v2: max_total_len 512 (~20% faster), fp16 embedding (-622 MB), add int4 LM tier

c5002c0

verified

Masterx commited on about 22 hours ago