bigscience-workshop/petals ? reverse-engineered prompt

Build me a Python project that lets me use really big language models like Llama 3 or Mixtral, even without a huge GPU. It should be based on the Petals library, which works by distributing the model across a network of people, kind of like BitTorrent for AI.

The main goal is to run inference and basic fine-tuning on these models by connecting to the public 'swarm' of users. Please set up a simple demo script that shows how to load a model—let's use Llama—and generate some text. Also include a second example for how I could do a basic fine-tuning task, like the ones shown in the Petals tutorials. I'd eventually want to run this in a private network for sensitive data, so please structure the code in a way that makes it easy to switch from the public swarm later on.