Skip to content

Comment on MLC-LLM: GPT/Llama on consumer-class GPUs and phonesparent

Comments

Llama.cpp recently added partial GPU acceleration. Model dequantization as well as some BLAS operations have been moved to GPU.

It runs a lot faster if you compile with cuBLAS (nvidia) or clblast (other). GPU vram doesn't matter much since it doesn't offload the model to vram.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.