Dangerous soft memory leak resulting from closing alpaca when using ollama instance

[FuwwyCodingCOwO]

When alpaca dies, either by crash or termination, it causes the running ollama instance to be adopted by systemd --user. I am referring to it as a soft memory leak because it can still be freed via passing a signal to the running ollama instance. Alpaca will also start a new ollama instance as a child of itself, which makes the memory usage worse. I am currently not aware of anyone else experiencing this issue.

https://www.youtube.com/watch?v=jcJv12JPbFA

[FuwwyCodingCOwO]

This wording is arguably bad

Alpaca will also start a new ollama instance as a child of itself, which makes the memory usage worse

I meant to say that starting new instance of alpaca will result in a new ollama instance being started as a child process. This makes the memory issue worse, especially with larger models.

[FuwwyCodingCOwO]

I had a thought about possible solutions. Instead of ollama being a child process, start it as a daemon. If ollama doesn't have contact with alpaca after n minutes it would cause the daemon to stop. Another idea is that ollama watches the parent process to see if it has changed or not.