// 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
  // than a single list of threads so aliased nested parallel devices don't
  // re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;

TF_CAPI_EXPORT extern void TF_DeleteShapeAndTypeList(
    TF_ShapeAndTypeList* shape_list);
TF_CAPI_EXPORT extern void TF_DeleteShapeAndTypeListArray(
    TF_ShapeAndTypeList** shape_list_array, int num_items);

// Infer shapes for the given `op`. The arguments mimic the arguments of the
// `shape_inference::InferenceContext` constructor. Note the following:
//   - The inputs of the `op` are not used for shape inference. So, it is