llama-int8 VS llama-cpu

If anyone is interested in running this at home, please follow the llama-int8 project [1]. LLM.int8() is a recent development allowing LLMs to run in half the memory without loss of performance [2]. Note that at the end of [2]'s abstract, the authors state "This result makes such models much more accessible, for example making it possible to use OPT-175B/BLOOM on a single server with consumer GPUs. We open-source our software." I'm very thankful we have researchers like this further democratizing access to this data and prying it out of the hands of the gatekeepers who wish to monetize it.

[1] https://github.com/tloen/llama-int8

[2] https://arxiv.org/abs/2208.07339

To add a bit more context, the code other people linked (https://github.com/tloen/llama-int8) assumes single GPU. So if you want to run it on 2x3090, you'll need to modify it a bit:

You've probably heard, but LLaMA just released, and its 13B parameter model outperforms GPT-3 on most metrics (because they trained it on a lot more data). Someone's already quantized it to 4 and 3 bits and it performs virtually the same. It also apparently performs well on CPUs (several words per second on a 7900X). Running something equivalent to GPT3.5 on a phone is not out that far out.

I tried to port the llama-cpu version to a gpu-accelerated mps version for macs, it runs, but the outputs are not as good as expected and it often gives "-1" tokens. Any help and contributions on fixing it are welcome!

You can run it with only a CPU and 32 gigs of RAM: https://github.com/markasoftware/llama-cpu

I was able to run 7B on a CPU, inferring several words per second: https://github.com/markasoftware/llama-cpu