TF_Status* status);

// Create and modify a variable placed on a parallel device which composes
// `first_device` and `second_device`.
void BasicTestsForTwoDevices(TFE_Context* context, const char* first_device,
                             const char* second_device);

// Implementations of templated functions ******************************

template <std::size_t num_replicas>

// two component devices, running `x = TPUReplicatedInput(inputs=[a, b])` on the
// parallel device creates a parallel tensor `x` with `a` on the first of
// `underlying_devices` and `b` on the second. Running `a_unpacked, b_unpacked =
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
// The filled `device` struct and the allocated `device_info` struct may be

TF_CAPI_EXPORT extern uint64_t TF_NowNanos(void);

// Returns the number of microseconds since the Unix epoch.
TF_CAPI_EXPORT extern uint64_t TF_NowMicros(void);

// Returns the number of seconds since the Unix epoch.
TF_CAPI_EXPORT extern uint64_t TF_NowSeconds(void);

// Populates a TF_ThreadOptions struct with system-default values.
TF_CAPI_EXPORT extern void TF_DefaultThreadOptions(TF_ThreadOptions* options);

//       .Attr("T: type");
// that defines two inputs, "concat_dim" and "values" (in that order).
// You must use TF_AddInput() for the first input (since it takes a
// single tensor), and TF_AddInputList() for the second input (since
// it takes a list, even if you were to pass a list with a single
// tensor), as in:
//   TF_OperationDescription* desc = TF_NewOperation(graph, "Concat", "c");
//   TF_Output concat_dim_input = {...};

    if (it == cache_.end()) {
      return false;
    }
    lru_list_.erase(it->second.lru_iterator);
    if (timer_seconds_() - it->second.timestamp > max_age_) {
      cache_.erase(it);
      return false;
    }
    *value = it->second.value;
    lru_list_.push_front(it->first);
    it->second.lru_iterator = lru_list_.begin();
    return true;
  }

    if (grad_it == gradients.end()) {
      result[i] = nullptr;
    } else {
      if (grad_it->second.size() > 1) {
        Gradient* grad = vspace.AggregateGradients(grad_it->second);
        grad_it->second.clear();
        grad_it->second.push_back(grad);
      }
      result[i] = grad_it->second[0];
      used_gradient_ids.insert(tensor_id);
    }
  }
  VLOG(1) << "Final gradients size: "

  // [Optional]
  // Create function for optimizer.
  void* (*create_func)();

  // Optimizer function for optimizer. The first param is an optimizer created
  // by create_func. The second param is input graph. The third param is
  // GrapplerItem. The fourth param is output graph.
  void (*optimize_func)(void*, const TF_Buffer*, const TF_GrapplerItem*,
                        TF_Buffer*, TF_Status*);

#ifndef TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_
#define TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.

                    BlockFetcher block_fetcher,
                    std::function<uint64_t()> timer_seconds = TF_NowSeconds)
      : block_size_(block_size),
        max_bytes_(max_bytes),
        max_staleness_(max_staleness),
        block_fetcher_(block_fetcher),
        timer_seconds_(timer_seconds),
        pruning_thread_(nullptr,
                        [](TF_Thread* thread) { TF_JoinThread(thread); }) {

    TF_Status* status);

TF_CAPI_EXPORT extern TF_Buffer* TF_CoordinationServiceGetKeyValueWithTimeout(
    const char* key, int64_t key_size, int64_t timeout_seconds,
    TF_CoordinationServiceAgent* agent, TF_Status* status);

TF_CAPI_EXPORT extern TF_Buffer* TF_CoordinationServiceTryGetKeyValue(
    const char* key, int64_t key_size, TF_CoordinationServiceAgent* agent,