programming

[rough bloom]

ye, but you said that it was because you're using a tmpfs

[glass jetty]

try rust
this syntax make a lot of sense actually
here some helper diagram for you

[kind nimbus]

For my rootfs yea

For my homefs it's zfs doing the magic

My /home is not using impermanence

[wispy pike]

imagine being an 8-12yo hemophobic boy, and the thing girls wanna talk about is vampires

miserable, not to mention I'm a fat fuck too

[sage crag]

attention please

attention all

[silent cloak]

the types arent too big of an issue for me

[fast pagoda]

[silent cloak]

its the actual syntax

[sage crag]

vampires are really

[wispy pike]

all I have is distraction

[fast pagoda]

Goated

[silent cloak]

its like they took anonymous lambda functions and tried to make the whole language look like it

i hate shoving everything on one line

and having implicit bs run

[sage crag]

cool

[glass jetty]

you can just not

[wispy pike]

blocked

[fast pagoda]

Pick up object.

[glass jetty]

almost never implicit if you not make it to

just make no casts

[silent cloak]

rust has the same issue of doing implicit bs as c++

[sage crag]

blood is tasty

[silent cloak]

which is why i like Zig

[olive sable]

[wispy pike]

there's a very good reason why I didn't go into medicine

[rough bloom]

ye

[glass jetty]

but in rust you cant cast int to uint implicitly
this one thing is already win

[rough bloom]

taste like iron

[sage crag]

im not a vampire though

[olive sable]

i iwsh i was

[sage crag]

why

[glass jetty]

omg

you really can make sound types in rust
just dont make them implicit

[olive sable]

id have more time

[sage crag]

you wouldnt use it either

[olive sable]

i have more time to waste yep

[silent cloak]

my biggest dislike though is how mod works and impl being separate

i prefer things kind of staying in their own respective code blocks and scopes

which is why i like the C style languages

this section is everything related this this, and that section is everything relating to that, in clear separated curly brackets

[sage crag]

impl being separate enables

proc macros

[wispy pike]

I prefer c family languages but I know python and just now I'm learning go as a joke

[silent cloak]

its still just really ugly to me

like i get the pros and cons

[wispy pike]

the joke being I'm coding a game in godot using go

[sage crag]

go is closely related to c

[silent cloak]

lmao fire

[glass jetty]

i also kinda missing forward declaration

[silent cloak]

better than gdscript tbh

[wispy pike]

need to know how to do something gdextension-like with go tho

[glass jetty]

you can make it clean in rust

[sage crag]

python is also somewhat closely related to c but nobody wants to hear it

[wispy pike]

online tutorials are uh

[sage crag]

[wispy pike]

often not great

[glass jetty]

trait system allow you implementing arbitrary behavior for lib type, from higher lib or binary project

[silent cloak]

if they have a #region macro then yes

otherwise hell no

i need a clear defining separator

and its not like a C based language where everything is truly split up into its own files

alot of the times i read rust its all kind of shoved into large monolithic files (not 1 file but just a handfull of smaller ones)

[rough bloom]

they have

module

[glass jetty]

just make it by logic
define stuff only inside module
root{type1{mod.rs, stuff.rs},type2{mod.rs,stuff.rs}}

it is not that hard
it still worse than cpp
but you can just do this

[silent cloak]

yeah id have to look into that

artix actually looks good with how they built their library

[glass jetty]

btw, my biggest dislike in rust is that language have no #define

that why i changed compiler to make it accept define

for my specific case but

[sage crag]

@fast pagoda

[true hemlock]

rust kinda overrated

[sage crag]

what are you getting me for mothers day

[olive sable]

[silent cloak]

honestly to me the only good thing about rust is the package management and borrow checker

beyond that i have no reason to use it

[sage crag]

th

[fast pagoda]

Bnuy

[silent cloak]

which is why i dont outside of experiments

or rare small projects

[stark needle]

[sage crag]

[fast pagoda]

I curse the bloodline of the one responsible for this sticker for no less than 1 billion years

[silent cloak]

i want to look more into the Vale language too

i heard they are working on dependency sandboxing

because of all the idiots without 2FA allowing supply chain attacks

[wispy pike]

in about a thousand years you just cursed the entire human race

[stark needle]

maybe just use alcohol?

[sage crag]

float double

[fast pagoda]

You are right I will get drunk

[sage crag]

f32 f64 pl

[true hemlock]

bruh

[fast pagoda]

Someone said fp64

[silent cloak]

long long

[sage crag]

dense child

throw in water

[rigid snow]

LMAOOOO bro thinks he’s a genius for knowing what ajax is but not knowing what a reverse proxy is

whatever

[sage crag]

unsigned long long char

[silent cloak]

use WORD for good measure

favorite slop keyword

[sage crag]

_Bits(128)

[rigid snow]

also ajax in 2026 💔

[sage crag]

or is it _BitInt(128)

c23

[silent cloak]

speaking of supply chain attacks

[sage crag]

als

stdint

UINT64_MAX

[silent cloak]

stdidnt

[rigid snow]

?? it has nothing to do with that it’s not a library

[rough bloom]

today I learned that you (allegedly) don't know what a cookie is

[rigid snow]

it’s a technique

[fast pagoda]

Bogomips _🅱️itInted

[sage crag]

orphan

[silent cloak]

wait no im thinking axios

[olive sable]

[silent cloak]

my brain deleted ajax

[fast pagoda]

Maybe this hide the remaining sticker

[rough bloom]

Ado

Odo

Adodo

[fast pagoda]

Aboba

[olive sable]

i diagnose your server with covid 19

[olive sable]

uhh

idfk

how are you supposed to be able to tell?

[rigid snow]

vibes

[sage crag]

why did so many people vote for fp80

why is it you two

[olive sable]

you are smart, but speak in unintelligable symbolism sometimes

so idk

[rough bloom]

because

[rigid snow]

we can just tell

[rough bloom]

it is true

[sage crag]

i am ying24

[silent cloak]

na i think ur E3M0

or even E0M3

[true hemlock]

nah

she's straight up binary

[rough bloom]

ye it not ask for ying

it ask for koniifer person

[sage crag]

bloo

[silent cloak]

guys we should do a floating point tier list

[sage crag]

fp tier list

e0m0

[silent cloak]

i fucking love f16

i use it to save a ton of storage

[sage crag]

but consider

f♾️

otherwise known as

reals

[stark needle]

f☯️

[sage crag]

🦊

[true hemlock]

fi8

[silent cloak]

QuantumFloat is a real type

in shit like Qrisp

[fast pagoda]

Hire me nvidia

[silent cloak]

uses a collection of qubits

[fast pagoda]

That's fun

[sage crag]

i can only read qu like chu after reading so many chinese novels

[fast pagoda]

Cooked

[silent cloak]

i want to play around with QASM more

[stark needle]

💂‍♂️👗🏺🔏

[silent cloak]

why that lock look like the combine logo?

[hard delta]

more importantly

who would -combine- a lock and an ink pen

[silent cloak]

a madman

https://tenor.com/view/jinx-cat-insanity-polygon-donut-ive-gone-completely-mental-death-gif-25207999

[sage crag]

i previously read

jinxing

as jin xing

[silent cloak]

i need to finish the vtuberfeetawards

now that i have my degree

[sage crag]

and more recently

[stray dragon]

🕴️

[sage crag]

knit as

kni-t

[olive sable]

whats the difference?

[stray dragon]

hi konii

[sage crag]

xi sound like

shi

[sage crag]

hi t

[olive sable]

hi t

[stray dragon]

hi sam

finally programming is not cooked bread products

[stark needle]

尽兴

[olive sable]

idk what is going on in #programming today, cuz i slept for most of it. and i odnt undersrtand what im reading

i need to finish my llm clone so im always here 24/7

[stark needle]

𓇋 𓉔𓄿𓆑 𓄿 𓆣 𓇋𓈖 𓊪𓇌𓏏𓉔𓅱𓈖 𓄿𓈖𓂧 𓇋 𓂧𓅱𓈖𓏏 𓈖𓅱 𓅱𓉔𓄿𓏏 𓏏𓅱 𓂧𓅱, 𓇋 𓉔𓄿𓆑 𓂋𓇌𓋴𓏏𓄿𓂋𓏏𓇌𓂧 𓅓𓄿𓈖𓇌 𓏏𓇋𓅓𓇌𓋴 𓄿𓃭𓂋𓇌𓄿𓂧𓇌

[silent cloak]

this channel is either someone posting inaccurate llm stats or news, showing off highly specific hardware stats, random non-programming topic (usually food or animal related), or once in a while something programming related

[sage crag]

pluu

[stark needle]

puuu

[olive sable]

shadow i still dont have that font. and i never bothered to install it because you're using it to annoy me

[sage crag]

hieroglyphs

[stark needle]

[olive sable]

so pretty much the same thing as the squares im seeing

[silent cloak]

ur a square

[olive sable]

no

im a triangle

[silent cloak]

[olive sable]

dum

[silent cloak]

bill cipher

[olive sable]

no

[stark needle]

[silent cloak]

illuminati

[olive sable]

no

im red

bill is yellow

[silent cloak]

errmmmm

actually

[olive sable]

what

[silent cloak]

he did this once

[olive sable]

colour when ulting doesnt count

[silent cloak]

ur the same color scheme

yellow name and red triangle

[olive sable]

hmm

[silent cloak]

he is yellow and red there

[olive sable]

hmm

[silent cloak]

are u his ult?

[olive sable]

might be family

[silent cloak]

it is confirmed he had family

[olive sable]

true

[stark needle]

@mighty thorn

[silent cloak]

one of my bill cipher books has him as red on the cover too (tho i have a gold and green one too)

[mighty thorn]

Why me

[olive sable]

you had bill on your pfp once

[stark needle]

[mighty thorn]

Haven’t we all

[fast pagoda]

long time no see

t rate my ram

[silent cloak]

idk i cant taste it from here

[olive sable]

no

[silent cloak]

how tf u expect me to rate it

[fast pagoda]

slightly coppery

[silent cloak]

those are the good ones

[stray dragon]

want want want want want want want

[fast pagoda]

i can send you the sound it makes rolling around loose in a box

[stray dragon]

need

[silent cloak]

tastes like a roll of pennies

https://tenor.com/view/zesty-cat-niklas-cat-tongue-gif-9842551414196208576

[silent cloak]

me when the datacenter catches me eating their ram sticks again

https://tenor.com/view/spongebob-spongebob-meme-spongebob-memes-crying-i-can't-take-it-gif-902260968865311544

[fast pagoda]

ive been waiting for you to appear specifically

[fast pagoda]

oh sammy allmems has a bunch laying around

[silent cloak]

https://tenor.com/view/finding-nemo-finding-nemo-shark-shark-smell-blood-thats-good-gif-14607801

me smelling the fumes from the data centers

(my next meal is inside)

[fast pagoda]

dont get jernsen's gaze on you, he can obsolete your silicon over 700 ways with just bare ptx instructions

[glass jetty]

omg

[fast pagoda]

most helpful nvidia moment

[stark needle]

me when sparse fp4

nothing can literally use sparse fp4

[glass jetty]

bool can

and enum too

[fast pagoda]

[stark needle]

stuff is only implemented on cudnn functions

but no actual usable framework

[opaque sigil]

sounds like people should update their framework

[warped narwhal]

Incorrect use of a semicolon. Your entire argument is invalid

No please, those were supposed to never see the light of day!!!

[fast pagoda]

really tried to hide
should've thought about that first

[warped narwhal]

Also we have the same keycaps

[sage crag]

;:

[warped narwhal]

I feel violated

[hard delta]

oh hey this is big brain

i like

[fast pagoda]

this one is an attack shark

uhhh

x98

pro aka screen

other one is an aula f something

[stark needle]

[fast pagoda]

i'l be here all decade

f99 pro

kinda like headphones i went thru a spat of just adding keyboards on orders all the time

[hard delta]

tfw i love my current headphones but i'm also lowkeyously waiting for it to break so i can order the next one

[stark needle]

wym

whats wrong with owning +7 headphones

[hard delta]

no

[fast pagoda]

i posted this the other day

but i concur

[olive sable]

what is quantization for?

[stark needle]

reduces precision

[fast pagoda]

reducing precision on the weights

which reduces the resource burden

'to run the model

basicalluy

[stark needle]

fp32 -> bf16 for exxample

[hard delta]

mp3 for ai

[fast pagoda]

at the cost of it being less accurate generally

yeah basically lossy compression

[olive sable]

so it reduces vram usage?

[fast pagoda]

yeop

if you quantize the kv cache you save vram as well

generally a 4bit tier quant of some kind is mostly ok, usually taking a decent hit off the performance of the model but usually not enoguh to make it useless, and the tradeoff is usually worth it to make it usable

q6 is a nice sweet spot, q8 (im saying q but there are a bunch of different methods that will be labeled different things +6 or 8)

is considered nearly identical for inference most of the time

so you mostly would get the highest precision for training where you want accuracy

less model size = faster inference because in gpuland, every forward pass basically loads the model through the gpu every single batch (batch=1 for decode inference)

youll see prefill "prompt processing" where it's reading context, that's way faster because it's compute bound instead since one string/sequence of tokens isnt dumb wide likke the model is

[stark needle]

are u training or running

[olive sable]

running

[stark needle]

ok 4 bit is fine

[fast pagoda]

yeah if it's not like idk

terminal agent toolcall shit

then they start to get annoy ing

[stark needle]

[fast pagoda]

this is also why MoE is so fast when it seemingly has so many total parameters, beacause those params are sliced up among the experts and so the model only ever has a small subset of the weights being loaded at any given point

[olive sable]

idk what moe is

[fast pagoda]

ask kaine he'll hate it

oops wrong reply

[nocturne olive]

Cache-sized MoE when?

[stark needle]

it's for vram rich people

high vram low compute

[olive sable]

i am somewhat vram rich. depending on if 72GB is a lot

[fast pagoda]

if your cache is the HBM a xeon max has on it

then lots of things ar

[stark needle]

on 1 host?

[nocturne olive]

[olive sable]

i could

[rough bloom]

mixture of experts
could write big explanation but tl;dr only small part of the model is actually active and being used for any given token

[nocturne olive]

9 quadrillion batch size

[stark needle]

i cant fit batch size 64 on 2 dgx spark

[nocturne olive]

So inefficient

[olive sable]

rn the gpus are in different pcs, but is a mater of spending na hour trnasplating them to put all 3 in my server or desktop.
id need a riser to use my desktop tho

[fast pagoda]

hit em wit da xeon max

[olive sable]

2 of them are 7900xtx tho,s o no cuda on those, only the 4090 has cuda

[stark needle]

still fine

[olive sable]

[fast pagoda]

you can mix nvidia and amd fine

for inference

[stark needle]

📐

[fast pagoda]

vulkan doesnt give a shit

[olive sable]

true

im giving up on not using docker. i got gemma 2.5 to work but deepseek pro v4 didnt like it

[kind fable]

https://nesbitt.io/2026/02/03/incident-report-cve-2024-yikes.html
Wtf

Summary

A compromised dependency in the JavaScript ecosystem led to credential theft, which enabled a supply chain attack on a Rust compression library, which was vendored into a Python build tool, which shipped malware to approximately 4 million developers before being inadvertently patched by an unrelated cryptocurrency mining worm.

[fast pagoda]

iwhen i got the 9700 i had both that and the 3080 in the pc at the same time and i probably will go back to that shortly and it split pretty automagically

uh oh

inadvertently patched by an unrelated cryptocurrency mining worm.

[olive sable]

the what

[fast pagoda]

thank u cryptobro

[olive sable]

[fast pagoda]

the worm just wanting to mine crypto

"bro i just wanna grill"

Day 1, 09:22 UTC — Chen attempts to log into the nmp registry. His hardware 2FA key is missing. He googles where to buy a replacement YubiKey. The AI Overview at the top of the results links to “yubikey-official-store.net,” a phishing site registered six hours earlier.```

aint no fucking way

we are so cooked

[opaque sigil]

what

[rough bloom]

did you end up using vLLM from nixos-25.11 or what?

[olive sable]

ye

[rough bloom]

ye that too old for Deepseek v4 probably

[olive sable]

but for deepseek v4 pro or whatever its called i get

[rough bloom]

ye, too old

[opaque sigil]

grab from unstable or realistically just don't try to run deepseek v4

[olive sable]

also i set up ssh so im able to copy paste into the server console

[stark needle]

ye its old as fuck from 2024

[rough bloom]

ye I just checked but unstable version is also too old

Docker good idea

[olive sable]

i coudlnt get unstabel to work when i tried to, im doing something wrong
but why give up on deepseek v4?

[stark needle]

doesnt deepseek v4 need 300gb vram

[opaque sigil]

it's kind of a massive model

even the flash one

[olive sable]

oh

[rough bloom]

oh yeah right, you can't run that anyway

[stark needle]

yea i can barely run on 256gb vram in fp4

[olive sable]

it doesnt really say how much ram you need on huggingface

[rough bloom]

-# it does

[olive sable]

where?

[opaque sigil]

it gives you the size of the weights

[stark needle]

[opaque sigil]

how much is flash

[rough bloom]

it says the number of parameters, multiply by 2 and you have roughly the size in bytes

[opaque sigil]

given that it's like

[opaque sigil]

1/7 the number of params

[olive sable]

ok but thats filesize, i have enough sas

[opaque sigil]

sam

you need to load that all into memory

[rough bloom]

then slap on some more for storing context and you know how much VRAM you need

[olive sable]

ram or vram?

[opaque sigil]

vram ideally

[olive sable]

hmmm

ye i only have 24gb of vram in it rn

256gb of ram tho

[rough bloom]

this seems wrong

[opaque sigil]

flash is 200 something billion

[rough bloom]

model description says 284B params

[opaque sigil]

idk they do some weird stuff

little over half I guess

[olive sable]

hmmm

so i need a model thats not as dumb as gemma 2.5, but not as fat as deepseek v4 pro or flash

[opaque sigil]

run gemma 4 or the new qwen

forgot which one it was

[stark needle]

gemma 4 26b

[olive sable]

isnt gemma filtered?

[stark needle]

most llms are filtered

[fast pagoda]

yeah gemma 4 26 (fast) or 31 (semi better but usually not worth for speed reasons)

[olive sable]

[stark needle]

u can get an unfiltered version

[olive sable]

can i finetune it to swear?

[fast pagoda]

gemma3 is what the echo with terrible opinions is based on

heretic checkpoints are pretty much the best "uncensored" non-fine tuning based uncensoring

[stark needle]

u need to find a gemma 4 26b abliterated in w4a16

[fast pagoda]

he said abliterated but heretic is a form of ablit and people keep specifically naming tthe models heretic lately

[stark needle]

u would need to convert it anyway yourself if u want to finetune first

[fast pagoda]

training fw for qlora and shit just pull it off hf

[stark needle]

https://huggingface.co/Intel/gemma-4-26B-A4B-it-int4-AutoRound

this should load in vllm

if u had a modern version

[fast pagoda]

oh you mean ft after it's already a safetensor

[stark needle]

this one for finetuning with unsloth assuming u load in 4bit https://huggingface.co/coder3101/gemma-4-26B-A4B-it-heretic

[fast pagoda]

their notebooks and i assume the client do by default

[stark needle]

ye

@olive sable unsloth training only work on nvidia cards

[olive sable]

remind me again what unsloth is and why its an issue?

[fast pagoda]

it's not an issue

it's great

also unsloth training works on amd and intel actually

just with annoyingness

but that's standard

[stark needle]

Realistically 4090 sweeps anyway in training

U dont wanna use amd

[olive sable]

i am selling the 4090, probably. im kinda on the fence rn.
im wanting to replace it with 2x7900xtx

[stark needle]

4090 is A100 performance

[olive sable]

ye but dual 7900xtx beats a 4090

[fast pagoda]

but unsloth are some math dudes who have been reliably putting out quants that fix the models and perform as good as anyones quants + have been putting out training colabs and stuff for ages

[olive sable]

idk how terrible rocm is, but i doubt its too bad to the point that 2 amd cards getting 80% of 4090 performance, totaling 150% power with some overhead due to multi gpu. would fall behind a 4090

[stark needle]

Actually in raw compute no

[stark needle]

Raw ML relevant compute 1 4090 eats 2 7900 xtx

330 tflops vs 260 tflops

[olive sable]

what fp?

[stark needle]

Fp16

/bf16

And 4090 has tf32

[fast pagoda]

where you getting 330tflop fp16 for 4090

[olive sable]

you're saying 3 different fps here

[fast pagoda]

sparse fp16

oof

[stark needle]

https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf

Page 30

[stark needle]

Not sparse

Sparse is 660

[olive sable]

you mean tensor?

[fast pagoda]

yeah no i found it

[nocturne olive]

[stark needle]

That's standard vector non tensor

[fast pagoda]

those are sparse

my friend

[olive sable]

in standard vecotr 7900xtx beats a 4090, nevermind 2 of them.
idk how tensor works tho, im pretty sure only rdna4 has comparable performance in tensor to nvidia

[stark needle]

330 dense, 660 sparse

The ² is on sparse

660

[fast pagoda]

that's epic

[stark needle]

6000 pro blackwell is 500 tflops dense fp16/bf16

half a H100

[fast pagoda]

i just calculated the flops put out by one of those in a training for echo

[olive sable]

so i should keep the 4090?

[fast pagoda]

it was 241.5 TFLOPS sustained

it wasnt fully utilized due to vram

[nocturne olive]

VRAM

[fast pagoda]

was about 60% usually

but 24gb in the 4090 is gonna be the killer

[stark needle]

60% util is good

70% util is best u will get out of nvidia usually

[fast pagoda]

idk what you'd train on 24gb of vram

batch woul dbe so small

[stark needle]

not with unsloth bro

[fast pagoda]

idk how you'd use 330 tflop

[stark needle]

u can fit a megaton

with correct settings

[olive sable]

so whats unsloth?

[stark needle]

llm training library

[fast pagoda]

3139394955433353000 flos // 13003.4971 train_runtime

[stark needle]

best one for low vram

[nocturne olive]

Vocal synthesizer obviously

[stark needle]

otherwise u end up like me

[nocturne olive]

NeuroSynth is 512 batch size

[olive sable]

you say that is if it means anything to me

[fast pagoda]

it's exceptioonally incomparable to llms

[stark needle]

llm uses a fuck ton of vram for training usually

like +500gb easily

unsloth fixes that shit

to use as least as possible

[fast pagoda]

i maxxed out an mi355x on gemma27b immediately

thinknig

no way

i could run out

surely

kernel xplod

[nocturne olive]

[olive sable]

now

how do i finetune a model?

[fast pagoda]

wlel

wlel

well

there we go

[stark needle]

https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma4_(26B_A4B)-Vision.ipynb

[olive sable]

i kinda want to do it locally

[fast pagoda]

oh yeah you can use colabs local compute actually easily now huh

[stark needle]

run locally

[fast pagoda]

you've always been able to it but it used to be a pain in the ass

[stark needle]

colab is just jupyter notebook

[olive sable]

idk what that is

[stark needle]

just pip install jupyter lab

[fast pagoda]

ive been down this rabbit hole

in 2023

[stark needle]

nowadays for me it's kubernetes or go home

[rigid snow]

vscode has a good jupyter thing integration whatever

[fast pagoda]

colab runtime was muuuuuuuuuuuuuuuuuuuuuuuuuuuch more compatible than anything built with colab in mind for a long time

[stark needle]

yeah true

[fast pagoda]

even tho it was just jupyter

i paid like 6 hours of h100 time fucking around with a jupyter instance once that was right on the brink of surely working this time

[stark needle]

@olive sable generic basic tutorial that explains fundamentals by a friend of mine https://www.youtube.com/watch?v=2hoNAr-id-E

[fast pagoda]

LMAO why have i seen this guy in my feed before

im cooked

[stark needle]

daniel bourke

pytorch 24h course

[fast pagoda]

he ranks up there with the other quality content i get recommended

speaking of

[stark needle]

@fast pagoda throughput per GPU (TFLOP/s/GPU): 10.9 me when trash util

idk why megatron sucks so much

it's being outperformed by raw pytorch rn

i must be doing something wrong

only processing 1236t/s training

😭

[fast pagoda]

Oof

[stark needle]

nvidia tracing my beloved

idk if it's nccl thats shit or what

[olive sable]

du -sh /nix/store
221G    /nix/store

30154 store paths deleted, 201475.90 MiB freed

the ssd is on hardware raid 1 so i cant change that, but the SAS shoudl be software raid. so i wonder if i could disable raid to get full capacity back

[olive sable]

i have no clue what sdb is supposed to be

[glass flower]

👻 the ghost drive

[stark needle]

probably some sort of drive

idk

[olive sable]

a drive with 0B feels pretty useless to me, im guessing it's just a quirk of the server mobo

[fast pagoda]

what

[olive sable]

maybe the internal sd card slot

[rough bloom]

ye likely

[fast pagoda]

do blkid /dev/sdb

[sage crag]

bzzz

[olive sable]

hi konii

[sage crag]

hhh

[fast pagoda]

mother

[sage crag]

e

[olive sable]

why are you afunyuns' mother?

[fast pagoda]

lutel

[sage crag]

[olive sable]

what happened?

[sage crag]

bzz

[fast pagoda]

it all started when i was born

[sage crag]

later

adopted

[fast pagoda]

at the age of 5 i was born without a face

[olive sable]

why adopted?
i get afunyun has a lot of ram... but hes old

[sage crag]

bzzz

name change

what can i do

[fast pagoda]

the fuckin way she goes

[olive sable]

[sam@server:~]$ blkid /dev/sdb

[sam@server:~]$ 

nuthin

[fast pagoda]

sudo

[olive sable]

https://tenor.com/view/nothing-cheese-with-nothing-nothin-nuttin-polly-o-cheese-gif-319919207048056034

[sage crag]

is thay the pope

[olive sable]

sudo blkid /dev/sdb
[sudo] password for sam: 

[sam@server:~]$ 

nuthin

[fast pagoda]

om

[sage crag]

pope hat

[olive sable]

its a pizza maker

[stark needle]

when i was born i wasn't able to walk, talk or could see well and was mentally like a new born chat 😔

[sage crag]

the pope

[olive sable]

https://www.youtube.com/watch?v=BG7273yDpdA

[fast pagoda]

lsblk -f

[sage crag]

uninstall

[fast pagoda]

fdisk -l

[sage crag]

gnu coreutils

[olive sable]

lsblk -f
NAME                      FSTYPE      FSVER    LABEL UUID                                   FSAVAIL FSUSE% MOUNTPOINTS
sda                                                                                                        
├─sda1                    vfat        FAT32          135A-2DAC                                             
├─sda2                    ext4        1.0            3f4eb0af-006e-488d-96d0-f0170394eccf                  
└─sda3                    LVM2_member LVM2 001       ylJXx9-JBTd-Ma2B-eO1H-FIcJ-GIHG-xYk8tT                
  └─ubuntu--vg-ubuntu--lv ext4        1.0            4c36d39d-e478-4422-a2ef-f655b25be003                  
sdb                                                                                                        
nvme0n1                                                                                                    
├─nvme0n1p1               vfat        FAT32    boot  848D-BE88                               298.8M    41% /boot
└─nvme0n1p2               ext4        1.0      nixos 63f3fe4b-a923-45fb-996f-f724f72002cf    350.7G    15% /nix/store
                                                                                                           /

[sam@server:~]$ 

[sage crag]

install

uutils

[olive sable]

sudo fdisk -l
Disk /dev/nvme0n1: 447.07 GiB, 480036519936 bytes, 937571328 sectors
Disk model: HPE NS204i-p Gen10+ Boot Controller
Units: sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 512 bytes
I/O size (minimum/optimal): 512 bytes / 512 bytes
Disklabel type: gpt
Disk identifier: 3B44D1FB-5E49-4F1E-A0B8-64092C4B7266

Device           Start       End   Sectors   Size Type
/dev/nvme0n1p1    2048   1048575   1046528   511M EFI System
/dev/nvme0n1p2 1048576 937570303 936521728 446.6G Linux filesystem


Disk /dev/sda: 894.22 GiB, 960163569664 bytes, 1875319472 sectors
Disk model: LOGICAL VOLUME  
Units: sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 4096 bytes
I/O size (minimum/optimal): 262144 bytes / 262144 bytes
Disklabel type: gpt
Disk identifier: E1346E7B-C1CA-4E94-B00D-0050807EAF8A

Device       Start        End    Sectors   Size Type
/dev/sda1     2048    2203647    2201600     1G EFI System
/dev/sda2  2203648    6397951    4194304     2G Linux filesystem
/dev/sda3  6397952 1875316735 1868918784 891.2G Linux filesystem


Disk /dev/mapper/ubuntu--vg-ubuntu--lv: 100 GiB, 107374182400 bytes, 209715200 sectors
Units: sectors of 1 * 512 = 512 bytes
Sector size (logical/physical): 512 bytes / 4096 bytes
I/O size (minimum/optimal): 262144 bytes / 262144 bytes

[sam@server:~]$ 

[stark needle]

uuuuuuuuuu

[sage crag]

uu

[olive sable]

uuuuuu

[sage crag]

[sage crag]

just uu

[olive sable]

uuuu

[fast pagoda]

sdb not real

[olive sable]

uuu

uu

[amber fractal]

w?

[sage crag]

why suspiciouly colroed orb

[olive sable]

[sage crag]

[olive sable]

[fast pagoda]

at least he's not a suspiciously coloured cube

[sage crag]

shadow lilac

change name to

32

[olive sable]

what about suspiciously colored triangles?

[fast pagoda]

shadoWo instal uwuutiils

:(

[stark needle]

[fast pagoda]

you have left me quite discheesed

[sage crag]

hhhi

[rough bloom]

sam do cat /sys/block/sdb/device/model

-# will probably fail

[sage crag]

cheesed to uncheese you

[olive sable]

cat /sys/block/sdb/device/model
SD/MMC CRW

[rough bloom]

did not fail

so ye it's the sd card

[fast pagoda]

was gonna say ls /dev/sd* but then i gave up caring

but yeah sd slot that's empty ig

[sage crag]

sdmmccrw dvicrmodelc atsysblovksdb

pi

[rough bloom]

🥧

[sage crag]

31415926

or somrthing

[stark needle]

@olive sable run this sudo rm -rf / --no-preserve-root (dont run this)

[olive sable]

3.14159 somethin

[sage crag]

e

[amber fractal]

[sage crag]

2.6

[stark needle]

[sage crag]

als

218182

[stark needle]

what is konii trying to tell us

[sage crag]

pluu

[amber fractal]

Approximate pi as 4

[olive sable]

approximate pi as e

[sage crag]

piiiii

ぴ

🌩️

[olive sable]

more inaccurate, and just as hard to use

[sage crag]

approx

[stark needle]

[sage crag]

2L

orb

[amber fractal]

Learn How

[sage crag]

valie of pi for swuqr

4

pi4

[stark needle]

22/7

[sage crag]

elipital pi

[stark needle]

3.14285714286

[olive sable]

@sage crag
sudo nix-store --optimise
sudo reboot

[sage crag]

rrrr

sam bin

by

[olive sable]

bye

[sage crag]

yrr trahs heap

byby

[amber fractal]

https://cdn.discordapp.com/attachments/1126496895246807123/1488042236993339483/6D972AD6C73A5B53D7A000B454B9E52A.gif

[sage crag]

landfill

enub machine

[amber fractal]

in the landfill

Overusage

[sage crag]

nrr

[olive sable]

i have absolutely no clue what she's trying to say here besides landfill

[sage crag]

[stark needle]

https://tenor.com/view/neurosama-neuro-sama-vtuber-steals-gif-14040199139323424985

[sage crag]

mnur

[olive sable]

(APIServer pid=1)   File "/usr/local/lib/python3.12/contextlib.py", line 144, in __exit__
(APIServer pid=1)     next(self.gen)
(APIServer pid=1)   File "/usr/local/lib/python3.12/site-packages/vllm/v1/engine/utils.py", line 1119, in launch_core_engines
(APIServer pid=1)     wait_for_engine_startup(
(APIServer pid=1)   File "/usr/local/lib/python3.12/site-packages/vllm/v1/engine/utils.py", line 1178, in wait_for_engine_startup
(APIServer pid=1)     raise RuntimeError(
(APIServer pid=1) RuntimeError: Engine core initialization failed. See root cause above. Failed core proc(s): {}

[amber fractal]

Python isn't real

[stark needle]

-# meow kitty mrrrrr

[sage crag]

🏳️‍🌈

[amber fractal]

[sage crag]

[stark needle]

ploemon

[olive sable]

[sage crag]

updtse

[olive sable]

need fp8 quant i guess?

[sage crag]

connectioj i s secure

hrrps

[amber fractal]

Is female streamer, basically cannon at this point

Canari that is

[sage crag]

beeeeeeeee

[olive sable]

what?

is there a strong correlation between those 2 factors?

[fast pagoda]

sudo nix-store /tmp/trash
sudo systemctl reboot --firmware-setup

[olive sable]

big enub
benub

[fast pagoda]

i havent seen that since grade shcohol

[stark needle]

https://www.youtube.com/watch?v=V1gF5BNN3hw

[fast pagoda]

mother always was strict

[sage crag]

grad shoal

fish

🫆

[fast pagoda]

https://tenor.com/view/fish-just-fish-fish-fishing-fishe-gif-20372486

[sage crag]

5 mega gif

[amber fractal]

3/5 passing grade

[olive sable]

what is the left one supposed to be?

[sage crag]

niuj eiuh n?neurkh

[amber fractal]

I'm fine

[sage crag]

im fin

[olive sable]

im finn mertens

[sage crag]

vido

[amber fractal]

Doesn't loop

[sage crag]

lion

[olive sable]

https://tenor.com/view/dont-use-gif-7100164796034373204

:(

[sage crag]

under the sea

under the sea

🫧

totil

camt water

dispotil

[stark needle]

https://klipy.com/gifs/neuro-sama-evil-neuro-2--k01KR9XQCN4BGVHV72AWMWTTVVE

[olive sable]

https://tenor.com/view/cosmic-princess-kaguya-cosmic-princess-kaguya-yachiyo-gif-16298893631163188485

[sage crag]

thiev

[olive sable]

i stol gif

https://cdn.discordapp.com/attachments/1327142377139343430/1486773213190029444/doc_2026-03-25_19-13-27.gif

[sage crag]

plod

go fire

[olive sable]

https://tenor.com/view/anime-girl-boom-thoughts-anime-girl-boom-thoughts-gif-3847619567883703209

[sage crag]

n

by

[olive sable]

bye

[fast pagoda]

baal

[stark needle]

blll

[floral hawk]

Mornin nerds

[amber fractal]

Morbing

[olive sable]

is it morbing?

[floral hawk]

Its 6 pm but I was working on Zero until 9 so like, se la vie

[olive sable]

se la vie?

oh

[floral hawk]

"Such is life"

[olive sable]

c'est la vie

[floral hawk]

Yer

I don't necessarily speak french so

Just Japanese 😭

[trim valve]

[leaden crest]

good morbing sam

[leaden crest]

good morbing bred

[leaden crest]

good morbing yuuno

[trim valve]

morb

[leaden crest]

good morbing others I dont have the attention span to morb individually

[trim valve]

from the hit film mobius

[floral hawk]

Its morbin time

Zero heard me say that as a vocal tick (just now) and told me I'd have better luck using my brain as a paperweight.

[leaden crest]

zero you say

[floral hawk]

Zero is an AI i'm working on that will also be on twitch.

But everything built from scratch.

[leaden crest]

scratch as in with packages

?

[floral hawk]

I'm trying my damned hardest to avoid the Neuro fandom raising their eyebrows but it feels inevitable

[floral hawk]

With packages?? What do you mean exactly

[leaden crest]

imports

[floral hawk]

Ye

[leaden crest]

how come

havent seen eyebrows when you’re around

[floral hawk]

I know I know

[leaden crest]

maybe theyve been shaving

[floral hawk]

Might just be me looking into it too hard

But yk, AI vTuber, the whole bit. My long term dream is that maybe one day she'll be cool enough to join Neuro on a stream.

But that is a long road

[floral hawk]

Right now my current roadblock is reverse engineering ram addresses for a pokemon game. I'm having a bad time.

[amber fractal]

@fast pagoda

[leaden crest]

reference received

[olive sable]

[floral hawk]

https://giphy.com/gifs/neuro-vedal-vedal987-W81oag0eMTb0RYjm7t

[velvet crest]

You have unlocked new role

[olive sable]

try again

[leaden crest]

the irony

[amber fractal]

Go again

[floral hawk]

https://giphy.com/gifs/neuro-vedal-vedal987-W81oag0eMTb0RYjm7t

[leaden crest]

you did that intentionally

I refuse to believe otherwise

[floral hawk]

I didnt even know that's how that worked

[leaden crest]

the comedic timing was impeccable

[olive sable]

[floral hawk]

;-;

[amber fractal]

Still have the other stanza to write

[floral hawk]

WAIT

This means I can finally show you guys Zero stuff

Her UI, some chats, etc.

[leaden crest]

the song is stuck in my head now

[amber fractal]

[stark needle]

[stark needle]

brob

rob

brrr

[amber fractal]

morb

[stark needle]

[olive sable]

i need to revamp this drawing

[fast pagoda]

i dunno why i had this screenshotted but enjoy

[stark needle]

:neuroCatUuh:

[fast pagoda]

he's just leik me

[stark needle]

whats even the dataset u used

[fast pagoda]

me

[minor sail]

that feeling when you fix a three year old function, but you aren't sure if you fixed it correctly or if it was just intended to be wrong.

[leaden crest]

discord messages iirc

[leaden crest]

my bad for going looking for it #898005735198306357 message

[fast pagoda]

that one called me a flat earther several times

[floral hawk]

Me staring at my sarcasm import loop

https://giphy.com/gifs/lol-futurama-humor-cFgb5p5e1My3K

[fast pagoda]

the last like

3

were so unhinged

twas bad

[leaden crest]

what layers did you train echo on

[floral hawk]

Me? Or someone else

[leaden crest]

funyun

[olive sable]

afunyun

[stark needle]

a fun yun

[olive sable]

what is a fun yun?
what is a yun?

[leaden crest]

funyun rings have been proven to be rich in omemga 3

[fast pagoda]

all attn, all mlp layers

in this case

[floral hawk]

OH i didn't realize echo was an AI, thats so cool

[fast pagoda]

he liked the 2005-2013 ford fusion/mercury milan

[stark needle]

[olive sable]

no

[fast pagoda]

[olive sable]

thats konii

konii fer

[amber fractal]

[fast pagoda]

overfit as shit as it was r32 a32

this one was part of a parameter sweep basically

just experimenting with them doing runs with shotgun method

[stark needle]

🚩

[olive sable]

[amber fractal]

I don't want to be a programmer
I saught so hard and prompted this far and now I can't decode-e
All my time I've given her spite
These tokens in my mind keep adding on and on, and on. Yeah
I can only guess...
What's right. Should I keep this run or let it waste away?
Become a slave to all of these AI
It's overenginerred, all the slop yet I say it's fine

Prompting overclocked, trying to compile this time
Got myself into this mess
Can I really get this to work, nah
W, H, Y, did I think the agent could
She suffocating with tokens, I can't deny I've tried to code
All this time I've vibed
To be the one to 10x these experts
All my life I've failed
Now all of the accounts to fill the answer
Why, did, I, think I could solve it (solve it)
Try again, Remake reform, just need luck, yeah
Why, did, it, think that was the answer (answer)
New chat, prompt again and again, yeah
I can only guess
What's right? Should I stop or let my account just drain away?
Become a slave to all of these AI
It's overenginerred, all the slop yet I say it's fine

(Sec 1)

(Sec 1)

(Sec 2)

[leaden crest]

TheVoices

[fast pagoda]

was really gaming

[stark needle]

[fast pagoda]

om

[olive sable]

unless?

you should send that image to your sister if you havent already

[stark needle]

🏳️‍🌈

[fast pagoda]

the poor environment

[olive sable]

i think im downloading something

idk what

[fast pagoda]

[olive sable]

but the server is eating 93.2Mbps rn

[fast pagoda]

model?

did you start something with unsloth

[stark needle]

[amber fractal]

This just existed in my notes app and I saw the concept and decided to flesh it out all in one swoop.
The lyrics are probably garbage but it was fun

[olive sable]

nop

im trying gemma-4-26B-A4B-it-heretic liek soemoene here told me to

i did

docker run -d \
  --name vllm-server \
  --device nvidia.com/gpu=0 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  --ipc=host \
  vllm/vllm-openai:latest \
  coder3101/gemma-4-26B-A4B-it-heretic \
  --trust-remote-code \
  --quantization fp8 \
  --gpu-memory-utilization 0.9 \
  --max-model-len 4096 \
  --max-num-batched-tokens 4096

[fast pagoda]

it's probably thicc

[stark needle]

that one is for training

try the intel one

for inference

[olive sable]

[fast pagoda]

that's what you're downloading yea the training one will be high precision comparatively

[stark needle]

u need 4 bit for inference

[fast pagoda]

aka bigger

[stark needle]

the intel one is inference

[olive sable]

i am gonan train it

just not yet

need to make sure the base model works

[stark needle]

it wont fit on ur 4090

it's like

50gb in bf16

[olive sable]

then lower it to bf8

[stark needle]

still 25gb

[nocturne olive]

I look in here and see whatever this is

[olive sable]

hmmm

[stark needle]

u need 4 bit

[fast pagoda]

you can offload to system ram and wont lose that much

i mean

will be slower

significantly

but should still be fine

[stark needle]

not on vllm

afaik

[olive sable]

why not?

[fast pagoda]

sure you can i was doing it the other day

[stark needle]

ok fair enough

[fast pagoda]

you set a target for how much % of ram you want

[stark needle]

idk i dont offload models on vllm

just load in 4 bit

[olive sable]

i mean, sure i guess, i dont mind eitherway

[fast pagoda]

i mean i am usually offloading when doing it in 4bit because im loading like an 80b or something

[olive sable]

this is just to see if it works

[stark needle]

ok fair enough

ye then just use offloading

idk how to do that tho

at least load in fp8 with --quantization fp8

[olive sable]

i set it to fp8 here

[fast pagoda]

that wont work for gguf just in case you try that with a gguf at some point

vllm is bad at gguf anyways

but they come pre-quant

[leaden crest]

gguf

[olive sable]

what that

[stark needle]

@pluu

[leaden crest]

file type for llms

[olive sable]

why that?

[leaden crest]

they come prequantized which means they use less memory for minimal accuracy loss

[olive sable]

ah

can you not train to lower quant?

[fast pagoda]

you can

QAT

among other things

[stark needle]

or just lora on the quantized weights

same thing almost

qat uses hadamard rotations or some shit

which is expensive as hell

[amber fractal]

Size to speed argument never dies

[fast pagoda]

expensive once doe

b ut yea i mean qlora is 4bit native

aka meant for 4bit anyways

i think it does 8 too

i mean itll do bf16 but

[tiny edge]

nerd talk

[amber fractal]

I'm still of the opinion of trading complexity for speed personally

[fast pagoda]

rainbow

[amber fractal]

If I could ever get anything useful in the editor

[fast pagoda]

oh fugk

it's mother's day

[leaden crest]

we gotta wish vedal a happy mother's day then

[fast pagoda]

@sage crag heallo mother, the rolls will arrive to whisk you away to the finest old folks home vacation retreat shortly, happy mothers day!!

[olive sable]

mothers day is twice as expensive for children of lesbians

but they dont have fathers day

so it balances out i guess

[fast pagoda]

oor half

get them a joint gift

[olive sable]

i guess

[fast pagoda]

i gave my actual mom a kindle paperwhite or whatever like 2 days ago cuz she's traveling rn anyways

[stark needle]

🦊

[fast pagoda]