class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // (1) If type == TF_ATTR_STRING and is_list == 0
  //     then total_size is the byte size of the string
  //     valued attribute.
  // (2) If type == TF_ATTR_STRING and is_list == 1
  //     then total_size is the cumulative byte size
  //     of all the strings in the list.
  // (3) If type == TF_ATTR_SHAPE and is_list == 0
  //     then total_size is the number of dimensions
  //     of the shape valued attribute, or -1

// Returns a new thread that is running work_func and is identified
// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread
// starts. In this way, data may be passed from the thread back to the caller.
//
// Caller takes ownership of the result and must call TF_JoinThread on it
// eventually.

// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",
// "description", "label");
// and then incremented that counter for various values of label:
// streamz->GetCell("label-value")->IncrementBy(1);
//
// The test code can then read and test the value of that counter:
//
// auto* reader = TFE_MonitoringNewCounterReader("name");
// test();

//   "/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.
//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw

// A tensorflow.ServerDef specifies remote workers (in addition to the current
// workers name). Operations created in this context can then be executed on
// any of these remote workers by setting an appropriate device.
//
// If the following is set, all servers identified by the
// ServerDef must be up when the context is created.
TF_CAPI_EXPORT extern void TFE_ContextSetServerDef(TFE_Context* ctx,

class WorkerCacheInterface;

class ImmediateExecutionDistributedManager {
 public:
  virtual ~ImmediateExecutionDistributedManager() {}

  // Set up distributed execution environment on local and remote tasks.
  // When `reset_context` is true, initialize new cluster context state based
  // on cluster configurations provided in `server_def`; otherwise, update
  // existing context state with the provided `server_def`. Contexts created

#define TENSORFLOW_C_EXPERIMENTAL_FILESYSTEM_PLUGINS_GCS_CLEANUP_H_

#include <type_traits>
#include <utility>

namespace tf_gcs_filesystem {

// A move-only RAII object that calls a stored cleanup functor when
// destroyed. Cleanup<F> is the return type of gtl::MakeCleanup(F).
template <typename F>
class Cleanup {
 public:
  Cleanup() : released_(true), f_() {}

  template <typename G>

// 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.

//
// The API here is subject to changes in the future.
// --------------------------------------------------------------------------

#ifdef __cplusplus
extern "C" {
#endif

// When `enable` is true, set
// tensorflow.ConfigProto.OptimizerOptions.global_jit_level to ON_1, and also
// set XLA flag values to prepare for XLA compilation. Otherwise set
// global_jit_level to OFF.
//