diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc
index 16c612c45a37dd2ffd17f8d5f5946df30e9b3fe6..69fcffe9bc34006aef2e5a39227cf6d947e4615f 100644
--- a/paddle/fluid/operators/detail/sendrecvop_utils.cc
+++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc
@@ -82,7 +82,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
         platform::CPUPlace cpu;
         auto& gpu_dev_ctx =
             static_cast<const platform::CUDADeviceContext&>(ctx);
-        auto copy_size = tensor.memory_size();
+        auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type());
         payload = memory::Alloc(cpu, copy_size);
 
         memory::Copy(cpu, payload,
@@ -99,7 +99,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
       } else {
         payload = tensor.data<void>();
       }
-      payload_size = tensor.memory_size();
+      payload_size = tensor.numel() * framework::SizeOfType(tensor.type());
       e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
     } break;
     case framework::proto::VarType_Type_SELECTED_ROWS: {
@@ -118,7 +118,8 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
         platform::CPUPlace cpu;
         auto& gpu_dev_ctx =
             static_cast<const platform::CUDADeviceContext&>(ctx);
-        auto copy_size = tensor->memory_size();
+        auto copy_size =
+            tensor->numel() * framework::SizeOfType(tensor->type());
         payload = memory::Alloc(cpu, copy_size);
         memory::Copy(cpu, payload,
                      boost::get<platform::CUDAPlace>(tensor->place()),
@@ -133,7 +134,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
       } else {
         payload = slr->mutable_value()->data<void>();
       }
-      payload_size = tensor->memory_size();
+      payload_size = tensor->numel() * framework::SizeOfType(tensor->type());
       e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
     } break;
     default:
diff --git a/paddle/fluid/operators/split_byref_op.h b/paddle/fluid/operators/split_byref_op.h
index 9b54c7c74acb512b9493d603e83380b9a92ac91b..a3aad68ea736e223d3917607cca17f5cccfef630 100644
--- a/paddle/fluid/operators/split_byref_op.h
+++ b/paddle/fluid/operators/split_byref_op.h
@@ -32,7 +32,8 @@ class SplitByrefOpKernel : public framework::OpKernel<T> {
     for (size_t i = 0; i < outs.size(); ++i) {
       // NOTE: no need to call mutable_data here to allocate memory.
       auto* out = outs[i];
-      *out = std::move(in->Slice(row_offset, out->dims()[0]));
+      VLOG(3) << "spliting by ref: " << row_offset << " " << out->dims()[0];
+      *out = std::move(in->Slice(row_offset, row_offset + out->dims()[0]));
       row_offset += out->dims()[0];
     }
   }