Mozilla-Ocho / llamafile

Distribute and run LLMs with a single file.
https://llamafile.ai
Other
20.58k stars 1.04k forks source link

Feature Request: Support AVX-512 for Intel Rocket Lake #628

Open aluklon opened 1 week ago

aluklon commented 1 week ago

Prerequisites

Feature Description

I have Intel i9-11900 (Rocket Lake) and i5-12600 (Alder Lake) CPUs. Llamafile runs twice as slow on "Rocket Lake" as "Alder Lake". Although "Rocket Lake" supports AVX-512. I suspect this is because "Rocket Lake" uses AVX2 instead of AVX-512. Is it possible to support AVX-512 for "Rocket Lake"?

lscpu-11900.txt lscpu-12600.txt

llamafile-bench-11900.json llamafile-bench-12600.json

Motivation

This feature will increase the performance of Llamafile for "Rocket Lake".

Possible Implementation

No response