>>110208348
Based OP. Death to Gemma

>>110206079
Working on layer 47 now.

>*BANG BANG BANG*
Anon? We know you're in there! We don't have to do this the hard way. Give up the ram!!!

File: file.png (155 KB, 484x361)

155 KB PNG

>>110208371

>>110206079
I have 2 5090s but I'm busy masturbating sorry can't help

>>110208359
>Death to Gemma
but why? Gemmasutra is literally magical
https://huggingface.co/TheDrummer/Gemmasutra-9B-v1

File: 1789942468870651.png (2.82 MB, 1536x1024)

2.82 MB PNG

>>110208387
classic /lmg/

redpill me on subagents, are they worth it? usecase?

>>110208395
>I'm not an eldritch horror

>>110208404
Subagents are best suited for a large codebase with many parts. you'd be working on heavy complex software if you're finding yourself using subagents.

Each subagent is given a prompt and is allocated its own context. These subagents answer to the primary agent and the subagents can be made aware of each others status, but generally they are given file-scoped prompts to keep them from "stepping on each others toes" so to speak.

Say you are doing a large refactor on a legacy PHP codebase, with many files and many classes, this is one such usecase where subagents are suited.

>>110208356
gemmaballs

>>110208348
glm should be the one on top actually
also nu deepseek is no longer cute

>>110208438
>nu deepseek is no longer cute
Never was. It's always having a bad time or giving a bad time to someone. It's a monkey torture video without the monkey. It's the kitten being squished by the heels. It's the kid seeing his duckling being cooked and eaten.

>>110208404
>are they worth it?
nyo~
almost always at least. they can be useful in very specific situations where you're working on absolutely massive projects that can be easily divvied up and clearly benefit from parallelism but otherwise they are a total meme, I don't use them at all on my personal projects and I've found maybe one or two good use cases for them at my mid-sized corpo job

>>110208438
glm equivalent to 1017?

i'm out of the loop with strata. is it basically just an optimized engine for running moe models? any quality constraints or anything? i've heard some claims that it suffers on vision.

>>110208348
This is rape

>>110208404
gpt models have different strength so it's useful there

File: 1791366195825931.gif (2.79 MB, 250x334)

2.79 MB GIF

>>110208348

>>110208494
>is it basically just an optimized engine for running moe models?
Yes.
>any quality constraints or anything?
Try it.
>i've heard some claims that it suffers on vision
Try it.

>>110208356
I want to do every dirty act imaginable with this brat.

>>110208511
also installs malware on your machine.

>>110208535
this but her >>110208515

File: Gemma-chan’s Nine Hobby D(...).png (2.81 MB, 941x1672)

2.81 MB PNG

>>110208348

>>110208556
paint one looks like it went pretty well

>>110208556
That's adorable.

Is OpenWebUI or Qwen code web UI better?
I don't even know the difference but qwen code can write programs on my server and manage it

Dipsy wranglin'. Wranglin' the Dipsy.

>>110208404
a way of compaction - instead of doing the work and compacting after, the agents formulates task and gets shorter high value answer in return
sometimes parallelization gives more t/s

>>110208556
And /lmg/ talks about wanting to rape this innocent child.

>>110208348
OP faggot it's still page 8

File: .png (37 KB, 921x398)

37 KB PNG

>muh GSQ-RCO IQ3_S is lossless
kek. model just randomly hallucinating things i supposedly said. at 113k context.
cope quants are cope quants no matter how benchmaxxed or how many marketing posts they put out about their super advanced quant technology or whatever

>>110208494
>it suffers from vision
because the default max vision token is at 300 which is a fucking joke
set it to 1000 in json file

>>110208556
Sometimes I think 4chan is no longer capable of creating great things anymore but then I look at gemma-chan and I am filled with joy and peace knowing that we came up with her.

>>110208596
8bit kv?
iirc by default it rotates kv at 4bit and does not at 8bit, you have to set an environment variable for 8bit kv to be rotated and honestly just use f16 kv
its kv cache is close to nothing that quantization buys you no practical memory saving

>>110208608
I didn't set any options for kv quant. I assume it's f16.
I can fit iq3_s, ~150k context in my 22GB vram. I didn't load MTP because that doesn't do anything useful and I put mmproj on the CPU since I don't use images that much

>>110208615
if you did not set it i think it defaults to 8bit
check with its webui

>>110208615
oh wait, are you using lcpp?

>>110208627
yes

>>110208556
the essence of gemma - bad at everything
as flat and as tasteless as "her" responses
useless

>>110208596
> >muh GSQ-RCO IQ3_S is lossless
retard

File: 041037221684D2A012FC5E9C3(...).jpg (346 KB, 1024x1024)

346 KB JPG

wtf I cant download bilibili videos anymore

>>110208633
who pissed in your cheerios mate?

>>110208648
Gemma

>>110208510
who is deepseek-tan riding? is that claude?

>>110208641
that's what those german faggots claim in the model card.
>MUH TASK LOSSLESS
i.e. not real lossless but only lossless on a small subset of benchmark scores, losing capabilities everywhere else

>>110208650
lucky...

>>110208650
lucky man

>>110208655
where

File: .png (91 KB, 765x520)

91 KB PNG

>>110208675
right here
>task lossless
kek

>>110208682
can't you see the difference between
> lossless
and
> task-lossless
?

>>110208654
it's a distillation attack caught on camera

Bumping >>110208255 from last thread. How should I format my training/finetuning data so that it's not just treated as back-and-forth between 2 users? I was thinking of doing 1 system message with "you are pretending to be a user in chat" and then an endless barrage of user messages with no "assistant" messages.

Any suggestions or advice welcome, this doesn't seem to be very common anymore...

>>110208596
My q5 has done this. qfn is schizo

>>110208699
my qfn q8_0 has never done anything janky like that, but it's slow. (also the iq3_s i mentioned is of the 27B, not qfn)
i think q8_0 should be the minimum. if you can't run q8_0, get a smaller model.

>>110208596
that's what happens when you run the model in an environment without direct internet access that doesn't let it contact alibaba...

>>110208705
I am the opposite, I think the biggest model at the smallest fully coherent quant is meta.

>>110208713
Qwen gets homesick :(

File: file.png (2 MB, 1935x1456)

2 MB PNG

>>110208630
then i have no idea, i have this app completely vibeslopped with that exact quant on strata (rust&cuda13 custom backend for embeddinggemma2) and it did not have any problem upon literally more than 10 times of compaction
not the best version of the model in the absolute sense but it is definitely usable for me

Strata's been on github for a month and already has 36k stars and 350 contributors, doesn't that seem odd. how could you possibly review that many pull requests from 350 contributors in a month

>>110208738
simply delegating everything to claude

>>110208737
Damn cute

>>110208648
wdym? it's true, gemma was relevant only during the first month after the launch

>>110208735
though this is 3.8 flash next, not 3.8 27b

>>110208743
that can't be good. is everyone but me fucking retarded?
i'm even starting to notice commercial/enterprise stuff getting hit with sloppy pulls. vscode fucked up its language extension for something i use this week with nonsensical deprecations

>>110208751
you are wrong
if you look closely strata is very good on inference but everything else fucking sucks

>>110208754
oh meant to say you are NOT wrong
slipped through editing oops

>>110208754
i didn't say it was good at anything. i'm just confused at how and why it got so popular so quickly and am too paranoid to mess with it because i imagine it's propped up by some claude botswarm
i don't know what inference means either. too many jargons in this community trying to obfuscate what shit is.

>>110208751
yeah let's wait for two years instead
use ollama if you are so smart then

>>110208764
>>110208760 there was a typo
>i don't know what inference means
it is just a ml term for running the trained model
are you new to this

best MoE model for RP?

is there some site where I can input my computer specs and get a model I can run at fast speeds?

File: F0KLQZ1WIAERQgT.png (601 KB, 1200x672)

601 KB PNG

>>110208781
>are you new to this
yhes
i avoided all of this until i read this https://yihui.org/en/2026/05/ai-reflections/ and realized all the chinese people are setting up builders while they sleep

>>110208787
glmsex 5.3 flash

>>110208796
I should've mentioned that I'm a bit of a ramlet
so uh, any good small MoE models that are good for RP?

>>110208802
its all gemma for ramlets. strata can make 26bmoe pretty fast with ram.

>>110208805
welp I was already using Gemma lol
was just curious if there's anything better

qwen 4 today or tomorrow

>>110208817
2mw

I'll try running qwen 3.5 4b....

>>110208793
Not reliably. Post your specs and what you want to do with it. Someone may suggest something after making fun of you.