errors::InvalidArgument(
             "argument (position %d) must be "
@@ -211,7 +212,7 @@ std::vector<egr::EagerTensor> CastPyArg2VectorOfEagerTensor(PyObject* obj,
       if (PyObject_IsInstance(
               item, reinterpret_cast<PyObject*>(p_eager_tensor_type))) {
         result.emplace_back(
-            reinterpret_cast<EagerTensorObject*>(item)->eagertensor);
+            reinterpret_cast<EagerTensorObject*>(item)->eager_tensor);
       } else {
         PADDLE_THROW(platform::errors::InvalidArgument(
             "argument (position %d) must be "
@@ -258,6 +259,22 @@ platform::Place CastPyArg2Place(PyObject* obj, ssize_t arg_pos) {
   return place;
 }
 
+paddle::framework::proto::VarType::Type CastPyArg2ProtoType(PyObject* obj,
+                                                            ssize_t arg_pos) {
+  paddle::framework::proto::VarType::Type dtype;
+  if (PyObject_IsInstance(obj, reinterpret_cast<PyObject*>(g_vartype_pytype))) {
+    dtype =
+        ::pybind11::handle(obj).cast<paddle::framework::proto::VarType::Type>();
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "argument (position %d) must be "
+        "one of core.VarDesc.VarType, "
+        "but got %s",
+        arg_pos + 1, reinterpret_cast<PyTypeObject*>(obj->ob_type)->tp_name));
+  }
+  return dtype;
+}
+
 PyObject* ToPyObject(bool value) {
   if (value) {
     Py_INCREF(Py_True);
@@ -286,8 +303,8 @@ PyObject* ToPyObject(const egr::EagerTensor& value) {
   PyObject* obj = p_eager_tensor_type->tp_alloc(p_eager_tensor_type, 0);
   if (obj) {
     auto v = reinterpret_cast<EagerTensorObject*>(obj);
-    new (&(v->eagertensor)) egr::EagerTensor();
-    v->eagertensor = value;
+    new (&(v->eager_tensor)) egr::EagerTensor();
+    v->eager_tensor = value;
   } else {
     PADDLE_THROW(platform::errors::Fatal(
         "tp_alloc return null, can not new a PyObject."));
@@ -352,8 +369,8 @@ PyObject* ToPyObject(const std::vector<egr::EagerTensor>& value) {
     PyObject* obj = p_eager_tensor_type->tp_alloc(p_eager_tensor_type, 0);
     if (obj) {
       auto v = reinterpret_cast<EagerTensorObject*>(obj);
-      new (&(v->eagertensor)) egr::EagerTensor();
-      v->eagertensor = value[i];
+      new (&(v->eager_tensor)) egr::EagerTensor();
+      v->eager_tensor = value[i];
     } else {
       PADDLE_THROW(platform::errors::Fatal(
           "tp_alloc return null, can not new a PyObject."));
@@ -370,6 +387,12 @@ PyObject* ToPyObject(const platform::Place& value) {
   return obj.ptr();
 }
 
+PyObject* ToPyObject(const paddle::framework::proto::VarType::Type& dtype) {
+  auto obj = ::pybind11::cast(dtype);
+  obj.inc_ref();
+  return obj.ptr();
+}
+
 PyObject* ToPyObject(const void* value) {
   if (value == nullptr) {
     Py_INCREF(Py_None);
@@ -399,7 +422,7 @@ egr::EagerTensor GetEagerTensorFromArgs(const std::string& op_type,
     return emptytensor;
   }
 
-  return reinterpret_cast<EagerTensorObject*>(obj)->eagertensor;
+  return reinterpret_cast<EagerTensorObject*>(obj)->eager_tensor;
 }
 
 std::vector<egr::EagerTensor> GetEagerTensorListFromArgs(
@@ -430,7 +453,7 @@ std::vector<egr::EagerTensor> GetEagerTensorListFromArgs(
     for (Py_ssize_t i = 0; i < len; i++) {
       result.emplace_back(
           reinterpret_cast<EagerTensorObject*>(PyList_GetItem(list, i))
-              ->eagertensor);
+              ->eager_tensor);
     }
   } else if (PyTuple_Check(list)) {
     Py_ssize_t len = PyTuple_Size(list);
@@ -443,7 +466,7 @@ std::vector<egr::EagerTensor> GetEagerTensorListFromArgs(
     for (Py_ssize_t i = 0; i < len; i++) {
       result.emplace_back(
           reinterpret_cast<EagerTensorObject*>(PyTuple_GetItem(list, i))
-              ->eagertensor);
+              ->eager_tensor);
     }
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(

paddle/fluid/pybind/eager_utils.h

@@ -18,7 +18,7 @@ namespace paddle {
 namespace pybind {
 
 typedef struct {
-  PyObject_HEAD egr::EagerTensor eagertensor;
+  PyObject_HEAD egr::EagerTensor eager_tensor;
 } EagerTensorObject;
 
 int TensorDtype2NumpyDtype(pten::DataType dtype);
@@ -35,7 +35,8 @@ egr::EagerTensor CastPyArg2EagerTensor(PyObject* obj, ssize_t arg_pos);
 std::vector<egr::EagerTensor> CastPyArg2VectorOfEagerTensor(PyObject* obj,
                                                             ssize_t arg_pos);
 platform::Place CastPyArg2Place(PyObject* obj, ssize_t arg_pos);
-
+framework::proto::VarType::Type CastPyArg2ProtoType(PyObject* obj,
+                                                    ssize_t arg_pos);
 PyObject* ToPyObject(int value);
 PyObject* ToPyObject(bool value);
 PyObject* ToPyObject(int64_t value);
@@ -51,6 +52,7 @@ PyObject* ToPyObject(const std::vector<float>& value);
 PyObject* ToPyObject(const std::vector<double>& value);
 PyObject* ToPyObject(const std::vector<egr::EagerTensor>& value);
 PyObject* ToPyObject(const platform::Place& value);
+PyObject* ToPyObject(const paddle::framework::proto::VarType::Type& dtype);
 PyObject* ToPyObject(const void* value);
 
 template <typename Tuple, size_t N>

paddle/fluid/pybind/exception.h

@@ -19,6 +19,12 @@ limitations under the License. */
 #include "pybind11/pybind11.h"
 
 #define EAGER_TRY try {
+#define EAGER_SYNC_TRY                              \
+  try {                                             \
+    if (self->eager_tensor.Var().IsInitialized()) { \
+      self->eager_tensor.SyncToTensor();            \
+    }
+
 #define EAGER_CATCH_AND_THROW_RETURN_NULL             \
   }                                                   \
   catch (...) {                                       \

paddle/fluid/pybind/imperative.cc

@@ -29,6 +29,7 @@ limitations under the License. */
 #include <utility>
 #include <vector>
 
+#include "paddle/fluid/eager/api/all.h"
 #include "paddle/fluid/framework/scope_guard.h"
 #include "paddle/fluid/imperative/all_reduce.h"
 #include "paddle/fluid/imperative/amp_auto_cast.h"
@@ -868,9 +869,18 @@ void BindImperative(py::module *m_ptr) {
   m.def("_dygraph_debug_level", []() { return imperative::GetDebugLevel(); });
   m.def("_switch_tracer",
         [](const std::shared_ptr<imperative::Tracer> &tracer) {
-          imperative::SetCurrentTracer(tracer);
+          if (egr::Controller::Instance().InEagerMode()) {
+            egr::Controller::Instance().SetCurrentTracer(tracer);
+          } else {
+            imperative::SetCurrentTracer(tracer);
+          }
         });
-
+  m.def("_enable_eager_mode",
+        []() { egr::Controller::Instance().SetInEagerMode(true); });
+  m.def("_disable_eager_mode",
+        []() { egr::Controller::Instance().SetInEagerMode(false); });
+  m.def("_in_eager_mode",
+        []() { return egr::Controller::Instance().InEagerMode(); });
   py::class_<imperative::VarBase, std::shared_ptr<imperative::VarBase>> varbase(
       m, "VarBase", R"DOC()DOC");
   g_varbase_pytype = (PyTypeObject *)varbase.ptr();  // NOLINT

python/paddle/fluid/core.py

@@ -268,6 +268,9 @@ if avx_supported():
         from .core_avx import _is_dygraph_debug_enabled
         from .core_avx import _dygraph_debug_level
         from .core_avx import _switch_tracer
+        from .core_avx import _disable_eager_mode
+        from .core_avx import _enable_eager_mode
+        from .core_avx import _in_eager_mode
         from .core_avx import _set_paddle_lib_path
         from .core_avx import _create_loaded_parameter
         from .core_avx import _cuda_synchronize
@@ -321,6 +324,9 @@ if load_noavx:
         from .core_noavx import _is_dygraph_debug_enabled
         from .core_noavx import _dygraph_debug_level
         from .core_noavx import _switch_tracer
+        from .core_noavx import _disable_eager_mode
+        from .core_noavx import _enable_eager_mode
+        from .core_noavx import _in_eager_mode
         from .core_noavx import _set_paddle_lib_path
         from .core_noavx import _create_loaded_parameter
         from .core_noavx import _cuda_synchronize

python/paddle/fluid/framework.py

@@ -46,8 +46,6 @@ __all__ = [
     'Program',
     'default_startup_program',
     'default_main_program',
-    'eager_guard',
-    'in_eager_mode',
     'program_guard',
     'name_scope',
     'cuda_places',
@@ -79,46 +77,20 @@ _current_device = None
 global_prog_seed = 0
 _current_pipeline_stage = None
 _global_flags_ = core.globals()
-_eager_mode_ = False
+core._disable_eager_mode()
 
 
 @signature_safe_contextmanager
-def eager_mode_place_guard(place):
-    if place is not None:
-        expected_place = _get_paddle_place(place)
-    else:
-        expected_place = _current_expected_place()
-
-    global _global_expected_place_
-    tmp_place = _global_expected_place_
-    _global_expected_place_ = expected_place
-
-    _set_expected_place(expected_place)
-
-    try:
-        yield
-    finally:
-        _global_expected_place_ = tmp_place
-        _set_expected_place(tmp_place)
-
-
-@signature_safe_contextmanager
-def eager_guard(place=None):
-    global _eager_mode_
-    _eager_mode_ = True
+def _test_eager_guard():
+    core._enable_eager_mode()
     _C_ops.switch_to_eager_ops()
     try:
-        with eager_mode_place_guard(place):
-            yield
+        yield
     finally:
-        _eager_mode_ = False
+        core._disable_eager_mode()
         _C_ops.switch_to_core_ops()
 
 
-def in_eager_mode():
-    return _eager_mode_
-
-
 def require_version(min_version, max_version=None):
     """
         Check if the installed version of PaddlePaddle is in [min_version, max_version],
@@ -256,6 +228,10 @@ def in_dygraph_mode():
     return _dygraph_tracer_ is not None
 
 
+def _in_eager_mode():
+    return core._in_eager_mode() and in_dygraph_mode()
+
+
 def _dygraph_not_support_(func):
     def __impl__(*args, **kwargs):
         assert not in_dygraph_mode(
@@ -382,10 +358,9 @@ def _set_dygraph_tracer_expected_place(place):
 def _set_expected_place(place):
     global _global_expected_place_
     _global_expected_place_ = place
-    if in_eager_mode():
+    if _in_eager_mode():
         return core.eager._set_expected_place(place)
-    else:
-        _set_dygraph_tracer_expected_place(place)
+    _set_dygraph_tracer_expected_place(place)
 
 
 # TODO(zhiqiu): remove this function.
@@ -6441,14 +6416,17 @@ def _dygraph_place_guard(place):
     global _global_expected_place_
     tmp_place = _global_expected_place_
     _global_expected_place_ = place
-
+    if _in_eager_mode():
+        core.eager._set_expected_place(place)
     _set_dygraph_tracer_expected_place(place)
 
     try:
         yield
     finally:
         _global_expected_place_ = tmp_place
-        _set_dygraph_tracer_expected_place(tmp_place)
+        if _in_eager_mode():
+            core.eager._set_expected_place(_global_expected_place_)
+        _set_dygraph_tracer_expected_place(_global_expected_place_)
 
 
 def switch_device(device):

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

@@ -16,13 +16,13 @@ import paddle.fluid.core as core
 import paddle.fluid.eager.eager_tensor_patch_methods as eager_tensor_patch_methods
 import paddle
 import numpy as np
-from paddle.fluid import eager_guard
+from paddle.fluid.framework import _test_eager_guard
 import unittest
 
 
 class EagerOpAPIGenerateTestCase(unittest.TestCase):
     def test_elementwise_add(self):
-        with eager_guard():
+        with _test_eager_guard():
             paddle.set_device("cpu")
             np_x = np.ones([4, 16, 16, 32]).astype('float32')
             np_y = np.ones([4, 16, 16, 32]).astype('float32')
@@ -35,7 +35,7 @@ class EagerOpAPIGenerateTestCase(unittest.TestCase):
             self.assertTrue(np.array_equal(out_arr, out_arr_expected))
 
     def test_sum(self):
-        with eager_guard():
+        with _test_eager_guard():
             x_data = np.array(
                 [[0.2, 0.3, 0.5, 0.9], [0.1, 0.2, 0.6, 0.7]]).astype('float32')
             x = paddle.to_tensor(x_data, 'float32')
@@ -45,7 +45,7 @@ class EagerOpAPIGenerateTestCase(unittest.TestCase):
             self.assertTrue(np.array_equal(out_arr, out_arr_expected))
 
     def test_mm(self):
-        with eager_guard():
+        with _test_eager_guard():
             np_input = np.random.random([16, 32]).astype('float32')
             np_mat2 = np.random.random([32, 32]).astype('float32')
             input = paddle.to_tensor(np_input)
@@ -56,7 +56,7 @@ class EagerOpAPIGenerateTestCase(unittest.TestCase):
             self.assertTrue(np.allclose(out_arr, out_arr_expected))
 
     def test_sigmoid(self):
-        with eager_guard():
+        with _test_eager_guard():
             np_x = np.array([-0.4, -0.2, 0.1, 0.3]).astype('float32')
             x = paddle.to_tensor(np_x)
             out = paddle.nn.functional.sigmoid(x)

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

@@ -16,13 +16,14 @@ import paddle.fluid.core as core
 import paddle.fluid.eager.eager_tensor_patch_methods as eager_tensor_patch_methods
 import paddle
 import numpy as np
-from paddle.fluid import eager_guard
+from paddle.fluid.framework import _test_eager_guard
+from paddle.fluid.data_feeder import convert_dtype
 import unittest
 
 
 class EagerScaleTestCase(unittest.TestCase):
     def test_scale_base(self):
-        with eager_guard():
+        with _test_eager_guard():
             paddle.set_device("cpu")
             arr = np.ones([4, 16, 16, 32]).astype('float32')
             tensor = paddle.to_tensor(arr, 'float32', core.CPUPlace())
@@ -35,7 +36,7 @@ class EagerScaleTestCase(unittest.TestCase):
             self.assertEqual(tensor.stop_gradient, True)
 
     def test_retain_grad_and_run_backward(self):
-        with eager_guard():
+        with _test_eager_guard():
             paddle.set_device("cpu")
 
             input_data = np.ones([4, 16, 16, 32]).astype('float32')
@@ -55,33 +56,38 @@ class EagerScaleTestCase(unittest.TestCase):
 
 
 class EagerDtypeTestCase(unittest.TestCase):
-    def check_to_tesnsor_and_numpy(self, dtype):
-        with eager_guard():
+    def check_to_tesnsor_and_numpy(self, dtype, proto_dtype):
+        with _test_eager_guard():
             arr = np.random.random([4, 16, 16, 32]).astype(dtype)
             tensor = paddle.to_tensor(arr, dtype)
-            self.assertEqual(tensor.dtype, dtype)
+            self.assertEqual(tensor.dtype, proto_dtype)
             self.assertTrue(np.array_equal(arr, tensor.numpy()))
 
     def test_dtype_base(self):
-        self.check_to_tesnsor_and_numpy('bool')
-        self.check_to_tesnsor_and_numpy('int8')
-        self.check_to_tesnsor_and_numpy('uint8')
-        self.check_to_tesnsor_and_numpy('int16')
-        self.check_to_tesnsor_and_numpy('int32')
-        self.check_to_tesnsor_and_numpy('int64')
-        self.check_to_tesnsor_and_numpy('float16')
-        self.check_to_tesnsor_and_numpy('float32')
-        self.check_to_tesnsor_and_numpy('float64')
-        self.check_to_tesnsor_and_numpy('complex64')
-        self.check_to_tesnsor_and_numpy('complex128')
+        print("Test_dtype")
+        self.check_to_tesnsor_and_numpy('bool', core.VarDesc.VarType.BOOL)
+        self.check_to_tesnsor_and_numpy('int8', core.VarDesc.VarType.INT8)
+        self.check_to_tesnsor_and_numpy('uint8', core.VarDesc.VarType.UINT8)
+        self.check_to_tesnsor_and_numpy('int16', core.VarDesc.VarType.INT16)
+        self.check_to_tesnsor_and_numpy('int32', core.VarDesc.VarType.INT32)
+        self.check_to_tesnsor_and_numpy('int64', core.VarDesc.VarType.INT64)
+        self.check_to_tesnsor_and_numpy('float16', core.VarDesc.VarType.FP16)
+        self.check_to_tesnsor_and_numpy('float32', core.VarDesc.VarType.FP32)
+        self.check_to_tesnsor_and_numpy('float64', core.VarDesc.VarType.FP64)
+        self.check_to_tesnsor_and_numpy('complex64',
+                                        core.VarDesc.VarType.COMPLEX64)
+        self.check_to_tesnsor_and_numpy('complex128',
+                                        core.VarDesc.VarType.COMPLEX128)
 
 
 class EagerTensorPropertiesTestCase(unittest.TestCase):
     def test_properties(self):
-        with eager_guard():
+        print("Test_properties")
+        with _test_eager_guard():
             paddle.set_device("cpu")
             arr = np.ones([4, 16, 16, 32]).astype('float32')
-            tensor = paddle.to_tensor(arr, 'float32', core.CPUPlace())
+            tensor = paddle.to_tensor(arr, core.VarDesc.VarType.FP32,
+                                      core.CPUPlace())
             self.assertEqual(tensor.shape, [4, 16, 16, 32])
             tensor.name = 'tensor_name_test'
             self.assertEqual(tensor.name, 'tensor_name_test')
@@ -98,6 +104,25 @@ class EagerTensorPropertiesTestCase(unittest.TestCase):
             tensor.stop_gradient = True
             self.assertEqual(tensor.stop_gradient, True)
 
+    def test_global_properties(self):
+        print("Test_global_properties")
+        self.assertFalse(core._in_eager_mode())
+        with _test_eager_guard():
+            self.assertTrue(core._in_eager_mode())
+        self.assertFalse(core._in_eager_mode())
+
+    def test_place_guard(self):
+        core._enable_eager_mode()
+        if core.is_compiled_with_cuda():
+            paddle.set_device("gpu:0")
+            with paddle.fluid.framework._dygraph_place_guard(core.CPUPlace()):
+                self.assertTrue(core.eager._get_expected_place().is_cpu_place())
+        else:
+            paddle.set_device("cpu")
+            with paddle.fluid.framework._dygraph_place_guard(core.CPUPlace()):
+                self.assertTrue(core.eager._get_expected_place().is_cpu_place())
+        core._disable_eager_mode()
+
 
 if __name__ == "__main__":
     unittest.main()

python/paddle/tensor/creation.py

@@ -31,7 +31,7 @@ from ..fluid.framework import convert_np_dtype_to_dtype_, in_dygraph_mode, _varb
 from ..fluid.layers import linspace  # noqa: F401
 import paddle
 from paddle import _C_ops
-from ..fluid.framework import in_eager_mode
+from ..fluid.framework import _in_eager_mode
 
 __all__ = []
 
@@ -116,7 +116,7 @@ def to_tensor(data, dtype=None, place=None, stop_gradient=True):
             ) != _current_expected_place()._get_device_id():
         place = _current_expected_place()
 
-    if in_eager_mode():
+    if _in_eager_mode():
         if dtype is None:
             dtype = paddle.get_default_dtype()
         return core.eager.to_tensor(data,