Speed boost for NUMA aware systems #238
chhu
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Just FYI, after building on my aged 1st gen Threadripper system (debian) I got 26 t/s output, but when I added
--numa distribute
to the llama-cli call it jumped up to 40! Kind of makes sense but some may have NUMA systems and don't know / care about its potential. Be aware that if you try that you need to evict the cache first. With sudo bash:echo 3 > /proc/sys/vm/drop_caches
Not sure why there are no better defaults for numa awareness or how / if this works on windows.
Beta Was this translation helpful? Give feedback.
All reactions