- cross-posted to:
- hackernews@derp.foo
- technews@radiation.party
- technology@lemmy.ml
- cross-posted to:
- hackernews@derp.foo
- technews@radiation.party
- technology@lemmy.ml
Context: Falcon is a popular free LLM, this is their biggest model yet and they claim it’s now the best open model in the market right now.
Am I the only one that really hates huggingface? It’s such a confusing website to use, and most of the time the things just spit out errors. And I’m never sure if this is just a free thing, or a demo, or something that I have to pay.
Like what is the proper way to use that thing?!
Huggingface takes a bit of getting used to but it’s the place to find models and datasets, imo it may be one of the most important websites on the internet today.
But what exactly is a model? Can I download and run it? Do I need to access it through their API? Do I need to pay for some server that has all the needed software already running on it? It seems open and not open at the same time.
It’s a bit complex, and you can find a better answer elsewhere, but a model is a set of “weights” and “bias” that make up the pathways of the neurons in a neural network.
A model can include other features but at its core it gives users the ability to run an “ai” like gpt, though models aren’t limited to only natural language processing.
Yes, you can download the models and run them on your computer, generally there will be instructions in each repository, but in general it involves downloading the model which can be very large and running it using an existing ml framework like pytorch.
It’s not a place for the layman right now, but with a few hours of research you could make it happen.
I personally run several models that I got through huggingface on my computer, llama2 which is similar to gpt3 is the one I use the most.
The model is the brain, you use something like Kobold.cpp to load the model. You’ll have to work with the settings and try different models to get the right load on your GPU.
deleted by creator
It’s made for researchers and engineers. Nothing is packaged in a form to simply download and run on a stock PC. It assumes a high level of comfort configuring Python environments, GPU drivers, and GPU compute backends like CUDA.
If you don’t know what all of that means, you would be better off looking downstream to projects like GPT4All that package some of this stuff into a simple installer that anyone can run.
As for Falcon in particular, you will not be able to run this on any consumer hardware. It requires at least 160GB of memory, and that’s GPU memory ideally. The largest consumer GPU on the market only has 24GB.