Support local model files without duplicating into model store

[fozzee]

Feature request description

Currently when using ramalama run file://model the model file will be copied into the ramalama model store. Even when the user provides a symlink to a model file the symlink target will be followed and the actual model file will be copied into the model store.

Allow users to bring their own model files without needing to duplicate the model files.

Users have large models downloaded external to ramalama with lmstudio or other tools.

Support model files continuing to be used by external tools other than ramalama without duplicating these large model files.

Suggest potential solution

Add a command line argument to allow the user to specify whether the model file should be copied into the model store. Should result in a symlink targeting the model file being created in the model store instead of copying the model file. Something like --no-model-copy.

Other considerations:

• only for file://model type url or also for other urls using a local model cache external to ramalama model store
• support split model urls: file://models/model-00001-of-00003.gguf

Have you considered any alternatives?

Tried using a symlink to the model file with a one line change to not follow the symlink. See comment here. But that was a lazy implementation that would require the user to manually create symlinks.

Only other option I could think of would be lletting ramalama control where the model is stored and trying to get other tools to use the model in the ramalama model store.

Additional context

#1989 original related enhancement request in the context of split models

[github-actions]

A friendly reminder that this issue had no activity for 30 days.