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

  CheckTFE_TensorHandleHasFloats(retval_task0, {7, 10, 15, 22});

  TFE_DeleteTensorHandle(retval_task0);

  const char remote_device_name[] =
      "/job:localhost/replica:0/task:1/device:CPU:0";
  auto* h0_task1 =
      TFE_TensorHandleCopyToDevice(h0_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  auto* h1_task1 =
      TFE_TensorHandleCopyToDevice(h1_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

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

  // Copying on to a parallel device must be explicit.
  TensorHandlePtr failed_copy_on_result(TFE_TensorHandleCopyToDevice(
      cpu_value.get(), context.get(), device_name, status.get()));
  EXPECT_EQ(TF_GetCode(status.get()), TF_UNIMPLEMENTED);

  std::array<TFE_TensorHandle*, 2> components{cpu_value.get(), cpu_value.get()};

  std::vector<TensorHandlePtr> components;
  components.reserve(underlying_devices_.size());
  for (const std::string& underlying_device_name : underlying_devices_) {
    TFE_TensorHandle* t = TFE_TensorHandleCopyToDevice(
        tensor, context, underlying_device_name.c_str(), status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    components.emplace_back(t);
  }

          unwrapped_op,
          reinterpret_cast<tensorflow::ImmediateExecutionTensorHandle**>(
              retvals),
          num_retvals);
}

TFE_TensorHandle* TFE_TensorHandleCopyToDevice(TFE_TensorHandle* h,
                                               TFE_Context* ctx,
                                               const char* device_name,