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();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //

//   materialized in the address space of the calling process.
// * int is used as an index into arrays.
// * Deletion functions are safe to call on nullptr.
//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation

typedef enum TFE_ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  TFE_DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  TFE_DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default placement
  // policy.

  pop_backward_tape.release()();

  std::vector<int64_t> targets;
  std::vector<Gradient*> used_in_grads;
  // We may end up with slightly fewer elements than we reserve, but grad.size()
  // should be a reasonably tight upper bound.
  targets.reserve(grad.size());
  used_in_grads.reserve(grad.size());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;

// Creates a new eager Executor. Nodes in one executor are guaranteed to be
// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.

//   * `void* ext` is a free-form field that can be populated by
//     a plugin in `TP_*` structs or potential future extension points .
//
// Example usage:
//
//   /* Sample TensorFlow code below, exact implementation might differ. */
//   // Version checking uses `struct_size`. It should be set both by core
//   // and the plugin.
//   TP_OptimizerRegistrationParams params{
//       TP_OPTIMIZER_REGISTRATION_PARAMS_STRUCT_SIZE};

  //
  // For example some tensor handles may represent distributed values, in which
  // case placement information is lost when resolving the handle.
  //
  // If false, a caller might implement pretty-printing by resolving and
  // iterating over the resulting tensor. This may still be viable if resolving
  // the handle loses information, but `SummarizeValue` would be more precise.

enum ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.

  int64_t (*tell)(const TF_WritableFile* file, TF_Status* status);

  /// Flushes `*file` and syncs contents to filesystem.
  ///
  /// This call might not block, and when it returns the contents might not have
  /// been fully persisted.
  ///
  /// DEFAULT IMPLEMENTATION: No op.
  void (*flush)(const TF_WritableFile* file, TF_Status* status);

struct TFE_Executor {
  explicit TFE_Executor(bool async, bool enable_streaming_enqueue,
                        int in_flight_nodes_limit)
      : owned_executor(new tensorflow::EagerExecutor(
            async, enable_streaming_enqueue, in_flight_nodes_limit)) {}

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)
      : owned_executor(nullptr), unowned_executor(executor) {}