int num_retvals = 1;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_DeleteTensorHandle(retval);

  // Custom device: inputs in same custom device works.
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcustom0.get()));
  num_retvals = 1;
  executed = false;

// is done.
// TODO(agarwal): change num_retvals to int from int*.
TF_CAPI_EXPORT extern void TFE_Execute(TFE_Op* op, TFE_TensorHandle** retvals,
                                       int* num_retvals, TF_Status* status);

// Add a function (serialized FunctionDef protocol buffer) to ctx so
// that it can be invoked using TFE_Execute.
TF_CAPI_EXPORT extern void TFE_ContextAddFunctionDef(

      TFE_TensorHandle* dummy = nullptr;
      TFE_Execute(add_op_dummy, &dummy, &num_retvals, status);
      ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
      TFE_DeleteTensorHandle(dummy);
      TFE_DeleteOp(add_op_dummy);
    }
  }
  TFE_TensorHandle* retval = nullptr;
  TFE_Execute(add_op, &retval, &num_retvals, status);
  EXPECT_EQ(1, num_retvals);

  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  auto* retval_task0 =
      TFE_TensorHandleCopyToDevice(retvals[0], ctx, local_device_name, status);

  if (!device_name.empty()) {
    TFE_OpSetDevice(op, device_name.c_str(), status);
  }
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  TFE_Execute(op, &var_handle, &num_retvals, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_DeleteOp(op);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  CHECK_EQ(1, num_retvals);

  if (remote) {
    TFE_OpSetDevice(func, task1_name, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  }

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(func, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);
  TFE_DeleteOp(func);
  TFE_DeleteTensorHandle(packed_handle);

  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], 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;
    TFE_Execute(op, result.data(), &num_retvals, status);
    TFE_DeleteOp(op);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    ASSERT_EQ(num_retvals, 1);

    // The input handles should never change since they have been mirrored.
    ASSERT_FALSE(remote_arg->HasLocalMirror(nullptr));
  }

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  // TODO(gjn): Add support for waiting on async local mirrors
  if (!remote && !async && !remote_func_outputs) {

      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);
  TFE_DeleteOp(op);
  if (TF_GetCode(s) != TF_OK) return;
  std::vector<TFE_TensorHandle*> unwrapped_outputs;
  unwrapped_outputs.reserve(op_outputs.size());