GitLab

When CUDA peer access is enabled, multi-gpu inference will produce
garbage output. This is a known bug of llama.cpp (or nvidia). Until the
upstream bug is fixed, we can disable CUDA peer access temporarily
to ensure correct output.

See #961.

pin to 9e70cc03229df19ca2d28ce23cc817198f897278 for now since
438c2ca83045a00ef244093d27e9ed41a8cb4ea9 is breaking

- this makes it easier to see that the subprocess is associated with ollama

* subprocess improvements

- increase start-up timeout
- when runner fails to start fail rather than timing out
- try runners in order rather than choosing 1 runner
- embed metal runner in metal dir rather than gpu
- refactor logging and error messages

* Update llama.go

* Update llama.go

* simplify by using glob

* enable packaging multiple cuda versions
* use nvcc cuda version if available

---------
Co-authored-by: Michael Yang <mxyng@pm.me>

* linux gpu support
* handle multiple gpus
* add cuda docker image (#488)
---------
Co-authored-by: Michael Yang <mxyng@pm.me>

* remove c code
* pack llama.cpp
* use request context for llama_cpp
* let llama_cpp decide the number of threads to use
* stop llama runner when app stops
* remove sample count and duration metrics
* use go generate to get libraries
* tmp dir for running llm