What about Docker Desktop with GPU passthrough to a container running the LLM? That way you can be explicit on which files you share through volume mapping and the LLM is contained in the container otherwise.