auto-round

disable quantizing lm-head with tied weights as a workaround

#101

Merged