TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "AddV2", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, a, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, b, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);
  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(a));

  op.reset(TFE_NewOp(context.get(), "AssignVariableOp", status.get()));
  TFE_OpSetAttrType(op.get(), "dtype", TF_FLOAT);
  TFE_OpAddInput(op.get(), var_handle, status.get());
  TFE_OpAddInput(op.get(), one.get(), status.get());
  TFE_OpSetDevice(op.get(), name, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  executed = false;

  int num_retvals = 1;
  for (auto s : state) {
    TFE_OpReset(matmul, "MatMul", nullptr, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_OpAddInput(matmul, m, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_OpAddInput(matmul, m, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_Execute(matmul, &retvals[0], &num_retvals, status);

    matmul = TFE_NewOp(ctx, "MatMulFunction", status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
    TFE_OpAddInput(matmul, h0_task0, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
    TFE_OpAddInput(matmul, has_packed_input ? packed_handle : h1_task2, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  } else {

      LoggedTensor* t = reinterpret_cast<LoggedTensor*>(
          TFE_TensorHandleDevicePointer(input, s));
      if (TF_GetCode(s) != TF_OK) return;
      TFE_OpAddInput(op, t->tensor, s);
    } else {
      TFE_OpAddInput(op, input, s);
    }
    if (TF_GetCode(s) != TF_OK) return;
  }
  std::vector<TFE_TensorHandle*> op_outputs(*num_outputs);
  TFE_Execute(op, op_outputs.data(), num_outputs, s);

  TF_Status* status = TF_NewStatus();
  TFE_Op* op = TFE_NewOp(ctx, "ReadVariableOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpAddInput(op, var, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_retvals = 1;
  TFE_Execute(op, out_value, &num_retvals, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  TFE_Op* func = TFE_NewOp(ctx, "AddVariablesFunction", status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_OpAddInput(func, packed_handle, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  if (remote) {
    TFE_OpSetDevice(func, task1_name, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

        TFE_NewOp(context.get(), "TPUReplicatedOutput", status.get()),
        TFE_DeleteOp);
    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetAttrInt(op.get(), "num_replicas", 1);
    TFE_OpAddInput(op.get(), value_one.get(), status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetDevice(op.get(), device_name, status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;

    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    TF_DeleteTensor(t);

    TFE_Op* op = TFE_NewOp(ctx, "ident", status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    TFE_OpAddInput(op, h, status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);

    std::vector<TFE_TensorHandle*> result;
    result.push_back(nullptr);
    int num_retvals = 1;

TF_CAPI_EXPORT extern const char* TFE_OpGetDevice(const TFE_Op* op,
                                                  TF_Status* status);

TF_CAPI_EXPORT extern void TFE_OpAddInput(TFE_Op* op, TFE_TensorHandle* input,
                                          TF_Status* status);

TF_CAPI_EXPORT extern void TFE_OpAddInputList(TFE_Op* op,