The AI Horde is a project I started in order to provide access to Generative AI to everyone in the world, regardless of wealth and resources. The objective is to provide a truly open REST API that anyone is free integrate with for their own software and games and allows people to experiment without requiring online payment that is not always possible for everyone.
It is fully FOSS and relies on people volunteering their idle compute from their PCs. In exchange, you receive more priority for your own generations. We already have close to 100 workers, providing generations from stable diffusion to 70b LLM models!
Out of pure curiosity: isn't communicating via internet a bottleneck for distributed neural networks implementation? I mean the distribution of the work, not the API part
We're not doing distributed inference. We're doing distributed clustering The inference runs on individual PCs
So learning phase is distributed but then one prompt goes to one worker?
Yes, each prompt goes to a single worker. The AI Horde also doesn't do training. We only do inference.
Does that mean that with a big (parameter-wise) model and not that powerful worker it can take a long time to respond?
Or the distributed clustering prevents from choking a worker?
Yes, it could. But we have a timeout. If a worker is unreasonably slow to respond, they get put into maintenance, so we expect people to only serve what they can run.