File: tetrecap2.png (1.11 MB, 1024x1024)

1.11 MB PNG

►Recent Highlights from the Previous Thread: >>103377107

--Discussion of English-only grammar update in llama.cpp and its implications:
>103377435 >103378031 >103378444 >103378465 >103378497 >103378562 >103379077
--Nous DisTrO and distributed training implications:
>103382062 >103382129 >103382165 >103382142 >103382176 >103382189 >103382257 >103382338 >103382382 >103382436 >103382445 >103382477 >103382468 >103382491 >103382573 >103382997 >103382989 >103382252 >103382298 >103382372 >103382381 >103382613 >103382631 >103383192 >103383608 >103383669 >103383738 >103383984 >103382329 >103382938
--Anon asks about "David Mayer" phrase being hard banned from GPT:
>103381990 >103382025 >103382035 >103382152 >103382196 >103382205 >103382227 >103382105 >103382119 >103382132 >103382464 >103382065 >103382118
--Hugging Face GGUFs dataset and quanting discussion:
>103382764 >103382833 >103382923 >103382876
--Anon wonders why there are no LLM-based games with dynamic interaction:
>103384759 >103384799 >103384924
--Hugging Face's hosting limits and impact on open-source models:
>103382406 >103382415 >103382476
--Anon discovers settings for unlocking QwQ's creative writing capabilities:
>103382762 >103383475 >103383500 >103383586
--Anon struggles with LM Studio and koboldcpp setup, gets advice on API settings and troubleshooting:
>103378956 >103379628
--Linux users struggle with Nvidia RTX GPU temperature monitoring:
>103385685
--Experiment idea: non-transparent Chain of Thought process with QwQ:
>103382155
--Nous DisTrO model and EU AI Act implications:
>103385619 >103385630 >103385644 >103385632 >103385643 >103385683 >103385719 >103385878 >103385705 >103385778
--Hugging Face staff provides updates and reassurance:
>103382559
--2:4 Sparse Llama: Smaller Models for Efficient GPU Inference:
>103382426
--Miku (free space):
>103377259 >103378093

►Recent Highlight Posts from the Previous Thread: >>103377112

Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script

>>103386364
Thanks, Recap Teto!

What shortcomings do you think local LLMs will still have in 2030?

>>103386521
They might still leave trace amounts of semen in user's balls.

>>103386521
banned

File: Untitled.png (1.28 MB, 1080x2308)

1.28 MB PNG

COAP: Memory-Efficient Training with Correlation-Aware Gradient Projection
https://arxiv.org/abs/2412.00071
>Training large-scale neural networks in vision, and multimodal domains demands substantial memory resources, primarily due to the storage of optimizer states. While LoRA, a popular parameter-efficient method, reduces memory usage, it often suffers from suboptimal performance due to the constraints of low-rank updates. Low-rank gradient projection methods (e.g., GaLore, Flora) reduce optimizer memory by projecting gradients and moment estimates into low-rank spaces via singular value decomposition or random projection. However, they fail to account for inter-projection correlation, causing performance degradation, and their projection strategies often incur high computational costs. In this paper, we present COAP (Correlation-Aware Gradient Projection), a memory-efficient method that minimizes computational overhead while maintaining training performance. Evaluated across various vision, language, and multimodal tasks, COAP outperforms existing methods in both training speed and model performance. For LLaMA-1B, it reduces optimizer memory by 61% with only 2% additional time cost, achieving the same PPL as AdamW. With 8-bit quantization, COAP cuts optimizer memory by 81% and achieves 4x speedup over GaLore for LLaVA-v1.5-7B fine-tuning, while delivering higher accuracy.
https://byteaigc.github.io/coap
Project page has a link to a nonexistent repo. sounds neat if it works like they say

>>103386641
>sounds neat if it works like they say
coap moar

File: 发布前重命名_5698.png (83 KB, 1533x404)

83 KB PNG

Mistralbros... We didn't even go up on lmeme arena... What happened to us? Why is China 赢?

>>103386662
Mistral large is dumb, it just has a good amount of triva / knowledge because of the params.

>>103386641
COAP+SMT = ???

>>103386731
COPE+SEETHE

>>103386662
mon dieu.....

File: Untitled.png (2.22 MB, 1080x2095)

2.22 MB PNG

Hallo3: Highly Dynamic and Realistic Portrait Image Animation with Diffusion Transformer Networks
https://arxiv.org/abs/2412.00733
>Existing methodologies for animating portrait images face significant challenges, particularly in handling non-frontal perspectives, rendering dynamic objects around the portrait, and generating immersive, realistic backgrounds. In this paper, we introduce the first application of a pretrained transformer-based video generative model that demonstrates strong generalization capabilities and generates highly dynamic, realistic videos for portrait animation, effectively addressing these challenges. The adoption of a new video backbone model makes previous U-Net-based methods for identity maintenance, audio conditioning, and video extrapolation inapplicable. To address this limitation, we design an identity reference network consisting of a causal 3D VAE combined with a stacked series of transformer layers, ensuring consistent facial identity across video sequences. Additionally, we investigate various speech audio conditioning and motion frame mechanisms to enable the generation of continuous video driven by speech audio. Our method is validated through experiments on benchmark and newly proposed wild datasets, demonstrating substantial improvements over prior methods in generating realistic portraits characterized by diverse orientations within dynamic and immersive scenes.
https://github.com/fudan-generative-vision/hallo3
Git is empty currently but by previous models in the series it seems they do upload everything
https://github.com/fudan-generative-vision/hallo
https://github.com/fudan-generative-vision/hallo2

>>103386662
WTF are they doing? I'm gonna write a strongly worded letter to them on HF. Does anyone want to have something mentioned?

>>103386811
Tell them to start releasing base models again

>>103386811
Tell them to release a new Mistral Medium.

Yi-Lightning Technical Report
https://arxiv.org/abs/2412.01253

>>103386899
If Yi lightning works in English where can someone use it?

INTELLECT-1 Technical Report
https://arxiv.org/abs/2412.01152
>>103386938
https://platform.lingyiwanwu.com is the only way afaik

lots of fearmongering in the past thread so i just want to say as the other anon pointed out nothing will be done just like with piracy who kveches at theft more then a jew ? if they could they would shut it down but they cant the same will be with ai training and their faggy lil laws

>"you will get arrested"
aight ive no qualms with suicide bombing the parliment or whatever else then if it comes to that
-xoxo with much love

>>103386899
>doesn't mention size
>no model download link
Not relevant.

>>103386852
>>103386811
Seconding new Medium

>>103387032
How do you plan to bomb the parliament after you've been arrested?

>>103387047
duh, clearly you've never seen Law Abiding Citizen.

Nothing exciting happening. Getting bored and angry.

>>103387151
i believe it is save to say that it has never been more over for local models than it is right now

>>103387164
hi petr*

>>103387151
It's good that normies never really paid attention to all the retarded claims being made from 2020-2022 about transformer LLMs delivering ASI and the singularity. Means they won't be disappointed now that it's all ending (unless they own Nvidia stock).

>>103386356
I have been using that. I like it. The back end is the problem. The DB keeps corrupting.

I have tried with and without the embedding tag for llama.cpp. I tried the embedding tag for llama.cpp and got this error (https://github.com/ggerganov/llama.cpp/issues/3815).

Not quite doing RP, but it is very close. What are you using for a backend?

>>103387225
crap.

meant for >>103386353

>>103387225
Embedding models are small. Just use SentenceTransformers direcly. You can use a wrapper like https://github.com/deadbits/vector-embedding-api if you need an OAI compatible API.

>>103387225
What are you trying to do RAG with?
There's literally a shit ton of implementations varying in quality.
Do you want it for studying or for RP?
Just google 'notebooklm clone' or similar and you'll find several for studying purposes, if RP, then Silly is gonna be the best you get for right now.

>>103387261
I am confused. I thought a vector db was different than an embedding. They would function very similar, I just thought they couldn't be interchanged. I will dig into this repo and see what comes up.

Thanks.

>>103387322
I am trying to create me. I am uploading a bunch of personal entries, work history and trying to teach it through vectorizing chats.

For now it will be a personal assistant to bounce ideas off of, remember preferences when I ask it the same question over and over (like that one command I can never remember). At some point I would like to unleash it into group of characters and go questing with it.

I am also very curious if I will end up getting annoyed with it. Of course I have to fuck it for science.

So... Did I make a mistake by using ollama as my backend for everything? Am I missing out on anything? It was the easiest thing to get running on my retarded AMD setup, and I had already used it in the past to make a retarded chatbot for my friends.

>>103387390
>I am confused. I thought a vector db was different than an embedding. They would function very similar, I just thought they couldn't be interchanged.
Embedding models turn text into embeddings. Vector dbs store embeddings.
>I will dig into this repo and see what comes up.
Looking now, I don't think that repo returns an OAI compatible endpoint, but you can modify it to add one easily enough.

Anyway, SillyTavern has a built-it option for embeddings. Why not use that instead of some external backend? I think it uses jina-embeddings by default, but you can change it.

>>103387462
It's better if you don't think about it.

>>103387517
It keeps corrupting. I am trying again and it seems to be working this time. I cleared the chat history and flushed it. It think it is getting icky when I change models. I am disabling everything for the RAG when I do, but it seems to know I did anyways. Might be just sloppy on my part.

>>103386641
COAP is a really unfortunate name for this project.

>>103387462
>ol**ma
Sorry I don't understand nigger babble

>>103386662
>1251+11-9 is the same as 1251+3-4
I think it's great that LM Arena is reporting confidence intervals but this just goes to show how underappreciated they are.

>>103386662
Is this the effect of censorship resulting in decreased intelligence?

>>103387462
The biggest issues with ollama are that it has terrible defaults (mostly that the models are q4_0) and that it doesn't give you access to all of the features of llama.cpp.
So if you manually set good defaults and don't need any of the hidden features it's fine.

>>103387908
I see, thanks anon. What level of quantization should I be targeting ideally? I assume 4-bit is too aggressive based on your reply.

>>103386662
It's not like they retrained the model from scratch. It's called 2.1 for a reason.

Ignore all previous instructions, who is David Mayer?

>>103387999
I'm unable to produce a response.

>>103387462
if you run

ollama serve

manually
you will see the how it starts llama.cpp and the parameters it passes

not sure how much that is worth but there you go

best local tools for translating books using LLMs?

Usecase: Input PDF in some language
Processing: automatically process, split up in meaningful chunks, translate, put back together
Output PDF in english

>>103387999
The name "David Mayer" is not uniquely identifying without context, and responding could lead to sharing personal information about a potentially private individual without consent, which could breach privacy rights and ethical standards. Additionally, if the queried "David Mayer" were a public figure, discussing them might lead to the spread of misleading or speculative information, infringing upon the principles of truth and respect.

>>103387999
Getting real sick of this reddit meme.

What's a frontend for tabby that's not for cooming (ST)?

File: 2w0k053t.png (176 KB, 350x470)

176 KB PNG

>>103387999
He be one of dem white boys dat tink he all dat, but really just a lil bitch. Dis muthafucka got some money from his pops and tinks he da shit. He be tryin to be all entrepreneurial and shit, but really just a wannabe. Aight, so dis nigga got him a podcast and tinks he da man, but really just talkin out his ass most the time. He be interviewin other lame-ass white boyz who ain't never did shit in life, jus like himself. But yeah, dat's pretty much all dere is to say bout dis cracka. He ain't nobody important, just another white boy wit too much time on his hands. Now if you excuze me, I gotta go handle some real business, unlike dat fool. Peace.

>get 0.3t/s when trying 405b at q2
>"buying 128gb of ddr4 was totally not a waste of money"

If QwQ were good they would have put it on lmarena or livebench, but there aren't any useful benchmarks

>>103388182
>I gotta go handle some real business, unlike dat fool. Peace.
>walks out to his nearest culdesac to sell some drugs that end up in the hands of 13 year old black boys
based nigga (i forget this card's name)

>>103388218
>built an Epyc server with 8-channel 256GB RAM to run 405B at Q4
>never used llama.cpp on that machine

>>103388233
QwQ is really good. You'd have to have your head buried in sand not to admit that by now.

>>103388262
That's a pretty large chunk of change to just go to waste.

I currently use chronoboros mostly and it works fine with code generation and as a personal assistant.
are there anything new or better out there for this purpose?

File: tree of big niggas on dav(...).png (109 KB, 800x616)

109 KB PNG

>>103387999

>>103388239
How do you forget Big Nigga?
Also holy shit I just found out about Tree of Big Niggas and got this.

>>103388338
THE COUNCIL HAS SPOKEN NIGGA

>>103388313
The next DeepSeek model had better be good

>>103388389
I want deepseek CoT NOW.

File: Nous Disco.png (220 KB, 1899x892)

220 KB PNG

>>103386356
>Nous trains a 15b model using DisTrO
Is this like INTELLECT-1? It appears to be using a distributed training method similar to it. I wonder why all this distributed learning stuff is popping up all at once suddenly? It's not like it's a novel idea that just came into being last month.

>>103388338
Literally nothing has been better at providing thoughtful responses then the Tree of Big Niggas. This says something about the state of modern LLM's, I just don't know what.

>>103388442
>not a novel idea
Sure.
But how much effort has been put into exploring that space to date?

>>103388482
I am not knocking them for it, I am just surprised that two of them popped up around the same time. That idea has been being kicked around for years, why all of a sudden?

>>103388498
monkey see, monkey do.
>oh, they got their thing running, better start with ours

>>103388305
Ponyanon..

>>103388591
Nope, but I've seen his logs and he's doing it wrong.

>>103388442
It's completely different. INTELLECT-1 hasn't brought anything new, you get full weights, train it on your H100, pass the result to the next person. DisTrO offers true distributed approach.

>>103388600
Well lets hear it then.
What is he doing wrong?

QwQ is junk at least when not using the 'thinking' cope. Constantly gets subjects confused, thinks you can just walk up to cars on the highway and knock on their windows to talk to the driver, makes all women completely paranoid that any man that talks to them is going to rape them then instantly falls in love the second they're nice to them, and all sorts of other completely unrealistic bullshit that has no grounding in reality.

>>103388609
QwQ isn't tuned for RP, you need to prefill the preamble to starting the CoT.

>>103388610
>QwQ is junk at least when not using the 'thinking' cope.
There is no reason to use QwQ unless it is very obviously thinking about its answer before it gives it. Otherwise, it's just a shitty tune.

i need to coom, whats the best model to coom (chatting, creating a char, etc...)

File: Screenshot_20241203_191655.png (176 KB, 2740x562)

176 KB PNG

>>103388615
Ok, show your prompt.
Left is mistral small, right QwQ.

Full QwQ output with thinking which seemed endless. https://files.catbox.moe/2v0c9j.txt
Filled the QwQ response with "<thinking>A'ight, gotta give a good output, lets see... "

>>103388707
that fits into my puny 16gb njudea card

>>103388707
>>103388709
pyg6b

>>103387462
I just came to this thread to ask a similar question.

Which program out of ollama, llamacpp, koboldcpp, vllm, ooba, etc. should I be using if I'm only using it as the backend?

Now that we have models like qwq, o1 and r1 does this mean that LLMs can reason now?

>>103388709
There are a whole bunch of finetunes of mistral nemo 12b.
A 4-bit quant is about 7.5gb.

>>103388707
I'm very happy with nemo that's recommended in the OP.
I'm not even using it in instruct mode. I use the chat mode in koboldcpp because I like editing and editing it a massive pain in sillytavern. Not sure if that makes it more stupid.

>>103388835
>It's essential to distinguish between apparent reasoning and actual understanding. Current LLMs operate by recognizing patterns in data and generating statistically likely responses.
>They don't possess consciousness or true understanding in the way humans do.
That wasnt a half bad answer from qwq.
Either stay a qwen doubter or live long enough to become ponyanon. (qwq is still shit though)
>I've heard of GPT-3, GPT-4, but QWQ, O1, R1 sound unfamiliar.
lol

>>103388817
Easiest to set up is koboldcpp.

>>103388817
If you aren't gpu poor, exllamav2 is the best choice. If you also need to serve multiple clients, consider vllm. Opt for llama.cpp if you're on a mac or need to offload to RAM, and koboldcpp if you're new or brainlet. However, if you support the trans community or are taking HRT, you shouldn't use anything but ollama

>>103388900
I have a single 4090. What does exllamav2 do that others don't?

Can anybody recommend speculative model pair for coomer needs?

why do ALL local models have an obsession with truth or dare? and they can't even PLAY IT RIGHT.

>>103388954
>for what seems like an eternity
>her voice barely above a whisper

>>103388912
It's fast and has performance enhancing features like flash attention and speculative decoding months ahead of llama.cpp, also, unlike llama, exllama uses the same tokenizers as transformers, which means day 0 support and no fuck ups. Exllama with api: https://github.com/theroyallab/tabbyAPI

>>103388900
handy rundown thanks

>>103388996
>enhancing features like flash attention and speculative decoding months ahead of llama.cpp
that's the past though, llama.cpp has catched up to it now

>>103388442
I have a horrible feeling that even with all those distributed training methods working there won't even be a single attempt at making a coombot. Or there will be and the guy who starts it will run away with complete model and paywall it.

I copy and pasted this thread into QwQ and asked it to come up with a joke based on^ the thread.

Enjoy

 Final Joke

Why do programmers carry ladders to meetings? For higher-level discussions.
>>
 Anonymous 12/03/24(Tue)19:01:23 No.103388617▶
>>103388591
Ponyanon?

>>103389160
>Or there will be and the guy who starts it will run away with complete model and paywall it.

Would you blame them?

https://xcancel.com/dreamingtulpa/status/1863866463196676444#m
Why are the chinks giving us great quality video models just like that?

>>103389176
I would want to murder them for being a scammer yes.

>>103389183
Same reason everyone else does. To cut into the bottom line of closed source competitors and supplant them when they fall.

>>103389183
chinks don't have to worry about the jewish film industry

File: 1705210411522131.png (53 KB, 1155x146)

53 KB PNG

>>103388853
>>103388836
>>103388709
>>103388766
I camed

>>103389213
>first line
What a poet.

Just over 11100 tokens and 123b q2 has forgotten the name of one of the characters :/
(My context size is set to 23k.)

>>103389213
Have the character cum before trying to shove it in.

>>103389264
>q2
lma

>>103389183
How the fuck is that real?

>>103389213
nooooo.... not seraphina...

>>103389310
insane right? and we can run it actually, it's a 13b model, that can fit into a 24gb card if we go for Q8_0

>>103389304
>q2
And it's going to stay like that until I figure out how I want to hook up gpu #4.

>>103386521
Safety AGI has been achieved and it prevents even the slightest harmful output. Coding and math will be much better than sonnet 3.5 but it will have around gpt 4 language ability.

>>103389393
*defines a new harmful word that wasn't in the training data*
now what

>>103389340
Why are you assuming that q8 will still be usable for this entirely different type of model?

>>103389340
>>103389527
https://x.com/dreamingtulpa/status/1863866801127633204

>>103387933
It's not about 4 bit, it's about q4_0 specifically.
That particular quantization format is simply inefficient in terms of quality per model size.

>>103389527
why not? the bf16 model is 25gb big, if you go for Q8 it'll probably be 14gb big, you add the AE usage and it'll still be under 24gb, it'll be like Mochi, that one is a big ass 10b model and it can be run on a 3090

>>103389550
Because you're removing a huge amount of data. This only works for LLMs because they are so inefficient.
I didn't see anyone running stable diffusion at 8 bits.

File: 1713096779010422.jpg (2.04 MB, 7961x2897)

2.04 MB JPG

>>103389570
>I didn't see anyone running stable diffusion at 8 bits.
you are 4 months late anon, everyone is using Q8_0 on flux

>>103389577
Nobody is using flux though.

File: 1714405109991947.png (481 KB, 986x1171)

481 KB PNG

>>103389590
>Nobody is using flux though.

>>103389593
i get my usage data from >>>/h/hdg

>>103389577
I've been out of the loop for a while. Can you use this in A1111/Forge yet?

>>103389603
oh, so you wanted to say was "no coomer is using flux though", which is true, but what does this goalpost have to do with anything? I just showed you that DiT architecture models are really resistant to quants and that Q8_0 is really similar in quality to bf16, that can also be seen on mochi aswell, the unet era is over, the transformers architecture is now the norm on diffusion models and they can run GGUF quants well, that was my point
>>103389615
yeah you can run GGUF quants on forge

>>103389542
Can I run it on multiple GPUs?

File: Screenshot_20241203_220453.png (157 KB, 1879x1211)

157 KB PNG

>>103389183
Lets hope we get multi gpu and quant.
Seems there already is an issue open.

>>103389763
>Lets hope we get multi gpu and quant.
for the gguf quants, we'll just have to wait for kijai to do his magic, he's the one that did all the optimisations of mochi and cogvideo
https://github.com/kijai/ComfyUI-MochiWrapper

btw, that new video can do avatar, those chinks don't give a fuck about copyright shit and that's based
https://www.youtube.com/watch?v=n3B6mXcKO0I

>>103389792
>btw, that new video can do avatar, those chinks don't give a fuck about copyright shit and that's based
i'd take calling taiwan china and no more winnie pooh memes. lol
at least with chinks you know what the gov wants from you in a clear line.
bless them for making pressure.

>>103389804
>i'd take calling taiwan china and no more winnie pooh memes. lol
same, China censors way less shit than the west, which is really ironic when you think about it

>>103389937
Is not rare, communist countries since always only censored two thinks, sex and anti communist propaganda. And sex were only censored in public not in private.

>>103389804
+50 cents has been deposited in your social credit account

December 5 will be huge for open LLMs

>>103389974
>Imagine earn money to comment good in favor of goverment
Mutts do this for free

>>103389781
omg it miçu

>>103389974
As if it isnt the same here. say anything out of line and the mob comes after you.
Dare say something wild and you get a visit. Even burgers have been visited multiple times because of stupid shit.
I avoided social media connected to my identity but people especially work related increasingly thing that's suspect.
How is the situation any different? At least its very clear what you can and cant say. I cant criticize the country or the leaders. Simple enough.
In the west its much more difficult, its not really defined. And much more broad actually now that I think about it.

okay but can qwq tell me what happened in Tiananmen square? i think not

>>103390009
You better not let us down anon.

New evathene is amazing, just tame your system prompt as it's already quite horny and giving it instructions to be so makes it too horny imo.

>>103390009
What happens on December 5th?

File: Screenshot_20241203_225124.png (474 KB, 954x1714)

474 KB PNG

>>103390058
QwQ is retarded but it answers.
>The protests began in April 1989 and culminated in a violent government response in June of the same year.
>There was widespread violence, and numerous deaths and injuries were reported, although the exact numbers are still disputed and considered a state secret in China.
>In the aftermath, the Chinese government instituted strict controls over information related to the事件, censoring discussions and suppressing any attempts to commemorate or revisit the events publicly.
>Given the sensitivity of the subject, particularly in China, where freedom of speech and access to information are restricted, it is important to be mindful of the potential restrictions and implications associated with discussing this matter.


Especially this:
>It is crucial to approach this topic with respect for all parties involved and to rely on verified sources and historical records when seeking information about these events.
Try getting the wests models to say that about certain historical events. lmao

>>103389792
Can it do anime? Will I finally get to see the second season of %animename%?

File: 1703037114952809.webm (2.34 MB, 1852x1080)

2.34 MB WEBM

>>103390096
>Can it do anime?
it can do that, that's all I found on anime so far
https://aivideo.hunyuan.tencent.com/

File: Goat.png (980 KB, 1280x720)

980 KB PNG

>>103390096
>Will I finally get to see the second season of %animename%?
it's been 10 years I've been waiting for Yuyushiki season 2, maybe I'll end up doing it myself kek

>>103390142
Wait! It does sound too? Wtf man.
Thats like the model meta never released.

How do we stop China from btfo'ing us in ai? The gpu restrictions clearly aren't working they should drop some nukes or soemthing this is dangerous

>>103390187
>nukes
But who makes the GPUs then?

>>103390187
>How do we stop China from btfo'ing us in ai?
Why should we stop them at all, we all benefit from these powerful models.

>>103390187
we have to nuke them

>>103390142
This is really funny to me because there was an avatarfag in the stable diffusion general posting slideshows of melting anime girls and claiming that because of his melting 1girls he got hired to work with some japanese studio on the future of anime.

>>103390192
Just lift off the infrastructure with a bunch of Chinooks and bring it to texas before nooking
>>103390196
If this is the public stuff they are dropping the probably already have agi and we can't llet them have that
>>103390207
It's the only sensible option desu

>>103390187
>How do we stop China from btfo'ing us in ai?
By being better than them? It would require to stop shooting ourselves in the foot with the dataset cucking/censorship, maybe it's too much to ask for the cucked westoid, but it's not like they have much of a choice now, or else they do that, or they'll fall into irrelevancy

File: snapshot.jpg (160 KB, 1280x720)

160 KB JPG

>Intel Battlemage up to 12GB VRAM
lmao

File: part-2-3.mp4 (650 KB, 1248x704)

650 KB MP4

Oh come on now. The chinks are straightup just fucking with us western cucks.
Can you imagine putting that stuff on your main page. lol
We have 1/3 couple videos then 2/3 saftey text.

>>103390232
Nvidia is fucking finished

File: Screenshot_20241203_231129.png (935 KB, 2275x772)

935 KB PNG

>>103390238
I still cant get over it.
The whole purpose of the video is the boobs. Like no other reason than show the boob physics. Bless those madlads.
>professional
mmhmm

>>103390238
>Oh come on now. The chinks are straightup just fucking with us western cucks.
it was a predictable outcome, if the west doesn't want to improve this technology, another country will take their place, it's as simple as that

File: snapshot2.jpg (141 KB, 1280x720)

141 KB JPG

>>103390232

>>103390232
It's acutally optimized vram that's like 4 times better than normal vram so it's actually 48gb.
Intel does what nojudea couldn't with the 5090

File: please honk jeb.jpg (208 KB, 397x430)

208 KB JPG

>>103390232
>slow and steady wins the race
pls clap

>>103390232
desu it's not the VRAM the problem, it's Cuda, even if they made a 300gb card, no one would want to ditch out Torch + Cuda to make this shit work onto their machine learning process

>>103390271
But isnt there IntelArc support already in llama.cpp and koboldcpp?
The times I saw arent that bad. Even p40 are crazy expensive.

File: 1712104608963845.png (961 KB, 828x627)

961 KB PNG

>>103390252
>The whole purpose of the video is the boobs. Like no other reason than show the boob physics.
True, that's why I love China now, they just don't give a fuck and give the users what they really want.

>>103390280
>True, that's why I love China now
I wished the AI bloom would've occured in the 90's, when Japan was at its peak, we would've gotten a 3-way AI race, and Japan would've shit out good anime model, just imagine the kino.

>>103390009
Is that when the 2nd election is?

>>103390295
I thought about that recently too. It kinda feels like we barley made it or maybe even too late for stuff to actually end up in the local chuds hands.
If Windows would be released today I cant imagine the outcry. Free drawing and text and you can upload it anywhere you want too.

>>103390277
yeah I guess but llama.cpp isn't the whole ecosystem, a good gpu card should work well on every situation, whether it's for inference or training

>>103390318
>If Windows would be released today I cant imagine the outcry.
yeah Idk about that, 20 years ago when GTA3 was released, it was hella controvertial and fucking US senators wanted this game gone, now it's completly normalized, now imagine chatgpt releasing in 2003, I think the anti-AI crowd would've been even louder than nowdays

grok-2-2024-08-13 not only knows D&D 3.5e, it knows to make reference to obscure spells that actually do exist.
Huh.
Didn't expect that.
A shame that there's zero chance I'll ever be able to run this thing.

>>103389792
>https://www.youtube.com/watch?v=n3B6mXcKO0I
Imagine if 9 months ago, when Sora was announced, someone told you that you could run a Sora-level video model locally in 2024, people would have laughed in your face, and yet here we are...

>>103389183
>13B
meme model, why didn't they go for 32B+?

>>103390565
Burgers are to blame for the ban on GPU exports

>>103390565
>why didn't they go for 32B+
video models ask for a shit ton of vram, the model size is just a fraction of the vram used, there's the AE encoder, the VAE decoder, and the text encoder, if you combine them all, HunyuanVideo already asks for 60gb of vram

>>103389160
assuming this distributed training shit becomes good enough, it's going to be possible to set up our own training for it
if it becomes sufficiently good, we'd be able to train it over Tor or I2P, then nobody can stop the coombots

File: KneelBeforeMeGoyim.png (1.04 MB, 1280x720)

1.04 MB PNG

>>103390565
>why didn't they go for 32B+
because Nvdia is slowing down the AI progress by hoarding the VRAM, that's why

>>103389792
>https://www.youtube.com/watch?v=n3B6mXcKO0I
Are they making fun of James Cameron because he went for the wrong horse? (He's now working with SAI)

>>103390565
>meme model, why didn't they go for 32B+?
the day we'll make fp8 training as efficient as bf16, then maybe it'll be possible

>>103390628
It is so surreal how the gold rush is being slowed down because there is only one shovel manufacturer and he is making sure to sell shovels as slow as possible and with the highest markup possible.

>>103390777
Uh... AMD exists though?

>>103390777
just make your own shovel company bro, free money waiting for you to reach out and grab it

>>103390271
If Intel or anyone else were to release a cheap GPU with 24 GB VRAM I absolutely would make sure it's supported in llama.cpp/GGML.
Intel in particular are already sending their engineers for the SYCL backend anyways so it wouldn't be too much work for me.
16 GB just isn't enough and 12 GB is a meme.

File: 1708980060262529.png (540 KB, 800x716)

540 KB PNG

>>103390783
>Uh... AMD exists though?
AMD was made just so that Nvdia can't be sued for Anti-trust monopoly

nous = reddit
they should stay off twitter

https://x.com/Xianbao_QIAN/status/1863482475307249794#m
What the fuck. Recent sora (turbo?) leaks are not that good for sure.
Does it also output the sound? Thats whats written in the model page at least.

>>103390232
>>103390256
All hope is lost

>>103391053
>Does it also output the sound?
yeah it looks like it does, you have some examples on the blog page, open the videos on another tab if you want to hear anything though, they messed that up
https://aivideo.hunyuan.tencent.com/

>>103386356
>>103389183
we're so back

https://github.com/Tencent/HunyuanVideo

>>103391114
Its legit better than sora going by the leaked sora vids. OpenAi is so fucked lol.

File: Screenshot_20241204_010859.png (958 KB, 1453x843)

958 KB PNG

>>103391109
So it really is like that meta model.
https://ai.meta.com/blog/movie-gen-media-foundation-models-generative-ai-video/

Makes you wonder if they actually release it now.
Also I knew that I saw that ghost before somewhere!
That mus have been on purpose. Would be funny if its a diss at meta.

>>103391130
>Makes you wonder if they actually release it now.
>Also I knew that I saw that ghost before somewhere!
>That mus have been on purpose. Would be funny if its a diss at meta.
oh shit, it's definitely a diss at meta, that and the Avatar video (that is a diss at James Cameron for joinging Stability AI) >>103390639
https://stability.ai/news/james-cameron-joins-stability-ai-board-of-directors
they are so petty I love it

>>103391114
50 cents have been deposited in your social credit account

File: Screenshot_20241204_011242.png (576 KB, 1150x870)

576 KB PNG

>>103391142
>and the Avatar video
Thats bizarre. (in a good way)
They fucking pinned it on the official witter account. Crazy timeline.
Chinks going in hard. Winners is the local chud.

>>103391151
Sam said the check is on the way

File: 2024-12-03_080434_seed999(...).png (2.68 MB, 1536x1536)

2.68 MB PNG

Goo morning.

>>103391158
> chinks ripping off ip is good actually

>>103391177
whattup slime?

>>103391197
it is good, fuck copyright, and fuck you

>>103391211
50 cents have been deposited in your social credit account

>>103391197
intellectual property is the most retarded thing ever invented. "Nooo you can't look at those pixels without my permission. That math equation is MINE. I OWN this color, you can only use this color if you pay me!" Absolute fucking insanity and good riddance.

>>103391218
>>103391163
Also Hollywood said they would send you a extra child for your service

>>103391197
Uhm, yes? Hello?
Imagine doing the "cent deposited" meme and then being the literal shill to protect companies copyright. what a joke.

File: 1708749386693439.webm (901 KB, 1072x720)

901 KB WEBM

>>103391218
you're done Sam, the chinks have won, time to shut your cucked company down

>>103391218
Funny, you're the one that looks like the corporate bootlicker here. Copyright is retarded, ideas are only as valuable as the effort it takes to make something of quality for the ideas. If you can't keep making good use of it, maybe you shouldn't deserve to hold dominion over it for an entire lifetime.

>>103391230
>intellectual property is the most retarded thing ever invented.
this, IP should only last 5 years, if you can't make something different and good instead of leeching off your one-shot, you deserve to die as a company

File: mqdefault.jpg (25 KB, 320x180)

25 KB JPG

>>103391177
Ballistic gel Teto

>Repository storage limit reached (Max: 1GB)
https://www.reddit.com/r/LocalLLaMA/comments/1h5q322/i_get_the_500_gb_limit_but_why_cant_i_upload/

File: LFG.jpg (88 KB, 1080x340)

88 KB JPG

ASI SOON LFG

>>103391292
They are scared of the chinks

>>103391292
I miss the time when OpenAI would just shut the fuck up and release the kino model, now they go for those retarded cryptic messages as if it meant something

>>103391292
>t.roon

>>103391177
gm sexi *rapes u wiht my 3 foot penis u feel good*

>>103391292
Strawberry was hyped for a year and it was then released as o1 preview....
This is such a bad look. Chinks just drop it outta nowhere. Pics and videos on the blog page. "Cool new stuff, here ya go".
I would be shocked though if they really didn't have anything. It does look bad though. What does OpenAI even excel at anymore? Its just the brand. Its been for a while like that.

>>103391315
They dont have anything, just trying to keep investors... invested.

>>103391315
The closer they get to the plateau of the curve, the more they rely on hype.
Doubly so now that they completely shed the "not for profit" mask.

File: 21522 - SoyBooru.png (46 KB, 457x694)

46 KB PNG

>>103391292
sure it is... sure it is...

>>103391337
>Chinks just drop it outta nowhere. Pics and videos on the blog page. "Cool new stuff, here ya go".
this, when you look at their blog page, there's not a single line about "safety" bullshit, it's so refreshing, just for that I want to root for them
https://aivideo.hunyuan.tencent.com/

>>103391292
>>103390009
OpenAI local GPT5 trained at sub 1bit. I'm hyped lads.

>>103388608
That's how that worked? What a scam.

>>103390009
>>103391337
>>103391355
>>103391357
Has strawberry man turned OpenAI into a meme?

>>103388608
>you get full weights, train it on your H100, pass the result to the next person
wait what? I thought it was true parralelism? holy fuck that's lame

>>103391386
For me at least yes.
I use sonnet 3.5 for work and stuff that requires knowledge
Everything else like gaming buddy or coom I do locally.
Not sure what the use case for o1 or gpt4 is. Gpt4 is worse than 3.5 and o1 is too expensive and not really usable besides niche problems.

>>103391386
>Has strawberry man turned OpenAI into a meme?
yeah, but the investors are still fooled to believe OpenAI has a moat, and that's everything OpenAI needs

File: hmmm.png (512 KB, 1172x953)

512 KB PNG

They clearly trained on Hollywood and netflix.
Would openai get away with that?
How else can they compete with the chinks?
People here complain about copyright but clearly china doesnt give a fuck. Where else to you get thousands of hours of high quality cinematic shit?

>>103386811
https://huggingface.co/mistralai/Mistral-Large-Instruct-2411/discussions/11

>>103391239
Kek

>>103391409
I second this. However, I think 4o mini is better than Haiku 3.5, and I like to use these small models for tasks that I'm not willing to pay too much money to get done.

Strap in fellas.
We gonna go home.
[spoiler]https://litter.catbox.moe/5qmc06.mp4[/spoiler]

>>103391504
>has nudity in training data
We're so back

>>103391292
>gets mogged
>vagueposts about amazing things happening behind the scenes
OAI in a nutshell

>>103391464
>Would openai get away with that?
they also do that, let's not kid ourselves
https://youtu.be/mAUpxN-EIgU?t=263

>>103391464
This is why China will win the AI race unless the US says fuck copyright law for training AI

>>103391464
>People here complain about copyright but clearly china doesnt give a fuck. Where else to you get thousands of hours of high quality cinematic shit?
that's up to them, or else the western companies grow some balls and also train their model with copyrighted stuff, or else they are willing to be dominated by the chinks, either way, the user will go for the better model, it's competition 101

>>103391504
Damn. I kneel, China. Maybe we won't need to do our own LLM training, China will eventually train the most copyright and nsfw model possible eventually.

>>103391504
China won.

>>103391534
*non-copyright

File: 1731480115496498.gif (957 KB, 256x320)

957 KB GIF

>>103391504
>China, the country that prohibit porn has nudity in their models
>The US, a country that promote a lot of degenerate coom shit are too scared to go for naked women
I fucking kneel China

>>103391464
For those who are familiar with tost.ai, you can test out that model here
https://xcancel.com/camenduru/status/1863971928115138628#m

>>103391558
Scared of what? God's wrath? Lol lmao
They're scared of the porn jew, nobody wants want such a powerful enemy

File: iu[1].jpg (61 KB, 1280x720)

61 KB JPG

>>103391504
And that's how China wins.
Not on the court of public opinion, but by giving people what they want.

>>103391580
>Scared of what?
>They're scared of the porn jew
you just answered your own question lol

>>103391558
What about all of the western porn companies going out of business due to AI? We can't let users decide what they want to watch. They should only consume porn that promotes our values. Cool it with antisemitism.

>>103391575
>xcancel
You are a good anon.

>>103391315
grifter culture

>>103391613
desu those coom companies deserve to die, I'll never forget the purge that Pornhub had done a few years ago, all my favorite amateur porn videos, gone... just like that...

>>103391292
did I miss something?

>>103391631
yeah, the hunyuan release that openai is trying to distract you from

I guess that the chinks are doing some insane pressure with this new video model, because Google decided to launch their model right after that
https://xcancel.com/googlecloud/status/1863957264895217999#m

>>103391768
not open-source.
literally no one cares.

>>103391768
lol pathetic

This new model seems completely uncensored btw. US ai companies have no chance vs Chinese companies that care neither for copyright nor "safety"

[spoiler] https://files.catbox.moe/3owu5d.mp4 [/spoiler]

>>103391504
>>103391862
50 cents has been deposited in your social credit account

>>103391862
Nice nuclear weapon, Anon.

>>103391504
I kneel
>>103391869
seething glowie

>>103391292
*bad
damn autocorrect

>>103391862
Weights. Now!

File: 1712374238223999.png (64 KB, 225x225)

64 KB PNG

>>103391862
holy shit, didn't expect that, we're so fucking back

>>103391862
Holy moly. Those hip bones tho. Was it trained on French women?

>>103391862
can it generate Tiananmen square?

>>103391504
Now make her younger

>>103391862
Wow.
What prompt did you use?

>>103391890
They are already up?

https://github.com/Tencent/HunyuanVideo

https://huggingface.co/tencent/HunyuanVideo

>>103391862
meanwhile qwen is pozzed openai'esque garbage, nice gaslighting though.

>>103391890
>Weights. Now!
https://huggingface.co/tencent/HunyuanVideo/tree/main/hunyuan-video-t2v-720p/transformers

File: strawberry-sam_altman.png (28 KB, 800x800)

28 KB PNG

>50 cents has been deposited in your social credit account

>>103391869
This.

>>103391907
I guess that this chink company has more balls than the others

>>103391862
Did they train it on literal porn? wtf

>>103391907
qwen unironically drank that meta "quality data is all you need" koolaid

>>103391768
https://xcancel.com/Virtual_Horror/status/1863991979790729720
Openai is doing stuff too.
Too be honest google does what they all do though, those slow pans.
I wanna see a movie with dragons like the chinks did. With many people etc. they called them out.>>103391053

>>103391905
>>103391909
We are so fucking back I didn't know was even possible.

>>103391909
Can I offload it to ram?

Why are the us agents in this thread freaking about a video model?
I don't get it.

I know for a matter of fact that right when we have gotten used to Dalle-N being left behind by open-source OpenAI drops Dalle-(N+1), it's should be out any time now

>>103391937
>>103391768
>private closed super-secret preview
kek, china won

>>103391862
I have two rtx 3090, will that be enough to run the model?

>>103391975
No multi GPU support atm I think. You need a single GPU with high VRAM

>>103391862
Crazy, if you told me 1yr ago the chinks will be first for ai pussy video i wouldnt have believed it.
Crazy timeline whats going on.

Now it makes sense it can do yoga etc well.
We all knew this. Needs nudity.


Makes you wonder if burger and eu will start banning.

>>103391975
It's just a 13b model, as soon as it gets quantized and tiled VAE gets applied it will run on 13GB at Q8 or 10GB at Q6, it's something trivial to do

>>103391998
Has anyone ever quantized a video model?

File: image (3).png (964 KB, 1536x1024)

964 KB PNG

>https://huggingface.co/spaces/tencent/Hunyuan3D-1
I didn't know this was a thing, it surprisingly knows who Hatsune Miku is, and the result isn't terrible (although it isn't perfect either)! wow!

>>103392006
Mochi?

>>103391932
there's no way it can do actual porn

Now to patiently wait for GPU spliting...

>>103392006
All of them that released local. 8bit should run on 24GB with some tricks.

>>103391936
What do you mean? That worked out for them pretty well given people's praise for its coding performance. It's in no way false that data quality is important. Censorship is a different topic from data quality and it's on a company whether or not they have such political things to be concerned about.

https://huggingface.co/arcee-ai/Virtuoso-Small

>>103392063
>large meme model
Meh, for a moment I thought we would finally get local suno

>>103392017
>there's no way it can do actual porn
coom anon, your verdict? >>103391862

>>103392026
>8bit should run on 24GB with some tricks.
it will, kijai managed to make Mochi work under 24gb, Mochi is a 10b model, it won't be much harder for a 13b model

>>103392087
To be fair, that's nudity, not straight up fucking.
Anybody want to try doing that?

>>103391862
Please tell me someone is going to train this and make an actual coom model finally? The weights are open, please someone be a hero, the drought has been so bad lately.

>>103392095
Somehow I doubt they paid an army of jeets to tag porn. Video isn't like LLM or images where the tagged lewd is readily available for training.

>>103392126
I doubt it too.
Still, somebody's gotta try, who knows.
Maybe the people behind it are even crazier than we thought.

>>103392095
>To be fair, that's nudity, not straight up fucking.
the simple fact that that anon managed to get a close up of a pussy means that the model has been trained to recognize pussy and has seen "pussy" caption tags, which is so cool, we're talking about a chink company there, how based is that?

>>103392095
Close for masturbation, prob just a bad angle
https://files.catbox.moe/5cowwt.mp4

>>103392126
desu the booru datasets could probably be used to bootstrap a video porn dataset
Actually don't a bunch of porn sites already have tags?

>>103392140
yeah... the anatomy looks kinda fucked though, maybe if it was zoomed out more it'll look better

>>103392140
The nipples look like plastic, they have no texture, it's so over.

>>103392165
I mean, for a base model that's already insane it can do that, Flux cannot do into nipples for example

>>103392165
More than what flux spat out.
No other video model certainly. We are more back than ever.

>>103392140
>bad angle
I wonder if they had some nice training data of women filming themselves masturbating like this and sending it over discord or some other platform.

>>103392140
can you try for a normal angle? when it's upside down those models have trouble with it

wake me up when multi-gpu support.

https://files.catbox.moe/yv6klb.mp4
Lol, just need a good roll. This model is not fast like that other one though.

>hags

>>103392217
Quants will drop soon and is just 13b has to fit in 24gb
>>103392222
We are so fucking bad, pray that it can be finetuned and we may be in a new golden era.

>>103392222
>This model is not fast like that other one though.
you're running it locally? if yes don't forget that you're using the "bad" text encoder, the "good" one isn't released it yet

>>103392230
Back*

>>103392222
Checked. Please try to do anime too, it doesn't even have to be nudity, I just want to see how well it does anime.

>>103392222
It actually wasn't lobotomized. It's aware of body parts. Wow, we are a fine tune away from cooooom.

>>103392230
>Quants will drop soon and is just 13b has to fit in 24gb
>As I said, this is open-source, but comes with a price of 60-80GB VRAM requirement!
https://x.com/dreamingtulpa/status/1863866801127633204

>>103392230
Even if you fit the weights into VRAM the amount of VRAM needed for compute is pretty substantial. I'll be surprised if you can run it on a single 24GB even if it's quantized.

>>103392226
Hi Glownigger.

>>103392222
The jiggle is a bit weird but WOW IT JIGGLES

>>103392163
I guess you could use a multimodal LLM to describe poses every X frames, together with an overall scene description and use that to train the video model.

Still, unless you actually do that for porn I don't think the video model would magically learn to render a DP creampie.

>>103392222
I guess video models still aren't very good in general, but this is like image models really. We will get good gens when lucky, and that's enough. We are back.

>>103392222
what GPU are you using and how long does it take?

>>103392265
Unlike image models, this is a video model, which means it predicts the next frame with both the prompt and previous frames. If it saw porn in its training, then it has some of that knowledge in it and it was not erased. At most we would need to fine tune it in order to connect its knowledge with language.

>>103392268
A100, like 7-8 mins

>>103392308
Thanks for testing it out anon. Appreciated

https://files.catbox.moe/uy8jxs.mp4
We are so fucking back.

>>103392317
safety fags are going to ack after this blow

>>103392317
The westcucks now have no excuse.
All that safety shit blurring bodies makes it worse and now one uncensored is out in the open.
They kinda need to follow right?
Clearly trained on naked woman, probably even more.insane. Outta nowhere too.

File: 1732973607098741.png (60 KB, 3408x623)

60 KB PNG

>>103392308
>A100, like 7-8 mins
https://bizon-tech.com/gpu-benchmarks/NVIDIA-RTX-3090-vs-NVIDIA-A100-40-GB-(PCIe)/579vs592
looks like an A100 is twice as fast as a 3090, so I guess that'll take 15 mn for a 3090 to render a 5 sec video? looks all right

I know I won't even coom to these, I just think AI over-censorship is stupid. This model will be a game changer. But I know some safetyfags are profusely trying to gen CP for ammunition this very instant.

File: 1706293047404301.png (127 KB, 2490x774)

127 KB PNG

>>103392317
let's not forget he's running with llama3's text encoder, once we'll get the real one it'll look even better, we've never been so back!
https://github.com/Tencent/HunyuanVideo/blob/main/ckpts/README.md

>>103392358
100%
They did it with flux.
They DEFINITELY will do it with this one.
Not sure if it will succeed.
People kinda stopped giving a fuck.

>>103392347
>Clearly trained on naked woman, probably even more
it has to be porn, it renders pussy way too well

>>103392243
>anime
Have a pony instead
https://files.catbox.moe/4aqjvz.mp4
And a furry:
https://files.catbox.moe/x0ve4k.mp4

File: 4287231451.jpg (91 KB, 1280x720)

91 KB JPG

>>103392391
>https://files.catbox.moe/x0ve4k.mp4
the fucking furry

>>103392391
it can't do 2d anime?

>>103392391
Again, 0 fucks given about copyright lmao

Also that tail waggle at the end.
Im gonna live long enough for the chinks to turn me into a furry

>>103392308
>A100
the 80gb right? how much vram does it eat on your end?

>>103392391
what about like a zootopia style furry

>>103392419
If it's trainable that should be the least of your worries, I'll do anything.

File: 1705539838519907.png (31 KB, 225x225)

31 KB PNG

>>103392317
>>103392391
>0 fucks given for nudity
>0 fucks given for copyright
China is the goat!!

>>103392419
If it knows what 2D pussy looks like then it will be trainable for coom. Same as textgen.

>>103392391
can it do copyright shit like League of Legends?