/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>102731640 & >>102723173►News>(09/27) Emu3, next-token prediction multimodal models: https://hf.co/collections/BAAI/emu3-66f4e64f70850ff358a2e60f>(09/25) Multimodal Llama 3.2 released: https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices>(09/25) Molmo: Multimodal models based on OLMo, OLMoE, and Qwen-72B: https://molmo.allenai.org/blog>(09/24) Llama-3.1-70B-instruct distilled to 51B: https://hf.co/nvidia/Llama-3_1-Nemotron-51B-Instruct>(09/18) Qwen 2.5 released, trained on 18 trillion token dataset: https://qwenlm.github.io/blog/qwen2.5►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-build-guides►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksChatbot Arena: https://chat.lmsys.org/?leaderboardCensorship: https://hf.co/spaces/DontPlanToEnd/UGI-LeaderboardCensorbench: https://codeberg.org/jts2323/censorbenchJapanese: https://hf.co/datasets/lmg-anon/vntl-leaderboardProgramming: https://livecodebench.github.io/leaderboard.html►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/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/lmg-anon/mikupadhttps://github.com/turboderp/exuihttps://github.com/ggerganov/llama.cpp
►Recent Highlights from the Previous Thread: >>102731640--Paper: Differential Transformer paper discussion and potential applications:>102733185 >102733214 >102733290 >102733716 >102733282 >102733344 >102733249 >102733813 >102733853 >102734114 >102734156 >102734157 >102734358 >102734435 >102734961 >102735012 >102735040 >102735060 >102735046 >102735053--AGI sampler and ChatGPT updates discussed:>102735734 >102735812 >102735968 >102736030--Troubleshooting xtts2 installation issues with virtual environments and encoding settings:>102733343 >102733511 >102734022 >102734277 >102734371 >102734445 >102734528 >102734723 >102734757 >102736788 >102737035 >102734829--Glaiveai/Reflection-Llama-3.1-70B evaluation results and discussion:>102732372 >102732453 >102732507 >102733023--Diff Transformer performs better at lower bits than standard Transformers:>102734559 >102734573 >102734633 >102734601 >102734651 >102734620 >102734634 >102734828 >102734991 >102734663 >102734671 >102734681 >102734771 >102734745 >102734748--ST devs prioritize dedicated core over casual users and monetary compensation:>102734143 >102734165 >102734167 >102734190 >102734436 >102734494 >102734513 >102734687 >102734746 >102734812 >102734887 >102734711 >102734713--Rust-based web server for AI model frontend:>102732353 >102732399 >102732621 >102732651 >102732690 >102732657 >102732664 >102732695 >102732700 >102732718 >102732758 >102732843 >102732693--Leaked: ServiceTensor Agent, a neutral and friendly default bot for AI-interested users:>102731753--Chimping out and cleanup branch discussion:>102736128 >102736181--Miku (free space):>102735500 >102736657►Recent Highlight Posts from the Previous Thread: >>102731647Why?: 9 reply limit >>102478518Fix: https://rentry.org/lmg-recap-script
Is this the new imggen thread?
IT'S TETO TIME
tetters
'ocal 'ost
Is VRAM all you need? I can get AMD GPUs with more VRAM way cheaper than the NVIDIA ones.
even cohee knows it's over, why are you still in denial?
>>102737303has to be nvidia
Did anybody play around with Anole?How good is it?>>102737303>Is VRAM all you need?The more the better but AMD has the issue of software support.It works, thanks to hip/ROCM, but it's a second class citizen until AMD puts out something that has such amazing cost benefit that devs start developing for AMD directly..
SafeTavern
SafeTesnor
>>102737310It's only over when I say it's over.It's over...
>tool-callingwat is?
>>102737337Anole was only finetuned on 6000 images. Enough to get the image output working again, but not enough for it to be good.
>>102737392call tool. tool do. ai know.
>>102737392Look at this tool, he doesn't even know what the word means!
FALSE ALARM EVERYTHING IS FINE
>>102737337>until AMD puts out something that has such amazing cost benefit that devs start developing for AMD directly.Well, is the issue hardware (so, I guess AI and stuff are heavily CUDA dependant, aside from VRAM) or purely software?In my country AMD really is considerably cheaper in every GPU category, from the low end up to the high end, if the issue is software, I imagine it'll eventually get better.And I'm talking about a 4060 8GB being priced about 10% below a Radeon RX 7600 XT 16 GB.
>>102737449based tool-caller anon
>>102737214SAY HER NAME
>>102737481qrd on the ST stuff? been away for a while and i rarely update it
>>102737508WHORE
>>102737431>>102737449>>102737499thank you i now fully understand 100% completely
>>102737481I will not allow anything called "ServiceTesnor" onto my hdd.
>>102736128What happened to "I don't care if I lose 99% of users"? Don't let the chuds hold you back from your ServiceTensor dreams, Cohee!
>>102737361>>102737372If the community forks it, we need to name the new fork SexyTavern. Make it clear what it's about from the start.
>>102737508MADE FOR BBC
>>102737559SexyTavern was never about RP, that's a false history and really frustrating.
>>102737574You know what I do when I get frustrated? I head on over to SexyTavern and take a load off.
Microsoft Just started up their first Black Systems servers, as a reminder they have 2TB of vram, it's so over for local. https://x.com/Azure/status/1843637745186484406
>>102737524Fine.>You tell your model what tools it has available for what tasks.>You ask something to your model and it generates a "tool call". Typically, a json with the tool it needs to run and the parameters (your query, other params, whatever).>The client (frontend, little script, whatever) calls the tool (api call to get weather, news, compile a program, run a script, turn the toaster on...) and gives the model the data it fetched.>Model elaborates on the data it just received.It gives your model access to external tools. That's it.
>chimps out>proposes complete 180 shift in priorities>gets pushback>"why are you all overreacting"why are open source devs like this?
>>102737597*Blackwell Systems Server
you idiots come alive over the dumbest shit. l3 "updates". stupid drama. xitter bait.
>>102737605ty anon i love you whats the most interesting or complex tool-call anon has done thus far? if you'd indulge me further
>>102737597Why everyone is pouring absurd amounts of money into this? Are profits really that high?At this rate we’ll get AGI in 2 years
>>102737612they have soul unlike corpos
>>102737612hrt consumption
>>102737731Damage control activated!
Guess that 99% userbase leaving did matter.
>>102737514Majority of users are RP.Gay dev doesnt like that because "muh professionalism" and wants to make it more "neutral" aka gayGay dev pretends ST was never about coomers as if they werent directly supporting that shitGay dev says its fine if 99% of users leave because power users are better and were always the target anywayLook at the last few threads
>>102737635None, really. I don't have a use for it. I suppose i could have a model generate SD prompts+workflows and send the query directly to comfyui for example. The entire workflow would only work if the model is good enough, of course. Otherwise, a simple image prompt would be fine as well.I've seen anons wanting reaction images from their models. Not sure why they don't implement it with something like that.You don't even need a tool-call specific model to do it. Just tell the model to output the thing you want with marks you can filter/grep...I just find the tech interesting. I wish i had more use for them other than writing little stories.
>>102737508Snowbunny
>>102737833>Gay dev>Gay dev>Gay devCertainly the dev being gay cannot be such a big issue for RPers...
>>102737273I made a lucky pretty one.
What are we naming our ST fork? Who's the default character?
>>102737938Cooming tavern. 100% corporate hijack proof.
>>102737932weird mouth, enormous ears. what's with her ribs?
MikuTavern or if you want rid of "tavern" something like MikuLab
>>102737832?Stuff still is being pushed out to extensions, they still are going to rename stuff.
>>102737833So... what are we going to use now? IIRC the ComfyUI dev is a nice guy, I wonder if his skills could translate to something LLM related.I quite like his node-based UI.
>>102737392Regex
>>102737938SexyTavern gets my vote, because it retains and subverts "ST". As for our mascot(s), I say bring Seraphina back alongside the Konosuba characters.
>dev realizes he has a popular software and can make money from it is that what happened?
>>102738089Hard to know. But if you want to put "Core ST Developer" on your cv it's gonna get hard explaining its "use case" to the interviewer.On the other hand, moot got hired by google, so what do i know...
>>102738219Not really. In an interview, he just needs to say that ST is a frontend that lets users interface with LLMs. If the subject of RP even comes up, then he still doesn't even need to delve into ERP use.
>>102737597>72 GPUs>1.4 exaflopsJesus christ that's a lot of compute for not that many GPUs. I'm a boomer, I remember like 10+ years ago people were positing that we'd eventually have exascale GPU clusters for ML training, and it seemed so far away at the time. Now a single server rack with a couple dozen chips gives you an exaflop.
>>102737948Only if the default card is blacked miku.
>>102738219It's not hard at all. It's a frontend for recreational chats with LLMs.It would be better to say it's a very popular frontend, but he probably will still have a lot of installed users of old versions.
So, how's llama3.2 90b as a normal llm if you ignore the image stuff?
>>102738511The same as 3.1, by design.
sorry for these absolutely retarded questions, but i'm relatively new to local LLMs and trying to figure things out from lurking in these threads and reading the ST wiki... is ST trying to distance itself from coomer content, and can they even do that? if you're running a local LLM, can they even tell what you're generating?i've been experimenting with both erotic and not-erotic RP stuff and it's all pretty neat, i would be sad if they somehow limited it in any wayalso, this was on their wiki, does it still hold true today?>As of January 2024, Mixtral MOE 8x7B is widely considered the state of the art for local LLMs. If you have the 32GB of RAM to run it, definitely try it.ty for bearing with me anons
>>102738511I think I heard the original weights were frozen but I'm not sure or don't remember clearly.
>>102738553No, today's SOTA is Nemo.
>>102738528it's an extra 20b in size though
>>102738553>i would be sad if they somehow limited it in any waythe plan is to "organically" get rid of roleplayers and turn it into a productivity app
>>102737597So don't care. We don't need 2TB of vram when 123b largestral already rocks for RP. We just need enough vram for our individual usercase. We're reaching a point where just adding more vram has diminishing returns anyway. Compare 400b llama3,1 to 123b largestral. The 400b is better, but not by a high margin.There will always be a demand for local, because local users have fundamentally different priorities than cloudfags. I want models that are uncensored. I want models that are untainted by the worst aspects of woke politics. My models may not be 2TB in size, but they know what a man and a woman is. My models work when I'm offline. My non-windows-based activity is private, and not being watched by agent Smith.
>>102738607That's the image stuff
>>102737218>--ST devs prioritize dedicated core over casual users and monetary compensation:That's an unbelievably biased way of putting it, you shilling tranny. Never bake a thread again.
>>102738607It's literally 70B with a 20B image processing AI glued on top lol
>>102737567>>102737900
>>102738307Forgot to include you in that one >>102738678 you obese balding crossdresser.
don't interact
I'm getting real fed up with all the information about ServiceTensor in this thread. Grow the fuck up.
>>102738713watch instead, black gods CORRECTING your waifu!
>>102738734How do I edit posts on 4chan? I meant to say misinformation.
>>102738582>>102738623thanks anons. any suggestions for which Nemo model to play around with? is the base Mistral one okay for RP, or is it better to use a spin-off model like ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.1-GGUF?
Can I get a quick rundown on the sillytavern breaking changes? Did the tranny dev finally lose it?
>>102737978He is a complete fag. No noodleshit please.
>>102738770all RP features are being removed and changed to extensions, massive rebranding incoming, RPers to be "organically" squeezed out while they shift focus to productivity and tool-calling
>>102737214>>102737218Hey Recap Anon,I hope this message finds you well. I wanted to take a moment to address something that I think could be an important step for LMG moving forward.As the CEO of LMG, you have a unique opportunity to shape the direction and ethos of the community. There’s been a lot of talk lately about the nature of the content being generated, and I believe it’s time for LMG to take a clear stand on making the platform a family-friendly space.LMG has always been a place where people come together to share ideas, engage in discussions, and build a community. But over time, the tone and content have shifted in ways that might alienate younger users or those who are looking for a more welcoming environment for all age groups. This isn’t just about toning down certain posts—it’s about cultivating a culture of respect, inclusivity, and positivity.Here are two reasons why I think a move toward family-friendly content could benefit LMG:1. Wider Audience Appeal: By making LMG a space that’s safe for all ages, you open the doors to a broader audience. Families, young people, and even schools might be more inclined to engage with and contribute to the platform, knowing it’s a space they can trust.2. Brand Reputation: LMG has a growing influence, and being seen as a company that values family-friendly content could enhance the brand’s reputation. It shows a commitment to social responsibility and sets a positive example for other communities.I’m not suggesting a complete overhaul or censorship of content, but rather a strategic shift towards guidelines that promote a healthier, more inclusive community for everyone. This is about creating a space where anyone can contribute and feel comfortable without worrying about inappropriate content or harmful behavior.You’re in a position to lead this change, and I believe the community would rally behind a move that prioritizes positive
>>102738789... engagement.
>>102738784That's like... your opinion man...
>>102737596He is right but you are too. Here's the initial Readme.md for the project.https://github.com/SillyTavern/SillyTavern/blob/edd41989fd550a8d111fb7167d456c5614a3a610/.github/readme.mdIt was always as intended for powerusers, sure. But that also doesn't justify the removal of the anime and ERP friendly features given mentions of "VN-like Waifu Mode" and the example screenshots of which pic related literally has an interaction with Misaka Mikoto and a Brazilian catgirl that is borderline there.
>>102738693blacked miku is thread culture
>>102738833Your time is up, ERP chud. ServiceTesnor is for productivity chads now.
>>102738747I like Nemo instruct the best
>>102738833Kinda racist too, would love to see the transgender folks bitch about him being a neo-nazi or something.
>>102738865where's the lie though
>>102737833>Gay dev pretends ST was never about coomers as if they werent directly supporting that shit>Gay dev says its fine if 99% of users leave because power users are better and were always the target anywayI find the implied distinction between coomers and power users strange, who uses something constantly and doesn't eventually become a pro at it?
How slow would a 12B model be in an 8GB card?I'm using https://huggingface.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator to check it out and it says I can't do it, but I guess you can always offload to RAM no? What's the actual limit of usability for an 8GB card? 16B?
>>102738785So they want to become open webui to get some sponsorships? The whole point of this project was for cohee to pretend to be a woman and bring his tranny fantasies to life. This won't work.
>>102738926my mom owns a smartphone for about 10 years now, always the same samshit stuff too, and she still frantically taps the "back" button whenever she wants to switch appsyes, i did try to teach her, yes i showed that theres two other fucking buttons RIGHT BESIDE the back button, she simply doesnt learn.
>>102738933Have as many layers of the model in VRAM and the rest in RAM and you'll get some good speeds.Remember to use flash attention with q8 cache and to tweak the batch size to free make the best use of your VRAM.
>>102738959>Not teaching her gesturesngmi
>>102737282>>102737310>>102737567>>102737900>>102738307>>102738695>>102738740>>102738848>>102738850We're getting raided by whathever generic grooming discord channel is active tonight. The thread will be unusable until they all 48% themselves or jannies decide to do their "job", whatever comes next.Of course they decide to stop "cracking eggs" and come here right when something actually happens for once.
>>102738981what?>>102738979that's three gestures that dont even show on screen, when she cant tap the fucking button .5cm away from the one she always uses
>>102738986She is so valid!
>>102738970There's a 4-bit_cache option now for exl2. Saves even more space than a q8.How does flash attention help, though?
9 mentions limit doing it's job, nice work 4chan
>>102738986That hairline is uwu as heck
>>102738986You lost :3
>>102738782>>102739016It was funnier when you freaks were spamming scat hentai a few months ago whenever someone called you out on your non-stop BBC posting, but I guesss the permabans got too annoying so now it's Dall-E slop.Either that or the specific troon with the 15tb of loli scat finally ACK'd himself and you're the new guy.
>>102738933I use either of these two configurations:Q4_K_M GGUF, 32k context length, 35 layers on GPU, 4-bit KV cache quantizationQ4_K_M GGUF, 16k context length, 31 layers on GPU, no cache quantizationthe first one generates tokens faster, but cache quantization is incompatible with context shift, which means that once you exceed that 32k limit the context has to be reprocessed before every single reply, so I only use it when I'm sure I'm not going to use more than 32k tokensalso none of the nemo-based models can actually remember 32k tokens, the only reason I chose that is to postpone reprocessing as much as possible
>>102739081Not really, I've been on /lmg/ for about ayear and your raids typically only last for a day or two.
>>102739084>the specific troon with the 15tb of loli scat finally ACK'd himselfDoing friendly fire huh?
>>102739135Lower your temp
now that ST is over is it the end for local models? Can I delete my safetensors?
>>102739174unsafe sex with troons = hiv, I have to go safe
>>102739101>>102739174Go back >>>/pol/ chuds.
>>102738970>>102739086Thank you guys, I'm really having a hard time picking between the 3060 12GB and the 4060 8GB. I'm a massive poorfag so these are the only cards I can reasonably afford. Would a 12GB card make 20B models fast enough where a 8GB one wouldn't?
>addition is all you need>differential transformerwe are so back
>>102739198I was on /g/ while you were still but a novice troon, yet to groom even a single child.
>>102739199The 3060 is the better buy regardless.It isn't that much slower in games than the 4060, and the 12gb of vram means that you can both run better models (you can probably run Q4 mistral-small at decent speeds), but you'll also be able to get get better visuals in game by using higher resolution textures than with the 4060.With DLSS quality you can tun 1440p pretty well, I reckon.
Been away for a while, has there been ANY improvement at all since Largestral? Can't seem to find anything and I find it hard to believe that nothing better has come out...CYOA text adventure RPGs if that matters.
>>102739199the 4060 is faster but that's insignificant compared to the performance degradation caused by CPU offloading, so get the 12 GB one
>>102739278Oh yeah that's right, after the trannies stoped spamming scat porn due to getting permabanned on sight they started wpamming walls of irrelevant text. The mind of a tranny is a strange thing indeed.
>>102739199Get a 4060 ti with 16gb.
>>102739285>>102739233I don't plan on ever leaving 1080p, if anything I might just get a higher refresh monitor or something. I guess I'll just cave in and get a 3060 anyways...>>1027393144060Ti is 2x the price of a 3060.
remind me again why are we talking about healthcare on a llm thread on the technology board
>>102739343>4060Ti is 2x the price of a 3060it also has a narrower bus compared to the 3060it's a scam
>>102739376Raid.Just report and move on.
praying for lmg t. chad from other local bred
die spammer
>>102739376The Discord trannies started derailing the thread with these posts, >>102737282 >>102737310 >>102737567 probably because of the SillyTavern controversy. The lead ST dev, a tranny, has decided to 41% his own frontend, and everyone is understandably pissed. So now his fellow troons are raiding the thread, as they ghave dozens of times in the past.It'll have subisded in a day or two, nothing to do in the meantime except making fun of them.
Good heavens would you look at the time.png>i dont have it but i remember the image please post the lmg clock guy
>>102739513
>>102737214"Best rp model" for 24GB vramlets? Fucked around with Mixtral, MM and CMR, MM being the biggest bitch to handle oddly enough. Been lately fucking with Mistral-Small-Instruct and been happy, but feeling a bit limited in some aspects, or too basic and repetetive.
>>102739492>troon dev does something retarded>gets called out>suddenly /lmg/ gets buried under bot posts and BBC postingJust a coincidence.
>>102739542polskins do be like that.
>>102739545I was going to suggest mistral-small.What do your samplers look like?
>>102739542Thanks Anon
So that confirms it's literally one guy right?
>>102739585it's petra
>>102739589Hi Sao
What about pic related for default card in the ST fork we are totally doing now?
I miss Undi....
>>102737978whaaat??? in what planet is comfy a nice guy????? he's a complete piece of shit.
>>102739563I keep forgetting that those seem to matter, because I usually only fuck with formatting templates and be done with it.Think I haven't touched my completion thing in goddamn forever, only touching it for context and temps.Pic related is some "mixtral-roleplay" template from yesteryear.
>>102739585Probably a couple troons in a Discord channel.
>>102739545miqu if you have the ram to offload, it's not a terrible speednemo if you dont
>>102739619These are the casuals cohee is complaining about, by the way.
>>102739619>some "mixtral-roleplay" template from yesteryear.Suspicious.Save your current sampler config, and try with neutralized samplers.See how that behaves.Depending on if it's unhinged or it's accuracy seems low, lower tempo and increase minP to 0.05.Or try a higher temp with topK.For Nemo, something like temp 0.5, topK 10, minP 0.05 is pretty okay.
>>102739585It's the same autist that's been spamming this general for months, he just changes what he spams from time to time.For some time he spammed "Petra" while mass-replying (long before the mods filtered it), he spammed blacked Miku, he pretended to be ggerganov, the list goes on.
>>102739672Good. Retarded gooners aren't welcome on ServiceTesnor.
the lore is rich and deep i kneel
Should I get a second RTX 3090 with NVLink for better performance, or sell my 3090 now and wait for RTX 5090?
>>102739684You forgot when he spammed loli scat hentai too, the one time when jannies were quick on banning him.
>>102739714nvlink doesn't do shit unless you go for tensor parallelism
>>102739714I'd go with two 3090 if you aren't a gaymer.Maybe even if you are.
>>102737643>we close epistemic gaps by spending every dollar that exists on processing every possible outcome
>>102739645Screwed around with IQ2 as a joke way back to fully offload it, never found a model I had to offload fast enough to my liking.>nemo if you don'tGoing 12b nemo over 22b mistral... hmmmmm>>102739678>try with neutralized samplersJust to be sure, what do you mean with that exactly? Reset the order, or some option I didn't notice? Either way, thanks for the advice, gonna fuck around with it (and maybe nemo) when I find the time later.>>102739736>Maybe even if you areNot much point for SLI/NVLink with modern video games, or old for the matter.
>>102739750>Just to be sure, what do you mean with that exactly?The sampler order is on the bottom of the sampler panel.Near the top, there's a button called Neutralize Samplers that disables everything and puts temp at 1.>Not much point for SLI/NVLink with modern video gamesThe point was more that a single 3090 is still a pretty good gaming card if you aren't on the 4k highest settings treadmill.
slop = 1.0In the dimly lit room, Reiko sat across from her, legs crossed, fingers steepled. Her fiery red hair cascaded over one shoulder, long lashes casting shadows on her pale cheeks. She wore a smirk, half-lidded eyes daring a challenge."You're playing with fire," Reiko warned, knuckles turning white against her restraint.Her opponent, a woman with chestnut eyes and dusky nipples visible beneath her thin blouse, bit her lip nervously. Yet, she held Reiko's gaze, pupils blown wide with defiance. "The ball is in your court, Reiko."A rivulet of sweat trickled down Reiko's neck, her heart pounding like an audible pop in her ears. She wanted this game, needed it. Propriety be damned."The choice is yours," Reiko purred, voice huskier than usual. She leaned in, nails raking angry red lines down her own arm, grounding herself.Her adversary shivered, breath hitching. "I don't bite… unless you want me to." She grasped Reiko's chin, forcing their gazes to meet. Her cheeks were aflame, stars bursting behind her eyes.Reiko felt the world narrow, her focus solely on those chestnut irises. "Take your pleasure, then," she dared, feeling her own arousal pool in her belly.Her opponent's tongue darted out, wetting her lips. "With reckless abandon? Or wanton?"Reiko grinned wickedly. "Both. The night is still young."Their souls entwined, pleasures unspoken faded into the background. For now, they had each other, holding onto the hope that their bonds would only grow stronger.
Any card writing people itt? I'd love to see how people make them nowadays, I remember lots of people arguing over Alichat, XML, blablabla last year. Been thinking about rewriting my card to see if I can make it somewhat better.
slop = -1.0Facing her rival under the soft light, Reiko maintained a composed demeanor despite her racing heart. With her vibrant red hair and porcelain complexion, she cut an imposing figure, her gaze intent."You're playing dangerously," Reiko warned, fingers tightening around the table's edge.Her counterpart, a woman with warm brown eyes and rosy cheeks, held Reiko's stare unwaveringly. "It appears we've reached a stalemate," she noted calmly.Reiko's breath hitched, her focus narrowing on the woman before her. "What's your suggestion?" she inquired, reclining slightly.The woman pondered for a moment before responding, "Maybe we should investigate our choices together." Her tone remained steady throughout.Intrigued, Reiko agreed with a slow nod. "Very well. Let's discover where this leads us."
>>102739777Only if you promise not to use the card on ServiceTensor.
>>102739793I don't even know what that is, I just want to pretend I still have friends.
>>102739777I just do plaintext with concise definitoons like >character is, character has, character looks likleetc.Anything fancier is unnecessary in my opinion.What about you, yugi?
>>102739769>Their souls entwined, pleasures unspoken faded into the background. For now, they had each other, holding onto the hope that their bonds would only grow stronger.Peak slop, up there with shivers down the spine and "What the future held for them, tehy did not know, but they would face it together" and such.
>>102739777>>>/vg/aicg/
>>102739769>>102739783Yet another attempt at slop isolation has proven unsuccessful, resulting in damage to the model during the extraction process.
Has everyone forgotten agnai? It was supposed to be the next hot front end a year and a half ago.
>>102739777>>102739814This, plain text is good enough especially if you use adecently powerful model.If your model is shit you cna try separating the description into sections like:Char's persona: [blablablabla]Scenario: [blablabla]etc. Though it's probably more of awaste of contect than anything.
>>102739767Ahh, now I see. Thought you were talking about resetting the sampler order, the buttons at the didn't even register in my head. Thanks, again.>a single 3090 is still a pretty good gamingTrue enough. Thought you were telling him to go for SLI gayming.
>>102739777Interview-style cards are the best. Just some basic info and the rest of the card is dialogue written like an interview where the character talks about itself.
>>102739814>>102739857Well the thing is, the card is based in an existing character, that makes things a little harder since you can notice when their behavior begins to differ a little.I'll try pure plaintext anyways.>>102739826I'll ask there too, thanks.>>102739876Had very little luck with those, only tried twice but the results were very bad. Do you have any good interview-style card I can check out?
>>102739876What about examples? Is interview preferred now?
Damn, both 777 and 888 gets. I'm blessed.
>>102739888>the card is based in an existing character,Create a lorebook with keyowrds that would appear in certain situations that describe an exemplify how the character would behave.
>>102739750>Not much point for SLI/NVLinkSame for LLMs. Only useful for TP, yet exlama doesn't support it (yet) https://github.com/turboderp/exllamav2/issues/603
>>102739916lorebooks are completely pointless if you don't have ass for context, just put the information in the card
>>102740111go back >>>/g/aicg
where is qwen coder 32b? ACK
>>102739826>>102740160Shut up tranny. Constructive discussion will be had.
>>102740111The advantage is being able to control the depth of it's inserted in the chat and it being conditional.
I should put one of my 3090s in my gaming PC. Last time I did that Llama-3 dropped basically the next day. I will force a major model release through irony.
>>102740184Nice constructive reply there, chud. This is why gooners are being pushed out.
>>102740184>Constructive discussion will be had.Like:>Shut up tranny.
>>102739793You can service my tensor faggot.
>>102740213Welcome to modern day 4chinz, the basket weaving forum full of discussions around transexualism!
what model should i use if i have a 2080 ti? 11 gb vram
>have the best rp front end>destroy it to make souless fagtron corpodesign #500very high iq
>>102740274Just tried this, the (female) character told me that it's a sensitive topic, also that I shouldn't use a derogative term like that.
>>102740301It was never a roleplay frontend you fucking retard chud incel piece of shit.
>>102740301>souless fagtron corpodesign #500I hope they add a lot of pic related.
>>102740211>>102740213>>102740222>Shutting down requests for advice is not constructive, but getting called atranny over it is notFucking off yourself already you cunt, we know you're gonna do it soon anyway, why not now? So that we can have some peace and quiet.
>>102740326>Fucking off yourself already you cunt, we know you're gonna do it soon anyway, why not now? So that we can have some peace and quiet.This is not constructive discussion.
>>102740325for me its the ones with the huge bodies and tiny heads
>>102740211>>102740317>Using chud unironically
meta should buy mistral
>>102740361>meta should kill mistralas a large language model
>>102740361Please no.
blacked miku anon here. should I take a dump or are you guys doing fine without it?
All I see in this thread are a bunch of whiny, entitled gooners who are too stupid and lazy to contribute to ServiceTensor yet feel like the devs are OBLIGATED to give you everything you want. Keep bitching but you WILL download the roleplay extensions like a good bitch. Productivity chads and power users are taking ST back for good this time.
>>102740400dump em on the st discord
>>102740415>ServiceTensor
>>102740454Can anyone think of a name that is gayer than service tensor?
>>102740242Nemo based fine tunes.The official instruct and rocinante are my favorite currently.
>>102740438How about you make good posts instead of crying about trannies, retard?
>>102740415tensor this service FAGGOT
>>102740173is normal 32b not good with coding?Since 72b was good at it I expected the same of 32b.
>>102740514it is, but still
>>102740454ServiceTesnor*
>>102740533Yeah keep being an immature dickhead just because they had a typo but the future is here and you're not welcome anymore.
Sisters I can't take it, SafeTensor removed muh sarafina.... how could a tranny dev behave like this, this is totally unexpected......
Gonna start filtering the T-Word from now on frrfr this homophobia is getting real tiring senpai
>>102737481chickens for KFC lol
>>102740586>homophobiadont lump us faggots with troons
>>102740556This is not the first bait with this speech pattern where there's an obvious disconnect between the first and the second part, not fully understanding which side it's supposed to represent. Who bored are you?
>>102740604>Who bored are youThe absolute state of casual low-intelligent gooners. ServiceTensor is going to be so kino.
>>102740604who bored are you indeed
>>102740604Ignore him, he's mentally ill
>>102740604I am permanently bored
>>102740415>ServiceTensorkek nice one anon
What the fuck is all this servicetensor schizophrenia bullshit someone is trying to chill and why should I give a damn?
>>102737508she stands like a slut>>102737517she doesn't need money in the forest, your improper slur usage isn't DEI (Dark Side) approved
>>102740643>What the fuck is all this servicetensor schizophrenia bullshitSimple Tavern renamed themseves ServiceTensor I guess?
>>102740644Seraphina doesn't wear panties
>>102740644Is it morally correct to hire a hooker to attract people to your establishment and then fire her when your establishment gets popular enough?
>>102740664>Is it morally correct to hire a hooker to attract people to your establishment and then fire her when your establishment gets popular enough?no it's not, and that's the strategy of any big companies at this pointhttps://en.wikipedia.org/wiki/Enshittification
>>102740643SillyTavern decided they don't want coomer users anymore and are changing everything to full corposlop
>>102740654>simple tavern renamed themselvesI'm fucking sorry??
>>102740643see>>102738785
>>102740684You will be.
This too shall pass
>>102740677>>102740690>They don't want their main audience anymoreAHAHAHAHHAHAHAHAHAHHAHAHAHAHAHAHFunniest thing I've read all day, holy shit that is hilariously retarded, but expected of modern day open source ""developer"".
simple as
Don't worry the chuds(tm) will fork the project and maintain it for several days!!! We're saved!!
>>102740735cope
>>102740716The funniest part is they openly say they're fine with losing 99% of their audience (like it isn't straight up 100%), all to enter a market that's already so saturated that they have virtually zero chance of breaking in. They're legit fucked in the head.
>>102740774Where did they say that? I'm a mentally healthy heterosexual male and don't visit discord so I haven't seen that conversation.
i wish gemini 1.5 flash 8b was open
>>102740774>they're fine with losing 99% of their audienceWhere have I heard this line before... Oh yeah, the movie and TV industry!They (talentless hack fraud activists) are super okay with owning those chuds, while the people above them scream in agony because they're bleeding billions.>they have virtually zero chance of breaking inMany such cases, many such cases... Fucking adore modern days hyper retards, or how they keep manage to fail upwards in this wacky ass society.
>>102738785YES, FINALLY!
As a power user who accomplishes actual real shit with non toy models like GPT-4o1 preview I AM GLAD Sillytavern is finally changing. Finally I can code in peace.
I, a fellow pro coder, too, am using the widely known software SillyTavern for all of my GPT-4 professional adventures!
>>102740816To be honest, this is how I read this situation (trust me I know things, not from ST because I don't even use ST but from LIFE):>rossascends is a poorfag (I am too) and received an offer from some startup to rebrand the project in return of cash>coheeeeee the trans was probably frustrated with the project (internal discord feuds? people being annoying on github issues? who knows), and went along>suddenly they go from a reference project in the domain of roleplaying to a pariah project, criticized all the time with 1 or 2 autists to defend them and hundreds of people shitting on them daily>a downward spiral begins where the more they get shat on, the more they lose their motivation, and the less they code, the more they get shat on>the project dies in a few months
As a fellow coder I am also happy ST is finally taking out the trash.
>>102740874As an AI language model I...
>>102740844>>102740856It's ServiceTensor, chuds! Stop deadnaming the project!
Who the fuck cares about ST, talk about something else already
>>102740901No. =)
>>102740784Fucking hell anon, it's 1am and I'm fetching dicksword screeencaps just for you>I'm fine with losing the casuals to keep the power usersFuckers will have neither in the end at this rate
>>102740901discord general
>>102740907Thanks, that was amusing to read.
I guess this thread achieved ASI, judging by the posts…
>>102740907Hey pass it on to him: when you realize how bad of an idea it is and you come crawling back to us coomers the only way we will accept your apology is if you make a pissdrinking video. /aicg/ key begging rules of course.
>>102740930we have ACI (artificial chud intelligence)
>>102740955Nice headcanon, you'll be downloading those roleplay extensions like a good pig.
>>102737214>lastest news 09/27
>>102740907Addendum: he also has to suck off >>102740966on cam after drinking his piss. If he is >>102740966 sucking himself off is acceptable.
>>102740955Bold of you to assume I'd set foot anywhere near that cesspool
It's Tuesday, I'm in love
>>102740976it's unironically over.>>102741008>hyper generic AI artperfect
what is this retardation in lmarena sorting? lmao
>>102741026o1 on suicida watch!
Huh Rocinante 1.1 seems actually good, and with fresh prose. I thought finetunes were supposed to be memes
>>102741044You better not be roleplaying with that model.
>>102741112I'm using it for storytelling, the prose seems good enough that to maybe replace Command-R v01 for me after all this time. But out of curiosity, what's wrong about roleplaying with it?
>>102739245there are 0 good models besides largestral and miqu.
>>102741026>our strawberry o1 we hyped for a year is so sophisticated (CoT lol) and so good that we managed to beat it with gpt4-o anywaywtf are they doing
>>102741128Roleplaying reflects poorly on our professional power user coommunity. We do not accept roleplaying casuals here.
>>102741128I second this question even tho im mostly a storyfag.
its starting
sirs when 48gb gpu for the average joe?
>>102741283Being incompetent while destroying the company
>>1027413232 more weeks
>>102741323Q1... 2025Just stop being poor and pay up for that quadro 48gb goodness
>>102737214Bellatrix from DFO.
>>102741283LLMs are over...
What's the last unpozzed ST version?