LLaMA-13B-4bit-128g / README.md
Neko-Institute-of-Science's picture
add eval
a4ea139
13B  (act-order true-sequential groupsize)
wikitext2 5.0906524658203125 (stock 16bit)
wikitext2 5.153766632080078 (32g)
wikitext2 5.198880672454834 (128)
wikitext2 5.266944408416748 (128 no-act)
wikitext2 5.271687984466553 (128 no-t no-act)
ptb-new 9.080504417419434 (stock 16bit)
ptb-new 9.149489402770996 (32g)
ptb-new 9.268823623657227 (128)
ptb-new 9.45678424835205 (128 no-act)
ptb-new 9.497363090515137 (128 no-t no-act)
c4-new 6.798543930053711 (stock 16bit)
c4-new 6.866276264190674 (32g)
c4-new 6.910022735595703 (128)
c4-new 6.955390930175781 (128 no-act)
c4-new 6.956299781799316 (128 no-t no-act)