-
Notifications
You must be signed in to change notification settings - Fork 1.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add llama.cpp backend #2723
base: main
Are you sure you want to change the base?
Add llama.cpp backend #2723
Conversation
# Conflicts: # Cargo.lock
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks @mfuntowicz ! I can't get a correct generation using the Docker so I think something is wrong with tokenizer. I'll dig deeper later today!
Co-authored-by: Hugo Larcher <[email protected]>
Co-authored-by: Hugo Larcher <[email protected]>
… the size of the generated core allocation
* feat: Fix Cmakelist to allow building on Darwin platform * fix: Fix tokenizer in llama.cpp Dockerfile
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nice !
None => break, | ||
Some((ctx, sender)) => { | ||
if let Err(e) = backlog.send((ctx, sender)).await { | ||
todo!("What do we do") |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks like it can only happen id the channel is closed (all receivers are drop), so I guess you can break the loop and just return / exit
Co-authored-by: Corentin REGAL <[email protected]>
This PR is an initial implementation of llama.cpp as potential backend for TGI.
It mostly targets CPU inference in a single/multi stream scheduling fashion, potentially spawning multiple instances of the same model over a non-overlapping subset of the CPU cores.
The current implementation only allows a single request to be running on a working, this constraint will be removed later on.
The current implementation also dupplicate the weights for each worker, this constraint can potentially be removed later on.