What are the hardware requirements on these larger LLMs? Is it worth quantizing them for lower-end hardware for self hosting? Not sure how doing so would impact their usefulness.
The general rule of thumb that I’ve heard is that you need 1GB of memory for ever 1B parameters. In practice however I’ve found this to not be the case. For instance on a GH200 system I’m able to run Llama3 70b in about 50GB of memory. Llama3.1 405b on the other hand uses +90GB of GPU memory and spills over to using about another 100GB of system memory… but runs like a dog at 2 tokens per second. I expect inference costs will come down over time but for now would recommend Lambda Labs if you don’t have the need for a GPU workstation.
From what I've seen, it's definitely worth quantizing. I've used llama 3 8B (fp16) and llama 3 70B (q2_XS). The 70B version was way better, even with this quantization and it fits perfectly in 24 GB of VRAM. There's also this comparison showing the quantization option and their benchmark scores:
To run this particular model though, you would need about 45GB of RAM just for the q2_K quant according to Ollama. I think I could run this with my GPU and offload the rest of the layers to the CPU, but the performance wouldn't be that great(e.g. less than 1 t/s).
Wake me up when I can ran that locally on my potato Laptop.
LocalLLaMA
Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.
Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.
As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.
Rules:
Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.
Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.
Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.
Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.