HN discussion | Twitter announcement
Update (March 7, 3:35 PM CST): Looking to inference from the model? See shawwn#1 (comment) to use the improved sampler. (Facebook's sampler was using poor defaults, so no one was able to get anything good out of the model till now.)
Update (March 5, 12:52 PM CST): @anitakirkovska let us use their fabulous llama photo. If you happen to like the new header image as much as I do, be sure to check out their AI newsletter and their tweets about us.
Update (March 5, 9:51 AM CST): HN user MacsHeadroom left a valuable comment:
I'm running LLaMA-65B on a single A100 80GB with 8bit quantization. $1.5/hr on vast.ai
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.
This repository contains a high-speed download of LLaMA, Facebook's 65B parameter model that was recently made available via torrent. (Discussion: Facebook LLAMA is being openly distributed via torrents)
It downloads all model weights (7B, 13B, 30B, 65B) in less than two hours on a Chicago Ubuntu server.
real 98m12.980s
user 8m8.916s
sys 5m7.259s
This works out to 40MB/s (235164838073 bytes in 5892 seconds).
Personally, I just wanted to curl
the weights instead of dealing with a torrent. The fact that it's several times faster was just a nice bonus.
To download all model weights, cd
into the directory you want them, then run this:
Linux:
curl -o- https://raw.githubusercontent.com/shawwn/llama-dl/56f50b96072f42fb2520b1ad5a1d6ef30351f23c/llama.sh | bash
Mac:
brew install bash
brew install wget
curl -o- https://raw.githubusercontent.com/shawwn/llama-dl/56f50b96072f42fb2520b1ad5a1d6ef30351f23c/llama.sh | $(brew --prefix)/bin/bash
(Sorry mac users; they use some array syntax in the script that isn't supported on the version of bash that ships with Mac.)
Running random bash scripts generally isn't a good idea, but I'll stake my personal reputation on the fact that this link is safe. (It points to a specific SHA-1 hash rather than https://raw.githubusercontent.com/shawwn/llama-dl/main/llama.sh so that it's still safe even in the event that my repo or account got compromised.)
219G (235164838073 bytes) total. Here's a file list with sizes for each.
I ran this:
mkdir LLaMA
cd LLaMA
time curl -o- https://raw.githubusercontent.com/shawwn/llama-dl/56f50b96072f42fb2520b1ad5a1d6ef30351f23c/llama.sh | bash
cd ..
webtorrent 'magnet:?xt=urn:btih:b8287ebfa04f879b048d4d4404108cf3e8014352&dn=LLaMA&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'
Webtorrent began seeding immediately, which means every file is identical to what you would've gotten via the torrent. So this is just a faster version of the torrent.
Roughly 3.6x. As of March 4 2023, the torrent seems to download at around 11MB/s, which implies a download time of around 6 hours. (Help seed it, if you can.)
I doubt it. This is using the download link that was leaked in the original torrent. (i.e. the leaker accidentally leaked their own unique download link that Facebook sent them.)
Technically, it may be illegal to knowingly use a private download link that was intended for someone else. Realistically, Facebook would risk their ML reputation by going after people who are merely trying to use what they themselves advertise as "open source."
Update: Facebook shut off the link a couple hours after this repo went live. I mirrored everything to R2 and updated the script to point to that instead.
Note that LLaMA was released under a "non-commercial bespoke license". Interestingly, Nvidia had a similar arrangement for StyleGAN, but that didn't stop Artbreeder from using it anyway. Nvidia never seemed to care enough to go after them. But if you launch your own OpenAI API and start charging money, don't be surprised when Facebook's lawyers come knocking.
I was shocked that this script was distributed with the original torrent, and that no one seemed to notice (a) that it still works, and (b) is almost 20x faster than the torrent method. I was impatient and curious to try to run 65B on an 8xA100 cluster, so I didn't want to wait till tomorrow and started poking around, which is when I found this. I decided to just tweet it out and let you, fellow scientists and hackers, enjoy it before Facebook notices and shuts it off.
"Power to the people" is an overused trope, but as a research scientist, I feel it's important to let individual hackers be able to experiment with the same tools, techniques, and systems that professional ML researchers are fortunate to have access to. This is a tricky situation, because at some point between now and 10 years from now, this might become dangerous -- AI alarmists often ask "Would you want random people experimenting with nuclear weapons in their basement?" My answer is "No, but we're not there yet."
Word on Twitter is that LLaMA's samples seem worse than GPT-3 by a large margin, but then I realized no one has really been able to try the full 65B model yet, for a combination of reasons. (Mostly lack of access to 8xA100 hardware.) So I decided to try it out for myself and see.
Even if it's GPT-3 level, the fact is, LLaMA is already openly available. The torrent isn't going anywhere. So my own thoughts on this are mostly irrelevant; determined hackers can get it themselves anyway.
But for what it's worth, my personal opinion is that LLaMA probably isn't OpenAI-grade -- there's a big difference between training a model in an academic setting vs when your entire company depends on it for wide-scale commercial success. I wasn't impressed that 30B didn't seem to know who Captain Picard was.
People have already started decrying this leak as dangerous. But everyone used to say the same thing about 1.5B. (In fact, the allure of 1.5B's grandiose claims was what drove me to take ML seriously in 2019.) Turns out, four years later, no one really cares about 1.5B anymore, and it certainly didn't cause wide-scale societal harm. I doubt LLaMA will either.
2023 will be interesting. I can't wait for 2024.
Signed with love,
Shawn Presser
twitter: @theshawwn
HN: sillysaurusx