// multi worker collective ops to detect failures in the cluster.  If a peer is
// down, collective ops may hang.
TF_CAPI_EXPORT extern void TFE_CollectiveOpsCheckPeerHealth(
    TFE_Context* ctx, const char* task, int64_t timeout_in_ms,
    TF_Status* status);

// Information about the shape of a Tensor and its type.
struct TF_ShapeAndType {
  // Number of dimensions. -1 indicates unknown rank.
  int num_dims;

// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update
// the gradients of its inputs. Once we have computed all gradients for a single
// input we can mark this input as done, and this can trigger adding an entry to
// the stack if all outputs of that entry are now done.
//
// When the stack is empty we have gradients for all tensors we're interested
// in.

namespace {

    // During execution:
    //   * task:0 processes main function `VariableAddFunctionWithGraphError`
    //     and places the 'read0_maybe_with_graph_error' op on task:2
    //   * task:0 partitions the main function with a subgraph containing
    //     'read0_maybe_with_graph_error' sent to task:2
    //   * task:2 graph pass reports an error when it sees

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

  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  const char* first_device_name =
      "/job:localhost/replica:0/task:0/device:CPU:0";
  const char* second_device_name =
      "/job:localhost/replica:0/task:0/device:CPU:1";
  std::array<const char*, 2> underlying_devices{first_device_name,

// on each underlying device.
//
// For example if `device_name` is
//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"
// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",
//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//

  LoggingDevice* device = new LoggingDevice;
  device->arrived_flag = arrived_flag;
  device->executed_flag = executed_flag;
  device->device_name = name;
  device->underlying_device = "/job:localhost/replica:0/task:0/device:CPU:0";
  device->strict_scope_placement = strict_scope_placement;
  TFE_RegisterCustomDevice(context, custom_device, name, device, status);
}

                           int64_t init_timeout_in_ms);

tensorflow::ServerDef ReplaceTaskInServerDef(
    const tensorflow::ServerDef& server_def, int task_index);

void ReplaceTaskInServerDef(tensorflow::ServerDef* server_def, int task_index,
                            const std::string& host, int port);

std::vector<std::string> ListDeviceNames(TFE_Context* ctx);

      TF_Status* status) const;

  void AsyncWait(TFE_Context* context, TF_Status* status) const;

  // Device strings for component devices that only include a
  // worker/task/replica if any of those differ across components. Useful for
  // printing debug messages.
  std::vector<std::string> SummarizeDeviceNames() const;

 private:

  // Register the custom device
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};
  RegisterParallelDevice(context, device_name, underlying_devices,
                         status.get());

// A Server instance encapsulates a set of devices and a Session target that
// can participate in distributed training. A server belongs to a cluster
// (specified by a ClusterSpec), and corresponds to a particular task in a
// named job. The server can communicate with any other server in the same
// cluster.

// In-process TensorFlow server.
typedef struct TF_Server TF_Server;