const std::array<const char*, num_devices>& underlying_devices,
    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 ******************************

// TPUReplicatedInput and TPUReplicatedOutput respectively. For example, with
// 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.
//

  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s
  // return a bad status the caller must run all of the `Join`s or any future
  // `StartExecute`s will deadlock).
  //

//       .Output("output: T")
//       .Attr("N: int >= 2")
//       .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:

    const std::unordered_map<int64_t, int64_t>& op_missing_tensor) {
  std::vector<int64_t> result;
  for (auto& op_entry : op_tape) {
    if (op_missing_tensor.find(op_entry.first) == op_missing_tensor.end()) {
      result.push_back(op_entry.first);
    }
  }
  return result;
}

template <typename Gradient, typename BackwardFunction, typename TapeTensor>
Status InitialGradients(

    Entry entry{timer_seconds_(), value, lru_list_.begin()};
    auto insert = cache_.insert(std::make_pair(key, entry));
    if (!insert.second) {
      lru_list_.erase(insert.first->second.lru_iterator);
      insert.first->second = entry;
    } else if (max_entries_ > 0 && cache_.size() > max_entries_) {
      cache_.erase(lru_list_.back());
      lru_list_.pop_back();
    }
  }

// can correspond to the data contained in another dimension in on-host
// representation. The dimensions are indexed using the standard TensorFlow
// major-to-minor order (slowest varying dimension first),
// not the XLA's minor-to-major order.
// On-device dimensions can be padded. TFE_TensorDebugInfoOnDeviceDim returns
// the number of elements in a dimension after padding.

  size_t struct_size;
  void* ext;  // reserved for future use

  // [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_Operation* RandomUniform(TF_Operation* shape, TF_DataType dtype,
                            TF_Graph* graph, TF_Status* s);

// Split `input` along the first dimension into 3 tensors
TF_Operation* Split3(TF_Operation* input, TF_Graph* graph, TF_Status* s,
                     const char* name = "split3");

bool IsPlaceholder(const tensorflow::NodeDef& node_def);

/// `ReadOnlyMemoryRegion`) and one for all the operations a `Filesystem`
/// implements. Each of them is in a 1-to-1 correspondence with the wrapper
/// structures from the first section: these tables only contain function
/// pointers that operate on the corresponding data. Thus, the first argument of
/// each of these functions is a pointer to the paired struct and this argument
/// can be used to track state in between calls (from an object oriented point