Re:RedPajama (Open LLaMA Replication) 3B and 7B models released
by int_19h on Saturday May 06, 2023 @04:49AM
(#35839462)
It's not even that bad. Core i7-12700K with DDR5 gives me ~1 word per second on llama-30b - that is fast enough for real-time chat, with some patience. And things are even better on M1/M2 Macs.
Re:RedPajama (Open LLaMA Replication) 3B and 7B models released
by Joeri on Saturday May 06, 2023 @05:47AM
(#35839753)
The critical factor seems to be the ability to fit the whole model in RAM (--mlock option in oobabooga). With Apple's RAM prices most M1/M2 owners probably don't have the 32 GB RAM required to fit a 4bit 30B model.
Bigger ones as well, you just have to wait longer. Nothing for real time usage, but if you can wait 10-20 minutes, you can use them on CPU.