auto-round

82322ac9 - better support quant_lm_head for larger models (#263)