3 ms·
They're saying that likely, models of comparable size will perform worse (the paper claims as good) But since they are (optimized up to 8 or 10x if packing ter
by magnustesshu 3y ago
They're saying that likely, models of comparable size will perform worse (the paper claims as good)
But since they are (optimized up to 8 or 10x if packing terns beyond 2 bits, in practice it seems 3-5x considering larger other structures needed in memory) more memory efficient, the largest models can be that much larger.