// Creates a pass that lifts operations on external resource variables from
// device computation nested in `tf_device::LaunchOp` out so that resource
// variable load operations are all before device computation while resource
// variable store operations are all after device computation. After this pass,
// device computation no longer interacts with external resource variables.

  }

  if (output_to_input_alias.empty()) return absl::OkStatus();

  xla::HloModuleProto* module_proto =
      compilation_result->computation->mutable_proto();
  absl::StatusOr<xla::ProgramShape> program_shape_or_status =
      compilation_result->computation->GetProgramShape();
  TF_RET_CHECK(program_shape_or_status.ok());

  xla::ProgramShape& program_shape = program_shape_or_status.value();

  }
  return DeviceNameUtils::ParsedNameToString(tpu_device);
}

// Determine execution devices when topology and device assignment are not
// defined. This is a special case where a single core computation is replicated
// to every core in the mesh. TPU devices are simply added to
// `execution_devices` of one replica. `num_replicas` must be 1 or the total
// number of TPU devices available, and `num_cores_per_replica` must be 1.

 *       SMALL_DELAY_MS}, {@code MEDIUM_DELAY_MS}, {@code LONG_DELAY_MS}. The idea here is that a
 *       SHORT is always discriminable from zero time, and always allows enough time for the small
 *       amounts of computation (creating a thread, calling a few methods, etc) needed to reach a
 *       timeout point. Similarly, a SMALL is always discriminable as larger than SHORT and smaller