"Filesystem for ", dir, " does not support GetChildren()"));

  UniquePtrTo_TF_Status plugin_status(TF_NewStatus(), TF_DeleteStatus);
  std::string translated_name = TranslateName(dir);
  // Note that `children` is allocated by the plugin and freed by core
  // TensorFlow, so we need to use `plugin_memory_free_` here.
  char** children = nullptr;
  const int num_children =

  //
  // WARNING: The destructor of this class is marked as protected to disallow
  // clients from directly destroying this object since it may manage it's own
  // lifetime through ref counting. Thus this must be allocated on the heap and
  // clients MUST call Release() in order to destroy an instance of this class.
  virtual void Release() = 0;

  virtual Status Reset(const char* op, const char* raw_device_name) = 0;

// `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
// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.
void AllocateParallelDevice(const char* device_name,
                            const char* const* underlying_devices,

//
// Conventions:
// * We use the prefix TF_ for everything in the API.
// * Objects are always passed around as pointers to opaque structs
//   and these structs are allocated/deallocated via the API.
// * TF_Status holds error information.  It is an object type
//   and therefore is passed around as a pointer to an opaque
//   struct as mentioned above.

                                                TF_Status* status);

// Execute the operation defined by 'op' and return handles to computed
// tensors in `retvals`.
//
// 'retvals' must point to a pre-allocated array of TFE_TensorHandle* and
// '*num_retvals' should be set to the size of this array. It is an error if
// the size of 'retvals' is less than the number of outputs. This call sets
// *num_retvals to the number of outputs.

  //   Outputs:
  std::vector<TensorHandlePtr> op_outputs_ TF_GUARDED_BY(execution_mutex_);
  // TF_Status is an incomplete type and so can't be stack allocated. To avoid
  // unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);

///
/// Since memory that is allocated by the DSO gets transferred to core
/// TensorFlow, we need to provide a way for the allocation and deallocation to
/// match. This is why this structure also defines `plugin_memory_allocate` and
/// `plugin_memory_free` members.
///
/// All memory allocated by the plugin that will be owned by core TensorFlow
/// must be allocated using the allocator in this structure. Core TensorFlow

// The returned string is heap-allocated, and caller should call free() on it.
TF_CAPI_EXPORT extern const char* TF_GraphDebugString(TF_Graph* graph,
                                                      size_t* len);

// Returns the function content in a human-readable format, with length set in
// `len`. The format is subject to change in the future.
// The returned string is heap-allocated, and caller should call free() on it.
//

  std::vector<MaybeParallelTensorOwned> typed_outputs(
      std::move(maybe_typed_outputs.value()));

  if (typed_outputs.size() > *num_outputs) {
    TF_SetStatus(status, TF_INTERNAL,
                 "The allocated output buffer was too small.");
    return;
  }

  for (int i = 0; i < typed_outputs.size(); ++i) {
    MaybeParallelTensorOwned typed_output(std::move(typed_outputs[i]));

  for (int i = 0; i < len; ++i) {
    values[i] = TF_NewBuffer();
    status->status = MessageToBuffer(attr->list().shape(i), values[i]);
    if (!status->status.ok()) {
      // Delete everything allocated to far, the operation has failed.
      for (int j = 0; j <= i; ++j) {
        TF_DeleteBuffer(values[j]);
      }
      return;
    }
  }
}