I'm curious, where are you getting the GPUs?

in most cases the teams already had their own source or platform of choice!

Maybe build a FAQ for others on how to accomplish, or resources you can point people at?

A lot of people are using RunPod for experimental/small-scale workloads. They have good network and disk speeds and you can generally find availability for a latest-gen GPU like an L40 or 4090 if your workload can fit on a single GPU. One GPU is plenty for fine-tuning Llama 2 7B or 13B with LoRA or QLoRA. They also sometimes have availability for multi-GPU servers like 8xA100s, but that's more hit-or-miss.

If you want to go even cheaper vast.ai is a popular option. It's a P2P marketplace for individuals to rent out their GPUs. You can generally get a ~20-30% discount vs RunPod prices by using Vast, but network speeds and perf are much more variable and there's always the possibility that the host will just shut you off without warning. I also wouldn't recommend using it if you're training with proprietary data since they can't guarantee the host isn't logging it, but most of the OSS fine-tuning community publishes their datasets anyway.

Interesting, hadn’t thought of that, thank you! If you want to host end points Replicate is a great option, they also have a newer fine tuning api and solution! for raw VMs with GPUs right now it’s a bit situational and you have to try multiple different vendors tbh, also really depends on the capacity you need and which machines!!

I've done a version of this: https://news.ycombinator.com/item?id=36632397

Let me know what you'd want to see added!

That was great! thank you.

One thing I cant glean ; What GPu/kit are preferred for which type of output?

Like chat vs imaging...

Do locally run models/agents have access to the internet?

Whats the best internet connected crawler version on can use?

1. I've updated the section now: https://gpus.llm-utils.org/cloud-gpu-guide/#so-which-gpus-sh... - that should answer it. Basically 1x 3090 or 1x 4090 is an ideal set up for stable diffusion, 1x A100 80GB is an ideal setup for llama 2 70b GPTQ (and you can use much smaller GPUs or even CPUs if needed, for smaller llama 2 models)

2. No, they don't have access to the internet unless you build something that gives them access

3. I'm not sure what you're asking

Thank you!

Our pleasure :)

Hello, I think this is a great project!

If an individual is looking to contribute to the field with different training data ideas - would they need to first establish themselves and get your attention, or would there be a way to submit a proposal?

For myself compute is rather expensive, I could likely afford a few test runs for proofs of concept but beyond that it would be difficult. I've got a single 4090 so I can't run llama70b faster then 1it/s.

Thank you so much for your work and enthusiasm here!! Unfortunately, that’s roughly the best way to approach it and how most of these have worked so far, demonstrations on smaller models (e.g. 7/13B), with slightly smaller datasets, catching the eye of the community, etc. in general it’s not a bad approach to research too, prove out concepts at smaller scale before scaling up!

Totally understand! I'm sure you'd get swarmed with requests.

Are you interested in non-LLMs as well? Stable diffusion for example?

Definitely. This was kicked off in the last month or two when there just seemed to be a bit more happening in OSS LLMs, but we are prioritizing diversifying this for the next cohort!