} else {
        TF_SetStatus(
            status, TF_INVALID_ARGUMENT,
            absl::StrCat(
                "Got a non-parallel tensor ",
                tensorflow::unwrap(absl::get<TFE_TensorHandle*>(input))
                    ->DebugString(),
                " as input to a parallel operation. First pack non-parallel "
                "tensors for each device into a parallel tensor explicitly.")

AbstractContext* BuildFunction(const char* fn_name) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TF_ExecutionContext* graph_ctx = TF_CreateFunction(fn_name, status.get());
  return unwrap(graph_ctx);
}

Status CreateParamsForInputs(AbstractContext* ctx,
                             absl::Span<AbstractTensorHandle* const> inputs,
                             std::vector<AbstractTensorHandle*>* params) {

    TFE_OpSetDevice(matmul, cpu_device_name.c_str(), status);
    EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
    auto remote_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(h1_task2));
    // 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;

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

    auto cpu_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(hcpu));
    auto gpu_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(hgpu));
    auto gpu_device = gpu_arg->device();
    ASSERT_FALSE(cpu_arg->HasLocalMirror(gpu_device));

    TFE_Op* matmul = MatMulOp(ctx, hcpu, hgpu);

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

  // Test that full type information can be accessed.
  auto outs = unwrap(add_outputs);
  auto h = outs->outputs[0];
  ASSERT_NE(h, nullptr);
  ASSERT_EQ(h->FullType().type_id(), TFT_UNSET);
  ASSERT_EQ(unwrap(inputs[0])->FullType().type_id(), TFT_UNSET);

  // Clean up operation and inputs.
  TF_DeleteAbstractOp(add_op);

      context.get(), components, device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  ImmediateExecutionTensorHandle* unwrapped_handle =
      tensorflow::unwrap(device_value.get());
  std::string summarized;
  TF_ASSERT_OK(unwrapped_handle->SummarizeValue(summarized));
  EXPECT_THAT(summarized, HasSubstr("\"CPU:0\": 3"));
}

}  // namespace parallel_device

                                            absl::string_view result,
                                            size_t read) {
  // Result isn't a null-terminated string so we have to wrap it inside a
  // `string_view`
  if (length == read && content_view.substr(offset, length) ==
                            absl::string_view(result).substr(0, read))
    return ::testing::AssertionSuccess();
  else