/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>107643997 & >>107636165►News>(12/22) GLM-4.7: Advancing the Coding Capability: https://z.ai/blog/glm-4.7>(12/17) Introducing Meta Segment Anything Model Audio: https://ai.meta.com/samaudio>(12/16) MiMo-V2-Flash 309B-A15B released: https://mimo.xiaomi.com/blog/mimo-v2-flash>(12/16) GLM4V vision encoder support merged: https://github.com/ggml-org/llama.cpp/pull/18042>(12/15) llama.cpp automation for memory allocation: https://github.com/ggml-org/llama.cpp/discussions/18049►News Archive: https://rentry.org/lmg-news-archive►Glossary: https://rentry.org/lmg-glossary►Links: https://rentry.org/LocalModelsLinks►Official /lmg/ card: https://files.catbox.moe/cbclyf.png►Getting Startedhttps://rentry.org/lmg-lazy-getting-started-guidehttps://rentry.org/lmg-build-guideshttps://rentry.org/IsolatedLinuxWebServicehttps://rentry.org/recommended-modelshttps://rentry.org/samplershttps://rentry.org/MikupadIntroGuide►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksLiveBench: https://livebench.aiProgramming: https://livecodebench.github.io/gso.htmlContext Length: https://github.com/adobe-research/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsAlpha Calculator: https://desmos.com/calculator/ffngla98ycGGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-CalculatorSampler Visualizer: https://artefact2.github.io/llm-sampling►Text Gen. UI, Inference Engineshttps://github.com/lmg-anon/mikupadhttps://github.com/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/ggerganov/llama.cpphttps://github.com/theroyallab/tabbyAPIhttps://github.com/vllm-project/vllm
K-Kurisulove??
>rentry.org/jartedyou dropped this
>>107652781Too late. Now spit out miku's penis and worship the true redhead queen of /lmg/.
>>107652789Dammit. Merry Christinamas
>>107644945there were tests done similar to this with mixtral models, just dont do it
>>107652781I admire your dedication, migubaker.
>>107652821Not him but thanks
►Recent Highlights from the Previous Thread: >>107643997--Custom PC build challenges with dual-CPU motherboard and GPU development:>107648704 >107648802 >107648838 >107648947--Z.AI's upcoming model announcement generating speculation:>107645563 >107645577--Critique of uncensored release and adoption:>107649191 >107649258 >107649238--glm 4.7 template configuration issues causing excessive token usage:>107645140 >107645154 >107645187--Model 4.7 evaluation and comparison with previous versions:>107647330 >107647367 >107647441 >107647641 >107647542 >107647495--Finetuning scalability with web scraping and model distillation:>107644302 >107644406 >107644448 >107644493 >107644521 >107644775 >107644791 >107644798 >107644810 >107645002--GLM performance debates and benchmark analysis:>107647533 >107647803 >107647892 >107648045 >107648063--Training loss reduction strategies and model performance analysis:>107648741 >107648779--Critique of unsloth template and GGUF challenges:>107648117 >107648188 >107648189 >107648231 >107648262 >107648306 >107648403 >107648431 >107649512 >107649536--Scraping and roleplay prompting Opus 3 via Claude and OpenRouter:>107648858--Ethical concerns about Qwen3-TTS cross-species voice cloning:>107644576 >107644634--AI policy restrictions and model limitations in generating stylized Hatsune Miku SVGs:>107644661 >107644704 >107644718 >107644785 >107644738 >107644861 >107644870 >107645028 >107645058--GLM 4.7 cockbench and speculation on llama 4 scout behavior:>107644330 >107644369--GLM 4.7 shows unexpected reasoning capability about animal knowledge:>107645272 >107646055--Comparing GLM model perplexity across parameter sizes to analyze non-activated parameter impact:>107644945--Teto and Miku (free space):>107644176 >107644885 >107649257 >107651275 >107651326 >107651473 >107652578 >107652818►Recent Highlight Posts from the Previous Thread: >>107644002Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
>>107652827Thank you Recap Miku
>>107652819>>107644945https://xcancel.com/sbeastwindy/status/1735185274475524333
>>107652836https://github.com/ggml-org/llama.cpp/pull/4406#issuecomment-1855151885last one i swear
I have a half working pi zero W. How to get it working with a lm?
>>107652846What do you mean? Are you trying to host the LLM on the Pi, or just use it as like a frontend?
>>107652767
>>107652999didnt know miku was african american
>>107653005miku has always been a real black hood nigga
>>107653005apparently kfc for christmas is some sort of thing in Japan. I think it has to do with them having no real Christian tradition, some marketers swooping in, and the colonel kinda looking like santa clause.https://www.bbc.com/worklife/article/20161216-why-japan-celebrates-christmas-with-kfc
wew.. lads, just finished reading rhrough the last 3 threadswhat am i gonna do now?!
man
>Delivery AttemptedThis is the second time I've bought a GPU from Amazon and this has happened. Last time I had to buy a cheap GPU from China. I hope they don't screw me over again.By the way, is this GPU good enough to play around with or is it too weak? My initial goal is to generate some assets for my game, nothing professional, it would be more for prototyping, nothing 4k.
>>107653014I don't understand how they could make that connection. The colonel has a puny beard compared to the real claus.
>>107653018Wait for Z-Image base and edit, and for ubergarm to wake up
>>107653045maybe I just made up that part. Seems to be all marketing https://old.reddit.com/r/Tokyo/comments/1hmg29n/dont_people_eat_kentucky_fried_chicken_on/
>>107653041>My initial goal is to generate some assets for my gameiirc Hunyuan3D only needs 10GB so it should be fine. Definitely fine if your game is 2D.
>>107652978Everything.
>>107653047Models that will never be released:- WizardLM- Meta Movie Gen- Mistral Medium- GLM 4.6 Air- Z-Image base and edit
>>107653073
>>107653073512 MB of RAM isn't nearly enough to host most LLMs but I'm sure you could fit one in there if you really tried.
>>107653073you could do a q2 of gpt2
>>107653080you forgot llama2 34b
Does running an LLM cause your GPU and CPU to run hot?
>>107653129It can if it has to process a lot of tokens at once, but it shouldn't be much worse than running a demanding video game for a long time
>>107653129It causes my cock to run hot.
>>107653151use lube, baka
>>107653154I don't need to, I have a foreskin.
>>107653041Amazon drivers will unironically steal your GPU and give a bogus delivery or error code.
>>107653169Then it becomes Amazon's problem to replace or reimburse.
>>107653005KFC is unironically a Japanese tradition thanks to a very successful marketing campaign several decades ago.
>>107653211Don't forget to make your reservation, lines will be long
GLM 4.7 GGJTs when?
>>107653129it takes 100% of anything it can get even more so than the most demanding video games
bros.. i have to confess.. i had my first LLM coom today.around a year and a half ago i complained about it and some anon recommended twin cow girl erp cards, but i couldnt be serious in any kind of erpbut after not jerking off for two weeks (reasons), i talked about reasons with deepseek ( for like a week ), and then told it about not jerking offdipsy told me about joi, and then did itcrazy.that was absolutely crazy. insane. never knew my balls could store that much cum
>>107653257Nips are fucking weird. Meat on Christmas is literal blasphemy in the first place.
>>107653468They're not Christians thougheverbeit
>>107653466Cooming after weeks of not cooming always feels amazing but sometimes it can be annoying to clean up the sheer volume of cum, especially since it usually shoots out further as well.
>>107653177>The driver marked it as delivered>The driver took a picture on their phone to authenticate it as delivered before pocketing itEnjoy the hassle of convincing Amazon what happened. Niggers abusing the system have completely eroded Amazon's consumer-facing trust in incidents like this in recent years.
>>107653468>Meat on Christmas is literal blasphemy in the first place.what the fuck are you on about? that's for easter you TERTIARY.
>>107653468Christmas in December is blasphemy, you unorthodox fucks
>>107653479Why are you cumming all over the place instead of in a tissue or some toilet paper?
>>107653479>especially since it usually shoots out further as well.this is so true, the most annoying part is when the cum lands on your hair and then you have to wash it (easy part)the hard part is drying the hair, i bought a 2.6kW hair dryer and it still takes 30 minutes (best case) up to an hour :*
>>107653476Which is why them celebrating the birth of Christ makes no sense.
>>107653510nta but i literally just cummed on my hoodie and all over my thighs and the floor and the wallusually you cant exactly focus on aiming and feeling good at thd same time
>>107653493Filthy heretic.>>107653495I blame the pope.
ARC-AGI 2 has been finished.What are the implications of this and how fast do you assume ARC-AGI 3 (Which will by dynamic real time environments instead of static puzzles) will be solved by LLMs with agentic scaffolding?
>>107653466My BROTHER you have no idea what you've been missing.Think about every crush you've ever had. Every fetish you've got. Every single sexual fantasy and scenario you've ever thought about. Well now you can simulate them with a good model and system prompt. >scifi rouge AI who inhabits a space pirates body who wants to rape you (her captain)>big tiddied dark elf pampers you after a quest>cute femboy with AGP who dresses up like princess peach needs help with homework>loli childwife on a deserted island>kobold adventure party shortstacks with fat butts get stuck in the wall of a dungeon>rouge the bat JOI instructor>Pinkie Pie PISS simulator>busty anthro girl traps you in her trenchcoatYou are only limited by your imagination and the models SLOP. Truly the greatest coom tool in all of human history, iykyk.
>>107653579heh, i think i didnt make it cleari was erping, a lot of the time not seriousbut usually it was erp to get me hard and horny for regular pornman, to think we've come so far since vicuna-unlockedi wonder what gozfarb is up to nowadaysthese scenarios are sick, ill have to try them out
>>107653521Were you backed up for like a week or something?
>>107653641around 2 weeks >>107653466
>>107653614>but usually it was erp to get me hard and horny for regular pornFair, I do this quite often as well. I find that text based porn fills a different need. Sometimes I'll load up a model and start RP for a few turns before I realize I'm only in the mood for standard pornography. >man, to think we've come so far since vicuna-unlockedI remember my first COOM to pyg6b. Its easy to shit on the state of local right now but the fact is that if you can run at minimum Nemo you essentially have the ability to RP any scenario you want. Pretty powerful stuff.
>>107653671Damn, I feel antsy after just 2 days
>>107653521I'm gonna teach you a method I came up with in my teens.Take two joined squares of toilet paper and roll them up in a cylinder a bit wider than your dick.Fold the top third inwards. This keeps if from unrolling. Don't fold it all way, you want to form a closed end. Palm the folded then push your fingers through the open end and press the folded bits against your palm to bunch them up and form the closed end.It only takes a few seconds and now you have a disposable cup you can cum in.You can just plop it over your dick if it's wide enough or you can hold it with your other hand.You can even use it with one hand if you only grip the very bottom of it with your thumb and index finger and move it along as you stroke. I prefer this because it's a complete seal even though it restricts your stroke length a bit. I think it only works with a foreskin because otherwise you'd be rubbing your dick with paper as opposed to it gliding along with the foreskin.
>>107653692Toilet paper is thin and completely dissolves when exposed to moisture. If you edge then your pre-cum is going to make your dick just rip through it after a few strokes.Get a few tissues, ideally something soft like Kleenex™ and use 3-4 of them. Layer them on top of each other, then wrap it around your dick. It will feel great and your dick won't rip through it, and you won't end up with cum inside your belly button.
>>107653692>>107653705
>>107653692>>107653705thank you anons, this is amazingabsolutely amazing.another day thankful to have a foreskin
>>107653705It's not an issue. I only needs to hold for a few seconds.Also I find that it feels so much worse if the cumshot is impeded by something covering the tip. The cup shape is great because it leaves a few cm of free space in front.
>>107652999good girl, consume the goyslop
>>107653727>I only needs to hold for a few seconds.Oh, I like to go for a bit longer, and I tend to leak during.>Also I find that it feels so much worse if the cumshot is impeded by something covering the tip.My mother fell for the circumcision jew so this might also be different. Tapping a soft tissue with the tip feels nice for me.
>>107653556I'm no researcher or learned person so I could be completely wrong, but IMO the saturation of these benchmarks are less of the models being able to generalize better between releases and more of the researchers applying RL to specifically saturate the benchmarks. Impressive, but not the kind of improvements we need to increase the intelligence of the model. This doesn't mean the models are dumb, or that RL isn't good enough to make models better than humans at certain tasks. It does mean that something fundamental needs to improve outside of RL in order to give these things the spark needed to be considered "AGI". I think ARC-AGI 3 will follow a similar path, although I would expect it to take a little bit longer since the challenges are significantly harder. I don't think people will consider it AGI when ARC-AGI 3 gets saturated.If anyone here actually knows what they're talking about please correct me
>>107652767>"official" /lmg/ mascot>mentally ill tranny/tranny-adjacent retards trying to cram their generic tranime girl obsession into any and all online communities regardless of relevance, just like lgbt fags and pedo groomers>also purely coincidentally its one of the most generic and popular normie npc tranime girls every single time, like troonku, lain, makise or whatever was the tranime girl of the year in the groups they hanged out at the point in their life where their life peaked (highschool) which doesnt speak to their low IQ and unironic mental illnesss and autism at all of course, they are completely normal and well adjusted 30+ year old men spamming the same generic copy paste tranime girl in every conversation onlineAlso reminder for the newfags that the main terminally online troonku baker of /lmg/ was exposed as a 4chan moderator, who literally instantly deleted messages posted in the general that made him look bad while he himself posted miku porn on /ldg/ that stayed up after being reported for multiple hours, multiple times. He was also exposed as an autogynephilia fetishist, and a poopdick fetishist, completely coincidentally again of course.
>>107653692>>107653705I just finish jerking off on the toilet.All of my important files are on my desktop with an automatic sync to my NAS.So after edging for some time I can move seamlessly from my desktop to my laptop.One hand for stimulation, one hand to hold a bit of toilet paper, and one foot to control my laptop.
>>107653757ARC-AGI 2 is a "generalization benchmark". They are hard puzzles that need a sort of generalization in logic for you to solve. It's correlated with better reasoning in for example coding but NOT creative endeavors like roleplaying.ARC-AGI 3 will actually be worse because it will be dynamic puzzles for testing agents. So while that is probably a step towards AGI it isn't as correlated with generalization. For example an AI that is just better at being dynamic might score better than an AI that is actually smart purely because the dynamic shitty AI is better for real time environments. Humans would still consider the lower scoring one to be more general.
>107653768This is what you're spending Christmas Eve doing?
You just want to cum.I want a companion.We are not the same.
>>107653816I want both
For the deluded minds who think they can solo RP models, from the GLM 4.7 AMA:https://www.reddit.com/r/LocalLLaMA/comments/1ptxm3x/ama_with_zai_the_lab_behind_glm47/nvkhgjk/>I can analyze this from the perspective of post-training. At present, due to differences in compute reserves across organizations, the amount of compute invested in post-training also varies significantly. One clear trend we observe is that Chinese large model providers still invest substantially less compute in post-training compared with their U.S. counterparts, although this gap is gradually narrowing.>>For post-training, the compute consumed by experimentation is often much higher than that used in the final training runs. For example, during the post-training of GLM-4.7, the compute cost spent on post-training experiments was likely dozens of times higher than that of the final GLM-4.7 post-training run itself.>>Returning to the original question, in my view, building a reasonably strong model team for post-training requires at least a dozen highly talented researchers, along with compute resources equivalent to roughly 2,000 H100/H800 GPUs.
>>107653829A cumpanion, if you will
>>107653848heh
>>107653815It's always a good deed to publicly humiliate and shit on the disgusting mentally ill faggots like yourself.
>>107653884being mean on christmas? you're gonna get a lump of COAL
>>107653903what about a lump of CUM
>>107653884>mentally ill faggotsLook into a mirror. You went through the effort of posting your seethe in an otherwise friendly and on topic thread because you have an axe to grind. No one cares about your irrational grudge against the dumbest shit. Maybe reflect and become better instead of letting your anger dictate your actions. Hope you have a nice Christmas with your friends and family.
>>107653930>axe to grindheh, this is funny because trannies have an axe wound between their legs LOL
>>107653930miku ritual posts > seething posts.
Seethejeet is by far the brimmiest poster in these threads and it's not even close.
>>107653936Yeah, troons castrating themselves is quite funny, but not as funny as untermensch exposing their infantile ego and subhuman level emotional control for everyone to laugh at, kek.
OK, I got 4.7 working on my new setup but it's kinda slow (3.5 T/s) but holy fuck! Coming from nemo and mistral small, it's like a another world.Is there anyway to limit the amount of reasoning it does in sillytavern? Do I need it at all if I'm mainly using it for RP or will it drop the quality significantly?
>>107654025you can just disable it
>>107654033yeah, but does it kill the creativity?
>>107654054I don't know because I always turn it off
>>107654065>I always chop my peen off
.
>>107654193FUCK ME
>>107654199the point is hf has/had? a limit of 50gb per file
>>107654199there is no need to be upset. learn how2use hf cli
just cat them, bwo
>>107654222https://huggingface.co/blog/rearchitecting-uploads-and-downloadsgoddamit zer fucking juden every fucking time with their fucking hooknosed bullshit im fucking sick of it everything is a fucking pile of shit nothing fucking works easier to be a caveman and make everything from scratch then put up with all this bullshit anymore i hope aws goes offline again and deletes all theri fucking bullshit does modelscope have the same ? can someone reupload there ?>>107654280may your cock rot and maggots eat it from the inside you son of a whore
>>107654199I can't believe those evil jannies robbed us of your high-quality post.
>>107653512What fucking bubble do you live in where you think everyone is just celebrating christmas as the birth of christ or think that's weird
>>107652836>>107652840I was more thinking of comparing, say, running smaller MoE with the normal number of activated params vs running a larger MoE with a cut down number of activated experts to match the number of activated params with the smaller MoE.Also, Mixtral is so old. I can only imagine that the numbers look pretty different for modern models.I do remember that Qwen 3 30B got a nice bump in PPL when using 10 experts instead of 8.
>>107653492>Niggers abusing the system have completely eroded Amazon's consumer-facing trust in incidents like this in recent years.amazon itself has also eroded its own reputation for meI live in a country where delivery was never, and I mean it, never reliable (France) so I always get my packages delivered at pickup locations (think services like UPS Access Point if you're a burger) which 1/ gives me the time to inspect the package properly and reject it if I see signs of tempering - damage and 2/ avoid porch pirates because it's not uncommon for lazy delivery drivers to just leave the package in the building hall and it gets stolen by the neighbors, I've known people it kept happening toa couple years ago, amazon has made it impossible to order packages over something like 400 bucks through this method. They only allow ordering this shit to your home. I absolutely refuse! no relay pickup no order period. I haven't ordered a single thing from amazon since then. I don't know why they did this to valuable packages because issues are more likely to happen when you have them delivered to your home than to a retailer pickup location, but they certainly lost me forever as a customer.
>>107654401how much is the pickuper getting paid to deal with 1000 amazon packages a day?
