Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_multiGPU_dist_train

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -34,7 +34,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes,
+    const std::vector<Scope *> &local_scopes, bool skip_scale_loss,
     platform::NCCLContextMap *nccl_ctxs)
     : loss_var_name_(loss_var_name),
       places_(places),
@@ -45,7 +45,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes)
+    const std::vector<Scope *> &local_scopes, bool skip_scale_loss)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes) {
@@ -53,6 +53,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
   }
+  skip_scale_loss_ = skip_scale_loss;
 }
 
 void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
@@ -133,7 +134,9 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
     } else if (IsDistTrainOp(*op, send_op)) {
       CreateComputationalOps(&result, *op, 1);
     } else if (IsScaleLossOp(*op)) {
-      CreateScaleLossGradOp(&result);
+      if (!skip_scale_loss_) {
+        CreateScaleLossGradOp(&result);
+      }
       is_forwarding = false;
     } else {
       CreateComputationalOps(&result, *op, places_.size());

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -34,12 +34,14 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
+                          bool skip_scale_loss,
                           platform::NCCLContextMap *nccl_ctxs);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
-                          const std::vector<Scope *> &local_scopes);
+                          const std::vector<Scope *> &local_scopes,
+                          bool skip_scale_loss);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
@@ -57,6 +59,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
+  bool skip_scale_loss_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
 

paddle/fluid/framework/details/send_op_handle.cc

@@ -34,7 +34,10 @@ void SendOpHandle::RunImpl() {
     }
     in->generated_op_->Wait(dev_ctxes_[p]);
   }
-  this->RunAndRecordEvent([&] { op_->Run(*local_scope_, place_); });
+  auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead
+  // lock.
+  op_->Run(*tmp_scope, place_);
 }
 
 std::string SendOpHandle::Name() const { return "send"; }

paddle/fluid/framework/parallel_executor.cc

@@ -57,7 +57,8 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &params,
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
-    Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay)
+    Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
+    bool customize_scale_loss)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
 
@@ -90,12 +91,13 @@ ParallelExecutor::ParallelExecutor(
 // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
 // ncclOp
 #ifdef PADDLE_WITH_CUDA
-  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
+  details::MultiDevSSAGraphBuilder builder(
-                                           params, member_->local_scopes_,
+      member_->places_, loss_var_name, params, member_->local_scopes_,
-                                           member_->nccl_ctxs_.get());
+      customize_scale_loss, member_->nccl_ctxs_.get());
 #else
   details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
-                                           params, member_->local_scopes_);
+                                           params, member_->local_scopes_,
+                                           customize_scale_loss);
 #endif
   auto graph = builder.Build(main_program);
 

paddle/fluid/framework/parallel_executor.h

@@ -40,7 +40,7 @@ class ParallelExecutor {
                             const ProgramDesc& main_program,
                             const std::string& loss_var_name, Scope* scope,
                             const std::vector<Scope*>& local_scopes,
-                            bool allow_op_delay);
+                            bool allow_op_delay, bool customize_scale_loss);
 
   ~ParallelExecutor();
 

paddle/fluid/pybind/pybind.cc

@@ -502,11 +502,11 @@ All parameter, weight, gradient are variables in Paddle.
               const std::unordered_set<std::string> &bcast_vars,
               const ProgramDesc &main_program, const std::string &loss_var_name,
               Scope *scope, std::vector<Scope *> &local_scopes,
-              bool allow_op_delay) {
+              bool allow_op_delay, bool customize_loss_grad) {
-             new (&self)
+             new (&self) ParallelExecutor(num_threads, use_event, places,
-                 ParallelExecutor(num_threads, use_event, places, params,
+                                          params, bcast_vars, main_program,
-                                  bcast_vars, main_program, loss_var_name,
+                                          loss_var_name, scope, local_scopes,
-                                  scope, local_scopes, allow_op_delay);
+                                          allow_op_delay, customize_loss_grad);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
       // NOTE: even we return a vec<Scope*>* to Python use reference policy.

python/paddle/fluid/parallel_executor.py

@@ -29,7 +29,8 @@ class ParallelExecutor(object):
                  main_program=None,
                  num_threads=None,
                  allow_op_delay=False,
-                 share_vars_from=None):
+                 share_vars_from=None,
+                 customize_loss_grad=False):
         """
         ParallelExecutor can run program in parallel.
 
@@ -122,7 +123,8 @@ class ParallelExecutor(object):
             loss_name if loss_name else '',
             scope,
             local_scopes,
-            allow_op_delay)
+            allow_op_delay,
+            customize_loss_grad)
         self.scope = scope
 
     def run(self, fetch_list, feed=None, feed_dict=None):