// otherwise be the name of this custom device. Ops are placed onto a custom
  // device if any of their inputs are on that custom device, but custom devices
  // are free to set a bad status in order to require explicit placement.
  void (*execute)(const TFE_Op* op, int* num_outputs,
                  TFE_TensorHandle** outputs, TF_Status* s, void* device_info);

  // Method to delete a device.

  // parallel.
  //
  // Conceptually this is a thread pool with one thread per device. It requires
  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic
  // allocation/scheduling.
  //
  // TODO(allenl): Keep a map from outer thread to list of inner threads rather

  // from a backward gradient function. Fills `out_grads` corresponding to
  // `output_tensors`. `out_grads` must not be null.
  //
  // Executes the backward function in order to trace its gradient, which will
  // waste computation if executing eagerly (when graph building the unneeded
  // computation is pruned). Temporarily sets `backward_tape` so that

// 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.

  /// be accessed while holding the Block's mu lock. The data vector should only
  /// be accessed after state == FINISHED, and it should never be modified.
  ///
  /// In order to prevent deadlocks, never grab the block-cache-wide mu_ lock
  /// AFTER grabbing any block's mu lock. It is safe to grab mu without locking
  /// mu_.
  struct Block {
    /// The block data.