Masterx's picture
v2: max_total_len 512 (~20% faster), fp16 embedding (-622 MB), add int4 LM tier
c5002c0 verified