const DLTensor* dltensor_out = &dlm_out->dl_tensor;
  EXPECT_EQ(dltensor_out->device.device_type, dltensor_in->device.device_type);
  EXPECT_EQ(dltensor_out->device.device_id, dltensor_in->device.device_id);
  EXPECT_EQ(dltensor_out->ndim, dltensor_in->ndim);
  EXPECT_EQ(dltensor_out->dtype.code, dltensor_in->dtype.code);
  EXPECT_EQ(dltensor_out->dtype.bits, dltensor_in->dtype.bits);

      // [session->last_num_graph_nodes, num_nodes), that is the nodes
      // added since the last TF_SessionRun() call.
      for (auto id = session->last_num_graph_nodes; id < num_nodes; ++id) {
        Node* const node = graph.FindNodeId(id);
        if (node != nullptr && node->IsOp()) {
          NodeDef* const node_def = graph_def.add_node();
          *node_def = node->def();
        }
      }

  }
  return ret;
}

TF_Tensor* TF_DequeueNamedTensor(TF_Session* session, int tensor_id,
                                 TF_Status* status) {
  assert(session);
  {
    tensorflow::mutex_lock c(session->graph->mu);
    VLOG(1) << "Dequeuing named tensor with id " << tensor_id
            << ", with input graph: "
            << session->graph->graph.ToGraphDefDebug().DebugString();
  }

           {{"T", DT_FLOAT},
            {"group_size", static_cast<int>(devices.size())},
            {"group_key", 0},
            {"instance_key", 0},
            {"merge_op", "Add"},
            {"final_op", "Id"},
            {"subdiv_offsets", std::vector<int>()}},
           /*dep=*/{"assert"}},
      });
  TF_ASSERT_OK(ContextFromInterface(unwrap(context.get()))
                   ->AddFunctionDef(assert_and_collective));

                  .ok());
  ASSERT_TRUE(worker_server->Start().ok());

  // Update server def to replace the remote device with the device info on the
  // new worker (different incarnation ID).
  server_def.set_task_index(0);
  string serialized_update = server_def.SerializeAsString();
  TFE_ContextUpdateServerDef(ctx, 0, serialized_update.data(),
                             serialized_update.size(), status);

  tensorflow::DeviceNameUtils::ParseFullName(device_name, &parsed_name);
  std::string device_type = parsed_name.type;
  int device_id = 0;
  if (parsed_name.has_id) {
    device_id = parsed_name.id;
  }

  ctx.device_id = device_id;
  if (device_type == "CPU") {
    ctx.device_type = DLDeviceType::kDLCPU;
  } else if (device_type == "GPU") {
#if TENSORFLOW_USE_ROCM

  for (int i = 0; i < target_tensor_ids.size(); ++i) {
    int64_t target_id = target_tensor_ids[i];
    if (sources_set.find(target_id) != sources_set.end()) {
      auto tensor = targets[i];
      sources_that_are_targets.insert(
          std::make_pair(target_id, TapeTensor(tensor)));
    }
  }

  TF_RETURN_IF_ERROR(GradientTape::ComputeGradient(

  if (TF_GetCode(status) != TF_OK) return;

  // unwrap op_ and set step_id only if valid step id value was set.
  // Currently only required for non-TFRT use cases, e.g., EagerOp.
  if (step_id_.has_value()) {
    tensorflow::unwrap(op_.get())->SetStepId(step_id_.value());
  }

  TFE_Execute(op_.get(), unwrapped_results.data(), &real_num_outputs, status);

  TFE_OpSetAttrInt(op.get(), "instance_key", 0);
  const std::string merge_op("Add");
  TFE_OpSetAttrString(op.get(), "merge_op", merge_op.c_str(),
                      merge_op.length());
  const std::string final_op("Id");
  TFE_OpSetAttrString(op.get(), "final_op", final_op.c_str(),
                      final_op.length());
  TFE_OpSetAttrIntList(op.get(), "subdiv_offsets", nullptr, 0);

  TFE_OpAddInput(op.get(), input, status);

  auto* task_iter = static_cast<const tensorflow::CoordinatedTask*>(tasks.data);
  for (size_t i = 0; i < tasks.length; ++i) {
    task_vec[i].set_job_name(task_iter->job_name());
    task_vec[i].set_task_id(task_iter->task_id());
    ++task_iter;
  }
  auto results = coord_agent->GetTaskState(task_vec);
  if (!results.ok()) {
    status->status = results.status();
    return;
  }