add some debug infor

paddle/fluid/framework/details/async_ssa_graph_executor.cc

@@ -68,18 +68,21 @@ FeedFetchList AsyncSSAGraphExecutor::Run(
 
     if (pool_) {
       run_futures.emplace_back(pool_->enqueue(std::move(call)));
-      for (auto &f : run_futures) {
-        if (exception_holder_.IsCaught()) {
-          f.wait();
-        } else {
-          fetch_data.emplace_back(std::move(f.get()));
-        }
-      }
     } else {
       fetch_data.emplace_back(std::move(call()));
     }
   }
 
+  if (pool_) {
+    for (auto &f : run_futures) {
+      if (exception_holder_.IsCaught()) {
+        f.wait();
+      } else {
+        fetch_data.emplace_back(std::move(f.get()));
+      }
+    }
+  }
+
   if (exception_holder_.IsCaught()) {
     exception_holder_.ReThrow();
   }

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -249,6 +249,8 @@ void MultiDevSSAGraphBuilderBase::InsertScaleLossGradOp(
       break;
   }
 
+  VLOG(3) << "loss_scale: " << loss_scale;
+
   if (loss_scale) {
     // TODO(paddle-dev): Why is there no input for this op_handle?
     auto loss_grad_name = node->Op()->OutputArgumentNames()[0];

paddle/fluid/framework/scope.cc

@@ -271,16 +271,14 @@ std::string GenParentScopeTreeDebugInfo(Scope* leaf) {
   while (current_scope != nullptr) {
     scopes.push_back(current_scope);
     current_scope = current_scope->parent();
-    // end of a level
-    os << "\n------------------------------------------\n";
   }
 
-  os << "\nDetails:\n\n";
+  os << "\n--------------GenParentScopeTreeDebugInfo--------------\n";
 
-  for (auto* q : scopes) {
-    os << "====\n";
-    os << q << ":\n";
-    for (auto& var : q->LocalVarNames()) {
+  for (int i = scopes.size() - 1; i >= 0; --i) {
+    os << "=======level [" << i << "]=======\n";
+    os << scopes[i] << ":\n";
+    for (auto& var : scopes[i]->LocalVarNames()) {
       os << "  - " << var << "\n";
     }
   }