New sparse gradient is returned, ADDSD is optimized in place, we have to write a flag to turn on optimization in case of updates