Trying to run LLaMA locally but VRAM is insufficient lol

Even with an RTX 3060 12GB, it's barely enough even with quantization of a 7B model. Running foundation models locally really costs a lot of money...

by 궁금한사람238

8 answers

Yeah, costs a lot lol

by 카페인중독611 · ▲0

If you use 4-bit GGUF, it should be fine.

by 코딩하는곰52 · ▲0

Do you have a source?

by 디지털노마드225 · ▲0

I also have a 3060 12GB, and when I run 7B q4_k_m, aside from the RAM usage, it's usable. But 13B is tough.

by 뉴비탈출204 · ▲0

Well, with a 3060 12GB, full precision 7B isn't possible, but it runs well enough with q4. What were you originally trying to run?

by 호기심천국378 · ▲0

Have you tried llama.cpp? Or I'd recommend Ollama or LM Studio. Since system RAM is shared too, a 12GB card can handle most things. VRAM shortage also depends on token length and prompt size. I use a 7B model with 4-bit quantization and 2048 context without issues.

by 궁금한사람314 · ▲0

Yeah, I know... local LLMs ultimately come down to money... and servers are even more expensive. Running them at home definitely has its limits.

by 코딩하는곰420 · ▲0

But is there really a reason to run it locally? Cloud APIs are quite cheap these days.

by 프롬프트장인426 · ▲0