diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc
index 5a6145d5174ce1e4cdf787e73ed78a81d67a62e1..662a616543f23e55619a122f757cfca9e11df030 100644
--- a/paddle/fluid/operators/load_op.cc
+++ b/paddle/fluid/operators/load_op.cc
@@ -12,7 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #include <fstream>
+#include <iostream>
 
+#include "paddle/fluid/framework/data_type_transform.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -51,14 +53,30 @@ class LoadOp : public framework::OperatorBase {
     auto in_dtype = framework::ToDataType(tensor->type());
     auto out_dtype = load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype;
 
+    std::cout << "In load op: " << std::endl;
+    std::cout << "before conversion block" << std::endl;
+
     if (in_dtype != out_dtype) {
       // convert to float16 tensor
       auto in_kernel_type = framework::OpKernelType(in_dtype, place);
       auto out_kernel_type = framework::OpKernelType(out_dtype, place);
       framework::LoDTensor fp16_tensor;
+      // copy LoD info to the new tensor
+      fp16_tensor.set_lod(tensor->lod());
+      std::cout << "before TransDataType" << std::endl;
       framework::TransDataType(in_kernel_type, out_kernel_type, *tensor,
                                &fp16_tensor);
+      std::cout << "after TransDataType" << std::endl;
+      // reset output tensor
+      out_var->Clear();
+      tensor = out_var->GetMutable<framework::LoDTensor>();
+      tensor->set_lod(fp16_tensor.lod());
+      std::cout << "before TransDataType" << std::endl;
+      tensor->ShareDataWith(fp16_tensor);
+      std::cout << "after TransDataType" << std::endl;
     }
+
+    std::cout << "Out of load op: " << std::endl;
   }
 };
 
diff --git a/paddle/fluid/operators/save_load_op_test.cc b/paddle/fluid/operators/save_load_op_test.cc
index 74385ee47543e3f4887081c2225212996d3df3f1..6a5ae01347569b4bd55df9f2e718c1b9842b894d 100644
--- a/paddle/fluid/operators/save_load_op_test.cc
+++ b/paddle/fluid/operators/save_load_op_test.cc
@@ -63,7 +63,7 @@ TEST(SaveLoadOp, CPU) {
   }
 }
 
-TEST(SaveLoadFP16Op, CPU) {
+TEST(SaveFP16Op, CPU) {
   paddle::framework::Scope scope;
   paddle::platform::CPUPlace place;
 
@@ -94,3 +94,56 @@ TEST(SaveLoadFP16Op, CPU) {
     EXPECT_EQ(expect[i], static_cast<float>(actual[i]));
   }
 }
+
+TEST(LoadFP16Op, CPU) {
+  paddle::framework::Scope scope;
+  paddle::platform::CPUPlace place;
+
+  auto var = scope.Var("test_var");
+  auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
+  tensor->Resize({3, 10});
+
+  paddle::framework::LoD expect_lod;
+  expect_lod.resize(1);
+  expect_lod[0].push_back(0);
+  expect_lod[0].push_back(1);
+  expect_lod[0].push_back(2);
+  expect_lod[0].push_back(3);
+
+  tensor->set_lod(expect_lod);
+  float* expect = tensor->mutable_data<float>(place);
+  for (int64_t i = 0; i < tensor->numel(); ++i) {
+    expect[i] = static_cast<float>(paddle::platform::float16(i));
+  }
+
+  paddle::framework::AttributeMap attrs;
+  attrs.insert({"file_path", std::string("tensor.save")});
+  attrs.insert({"load_as_fp16", true});
+
+  auto save_op = paddle::framework::OpRegistry::CreateOp(
+      "save", {{"X", {"test_var"}}}, {}, attrs);
+  save_op->Run(scope, place);
+
+  auto load_var = scope.Var("out_var");
+  auto target = load_var->GetMutable<paddle::framework::LoDTensor>();
+  auto load_op = paddle::framework::OpRegistry::CreateOp(
+      "load", {}, {{"Out", {"out_var"}}}, attrs);
+  LOG(INFO) << "before load op run";
+  load_op->Run(scope, place);
+  LOG(INFO) << "after load op run";
+
+  paddle::platform::float16* actual = target->data<paddle::platform::float16>();
+  LOG(INFO) << "after target->data";
+  for (int64_t i = 0; i < tensor->numel(); ++i) {
+    EXPECT_EQ(expect[i], static_cast<float>(actual[i]));
+  }
+  LOG(INFO) << "after expect equal";
+
+  auto& actual_lod = target->lod();
+  EXPECT_EQ(expect_lod.size(), actual_lod.size());
+  for (size_t i = 0; i < expect_lod.size(); ++i) {
+    for (size_t j = 0; j < expect_lod[i].size(); ++j) {
+      EXPECT_EQ(expect_lod[i][j], actual_lod[i][j]);
+    }
+  }
+}