Merge pull request #11621 from panyx0718/bug_fx

Merge pull request #11608 from panyx0718/doc

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -200,6 +200,10 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         CreateScaleLossGradOp(&result);
       }
+      // This assumes the backward generating code will ensure IsScaleLossOp
+      // is true only for the op that scale the final scalar loss.
+      // It also assumes backward op will always follow the forward op in
+      // the block.
       is_forwarding = false;
     } else {
       int op_dev_id = GetOpDeviceID(*op);
@@ -244,6 +248,9 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
                     InsertAllReduceOp(&result, g_name);
                   }
                   break;
+                default:
+                  LOG(FATAL) << "Unknown reduce strategy ";
+                  break;
               }
             }
           } catch (boost::bad_get e) {
@@ -262,7 +269,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   }
   /*
     Dependency graph has been constructed. However, there are still data
-    harzaeds need to be handled.
+    hazards need to be handled.
    */
   PolishGraphToSupportDataHazards(&result);
 
@@ -447,6 +454,8 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
   return var;
 }
 
+// Find the first occurence of `prev_op_name` and make current `op` depend
+// on it.
 void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op,
                                         const std::string &prev_op_name) const {
   for (auto &prev_op : result->ops_) {
@@ -490,6 +499,7 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result,
   }
 }
 
+// Create RPC related op handles that connects its in ops and out ops.
 void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result,
                                           const OpDesc &op) const {
   int op_dev_id = -1;

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -96,6 +96,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
 
     if (timeout) {
+      std::lock_guard<std::mutex> l(exception_mu_);
       if (exception_) {
         auto exp = *exception_;
         exception_.reset();
@@ -199,6 +200,7 @@ void ThreadedSSAGraphExecutor::RunOp(
       ready_var_q->Extend(op->Outputs());
       VLOG(10) << op << " " << op->Name() << "Signal posted";
     } catch (platform::EnforceNotMet ex) {
+      std::lock_guard<std::mutex> l(exception_mu_);
       exception_.reset(new platform::EnforceNotMet(ex));
     } catch (...) {
       LOG(FATAL) << "Unknown exception catched";

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -56,6 +56,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
+  std::mutex exception_mu_;
   std::unique_ptr<platform::EnforceNotMet> exception_;
   std::atomic<int> running_ops_;