diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc
index 53ab8eb775442cc641c039c295c63406e96aa122..7a1b40c0b60a788b1f0a70e688f8fcbe427ad076 100644
--- a/paddle/fluid/framework/details/computation_op_handle.cc
+++ b/paddle/fluid/framework/details/computation_op_handle.cc
@@ -24,10 +24,10 @@ ComputationOpHandle::ComputationOpHandle(const OpDesc &op_desc, Scope *scope,
       place_(place) {}
 
 void ComputationOpHandle::RunImpl() {
-  auto *cur_ctx = dev_ctx_[place_];
+  auto *cur_ctx = dev_ctxes_[place_];
   for (auto *in : inputs_) {
     bool need_wait =
-        in->generated_op_ && in->generated_op_->dev_ctx_[place_] != cur_ctx;
+        in->generated_op_ && in->generated_op_->dev_ctxes_[place_] != cur_ctx;
     if (need_wait) {
       in->generated_op_->Wait(cur_ctx);
     }
diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc
index 4fc05b324897e9066576621a0264edd94e7dc497..9180903b864d03e59f55f41410b2240fa4199496 100644
--- a/paddle/fluid/framework/details/fetch_op_handle.cc
+++ b/paddle/fluid/framework/details/fetch_op_handle.cc
@@ -60,8 +60,8 @@ void FetchOpHandle::RunImpl() {
     auto &t = scope->FindVar(var_name)->Get<framework::LoDTensor>();
     if (platform::is_gpu_place(var->place_)) {
 #ifdef PADDLE_WITH_CUDA
-      TensorCopy(t, cpu, *dev_ctx_[t.place()], &tensors_[i]);
-      dev_ctx_[t.place()]->Wait();
+      TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i]);
+      dev_ctxes_[t.place()]->Wait();
 #endif
     } else {
       tensors_[i].ShareDataWith(t);
diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc
index 67987760764cd2c1fa0f11c290bee29e21120487..a1b913a863cc1853ea3a786d22e6e8baa8c98a02 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc
@@ -74,7 +74,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
 
       result.ops_.emplace_back(new ComputationOpHandle(*op, s, p));
       auto *op_handle = result.ops_.back().get();
-      op_handle->dev_ctx_[p] = const_cast<platform::DeviceContext *>(
+      op_handle->dev_ctxes_[p] = const_cast<platform::DeviceContext *>(
           platform::DeviceContextPool::Instance().Get(p));
 
       auto var_names = op->InputArgumentNames();
diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
index f77a4b55a172d740514bd97db2b11fdb371d6af6..5ddf331cfca39a4e81a42d9ff8efd5af7bcf6829 100644
--- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
+++ b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
@@ -23,7 +23,7 @@ NCCLAllReduceOpHandle::NCCLAllReduceOpHandle(
     const platform::NCCLContextMap &ctxs)
     : local_scopes_(local_scopes), places_(places), nccl_ctxs_(ctxs) {
   for (auto &p : places_) {
-    this->dev_ctx_[p] = nccl_ctxs_.DevCtx(p);
+    this->dev_ctxes_[p] = nccl_ctxs_.DevCtx(p);
   }
 }
 
@@ -34,7 +34,7 @@ void NCCLAllReduceOpHandle::RunImpl() {
     // Wait input done
     for (auto *in : inputs_) {
       auto &p = static_cast<VarHandle *>(in)->place_;
-      in->generated_op_->Wait(dev_ctx_[p]);
+      in->generated_op_->Wait(dev_ctxes_[p]);
     }
 
     auto &var_name = static_cast<VarHandle *>(this->inputs_[0])->name_;
diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc
index 63affb705424f847d531fc8c2f20d132c92784d8..e4194a7442f677ec8970dbc387bb01ebbbf579f1 100644
--- a/paddle/fluid/framework/details/op_handle_base.cc
+++ b/paddle/fluid/framework/details/op_handle_base.cc
@@ -42,7 +42,7 @@ OpHandleBase::~OpHandleBase() {
 void OpHandleBase::Run(bool use_event) {
 #ifdef PADDLE_WITH_CUDA
   if (events_.empty() && use_event) {
-    for (auto &p : dev_ctx_) {
+    for (auto &p : dev_ctxes_) {
       int dev_id = boost::get<platform::CUDAPlace>(p.first).device;
       PADDLE_ENFORCE(cudaSetDevice(dev_id));
       PADDLE_ENFORCE(
@@ -57,7 +57,7 @@ void OpHandleBase::Run(bool use_event) {
 
 #ifdef PADDLE_WITH_CUDA
   if (use_event) {
-    for (auto &p : dev_ctx_) {
+    for (auto &p : dev_ctxes_) {
       int dev_id = boost::get<platform::CUDAPlace>(p.first).device;
       auto stream =
           static_cast<platform::CUDADeviceContext *>(p.second)->stream();
@@ -70,7 +70,7 @@ void OpHandleBase::Run(bool use_event) {
 void OpHandleBase::Wait(platform::DeviceContext *waited_dev) {
 #ifdef PADDLE_WITH_CUDA
   if (platform::is_cpu_place(waited_dev->GetPlace()) || events_.empty()) {
-    for (auto &dev_ctx : dev_ctx_) {
+    for (auto &dev_ctx : dev_ctxes_) {
       dev_ctx.second->Wait();
     }
   } else {
@@ -81,7 +81,7 @@ void OpHandleBase::Wait(platform::DeviceContext *waited_dev) {
     }
   }
 #else
-  for (auto &dev_ctx : dev_ctx_) {
+  for (auto &dev_ctx : dev_ctxes_) {
     dev_ctx.second->Wait();
   }
 #endif
diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h
index 78f566c0356895f8a1828994b02f2e1a5f71755c..71672fd24c65ee654fb9f703ea5808c31ee8fbb0 100644
--- a/paddle/fluid/framework/details/op_handle_base.h
+++ b/paddle/fluid/framework/details/op_handle_base.h
@@ -31,7 +31,7 @@ class OpHandleBase {
   std::vector<VarHandleBase *> outputs_;
   std::unordered_map<platform::Place, platform::DeviceContext *,
                      platform::PlaceHash>
-      dev_ctx_;
+      dev_ctxes_;
 
 #ifdef PADDLE_WITH_CUDA
   std::unordered_map<int, cudaEvent_t> events_;
diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
index a6a67c9b1452392edbda60ebcdd1b4fb4311a37b..0a6f6129b812ca84db7573957b1ee0a32c1ef5c4 100644
--- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
+++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
@@ -21,7 +21,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(size_t num_dev, Scope *scope,
                                              platform::Place place,
                                              platform::DeviceContext *dev_ctx)
     : coeff_(static_cast<float>(1.0 / num_dev)), scope_(scope), place_(place) {
-  dev_ctx_[place_] = dev_ctx;
+  dev_ctxes_[place_] = dev_ctx;
 }
 
 ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
@@ -38,7 +38,7 @@ void ScaleLossGradOpHandle::RunImpl() {
   } else {
 #ifdef PADDLE_WITH_CUDA
     auto stream =
-        static_cast<platform::CUDADeviceContext *>(this->dev_ctx_[place_])
+        static_cast<platform::CUDADeviceContext *>(this->dev_ctxes_[place_])
             ->stream();
     memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
                  platform::CPUPlace(), &coeff_, sizeof(float), stream);
diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
index fc8403155625fdf21c467aa189c65b14d3487b41..105e21cab600b642aafc2eb3c619a801fb4c40d7 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
@@ -96,7 +96,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
 
     // FIXME: Use new device context
     for (auto &p : places_) {
-      op->dev_ctx_[p] = fetch_ctxs_.Get(p);
+      op->dev_ctxes_[p] = fetch_ctxs_.Get(p);
     }
 
     for (auto *var : vars) {