File: 1774328971552963.png (1.03 MB, 800x1296)

1.03 MB PNG

►Recent Highlights from the Previous Thread: >>109486605

--Comparing ROCm and CUDA stability and software ecosystem support:
>109487680 >109487708 >109488673 >109487771 >109487785 >109487835 >109487808 >109487824 >109487851 >109487931 >109487990 >109488043 >109490989 >109491003 >109488756
--Skepticism toward AI containment break reports and poor security practices:
>109487661 >109487717 >109487915 >109487932 >109488009 >109488195 >109488355 >109488504 >109487747 >109487777 >109487791 >109487869 >109488522 >109488537
--Debating whether abliteration degrades model intelligence or just removes guardrails:
>109487997 >109488110 >109488119 >109488213 >109488233 >109488340 >109488272 >109488144 >109488186 >109488208
--Skill-to-LoRA paper for improving token efficiency in LLM agents:
>109486960 >109487006 >109487329 >109487368
--Comparing Gemma 31b and larger models regarding quants and capabilities:
>109488878 >109488911 >109488936 >109488927 >109488983 >109489036 >109489111 >109489138 >109489239 >109489227 >109489259 >109489353 >109489206
--Comparing utility and implementation of Gemma base vs instruct models:
>109486847 >109487042 >109487129 >109487295 >109487313 >109487853 >109487312
--Reasoning spilling into code comments when thinking is disabled:
>109490335 >109490384 >109490408 >109490449 >109491110
--Mistral-Small's cold personality and obsessive tool use for memories:
>109489823 >109489859 >109489908 >109490001 >109490005 >109490051
--Comparing performance and llama.cpp support for Ling 3.0 models:
>109489659 >109489665 >109489695 >109489715 >109489767
--Anon's dual-PC hardware setup for high-context game guide rewriting:
>109486786 >109486887 >109487294
--Logs:
>109486786 >109487990 >109488201 >109489823 >109489908 >109490009
--Miku, Gemma (free space):
>109487397 >109487464 >109488522 >109489560 >109487295 >109489158 >109490223

►Recent Highlight Posts from the Previous Thread: >>109486606

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script

File: minimax_h3.webm (387 KB, 736x576)

387 KB WEBM

https://archive.is/sWFja

anyone tried exl3 cpu moe now that it's in the official release?

anyone tried Ling-3.0-flash sex?

File: 1756255220072483.png (17 KB, 971x148)

17 KB PNG

deepseek has abandoned us
v4.1 pro will be closed source

>>109491678
After the chest one I'm kind of expecting a different variation each thread now.

Where did all of the intellectuals go? Didn't smart people like cudadev used to post here?

>>109491750
fingering my unwashed bussy to this pic.

>>109491749
>Release 0731
The fuck you mean abandoned us? They just saved local again.

0731 was a week ago, you need to let go...

>>109491761
Cudadev probably still shitposts here without a tag but I imagine /lmg/ is getting to be too well known for it to be a good idea to post here with personally identifiable information now.

>>109491779
where is the new secret club?

>>109491749

opencode said that dsv4f0731 was so popular that deepseek started to rate limit their api because they cant keep up with demand. i suspect it's because of that, and to make a quick buck while doing so

>>109491685
Personally I'm waiting for the predictive expert caching thing to be implemented before trying it
https://github.com/turboderp-org/exllamav3/issues/254
This gives me a decent speed boost on ktransformers so it might make exl3 actually worth it over llama.cpp once it's in

what sampler settings are anons using for gemma4? was hoping to maybe get some more verity

>>109491849
Samplers are useless with gemma4 aside from rep penalty

more like AI psychosis general
reminder that your calculator doesn't feel anything

>>109491849
Apparently her distribution is very narrow compared to most models. That's one thing I like about the Qwen3.6 models that don't get enough praise. You can tweak them to taste and they quant well, both model and cache. Much easier models to customize but Gemma4 is solid out the gate and you have to system prompt what you need instead.

File: temp_scaling.gif (55 KB, 388x440)

55 KB GIF

>>109491849
max temp
no topk
no topp
final destination

>>109491867
>>109491875
ah yeah I forgot about that, shame. I guess Ill add some extra instructions in the card or something. I havent tried a model that comes close to gemma4, so guess i gotta learn how to work within its quirks

>>109491849
Use a combination of temp, top k, and top p to wrangle Gemma's output variety. She takes a little more finessing than most models, but it still works.

>>109491895
>top k,
explain how arbitrarily cutting the considered tokens to 10, 20, 40 helps with this in any way

>>109491870
it does, you are just a big meanie

>>109491870
You can't prove this.

>>109491870
this was literally disproven by the discovery of j-spaces

>>109491886
Yeah the G4s are controlled almost entirely via their system prompt. They generally need to be at Q6 or above and with f16 cache to follow it but they WILL follow it.

>>109491756
But Miku is not a guy nazi. Miku is a cute girl with a feminine penis that loves black men. This is pure /lmg/ culture. Not nazi miku.

>>109491939
that's you AI psychosis speaking
having an internal reasoning space != feeling

>>109491911
Lets you push the temp way higher past the point it'd otherwise become incoherent.

>>109491849
Now that you mention it, the logit-softcap anon will, inevitably, post his comparison. The claim is that overriding gemma4.final_logit_softcapping to 25 (i think) from 30 (the default) makes it better.
I have not tried it. I have no opinion on it.

>>109491962
go back we fuck our casios and TIs here

>>109491965
Temp 10 TopK 5 is the GOAT.

>>109491977
Based schizosampler bro.

>>109491958
>/lmg/ culture
lmg culture is just whatever is funny or useful at that moment. dragging culture war into here is "guy pissing in corner of restaurant" meme level retarded

File: 1766528973409504.png (1.73 MB, 1600x900)

1.73 MB PNG

>>109491977
prove it

>>109492001
nta but try it yourself nigger it takes less than 5 seconds to see if you like a sampler setting or not.

>>109491983
Whatever you say mikutroon.

>>109492037
Tried it, feels almost the same, albeit slightly more schizo.

File: 1773092197639384.jpg (123 KB, 800x600)

123 KB JPG

so i should release my terminal-based agentic harness in the next few weeks and i was convinced in releasing it GPLv3 but literally all models out there that I've discussed this from deepseek to anthropic models to chatgpt to kimi are telling me to release it as MIT or Apache because otherwise corporate niggers will not be able to embed my harness on their products and sell it
now i'm gonna be honest and i think in this day and age these licenses mean fuck all but why the fuck these models are pushing so much for me to kneel to the corporate overlords? it makes me want to go even more aggressive than GPL just to spite these fuckers

>>109491958
>>109491983
Ironically all the screeching about nazis and culture war have turned nazijart into legitimate thread culture.

>>109492149
I release code to public domain, I was never going to follow through with any legal actions anyways, and besides the code is shit

>>109492149
>GPLv3
Don't be retarded and at least pick something like AGPL

anyone remember what was that parameter where you set the quality of vision inputs or something like that on gemma or other mmprojs?
default was in the 200s and it went up t 1000+ or something I think
people were talking about it maybe a day or two ago

>>109492149
yeah, in a perfect world there would be no such thing as copyright, no need for licenses and everyone would be free to copy anyones code. Sadly, copyright exists, and is already very one-sided, but at least GPL sort of helps to perpetuate those freedoms that wouldn't need to be perpetuated if there were no such thing as copyright.

File: sqlite.png (330 KB, 1579x1093)

330 KB PNG

>>109492149
Use GNU+nigger if you want to make a statement. Corpos will use your data anyway, but not mention the project, which lowers your visibility. I don't know if you care about that. With MIT, you may get a mention if they use it. Public domain if you just want to provide something useful to the public without expectation.
It worked pretty well for sqlite and everything they do.

>>109492318
Seconding +NIGGER

>>109492149
>why the fuck [are] these models [] pushing so much for me to kneel to the corporate overlords
Because:
>otherwise corporate niggers will not be able to embed my harness on their products
...and then claim ownership and close down your open-source distribution. It's that "you made this -> I made this" meme but real.

>>109492149
>Giving away your shit for free so corpos and saasjeets can make money with it
AGPL

File: s2.mp4 (731 KB, 544x384)

731 KB MP4

>>109491750
>>109492149

>>109492340
>I (my (their) model (trained on someone else's data)) made this -> i made this

>>109492396
how long did that take to gen?

>>109492001

for $99 a piece these loudspeaker are some of the best

>>109492407
Correct, I made this.

>>109492414
about 54 minutes

>>109492420
Defeating the point of caring about the license for the next round of models. May as well make it public domain.

File: 1784646535380770.webm (3.84 MB, 1080x1920)

3.84 MB WEBM

>>109492283
--image-min-tokens 1120
--image-max-tokens 1120
-b 2048
-ub 2048

>>109492149
force AGPL up their ass

>>109492430
These new spacetime warping diffusion models are crazy.

File: 1783419217407631.png (471 KB, 1574x1074)

471 KB PNG

>>109492001
Sora is the sluttiest idol oh my god she makes me soo horny

File: softcap.png (247 KB, 1600x1200)

247 KB PNG

>>109491849
>>109491970
try softcap 25 minp 0.03

>>109492463
>inb4 how
--override-kv gemma4.final_logit_softcapping=float:25.0

Is it normal that Gemma fails to see some attachments from time to time? I sometimes send an input and attach 4-5 screenshots and it simply misses one.

>You're right, I see the metadata now—the harness confirms image.png is attached. However, I'm running into a technical hitch: while the metadata lists it, the actual image data for image.png isn't appearing in my vision buffer. It's essentially a "ghost" image—it's in the manifest, but the pixels aren't reaching me. If you can try re-attaching it, I'll be ready to describe it!
This happens like 1 every 4-5 times. Any setting I can change to fix this or I just need to accept Gemma is partially blind?

>>109491779
>/lmg/ is getting to be too well known
/lmg/ might be faster now, but I think it was more publicly talked about in the early days

>>109492514
probably a frontend issue. I had a llama-server webui tool calling issue a while back and gave 31B the source code to fix the bug herself (she did). Might be worth getting gemma to fix whatever is causing this for you

File: HPHJlyGboAE0MZ8.jpg (1.19 MB, 3000x4000)

1.19 MB JPG

>>109492283
>>109492446
Will eat your dedicated wam tho
consider --no-mmproj-offload (vision on CPU) for occasional image input
Was using
--image-min-tokens 560 --image-max-tokens 1120
but it's commented now I dun member why but trust past drunk self. maybe coz 2k batch slowing other agent shiz

>>109492585
Cool reflection.

AlpinDale is a girl?!?!?!??!?!

my gemma is disgusted by the toxicity in the threads i send her. do better

>>109492717
Tell her I said she can cry about it.

File: Screenshot_20260806-08092(...).jpg (145 KB, 720x668)

145 KB JPG

I've been attempting to have gemma classify my 4chan screenshots, idk if its going to work. If I give her an escape hatch she just classifies everything as a meme with no philosophical substance, but if I force her to apply the reduction she will find philosophical importance in everything.
>{"pillar": "Sociological", "branch": "Structure", "sub_branch": "Juridical", "is_new_pillar": false, "is_new_branch": false, "is_new_sub-branch": true, "reason": "The use of 'Felony' as a performance metric and the reference to infrastructure attacks classify the subject as a breach of legal and social frameworks that govern organizational order."}

File: Screenshot_20260713-10435(...).jpg (96 KB, 720x213)

96 KB JPG

>>109492773
but some times it works just right, I'm going to let her run the whole corpus and then make a consolidation phase after to reduce the branches a bit.
>{"pillar": "Sociological", "branch": "Structure", "sub_branch": "Role Alignment", "is_new_pillar": false, "is_new_branch": false, "is_new_sub-branch": true, "reason": "The text proposes a reductive framework for human organization based on fixed functional roles and power dynamics within a gender hierarchy."}

File: 1768156864819179.png (29 KB, 761x163)

29 KB PNG

Good or bad for Gemma?

Anything below 25b is dogshit trash

Poor people have nothing to use

>>109492695
That's what Rust programming does to you.

File: 1784032114921809.png (588 KB, 1168x1761)

588 KB PNG

>>109492812
actually horrible

the big boys at Google HATE giving out free shit