llama VS llama-mps

> All the magic of "7B LLaMA running on a potato" seems to involve lowering precision down to f16 and then further quantizing to int4.

LLaMa weights are f16s to start out with, no lowering necessary to get to there.

You can stream weights from RAM to the GPU pretty efficiently. If you have >= 32GB ram and >=2GB vram my code here should work for you: https://github.com/gmorenz/llama/tree/gpu_offload

There's probably a cleaner version of it somewhere else. Really you should only need >= 16 GB ram, but the (meta provided) code to load the initial weights is completely unnecessarily making two copies of the weights in RAM simultaneously.

My code for this is very much not high quality, but I have a CPU + GPU + SSD combination: https://github.com/gmorenz/llama/tree/ssd

Usage instructions in the commit message: https://github.com/facebookresearch/llama/commit/5be06e56056...

At least with my hardware this runs at "[size of model]/[speed of SSD reads]" tokens per second, which (up to some possible further memory reduction so you can run larger batches at once on the same GPU) is a good as it gets when you need to read the whole model from disk each token.

At a 125GB and a 2MB/s read (largest model, what I get from my ssd) that's 60 seconds per token (1 day per 1440 words), which isn't exactly practical. Which is really the issue here, if you need to stream the model from an SSD because you don't have enough RAM, it is just a fundamentally slow process.

You could probably optimize quite a bit for batch throughput if you're ok with the latency though.

I don't know about this fork specifically, but in general yes absolutely.

Even without enough ram, you can stream model weights from disk and run at [size of model/disk read speed] seconds per token.

I'm doing that on a small GPU with this code, but it should be easy to get this working with the CPU as compute instead (and at least with my disk/CPU, I'm not even sure that it would run even slower, I think disk read would probably still be the bottleneck)

https://github.com/gmorenz/llama/tree/ssd

There are currently at least 3 ways to run llama on m1 with GPU acceleration. - mlc-llm (pre-built, only 1 model has been ported) - tinygrad (very memory efficient, not that easy to integrate into other projects) - llama-mps (original llama codebase + llama adapter support)

There is also a gpu-acelerated fork of the original repo

https://github.com/remixer-dec/llama-mps

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!