Merge pull request #8659 from Yancey1989/fix_dist_bug

Registry var type infer in split_selected_rows op

paddle/fluid/operators/split_selected_rows_op.cc

@@ -59,6 +59,16 @@ class SplitSelectedRowsOp : public framework::OperatorWithKernel {
   }
 };
 
+class SplitSelectedRowsOpInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    for (auto &out_var : op_desc.Output("Out")) {
+      block->Var(out_var)->SetType(framework::proto::VarType::SELECTED_ROWS);
+    }
+  }
+};
+
 class SplitSelectedRowsGradMaker : public framework::SingleGradOpDescMaker {
  public:
   using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
@@ -80,7 +90,8 @@ class SplitSelectedRowsGradMaker : public framework::SingleGradOpDescMaker {
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(split_selected_rows, ops::SplitSelectedRowsOp,
                   ops::SplitSelectedRowsOpMaker,
-                  ops::SplitSelectedRowsGradMaker);
+                  ops::SplitSelectedRowsGradMaker,
+                  ops::SplitSelectedRowsOpInferVarType);
 REGISTER_OP_CPU_KERNEL(
     split_selected_rows,
     ops::SplitSelectedRowsOpKernel<paddle::platform::CPUPlace, float>);

python/paddle/fluid/distribute_transpiler.py

@@ -276,6 +276,7 @@ class DistributeTranspiler:
             pserver_program.global_block().create_var(
                 name=orig_var_name,
                 persistable=True,
+                type=v.type,
                 dtype=v.dtype,
                 shape=v.shape)
             print("create origin var: ", orig_var_name)
@@ -283,6 +284,7 @@ class DistributeTranspiler:
                 var = pserver_program.global_block().create_var(
                     name="%s.trainer_%d" % (orig_var_name, trainer_id),
                     persistable=False,
+                    type=v.type,
                     dtype=v.dtype,
                     shape=v.shape)
                 recv_inputs.append(var)
@@ -551,6 +553,7 @@ class DistributeTranspiler:
                         type="sum",
                         inputs={"X": vars2merge},
                         outputs={"Out": merged_var})
+                    if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS:
                         optimize_block.append_op(
                             type="scale",
                             inputs={"X": merged_var},