Show HN: Llama-dl – high-speed download of LLaMA, Facebook's 65B GPT model

It's been enough time since this leaked, so my question is why aren't there blog posts already of people blowing their $300 of starter credit with ${cloud_provider} on a few hours' experimentation running inference on this 65B model?

Edit: I read the linked README.

> I was impatient and curious to try to run 65B on an 8xA100 cluster

Well?

The compute necessary to run 65B naively was only available on AWS (and perhaps Azure, I don't work with them) and the required instance types have been unavailable to the public recently (it seems everyone had the same idea to hop on this and try to run it). In my other post here [1], the memory requirements have been lowered through other work, and it should now be possible to run the 65B on a provider like CoreWeave.

[1] https://news.ycombinator.com/item?id=35028738

I'm running LLaMA-65B on a single A100 80GB with 8bit quantization. $1.5/hr on vast.ai

Careful though — we need to evaluate llama on its own merits. It’s easy to mess up the quantization in subtle ways, then conclude that the outputs aren’t great. So if you’re seeing poor results vs gpt-3, hold off judgement till people have had time to really make sure the quantized models are >97% the effectiveness of the original weights.

That said, this is awesome — please share some outputs! What’s it like?

The output is at least as good as davinci.

I think some early results are using bad repetition penalty and/or temperature settings. I had to set both fairly high to get the best results. (Some people are also incorrectly comparing it to chatGPT/ChatGPT API which is not a good comparison. But that's a different problem.)

I've had it translate, write poems, tell jokes, banter, write executable code. It does it all-- and all on a single card.

That's great to hear. Thank you very much, both for reporting this, and especially for the crucial note about temperature.

In fact, sampling settings are so important and so easily underestimated that I should just pester you to post your exact settings. If you get a moment, would you mind sharing your temperature, repetition penalty, top-k, and anything else? I'll be experimenting with those today, but having some known working defaults would be wonderful. (You're also the first person I've seen that got excellent outputs from llama; whatever you did, no one else seems to have noticed yet.)

If you're busy or don't feel like it, no worries though. I'm just grateful you gave us some hope that llama might be really good. There were so many tweet chains showing universally awful outputs that I wasn't sure.

EDIT: I added your comments to the top of the README and credited you. Thanks again.

Would you mind publishing your notes/learnings once you gain enough understanding of this model?

Absolutely! I'll make sure to leave a comment here for you whenever something gets written up so you don't miss it.

Getting "as good as davinci" on a single A100 is groundbreaking work. Facebook and the community should both be credited here -- maybe llama-int8 would've been created even if the model hadn't leaked, but I don't think it would've happened so quickly. Everyone is doing phenomenal work, and it's so amazing to see it all come together.

But, we'll see. Going to try it myself soon.

Long ago, I cloned OpenAI's API: https://github.com/shawwn/openai-server -- my plan is, once I get it running, I'll try to host it somewhere so that anyone can play with it. I assume it'll be quickly swamped, but it's still an interesting challenge; some basic load balancing should make it scalable across several A100 instances, so there's no reason we can't just roll our own OpenAI API.

Seconded. Do write it up.

I see vast.ai listing interruptible instance with a single A100 80GB at $1/hour which is pretty reasonable. ChatGPT plus is $20/month which would be roughly 20 hours of use and I wont't be lectured like I'm in a kindergarten or something.

A bonus point would be to make the writeup accessible for AI challenged developers. Asking for a friend.

I would like to support this request for AI challenged developers :)

For things like these, I always wonder: How much slower would it be to run such a model on a CPU? I mean, clearly a lot less interactive, but is it possible at all? Could it be chopped up and "streamed" to a GPU with less memory halfway efficiently? What is the bottleneck currently on GPUs, memory bw or compute?

On a CPU I'd estimate it would get a maximum of around 5 tokens per second (a token being a sub-word token, so generally a couple of letters). I suspect it'd be more like 1 token per second on the large model without additional optimisation.

Yes models can be split up. See eg Hugging Face Accelerate.

That's actually a lot better than I would have thought. Almost usable, and a good exercise in patience.