torrent link on their official github:
https://github.com/facebookresearch/llama/pull/73/files
coomers rejoice, globohomo can suck a fat one
torrent link on their official github:
https://github.com/facebookresearch/llama/pull/73/files
coomers rejoice, globohomo can suck a fat one
Too bad you need a 4090 to run it and even then it's basically useless out of box
>4090
3060
>out of box
wrong box
https://github.com/oobabooga/text-generation-webui/issues/147#issuecomment-1454987216
can I do anything with an old 1080TI? I have 64gb of regular memory of that helps
don't believe your GPU will help, but you can still run it in RAM with your CPU, it'll just be much slower.
Can I do anything with an old RX580?
>need a 4090 to run it
You don't own a CPU, anon?
you can rent that gpu for 10cent hour or use google colab for free.
LLaMA-7B runs at great speeds on 8GB cards. Better than any other model that could run on those cards.
I'm running LLaMA-13B on 16GB of VRAM right now.
*Multicore CPU and some RAM
It took me 5 hours to download the 220gb file and 3 hours of waiting for seeders... It's just gpt3
>it's just gpt3
Is that bad?
It's continuations of prompts granted I only got the 7B model going. It's pretty decent at few shot learning and gpt3 isn't bad, but ya it's not the chatbot ur looking for
The model is pure garbage. Not worth the bandwidth it takes to download. Never trust the zuck.
>he didn't convert it
bro your repetition penalty?
It's a bare LLM. It needs further training before it can be your waifu. Think of it as a "how to brain" rather than "what to brain"
It's not garbage at all.
This is 13b with topk and repitiion_penalty
And this is princess smut with 35b.
10gb for 7b in 8bit mode and 16gb for 13b in 8bit mode.
For 35b in 8bit mode, you need 35gb vram
can you run 35B on 2x4090s?
Yes, people are doing it with 2x3090 and getting good inference speeds.
How? Just SLI it?
On my system it's closer to 39GB in total.
>topk and repitiion_penalty
What's that?
>"coomers rejoice, globohomo can suck a fat one"
>BOT user
>close to 0 knowledge in pytorch, coding, anything ml
>confirmed retard
congratulations, you downloaded llama, now what you gonna do, use it on your koboldai gui? oh you gonna generate some text with the 7B model? can't use cli because all your retarded hands can do is press buttons and thats it?
wannabe researcher go have a nice day gay
infecting the ai space with retards since august, 2022.... SD should have never became a thing...
oobabooga
>chimp noises
https://rentry.org/llama-tard
>while you still can
>GLP-3 License
>712 Forks
?
the github forks don't have the model weights in them. however the torrent basically can't be stopped and the weights are on huggingface now
Good morning Sirs can we use this to create anime girl sex story?
Yes but even in the story she will be repulsed by our stench,,, indianbros i don't feel so good:(
How much VRAM needs each models?
I fucking hate when they don't specify the hardware requirement, data "scientist" my ass
At this point Meta should publish the original checksums so that we know if it's the real model. It appears that PyTorch models can execute arbitrary code, so this might as well be a clever attack..
>anon downloads
>ai.exe
>screen instantly goes dark
>bobs and vagene
The torrent referenced in OP's post has a different infohash than the first torrent, it's also missing the llama.sh file that the first torrent has, have not checked individual file hashes yet
first torrent b8287ebfa04f879b048d4d4404108cf3e8014352
i have it working
I'll wait for improvements where I'll be able to run on 8GB VRAM quite fast and without worries. Thank you very much and see you in a month.
Your improvements are here. 7B runs on 8GB cards getting up to 30it/s depending on the card. https://github.com/oobabooga/text-generation-webui/issues/147
This is the best faceberg can do? Oy vey!!!
It's over, bro. Big Tech has it all. We are doomed. We'll have to use Chinese Meyyg4n models.
https://news.ycombinator.com/item?id=35026902
What’s “it/s”?
Tokens per second. A token is ~4 text characters.
pronouns
Iterations per second.
How good is compared to GPT-J (6B)?
I second this question, this is the big question here, and in practice, not on meme metrics used to make papers look good
Will it run on 1060 3GB?
not yet
So what I'm getting right now is that this is basically useless for now unless you specifically train it. Is this any different from anything we already have like chatgpt/bing ai or is this just hype unless you're arsed to train it out of box
2JTJTJ
>unless you're arsed to train it out of box
how is that not appealing? you can train it how you please make it into a BOT shitposter or just an erotica machine.
Training is half the fun
it's really not
>change a bit, click submit
>As an AI...
>change a bit, click submit
>As an AI...
>change a bit, click submit
>As an AI...
>change a bit, click submit
>Error, too many requests in 1 hour
it sure works fine when it finally decides to give what you want, but getting there is just fucking pain
Meta used 2048 a100s to train the model so good luck training champ
I don't know what I would even do with it so I'll pass
Christ almighty super mario