Merge pull request #12595 from reyoung/fix_scale_loss_with_memopt

Fix bug when memopt optimize loss.grad and use ParallelExecutor

Author: reyoung

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -275,7 +275,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
       if (strategy_.gradient_scale_ !=
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         // TODO(paddle-dev): Why is there no input for this op_handle?
-        CreateScaleLossGradOp(&result);
+        auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
+        CreateScaleLossGradOp(&result, loss_grad_name);
       }
       // This assumes the backward generating code will ensure IsScaleLossOp
       // is true only for the op that scale the final scalar loss.
@@ -535,7 +536,8 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
   return got == sharded_var_device.end() ? -1 : got->second;
 }
 
-void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
+void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
+    ir::Graph *result, const std::string &loss_grad_name) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle
 #ifdef PADDLE_WITH_CUDA
@@ -558,10 +560,10 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
     // loss->pending_ops_.emplace_back(op_handle);
     // op_handle->inputs_.emplace_back(loss);
 
-    CreateOpOutput(result, op_handle,
-                   result->CreateEmptyNode(GradVarName(loss_var_name_),
-                                           ir::Node::Type::kVariable),
-                   places_[i], i);
+    CreateOpOutput(
+        result, op_handle,
+        result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable),
+        places_[i], i);
   }
 }
 

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -75,7 +75,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
   void CreateComputationalOps(ir::Graph *result, ir::Node *node,
                               size_t num_places) const;
 
-  void CreateScaleLossGradOp(ir::Graph *result) const;
+  void CreateScaleLossGradOp(ir::Graph *result,
+                             const std::string &loss_grad_name) const;
+
   VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
                             int dst_dev_id) const;
   void CreateComputationalOp(ir::Graph *result, ir::Node *node,

paddle/fluid/framework/operator.cc

@@ -778,6 +778,7 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
     const ExecutionContext& ctx) const {
   auto& scope = ctx.scope();
   int data_type = -1;
+  std::string last_input_name;
   for (auto& input : this->inputs_) {
     for (auto& ipt_name : input.second) {
       auto* var = scope.FindVar(ipt_name);
@@ -794,9 +795,10 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
           int tmp = static_cast<int>(ToDataType(t->type()));
           PADDLE_ENFORCE(
               tmp == data_type || data_type == -1,
-              "DataType of Paddle Op %s must be the same. Get %d != %d", Type(),
-              data_type, tmp);
+              "DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)",
+              Type(), last_input_name, data_type, ipt_name, tmp);
           data_type = tmp;
+          last_input_name = ipt_name;
         }
       }
     }