Recompute Offload (#30233) (#30372)

paddle/fluid/framework/distributed_strategy.proto

@@ -22,7 +22,11 @@ enum Mode {
   HETER = 4; // support XPU and GPU computing server
 }
 
-message RecomputeConfig { repeated string checkpoints = 1; }
+message RecomputeConfig {
+  repeated string checkpoints = 1;
+  optional bool enable_offload = 2 [ default = false ];
+  repeated int32 checkpoint_shape = 3;
+}
 
 message ShardingConfig {
   optional float fuse_broadcast_MB = 1 [ default = 32.0 ];

paddle/fluid/framework/ir/squared_mat_sub_fuse_pass.cc

@@ -394,5 +394,5 @@ REGISTER_PASS_CAPABILITY(squared_mat_sub_fuse_pass)
             .EQ("square", 0)
             .LE("elementwise_mul", 1)
             .LE("elementwise_sub", 1)
-            .EQ("fill_constant", 1)
+            .LE("fill_constant", 2)
             .EQ("fusion_squared_mat_sub", 0));

paddle/fluid/operators/fill_constant_op.cc

@@ -116,6 +116,15 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
                   "memory. Otherwise, fill output variable to the running "
                   "device")
         .SetDefault(false);
+    AddAttr<int>("place_type",
+                 "(int, default -1) allow mamually setting place where the "
+                 "variable should be hold. "
+                 "-1: not set manually, determine the place by executor. "
+                 "0: CPUPlace. "
+                 "1: CUDAPlace. "
+                 "2: CUDAPinnedPlace. "
+                 "3: XPUPlace. ")
+        .SetDefault(-1);
     AddOutput("Out",
               "(Tensor) Tensor of specified shape will be filled "
               "with the specified value");
@@ -154,4 +163,11 @@ REGISTER_OP_VERSION(fill_constant)
     )ROC",
         paddle::framework::compatible::OpVersionDesc().NewInput(
             "ValueTensor",
-            "In order to support new feature tensor support of Value"));
+            "In order to support new feature tensor support of Value"))
+    .AddCheckpoint(
+        R"ROC(
+      Upgrade fill_constant to add a new attribute [place_type].
+    )ROC",
+        paddle::framework::compatible::OpVersionDesc().NewAttr(
+            "place_type",
+            "In order to support tensor in CUDAPinnedPlace and XPUPlace", -1));

paddle/fluid/operators/fill_constant_op.h

@@ -39,6 +39,7 @@ class FillConstantKernel : public framework::OpKernel<T> {
     auto str_value = ctx.Attr<std::string>("str_value");
     auto float_value = ctx.Attr<float>("value");
     auto force_cpu = ctx.Attr<bool>("force_cpu");
+    auto place_type = ctx.Attr<int>("place_type");
     framework::Tensor *tensor = nullptr;
 
     framework::Variable *out_var = ctx.OutputVar("Out");
@@ -101,29 +102,59 @@ class FillConstantKernel : public framework::OpKernel<T> {
 
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(ctx.GetPlace());
-    bool cpu_place = force_cpu || ctx.GetPlace() == platform::CPUPlace();
-    if (cpu_place) {
+    int actual_place = place_type;
+
+    if (actual_place == -1) {
+      bool cpu_place = force_cpu || ctx.GetPlace() == platform::CPUPlace();
+      if (cpu_place) {
+        actual_place = 0;
+      } else if (platform::is_gpu_place(ctx.GetPlace())) {
+        actual_place = 1;
+      } else if (platform::is_xpu_place(ctx.GetPlace())) {
+        actual_place = 3;
+      }
+    }
+
+    if (actual_place == 0) {
       tensor->mutable_data(platform::CPUPlace(), data_type);
       math::SetConstant<platform::CPUDeviceContext, T> functor;
       functor(reinterpret_cast<const platform::CPUDeviceContext &>(dev_ctx),
               tensor, static_cast<T>(value));
-    }
+    } else if (actual_place == 1) {
 #ifdef PADDLE_WITH_CUDA
-    if (!cpu_place) {
       tensor->mutable_data(ctx.GetPlace(), data_type);
       math::SetConstant<platform::CUDADeviceContext, T> functor;
       functor(reinterpret_cast<const platform::CUDADeviceContext &>(dev_ctx),
               tensor, static_cast<T>(value));
-    }
+#else
+      PADDLE_THROW(platform::errors::PreconditionNotMet(
+          "PaddlePaddle should compile with GPU."));
 #endif
+    } else if (actual_place == 2) {
+#ifdef PADDLE_WITH_CUDA
+      tensor->mutable_data(platform::CUDAPinnedPlace(), data_type);
+      math::SetConstant<platform::CPUDeviceContext, T> functor;
+      functor(reinterpret_cast<const platform::CPUDeviceContext &>(dev_ctx),
+              tensor, static_cast<T>(value));
+#else
+      PADDLE_THROW(platform::errors::PreconditionNotMet(
+          "PaddlePaddle should compile with GPU."));
+#endif
+    } else if (actual_place == 3) {
 #ifdef PADDLE_WITH_XPU
-    if (!cpu_place) {
       tensor->mutable_data(ctx.GetPlace(), data_type);
       math::SetConstant<platform::XPUDeviceContext, T> functor;
       functor(reinterpret_cast<const platform::XPUDeviceContext &>(dev_ctx),
               tensor, static_cast<T>(value));
-    }
+#else
+      PADDLE_THROW(platform::errors::PreconditionNotMet(
+          "PaddlePaddle should compile with XPU."));
 #endif
+    } else {
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "Could NOT determine the place of variable, place_type = %d .",
+          actual_place));
+    }
   }
 };
 }  // namespace operators

paddle/fluid/operators/memcpy_op.cc

+/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/memcpy_op.h"
+
+#include <string>
+
+namespace paddle {
+namespace framework {
+class OpDesc;
+class Variable;
+}  // namespace framework
+namespace imperative {
+class OpBase;
+}  // namespace imperative
+namespace platform {
+struct CPUPlace;
+struct CUDAPlace;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
+namespace paddle {
+namespace operators {
+
+class MemcpyOp : public framework::OperatorWithKernel {
+ public:
+  MemcpyOp(const std::string &type, const framework::VariableNameMap &inputs,
+           const framework::VariableNameMap &outputs,
+           const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    auto type = ctx->GetInputsVarType("X")[0];
+    if (type == framework::proto::VarType::SELECTED_ROWS ||
+        type == framework::proto::VarType::LOD_TENSOR) {
+      ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+      if (type == framework::proto::VarType::LOD_TENSOR) {
+        ctx->ShareLoD("X", /*->*/ "Out");
+      }
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetKernelTypeForVar(
+      const std::string &var_name, const framework::Tensor &tensor,
+      const framework::OpKernelType &expected_kernel_type) const override {
+    return framework::OpKernelType(expected_kernel_type.data_type_,
+                                   expected_kernel_type.place_,
+                                   tensor.layout());
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
+        ctx.device_context());
+  }
+};
+
+class MemcpyInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(framework::InferVarTypeContext *ctx) const override {
+    ctx->SyncTypeAndDataType("X", "Out");
+  }
+};
+
+class MemcpyKernel {
+ public:
+  void operator()(const framework::ExecutionContext &ctx) const {
+    auto *x = ctx.InputVar("X");
+    if (x == nullptr) {
+      return;
+    }
+    PADDLE_ENFORCE_EQ(
+        ctx.HasOutput("Out"), true,
+        platform::errors::NotFound("Output(Out) of memcpy_op is not found."));
+    auto *out = ctx.OutputVar("Out");
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(ctx.GetPlace());
+    auto dst_place_type = ctx.Attr<int>("dst_place_type");
+    framework::VisitVarType(*x, MemcpyFunctor(out, dev_ctx, dst_place_type));
+  }
+};
+
+class MemcpyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(LoDTensor) The input variable ");
+    AddOutput("Out",
+              "(LoDTensor) The type of output "
+              "is the same as input X.");
+    AddAttr<int>("dst_place_type",
+                 "Determine the dst place of tensor copy. "
+                 "By Now it ONLY support CUDAPlace and CUDAPinnedPlace. Other "
+                 "place type is Unimplemented and will cause ERROR."
+                 "0: dst is on CPUPlace. "
+                 "1: dst is on CUDAPlace. "
+                 "2: dst is on CUDAPinnedPlace. "
+                 "3: dst is on XPUPlace. ");
+    AddComment(R"DOC(
+    Memcpy Operator.
+    By now, it ONLY supports the memcopy between CUDAPinnedPlace and CUDAPlace,
+    and used as an internal op by Recompute-Offload.
+    You would have to update it if you want other more capacities.
+
+Out = X,  when type in [LoDTensor]
+raise error if the type is not listed above.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+REGISTER_OPERATOR(
+    memcpy, ops::MemcpyOp, ops::MemcpyOpProtoMaker, ops::MemcpyInferVarType,
+    paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,
+    paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);
+
+REGISTER_OP_CPU_KERNEL_FUNCTOR(memcpy, float, ops::MemcpyKernel, double,
+                               ops::MemcpyKernel, int, ops::MemcpyKernel,
+                               int64_t, ops::MemcpyKernel, bool,
+                               ops::MemcpyKernel, plat::float16,
+                               ops::MemcpyKernel);
+
+#ifdef PADDLE_WITH_CUDA
+REGISTER_OP_CUDA_KERNEL_FUNCTOR(memcpy, float, ops::MemcpyKernel, double,
+                                ops::MemcpyKernel, int, ops::MemcpyKernel,
+                                int64_t, ops::MemcpyKernel, bool,
+                                ops::MemcpyKernel, plat::float16,
+                                ops::MemcpyKernel);
+#endif

paddle/fluid/operators/memcpy_op.h

+/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/fluid/framework/data_type.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/var_type.h"
+#include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
+namespace paddle {
+namespace operators {
+class MemcpyFunctor {
+ public:
+  MemcpyFunctor(framework::Variable *out,
+                const platform::DeviceContext &dev_ctx,
+                const int dst_place_type)
+      : out_(out), dev_ctx_(dev_ctx), dst_place_type_(dst_place_type) {}
+
+  void operator()(const framework::LoDTensor &lod_tensor) const {
+    auto &out_tensor = *out_->GetMutable<framework::LoDTensor>();
+
+    if (dst_place_type_ == 3) {
+      framework::TensorCopy(lod_tensor, platform::CUDAPinnedPlace(), dev_ctx_,
+                            &out_tensor);
+    } else if (dst_place_type_ == 2) {
+      framework::TensorCopy(lod_tensor, dev_ctx_.GetPlace(), dev_ctx_,
+                            &out_tensor);
+    } else {
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "memcpy dst_place_type: %d is not supported yet.", dst_place_type_));
+    }
+    out_tensor.set_lod(lod_tensor.lod());
+  }
+
+  void operator()(const framework::SelectedRows &rows) const {
+    // (JZ-LIANG) to support SelectedRows
+    PADDLE_THROW(platform::errors::Unimplemented(
+        "Memcpy for SelectedRows is NOT support yet."));
+  }
+
+  template <typename T>
+  void operator()(const T &v) const {
+    PADDLE_ENFORCE_EQ(
+        true, false,
+        platform::errors::PermissionDenied(
+            "Not support type for Memcpy  op with type %s", typeid(T).name()));
+  }
+
+ private:
+  framework::Variable *out_;
+  const platform::DeviceContext &dev_ctx_;
+  const int dst_place_type_;
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/distributed/fleet/base/distributed_strategy.py

@@ -632,8 +632,20 @@ class DistributedStrategy(object):
     @property
     def recompute_configs(self):
         """
-        Set recompute configurations. In general, the recompute strategy of current
-        implementation should have some manually assign checkpoints
+        Set recompute configurations. 
+        
+        **Note**:
+        checkpoints(list): list of string name of checkpoints. In general, the recompute
+        strategy of current implementation should have some manually assign checkpoints.
+
+        enable_offload(bool): enable recompute checkpoints offload feature. this feature 
+        will offload the checkpoint to host memory to allow even larger batch size. since
+        the memcpy from host to device takes time, it is a trade off between larger batch
+        size and training speed.
+
+        checkpoint_shape(list): list of int that specific the shape of checkpoint. so far
+        recompute-offload requires that all checkpoint to be same shape, and every dimension
+        specific here should be determined ("-1" is not allowed). 
 
         Examples:
 
@@ -642,7 +654,10 @@ class DistributedStrategy(object):
             import paddle.distributed.fleet as fleet
             strategy = fleet.DistributedStrategy()
             strategy.recompute = True
-            strategy.recompute_configs = {"checkpoints": ["x", "y"]}
+            strategy.recompute_configs = {
+                "checkpoints": ["x", "y"],
+                "enable_offload": True,
+                "checkpoint_shape": [100, 512, 1024] }
 
         """
         return get_msg_dict(self.strategy.recompute_configs)
@@ -692,6 +707,14 @@ class DistributedStrategy(object):
             This configuration will affect the communication speed in sharding training, 
             and should be an empirical value decided by your model size and network topology.
 
+            hybrid_dp(bool): enable hybrid data parallelism above the sharding parallelism. 
+            you are supposed to have at least double the number of gpu you have in normal sharding 
+            training to enable this feature.
+
+            sharding_group_size(int): attribute of hybrid_dp. specific the the number of gpus within
+            each sharding group; and therefore, the number of hybrid data parallelism ways will be equal
+            to (global_size / sharding_group_size).
+
         Examples:
 
           .. code-block:: python
@@ -699,7 +722,10 @@ class DistributedStrategy(object):
             import paddle.distributed.fleet as fleet
             strategy = fleet.DistributedStrategy()
             strategy.sharding = True
-            strategy.sharding_configs = {"fuse_broadcast_MB": 32}
+            strategy.sharding_configs = {
+                "fuse_broadcast_MB": 32,
+                "hybrid_dp": True,
+                "sharding_group_size": 8}
         """
         return get_msg_dict(self.strategy.sharding_configs)
 

python/paddle/distributed/fleet/meta_optimizers/recompute_optimizer.py

@@ -39,9 +39,13 @@ class RecomputeOptimizer(MetaOptimizerBase):
             return
 
         configs = self.user_defined_strategy.recompute_configs
-
         self.wrapped_opt = RO(self.inner_opt)
         self.wrapped_opt._set_checkpoints(list(configs["checkpoints"]))
+        if configs["enable_offload"]:
+            self.wrapped_opt._enable_offload()
+            # TODO(JZ-LIANG) might found a way to infer the checkpoint shape automatically
+            checkpoint_shapes = list(configs["checkpoint_shape"])
+            self.wrapped_opt.checkpoint_shape = checkpoint_shapes
 
     def _can_apply(self):
         if not self.role_maker._is_collective:

python/paddle/fluid/backward.py

@@ -99,8 +99,32 @@ class ProgramStats(object):
                 max_op_idx = max(max_op_idx, idx)
         if min_op_idx >= max_op_idx:
             return False, min_op_idx, max_op_idx
+
         return True, min_op_idx, max_op_idx
 
+    def _update_segment_start(self, min_idx, pre_segment_end_idx):
+        """
+        persist vars of amp-related cast should be included in recompute segment
+        """
+
+        def is_amp_cast(op):
+            return op.desc.type() == 'cast' and self.block.var(
+                op.desc.input_arg_names()[0]).persistable
+
+        idx_ = min_idx - 1
+        updated_min_idx = min_idx
+        while idx_ > pre_segment_end_idx:
+            if is_amp_cast(self.ops[idx_]):
+                _logger.debug("found amp-cast op: {}, : {}".format(self.ops[
+                    idx_].desc.type(), self.ops[idx_].desc.input_arg_names()[
+                        0]))
+                updated_min_idx = idx_
+                idx_ -= 1
+            else:
+                break
+
+        return updated_min_idx
+
     def build_stats(self):
         for i, op in enumerate(self.ops):
             self.op_deps[i] = {"in_ops": [], "out_ops": []}
@@ -751,20 +775,29 @@ def _append_backward_ops_with_checkpoints_(
             if name not in program_stat.var_op_deps:
                 break
             op_idx = program_stat.var_op_deps[name]["var_as_output_ops"]
+            # only count the last generate op
             for idx in op_idx:
                 max_op_idx = max(max_op_idx, idx)
         if max_op_idx > 0:
             segments.append([0, max_op_idx + 1])
     else:
         start_idx = 0
+        pre_segment_end_idx = -1
         while True:
+            _logger.debug("FW op range[0] - [{}]".format(len(ops)))
             if start_idx >= len(checkpoints_name) - 1:
                 break
+            # min_idx: checkpoint_1' s input op
+            # max_idx: checkpoint_2' s output op
             flag, min_idx, max_idx = program_stat.is_subgraph(
                 [checkpoints_name[start_idx]],
                 [checkpoints_name[start_idx + 1]])
             if flag:
+                # max_idx + 1 since the exact and used segment end idx is max_idx
+                min_idx = program_stat._update_segment_start(
+                    min_idx, pre_segment_end_idx)
                 segments.append([min_idx, max_idx + 1])
+
             start_idx += 1
 
     if segments != [] and segments[0][0] != 0:
@@ -772,12 +805,31 @@ def _append_backward_ops_with_checkpoints_(
     else:
         recompute_segments = segments
 
+    for i, (idx1, idx2) in enumerate(recompute_segments):
+        _logger.debug("recompute segment[{}]".format(i))
+        _logger.debug("segment start op: [{}]: [{}]".format(ops[idx1].desc.type(
+        ), ops[idx1].desc.input_arg_names()))
+        _logger.debug("segment end op: [{}]: [{}]".format(ops[
+            idx2 - 1].desc.type(), ops[idx2 - 1].desc.input_arg_names()))
+        _logger.debug("recompute segment[{}]".format(i))
+        _logger.debug("segment start op: [{}]: [{}]".format(ops[idx1].desc.type(
+        ), ops[idx1].desc.input_arg_names()))
+        _logger.debug("segment end op: [{}]: [{}]".format(ops[
+            idx2 - 1].desc.type(), ops[idx2 - 1].desc.input_arg_names()))
+
     # 2) go through all forward ops and induct all variables that will be hold in memory
     vars_should_be_hold = []
     # a. variables that are used across segments will be held in memory
     for segment in recompute_segments:
         vars_should_be_hold.extend(
             program_stat.get_out_of_subgraph_vars(segment[0], segment[1]))
+
+    cross_vars = set(vars_should_be_hold) - set(checkpoints_name)
+    _logger.debug("found [{}] vars which cross recompute segment: [{}], better checkpoints might be set to reduce those vars".format( \
+    len(cross_vars), cross_vars))
+    _logger.debug("found [{}] vars which cross recompute segment: [{}], better checkpoints might be set to reduce those vars".format( \
+    len(cross_vars), cross_vars))
+
     # b. output of seed op should be kept in memory
     vars_should_be_hold.extend(program_stat.get_reserved_vars())
     # c. input variables are checkpoints
@@ -792,8 +844,6 @@ def _append_backward_ops_with_checkpoints_(
 
     max_calculated_op_position = len(ops)
     if recompute_segments == []:
-        # if there is no recompute segment, add backward ops like
-        # _append_backward_ops_ function
         gap_ops = ops[0:max_calculated_op_position]
         for op in reversed(gap_ops):
             if op.has_attr("sub_block"):
@@ -807,7 +857,6 @@ def _append_backward_ops_with_checkpoints_(
             grad_to_var.update(op_grad_to_var)
 
     for i, segment in enumerate(recompute_segments[::-1]):
-        # add grad op for ops not in any segments
         gap_ops = ops[segment[1]:max_calculated_op_position]
         max_calculated_op_position = segment[0]
         for op in reversed(gap_ops):
@@ -851,7 +900,7 @@ def _append_backward_ops_with_checkpoints_(
         # added_descs should be in grad_op_descs because it is backward op desc
         grad_op_descs.extend(buffer_descs)
 
-        # 3.c. add backward ops of current recomputation ops
+        # 3.c. add backward ops for all ops in current segment 
         for op_desc in reversed(added_descs):
             grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
                 op_desc, cpt.to_text(no_grad_dict[block.idx]), [])
@@ -1480,9 +1529,11 @@ def append_backward(loss,
 
         # TODO: support _append_backward_ops_with_checkpoints_ in
         #  sub-block (control flow)
+        is_recompute = False
         if checkpoints != None and \
                 isinstance(checkpoints, list) and \
                 len(checkpoints) > 0:
+            is_recompute = True
             program_stat, checkpoint_names, \
                 vars_should_be_hold, \
                 recompute_segments = \
@@ -1577,7 +1628,10 @@ def append_backward(loss,
             attr_val.extend(g.op.attr(op_role_var_attr_name))
         g.op._set_attr(op_role_var_attr_name, attr_val)
 
-    return params_and_grads
+    if is_recompute:
+        return params_and_grads, checkpoint_names
+    else:
+        return params_and_grads
 
 
 def _as_list(x):

python/paddle/fluid/optimizer.py

@@ -4593,6 +4593,7 @@ class RecomputeOptimizer(Optimizer):
         self._checkpoints = None
         self._learning_rate = self._optimizer._learning_rate
         self._learning_rate_map = self._optimizer._learning_rate_map
+        self.enable_offload = False
 
     def _set_checkpoints(self, checkpoints):
         """
@@ -4608,6 +4609,10 @@ class RecomputeOptimizer(Optimizer):
             ), "_checkpoints should be a list of Variable or a list of String"
         self._checkpoints = checkpoints
 
+    # should enable offload before calling backward 
+    def _enable_offload(self):
+        self.enable_offload = True
+
     @framework.deprecate_stat_dict
     def load(self, state_dict):
         """
@@ -4696,6 +4701,358 @@ class RecomputeOptimizer(Optimizer):
 
         return self._optimizer.apply_gradients(params_grads=params_grads)
 
+    def _creat_vars(self, varname):
+        pinned_var_name = unique_name.generate(varname + "@Pinned")
+        fetched_var_name = unique_name.generate(varname + "@Fetch")
+
+        pinned_var = self._main_program.global_block().create_var(
+            name=pinned_var_name,
+            shape=self.checkpoint_shape,
+            dtype=self._main_program.global_block().var(varname).dtype,
+            persistable=False,
+            stop_gradient=True)
+
+        fetch_var = self._main_program.global_block().create_var(
+            name=fetched_var_name,
+            shape=self.checkpoint_shape,
+            dtype=self._main_program.global_block().var(varname).dtype,
+            persistable=False,
+            stop_gradient=False)
+
+        return pinned_var_name, fetched_var_name
+
+    def _append_fill_constant_ops(self, startup_program):
+        """
+        add fill_constant_ops to the end of the prog
+
+        we should fill the pinned vars before runing the main_prog
+        to instantiate their tensor hold_, which could tell us whether 
+        the host memory could hold all the checkpoints from all the 
+        GPU devices in this node. 
+        """
+        op_role = 0
+        block = startup_program.global_block()
+        fill_constant_vars = self.checkpoint_name2pinned_name.values()
+        OP_ROLE_KEY = core.op_proto_and_checker_maker.kOpRoleAttrName()
+        for varname in fill_constant_vars:
+            var = self._main_program.global_block().var(varname)
+            # NOTE (JZ-LIANG) to pre-allocate the CUDAPinned MEM
+            pinned_var = block.create_var(
+                name=varname,
+                shape=self.checkpoint_shape,
+                dtype=self._main_program.global_block().var(var.name).dtype,
+                persistable=False,
+                stop_gradient=True)
+            block.append_op(
+                type='fill_constant',
+                outputs={'Out': varname},
+                attrs={
+                    "shape": var.shape,
+                    "dtype": var.dtype,
+                    "value": 0.0,
+                    "place_type": 2,
+                    OP_ROLE_KEY: op_role,
+                })
+
+        return
+
+    def _insert_async_memcpy_op(self, insert_idx, src_varname, dst_varname,
+                                op_role, kind):
+        OP_ROLE_KEY = core.op_proto_and_checker_maker.kOpRoleAttrName()
+        self.block._insert_op_without_sync(
+            insert_idx,
+            type='memcpy',
+            inputs={'X': [self._main_program.global_block().var(src_varname)]},
+            outputs={
+                'Out': [self._main_program.global_block().var(dst_varname)]
+            },
+            attrs={"dst_place_type": int(kind),
+                   OP_ROLE_KEY: op_role})
+
+    def _insert_fetch_op(self, idx, varname):
+        assert varname in self.checkpoint_name2pinned_name, "Try to fetch {} from Pinned Memory, but it is NOT a checkpoint".format(
+            varname)
+
+        pinned_varname = self.checkpoint_name2pinned_name[varname]
+        fetch_varname = self.checkpoint_name2fetch_name[varname]
+        self._insert_async_memcpy_op(idx, pinned_varname, fetch_varname, 1, 2)
+
+    def _insert_offload_op(self, idx, varname):
+        assert varname in self.checkpoint_name2pinned_name, "Try to offload {} to Pinned Memory, but it is NOT a checkpoint".format(
+            varname)
+        pinned_varname = self.checkpoint_name2pinned_name[varname]
+        self._insert_async_memcpy_op(idx, varname, pinned_varname, 0, 3)
+
+    def _insert_sync_op(self, op_idx, checkpoint_name):
+        # single stream offload no need sync 
+        pass
+
+    def _record_fetch_op(self, idx):
+        assert len(self.un_fetch_checkpoint_names
+                   ) > 0, "Could NOT found checkpoint to fetch"
+        checkpoint_name = self.un_fetch_checkpoint_names.pop(-1)
+        logging.debug("Record fetch [{}]".format(checkpoint_name))
+        self.idx2insertions[idx] = ("fetch", checkpoint_name)
+
+        return checkpoint_name
+
+    def _record_offload_op(self, idx, checkpoint_name):
+        expected_checkpoint_name = self.un_offload_checkpoint_names.pop(0)
+        assert checkpoint_name == expected_checkpoint_name, "expected to offload [{}] but got [{}]".format(
+            expected_checkpoint_name, checkpoint_name)
+        logging.debug("Record offload [{}]".format(checkpoint_name))
+        self.idx2insertions[idx] = ("offload", checkpoint_name)
+
+    def _record_sync_op(self, idx, checkpoint_name):
+        assert checkpoint_name not in self.synced_checkpoints, "Try to sync the checkpoint [{}] twice".format(
+            checkpoint_name)
+        self.synced_checkpoints.add(checkpoint_name)
+        logging.debug("Record offload sync [{}]".format(checkpoint_name))
+        self.idx2insertions[idx] = ("sync", checkpoint_name)
+
+    def _parse_backward(self):
+
+        self.idx2insertions = {}
+        # don't offload the last checkpoints, to favor throughput        
+        self.un_fetch_checkpoint_names = self.sorted_checkpoint_names[:]
+        self.un_fetch_checkpoint_names.pop(-1)
+        need_fetch_checkpoint_names = self.un_fetch_checkpoint_names[:]
+        self.checkpoint_usage_count = {}
+        for checkpoint_name in self.un_fetch_checkpoint_names:
+            self.checkpoint_usage_count[checkpoint_name] = 0
+
+        self.bw_strart_op_idx = len(self.block.ops)
+        for idx, op in enumerate(self.block.ops):
+            if int(op.desc.attr("op_role")) == 1:
+                self.bw_strart_op_idx = idx
+                break
+
+        assert self.bw_strart_op_idx < len(
+            self.block.ops), "Could NOT found backword op in prog"
+
+        # fetch second to last checkpoint at the beginning of BW
+        fetched_checkpoint_varname = self._record_fetch_op(
+            self.bw_strart_op_idx)
+        last_last_fetch_checkpoint = None
+
+        for i, op in enumerate(self.block.ops[self.bw_strart_op_idx:]):
+            idx = self.bw_strart_op_idx + i
+            input_vars = op.desc.input_arg_names()
+
+            for input_var in input_vars:
+                if input_var in need_fetch_checkpoint_names:
+                    if input_var not in self.un_fetch_checkpoint_names:
+                        # fetch the  offloade checkpoint when the first usage of its previous one
+                        if self.checkpoint_usage_count[input_var] == 0:
+                            # TODO (JZ-LIANG) sync memcpy_stream if extra stream for memcpy
+                            second_to_last_fetch_checkpoint = fetched_checkpoint_varname
+                            # there is NO fetch ahead the first checkpoint 
+                            if input_var != self.sorted_checkpoint_names[0]:
+                                fetched_checkpoint_varname = self._record_fetch_op(
+                                    idx)
+
+                        # should check the current used checkpoint is ths last fetch one 
+                        assert second_to_last_fetch_checkpoint == input_var, "Current recompute segment should use [{}] BUT got [{}]".format(
+                            second_to_last_fetch_checkpoint, input_var)
+                        # rename
+                        self.block.ops[idx]._rename_input(
+                            input_var,
+                            self.checkpoint_name2fetch_name[input_var])
+                        self.checkpoint_usage_count[input_var] += 1
+                    else:
+                        raise ValueError(
+                            "use checkpoint [{}] before fetch in BW".format(
+                                input_var))
+
+        assert len(self.un_fetch_checkpoint_names
+                   ) == 0, "{} checkpoints have NOT been Recorded".format(
+                       self.un_fetch_checkpoint_names)
+
+    def _update_backward(self):
+        if len(self.idx2insertions) == 0:
+            return
+        total_op = len(self.block.ops)
+        for op_idx in reversed(range(self.bw_strart_op_idx, total_op)):
+            if op_idx in self.idx2insertions:
+                operation, checkpoint_name = self.idx2insertions[op_idx]
+                if operation == "fetch":
+                    self._insert_fetch_op(op_idx, checkpoint_name)
+                    logging.debug("Insert [{}] fetch op.".format(
+                        checkpoint_name))
+                    del self.idx2insertions[op_idx]
+                elif operation == "sync":
+                    self._insert_sync_op(op_idx, checkpoint_name)
+                    logging.debug("Sync [{}] fetch op.".format(checkpoint_name))
+        self.block._sync_with_cpp()
+        assert len(
+            self.idx2insertions) == 0, "{} checkpoints left un-Fecthed".format(
+                [ele[1] for ele in self.idx2insertions.values()])
+
+    def _parse_forward(self):
+
+        self.idx2insertions = {}
+        # don't offload the last checkpoints, faster, less memory saving       
+        self.un_offload_checkpoint_names = self.sorted_checkpoint_names[:]
+        last_checkpoint = self.un_offload_checkpoint_names.pop(-1)
+        need_offload_checkpoint_names = self.un_offload_checkpoint_names[:]
+        self.checkpoint_usage_count_and_idx = {}
+        for checkpoint_name in self.un_offload_checkpoint_names:
+            self.checkpoint_usage_count_and_idx[checkpoint_name] = {
+                'count': 0,
+                'idx': -1
+            }
+        self.synced_checkpoints = set()
+        self.fw_strart_op_idx = len(self.block.ops)
+        for idx, op in enumerate(self.block.ops):
+            if int(op.desc.attr("op_role")) == 0:
+                self.fw_strart_op_idx = idx
+                break
+
+        assert self.fw_strart_op_idx < len(
+            self.block.ops), "Could NOT found Forward op in prog"
+        last_offload_checkpoint = None
+
+        for i, op in enumerate(self.block.ops[self.fw_strart_op_idx:
+                                              self.bw_strart_op_idx]):
+
+            idx = self.fw_strart_op_idx + i
+            output_vars = op.desc.output_arg_names()
+            input_vars = op.desc.input_arg_names()
+
+            for output_var in output_vars:
+                if output_var in need_offload_checkpoint_names:
+                    assert len(
+                        output_vars
+                    ) == 1, "chekpoint should be the only Output of a certain op, but [{}] is from [{}]".format(
+                        output_var, op)
+
+                    if output_var in self.un_offload_checkpoint_names:
+                        # insert sync op if last checkpoint has not been sync
+                        if last_offload_checkpoint != None:
+                            if self.checkpoint_usage_count_and_idx[
+                                    last_offload_checkpoint]['count'] == 0:
+                                self._record_sync_op(idx,
+                                                     last_offload_checkpoint)
+                            else:
+                                last_usage_idx = self.checkpoint_usage_count_and_idx[
+                                    last_offload_checkpoint]['idx']
+                                assert last_usage_idx > 0, "last_usage_idx of checkpoint [{}] should large than 0".format(
+                                    last_offload_checkpoint)
+                                self._record_sync_op(last_usage_idx + 1,
+                                                     last_offload_checkpoint)
+                        # insert offload op after the checkpoint's generation op
+                        self._record_offload_op(idx + 1, output_var)
+                        last_offload_checkpoint = output_var
+                    else:
+                        raise ValueError(
+                            "There should be just ONE op that output checkpoint [{}]".
+                            format(output_var))
+                # need to sync the last need to offload checkpoint before the last checkpoint as output op
+                if output_var == last_checkpoint:
+                    assert len(
+                        output_vars
+                    ) == 1, "chekpoint should be the only Output of a certain op, but [{}] is from [{}]".format(
+                        output_var, op)
+                    assert last_offload_checkpoint == self.sorted_checkpoint_names[
+                        -2], "the last offload chekpoint before [{}] is suppose to be [{}], but got [{}]".format(
+                            last_checkpoint, self.sorted_checkpoint_names[-2],
+                            last_offload_checkpoint)
+                    # sync if last checkpoint has not been sync
+                    if self.checkpoint_usage_count_and_idx[
+                            last_offload_checkpoint]['idx'] == 0:
+                        self._record_sync_op(idx, last_offload_checkpoint)
+                    else:
+                        last_usage_idx = self.checkpoint_usage_count_and_idx[
+                            last_offload_checkpoint]['idx']
+                        assert last_usage_idx > 0, "last_usage_idx of checkpoint [{}] should large than 0".format(
+                            last_offload_checkpoint)
+                        self._record_sync_op(last_usage_idx + 1,
+                                             last_offload_checkpoint)
+            # record checkpoint usage  
+            for input_var in input_vars:
+                if input_var in need_offload_checkpoint_names:
+                    assert input_var not in self.synced_checkpoints, "checkpoint [{}] used after sync".format(
+                        input_var)
+                    self.checkpoint_usage_count_and_idx[input_var]['count'] += 1
+                    self.checkpoint_usage_count_and_idx[input_var]['idx'] = idx
+
+        assert len(self.un_offload_checkpoint_names
+                   ) == 0, "{} checkpoints have NOT been Recorded".format(
+                       self.un_fetch_checkpoint_names)
+        assert len(self.synced_checkpoints) == len(
+            need_offload_checkpoint_names
+        ), "{} checkpoints have NOT been Recorded".format(
+            set(need_offload_checkpoint_names) - set(self.synced_checkpoints))
+
+    def _update_forward(self):
+        if len(self.idx2insertions) == 0:
+            return
+        for op_idx in reversed(
+                range(self.fw_strart_op_idx, self.bw_strart_op_idx)):
+            if op_idx in self.idx2insertions:
+                operation, checkpoint_name = self.idx2insertions[op_idx]
+                if operation == "offload":
+                    self._insert_offload_op(op_idx, checkpoint_name)
+                    logging.debug("Insert [{}] offload op.".format(
+                        checkpoint_name))
+                    del self.idx2insertions[op_idx]
+                elif operation == "sync":
+                    self._insert_sync_op(op_idx, checkpoint_name)
+                    logging.debug("Insert [{}] offload_sync op.".format(
+                        checkpoint_name))
+                    del self.idx2insertions[op_idx]
+
+        self.block._sync_with_cpp()
+        assert len(self.idx2insertions
+                   ) == 0, "{} checkpoints left un-Offloaded".format(
+                       [ele[1] for ele in self.idx2insertions.values()])
+
+    def _check_offload_fetch(self):
+        # TODO(JZ-LIANG) the single stream offload need no sync
+        pass
+
+    def _offload(self, loss, startup_program=None):
+        """
+        core steps for recompute offload
+        1. create pinned vars and temp vars 
+        2. parse & update Forward pass: offload, sync
+        3. parse & update Backward pass: rename, fetch, sync
+        4. verify the correctness
+        """
+        self._main_program = loss.block.program
+        self.block = loss.block
+        if startup_program == None:
+            startup_program = fluid.default_startup_program()
+
+        with program_guard(self._main_program, startup_program):
+            assert len(self.checkpoint_shape) > 0, (
+                "checkpoints shape {} should be an non empty list like: [12, 512, 1024]".
+                format(self.checkpoint_shape))
+            assert all([ele > 0 for ele in self.checkpoint_shape]), (
+                "all ele in checkpoints shape {} should be a determined integer larger than 0".
+                format(self.checkpoint_shape))
+            self.checkpoint_name2pinned_name = dict()
+            self.checkpoint_name2fetch_name = dict()
+            for checkpoint_varname in self.sorted_checkpoint_names:
+                pinned_var_name, fetch_var_name = self._creat_vars(
+                    checkpoint_varname)
+                self.checkpoint_name2pinned_name[
+                    checkpoint_varname] = pinned_var_name
+                self.checkpoint_name2fetch_name[
+                    checkpoint_varname] = fetch_var_name
+            self._append_fill_constant_ops(startup_program)
+            # TODO (JZ-LIANG) to provide two offload stragtegy in future
+            # step 2. parse & update FW: rename, offload, sync
+            self._parse_backward()
+            self._update_backward()
+            # step 3. parse & update BW: rename, offload, sync
+            self._parse_forward()
+            self._update_forward()
+            # step 4. verify the correctness
+            self._check_offload_fetch()
+
+        return
+
     def backward(self,
                  loss,
                  startup_program=None,
@@ -4760,8 +5117,24 @@ class RecomputeOptimizer(Optimizer):
                 else:
                     checkpoint_vars.append(loss.block.var(ckpt))
 
-            params_grads = append_backward(
-                loss, parameter_list, no_grad_set, checkpoints=checkpoint_vars)
+            # allow return to non-recompute when checkpoints is empty
+            if len(checkpoint_vars) > 0:
+                params_grads, sorted_checkpoint_names = append_backward(
+                    loss,
+                    parameter_list,
+                    no_grad_set,
+                    checkpoints=checkpoint_vars)
+            else:
+                params_grads = append_backward(
+                    loss,
+                    parameter_list,
+                    no_grad_set,
+                    checkpoints=checkpoint_vars)
+
+        if self.enable_offload:
+            self.sorted_checkpoint_names = sorted_checkpoint_names
+            self._offload(loss, startup_program=startup_program)
+
         return params_grads
 
     def apply_optimize(self, loss, startup_program, params_grads):

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -84,6 +84,7 @@ if(NOT WITH_GPU OR WIN32)
     LIST(REMOVE_ITEM TEST_OPS test_collective_allreduce_api)
     LIST(REMOVE_ITEM TEST_OPS test_collective_broadcast_api)
     LIST(REMOVE_ITEM TEST_OPS test_collective_allgather_api)
+    LIST(REMOVE_ITEM TEST_OPS test_memcpy_op)
 endif()
 
 if(WIN32)

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

@@ -132,5 +132,12 @@ class TestFleetMetaOptimizer(unittest.TestCase):
         elif name == "sharding":
             strategy.sharding = True
             strategy.sharding_configs = {"fuse_broadcast_MB": 0.2}
+        elif name == "recompute-offload":
+            strategy.recompute = True
+            strategy.recompute_configs = {
+                "checkpoints": ["fc_0.tmp_2", "fc_1.tmp_2"],
+                "enable_offload": True,
+                "checkpoint_shape": [256]
+            }
         else:
             raise NotImplementedError()

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

@@ -153,6 +153,20 @@ class TestFleetRecomputeMetaOptimizer(TestFleetMetaOptimizer):
         self.assertIn('subprog', ''.join(outs))
         self.assertIn('lamb', ops)
 
+    def test_recompute_offload(self):
+        train_prog, startup_prog = fluid.Program(), fluid.Program()
+        avg_cost, strategy = self.net(train_prog, startup_prog)
+        self.set_strategy(strategy, 'recompute-offload')
+        self.optimizer(avg_cost, strategy, train_prog, startup_prog)
+        ops = [op.type for op in avg_cost.block.ops]
+        outs = [
+            op.output('Out')[0] for op in avg_cost.block.ops
+            if op.type == 'memcpy'
+        ]
+        self.assertIn('memcpy', ops)
+        self.assertIn('@Pinned', ''.join(outs))
+        self.assertIn('@Fetch', ''.join(outs))
+
 
 if __name__ == "__main__":
     unittest.main()

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

@@ -170,19 +170,19 @@ class TestFleetShardingMetaOptimizer(TestFleetMetaOptimizer):
 
         self.assertEqual(ops, [
             'cast', 'cast', 'cast', 'fill_constant', 'fill_constant',
-            'fill_constant', 'fill_constant', 'fill_constant',
+            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
             'c_sync_calc_stream', 'c_broadcast', 'c_broadcast', 'c_broadcast',
             'c_broadcast', 'c_broadcast', 'c_broadcast', 'c_broadcast',
-            'c_broadcast', 'c_broadcast', 'c_sync_comm_stream', 'cast', 'cast',
-            'mul', 'cast', 'elementwise_add', 'cast', 'tanh', 'cast', 'mul',
-            'elementwise_add', 'cast', 'tanh', 'cast', 'mul', 'elementwise_add',
-            'softmax', 'cast', 'cross_entropy2', 'mean', 'elementwise_mul',
-            'fill_constant', 'scale', 'elementwise_mul_grad', 'mean_grad',
-            'cross_entropy_grad2', 'cast', 'softmax_grad',
-            'elementwise_add_grad', 'mul_grad', 'cast', 'cast', 'mul', 'cast',
-            'elementwise_add', 'cast', 'tanh_grad', 'cast',
-            'elementwise_add_grad', 'mul_grad', 'cast', 'cast', 'mul', 'cast',
-            'elementwise_add', 'cast', 'tanh_grad', 'cast',
+            'c_broadcast', 'c_broadcast', 'c_broadcast', 'c_sync_comm_stream',
+            'cast', 'cast', 'mul', 'cast', 'elementwise_add', 'cast', 'tanh',
+            'cast', 'cast', 'mul', 'elementwise_add', 'cast', 'tanh', 'cast',
+            'mul', 'elementwise_add', 'softmax', 'cast', 'cross_entropy2',
+            'mean', 'elementwise_mul', 'fill_constant', 'scale',
+            'elementwise_mul_grad', 'mean_grad', 'cross_entropy_grad2', 'cast',
+            'softmax_grad', 'elementwise_add_grad', 'mul_grad', 'cast', 'cast',
+            'cast', 'mul', 'cast', 'elementwise_add', 'cast', 'tanh_grad',
+            'cast', 'elementwise_add_grad', 'mul_grad', 'cast', 'cast', 'mul',
+            'cast', 'elementwise_add', 'cast', 'tanh_grad', 'cast',
             'elementwise_add_grad', 'mul_grad', 'c_sync_calc_stream',
             'c_allreduce_sum', 'c_allreduce_sum', 'c_allreduce_sum',
             'c_allreduce_sum', 'c_allreduce_sum', 'c_allreduce_sum',

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import op_test
+import numpy as np
+import unittest
+import paddle
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+import paddle.fluid as fluid
+from paddle.fluid import compiler, Program, program_guard
+from paddle.fluid.backward import append_backward
+
+
+class TestMemcpy_FillConstant(unittest.TestCase):
+    def get_prog(self):
+        paddle.enable_static()
+        main_program = Program()
+        with program_guard(main_program):
+            pinned_var_name = "tensor@Pinned"
+            gpu_var_name = "tensor@GPU"
+            pinned_var = main_program.global_block().create_var(
+                name=pinned_var_name,
+                shape=[10, 10],
+                dtype='float32',
+                persistable=False,
+                stop_gradient=True)
+            gpu_var = main_program.global_block().create_var(
+                name=gpu_var_name,
+                shape=[10, 10],
+                dtype='float32',
+                persistable=False,
+                stop_gradient=True)
+            main_program.global_block().append_op(
+                type="fill_constant",
+                outputs={"Out": gpu_var_name},
+                attrs={
+                    "shape": [10, 10],
+                    "dtype": gpu_var.dtype,
+                    "value": 1.0,
+                    "place_type": 1
+                })
+            main_program.global_block().append_op(
+                type="fill_constant",
+                outputs={"Out": pinned_var_name},
+                attrs={
+                    "shape": [10, 10],
+                    "dtype": gpu_var.dtype,
+                    "value": 0.0,
+                    "place_type": 2
+                })
+        return main_program, gpu_var, pinned_var
+
+    def test_gpu_cpoy_to_pinned(self):
+        main_program, gpu_var, pinned_var = self.get_prog()
+        main_program.global_block().append_op(
+            type='memcpy',
+            inputs={'X': gpu_var},
+            outputs={'Out': pinned_var},
+            attrs={'dst_place_type': 3})
+        place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        gpu_, pinned_ = exe.run(main_program,
+                                feed={},
+                                fetch_list=[gpu_var.name, pinned_var.name])
+        self.assertTrue(np.allclose(gpu_, pinned_))
+        self.assertTrue(np.allclose(pinned_, np.ones((10, 10))))
+
+    def test_pinned_cpoy_gpu(self):
+        main_program, gpu_var, pinned_var = self.get_prog()
+        main_program.global_block().append_op(
+            type='memcpy',
+            inputs={'X': pinned_var},
+            outputs={'Out': gpu_var},
+            attrs={'dst_place_type': 2})
+        place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        gpu_, pinned_ = exe.run(main_program,
+                                feed={},
+                                fetch_list=[gpu_var.name, pinned_var.name])
+        self.assertTrue(np.allclose(gpu_, pinned_))
+        self.assertTrue(np.allclose(gpu_, np.zeros((10, 10))))
+
+
+class TestMemcpyOPError(unittest.TestCase):
+    def get_prog(self):
+        paddle.enable_static()
+        main_program = Program()
+        with program_guard(main_program):
+            pinned_var = main_program.global_block().create_var(
+                name="tensor@Pinned_0",
+                shape=[10, 10],
+                dtype='float32',
+                persistable=False,
+                stop_gradient=True)
+            main_program.global_block().append_op(
+                type="fill_constant",
+                outputs={"Out": "tensor@Pinned_0"},
+                attrs={
+                    "shape": [10, 10],
+                    "dtype": pinned_var.dtype,
+                    "value": 0.0,
+                    "place_type": 2
+                })
+        return main_program, pinned_var
+
+    def test_SELECTED_ROWS(self):
+        main_program, pinned_var = self.get_prog()
+        selected_row_var = main_program.global_block().create_var( \
+            name="selected_row_0", dtype="float32", persistable=False, \
+            type=fluid.core.VarDesc.VarType.SELECTED_ROWS, stop_gradient=True)
+        main_program.global_block().append_op(
+            type="fill_constant",
+            outputs={"Out": selected_row_var},
+            attrs={
+                "shape": selected_row_var.shape,
+                "dtype": selected_row_var.dtype,
+                "value": 1.0,
+                "place_type": 1
+            })
+        main_program.global_block().append_op(
+            type='memcpy',
+            inputs={'X': selected_row_var},
+            outputs={'Out': pinned_var},
+            attrs={'dst_place_type': 3})
+        with self.assertRaises(NotImplementedError):
+            place = fluid.CUDAPlace(0)
+            exe = fluid.Executor(place)
+            selected_row_var_, pinned_ = exe.run(
+                main_program,
+                feed={},
+                fetch_list=[selected_row_var.name, pinned_var.name])
+
+    def test_OTHER_PLACE_NotImplementedError(self):
+        main_program, pinned_var = self.get_prog()
+        lod_tensor_var = main_program.global_block().create_var( \
+            name="lod_tensor_0", dtype="float32", persistable=False, stop_gradient=True)
+        main_program.global_block().append_op(
+            type="fill_constant",
+            outputs={"Out": lod_tensor_var},
+            attrs={
+                "shape": lod_tensor_var.shape,
+                "dtype": lod_tensor_var.dtype,
+                "value": 1.0,
+                "place_type": 0
+            })
+        main_program.global_block().append_op(
+            type='memcpy',
+            inputs={'X': pinned_var},
+            outputs={'Out': lod_tensor_var},
+            attrs={'dst_place_type': 0, })
+        with self.assertRaises(NotImplementedError):
+            place = fluid.CUDAPlace(0)
+            exe = fluid.Executor(place)
+            lod_tensor_var_, pinned_ = exe.run(
+                main_program,
+                feed={},
+                fetch_list=[lod_tensor_var.name, pinned_var.name])
+
+
+if __name__ == '__main__':
+    paddle.enable_static()
+    unittest.main()