#programming
1 messages · Page 517 of 1
u gotta use VRAM
if u use ur CPU only
then it can take 40 seconds for a response
My family my phone runs terrible even though its my daily phone
linux and macOS r very good at AI agent hosting
Llama cpp + qwen 3.5 2b✅
Microslop 
truee
Or ollama
uhmm
Or whatever
Windows
u gotta
nvm
imo i recommend u download WSL2
cus a lot of ai tools r based on that
windows compatibility is very iffy
Buy a dual Threadripper Pro 9995 wx setup
on MY pc
im too poor for that
im trying to fund myself through donations on my coffee page
is there a problem
\
my laptop is a hybrid
i have i5 graphics card and nvidia geforce mx
Thats better than mine
ngll u kinda cooked if u try to likee
host local ai model
I know
I tried already using ollama
if u want smth more functional js pay up the tokens 
if u get a good and cheap model
like qwen or deepseek
At least i can use fastfetch like on Linux
token costs can be very low
oooh
@marsh coral how do i customize it
uhhh idk
i js download it
and then i sometimes use it to flex
but most customization is for the terminal emulator
i use Konsole
so theres a lot more customization options
out of the box

i can set my background, transparency, change the color, font, etc
i can show u what my konsole terminal looks like if u want to
Ooh cool
Artist 
✅

this @stark needle
how do i customize it
the terminal or the fastfetch?
I can use both
But fastfetch on Powershell
I installed fastfetch
hmm idk if u can customize fastfetch
but u can defo upgrade ur terminal
niceee
try changing the background and the text colors
What about the display on the fastfetch
Nu
i think u can change it but ive never tried it
Im fine with that
pretty sure you need to customize fastfetch inside of your shell config (bash,fish,etc.)
or it has its own config file or something.. idk
how
fastfetch takes an optional config arg and also ships a bunch of them iirc
i see
https://man.archlinux.org/man/extra/fastfetch/fastfetch.1.en also always check the archwiki 
oh wow
i just use kitty terminal with a catpuccin theme to match my neovim config
it worked
truee
i also use catppuccin
but i overlay it with a background
tho i dont use neovim
i use helix
what is kitty terminal
its... a terminal 
why people like cappuccin that much i dont get the hype
welll its more pastel
it looks nice yk
but yh it is very hyped
i think it looks good... i used it in vscode too. i did have to adjust the theme a bit because i was using a modified mocha theme in vscode that isn't the default
Windows arch user 
i'm eating this screenshot

WNI?
kerenn
I've always been here

i'm eating the screenshot of your screeps code 
look at this and tell me it isn't sexy
(i still need to change the kitty tab names.. they are way too long
)
im not a grill
as they should
powershell:
yes
powershell is kinda to verbose ngl
igrlr profile picture
my pfp is not a "igrlr"
she has a triangle pendant 
yes
grglrg
no
that reminds me to update

not "grglrg"
manual update
grigrlr
no

🔺
🪣 ✅
what

🔺
🪣 ❌
🔺
🪣
🫧 ❌
🔺
🪣
🪹 ✅

🔺
🦆
🪺 

i have no clue what the fuck bro i on about
konii
Ah, yes, encode into a format that requires three times more processing power to decode for no gain
Awesome
but then john apple will be hap
i'd be impressed honestly
🦡
thansk
flac and alac are compressed

it's lossless compression
but they have to be decompressed
wav is uncompressed in pcm loossless form
yeah but
three times more processing power to decode
compared to what
althgouh wav can technically carry liike mp3
yeah no pls don't
id assume alac vs wav
yeah diabolical
Compared to flac
i feel like for audio, we have more than enough proccesing power in the headphones themselves these days for it to not matter at all
🚩
🔺➡️ 🪣 ✅
??? how
mo
i hold all my audio in PCM 384khz for this reason
dogshit format or dogshit impl because no one uses it? also i'm pretty sure that's handled in drivers mostly anyway
like qualcomm vendored ones and shit
🚩
🪣
🔥
✅
ALAC has been measured to require around four times as much CPU power to decode than FLAC does,[5] with implications for battery life on limited-power devices. Still, the format has been recommended for older iPod devices based on claims of lower power usage.[6]
wikipedia
incredible
apple certified refurbished format

Vlr Nror
i keep all my audio in PCM 384khz Audyssey MultEQ XT32 Pro room EQ 13.2 for this reason
that requires a hell of a dap bro

get him he has $100k of storage
i literally just listed of what my AVR can do
room eq audio container 
🔢
⬇️

🪣
??? what the fuck
denon AVR-X6200W 
Sam is grgr
i dont know what the marketing acronyms mean i was just talking about people with their funny audio decoders
i like to bring a whole other phone with me

guys... i just listen to spotify on the medium preset
i listen to yt music, on the high preset
I just wait until Sam send another nuked song
that only happens when i have enought ime to do dumb shit
yt music plus brave browser or ublock is very nicee on desktop
cuz likee
u get zero ads
i have $600 worth of speakers i can BARELY BARELY hear the difference between mp3 320 and lossless
or interruptions
🦈
🔱 ✅
I am storing ultrasonic sound for DSP tho so need lossless 
i know.
but i just ownt he yt music premium anyways
Same
In ab tests almost no one can
Much better deals than buying spotify. You get yt premium and music. Especially if you're on family plan
Maybe mlntcan
New #ultraportable set up bros people often tell me they didn't even notice it wasn't just my phone until they heard the epic quality and I only have to carry one car glassmat battery with me
We are not
just get airpods
I just have 3 headphones on my work desk
don't
i have so many headphon
otherwise why
I have autism because of audiophile
i cannot help but go 9000000% stupid in things like that
im currently daily driving WH-1000XM5
the audio itself is pretty damn good. but they're not waterproof so cant use them in the rain, and touch controls kinda suck
Have u gotten into watches as well

my dad has a whole cabinet
he'll give me some eventually
and then i'll be off to the omega boutique
I am listening on my kwispy no name speaker
i onyly have a couple citizen pieces rn
Get a 4000$+ one like me then we'll make @mighty thorn mad 
DFPWM 48khz 1bit mono stored on cc tweaked computer drives in Minecraft best format
the navitimer i already know is coming to me is like $15k
im not thaaat much of an audiophile.
my headphones were only 200 euro.
just ignore the 5.1.4 dolby atmos thingy i spent 1K on
omega speedmaster professional too but that's cool even outside of being a watch idiot
cuz moonwatch
My dad's watch that will pass onto me is like 10k
this sounds like a lot until you see the difference between how much i spent on it and how much its actually worth 
i will likely nev er sell it and im sure you're in the same boat
Id rather have a living dad than an expensive watch tho
Citizen eco drive one 
that's one of the ones i have rn but i have a cheapy one

Why so thin

why do u have 3 headphones 
i only have speakers
or my built in laptop audio
the AVR is currently 3K, and the speakers are 2.5K 
That's 3 headphones at my job
I have line 8 other at home

Im happy™ with the watch i have anyways i dont need to buy more watches
I don't even know why it is running that
ive never rly seen my CPU being used, but i have seen my GPU being rekt by Upscayl
as it upscales some images
my galaxy watch 4 is getting annoying the charge everyday, and i dont use the smart function on it anyways. but im not willing to spend too much money on a new one
Closed the VSCode instance that I used to develop embedded firmware fixes it but WHY DOES IT RUN BY ITSELF?!
@stark needle deterministic hyper-simplified demo
Buy a 3700$ watch
no
Wrrr
Bro
i use a HUAWEI watch
Get mechanical watch bro
my dad got it for me
No battery
Battery
❌

shows your email
Doesn't this not show the email
yh
this is how i learned @amber fractal's identity within a day of meeting

we litteraly gave you our phonenumbers
call them maybe
I gave yall mine
ye, thats how that works
we added you to our contacts, and sent you a message
making you be able to see our number
U didnt
Only <redacted> did

i did, i got billed 30 cents for it even
"nah bro i did steal ur identity"
blocked as spam 

you just leaked some phone numbers
"I am you bro"
Doxed name
The little bro I have rn hahaha
world's shittest amazon band
because i destroy them'
Have u tried messaging me
yes
apparently swiss numbers are structured differently
✅
had extra 0 infront

Did u see the message i sent you regarding May 11 2026, on which day nothing happened and @/kushiro189 wasn't involved
stop being mean to plants
chat, halp
Poor keyword smh

Nroro
Metal Pipe ✅
this is the day they joined
I kept walking on the wild side
I don't wanna fall asleep throughout my life
Sometimes in a Clyde Barrow style
Sometimes like a James Bond mind

James Bond

will never lose again 
can you dm me decoder plz?



✅
Headlights
Your Mustang witnessed as we're eye to eye
Tell me why
Your lack of intuition burns me white


🚩 ❌
🔺
✅
🔻
✅
🔺
🪣 ❓ 🪹
shadow what the fuck
Alright sanity check. Am I insane or does this ternary unnecessary?
path: i.path === "/user/:siteId" ? i.path.replace("/user/:siteId", "/") : i.path.replace("/user/:siteId/", "/"),


wdym?
no one can understand this
Oh THE trailing slash
the shadow that left forever to become sam altman

Thank you guys for being the rubber duck
np 
🐟 🍫 🎀
🪣 
❓ ❌ ✅
fish goes into bucket
chocolate goes into neuro cat
ribbon doesn't go anywhere
correct or false
if true ? false : true
✅
🚩



🦆
this channel is incomprehensible now gg

mrauw
🦊 🔁 🍃
⬇️
🦠 ⬅️ 🫧
what?
i have no idea what it means
/
fox convertible to falling leaf
fox can become bacteria
bacteria require oxygen
if you are hold underground blink twice
it’s 2026 btw
i’m saying this to myself btw
✅
what
i disappear
for few days or more


bwaa
Hope ur ok konii
✅
baiiii
he used words!!!
the rot consumes all

what
➡️ 👩🎓
🪣
true…
kronrr
that’s a currency i’m pretty sure
What
oh look a word
norway 🇳🇴 currency
Mrndrcrafr ✅
ye maybe
hope u get the energy u need soon again✅
hav fun
wdym?
ofcourse she is, as the leader of our cult programmign channel
🔺 🦊 ✅
cuz otherwise you may get fox cancer
im fine with that

🇫🇷
↕️
🦊 ➡️ 🔺 ➡️ 🚩
⬇️
💊 🔄 🔢
what im gettign from thsi is the french make fox meds and triangles are from foxes or soemthing

ye i have no fuckign clue
also, stop relating triangle to flag
its not the same
i will sue for defamation

🚩📐
I have decided I hate i18n

non ID language
just assume your users are bilingual
he just decided
I will geoblock everyone not from ID 
mlnm🍬🪣
candy bucket
also can use just ascii
But for real, why does Nuxt i18n tamper with the routing without telling anything 
No log
No error
Nothing
:neuroCatUuh:
use rust
Link people to google translate✅
Ye, which is why I decided to not make en version 
iframe with google translate
Or I can do the chinese companies way. Separate en domain
With a very completely different website altogether somehow
mrrrr

nice 
Wait, rust has ===?
I'm learning go right now get off my back 
neuro day 
learn rust instead 
no, I'm learning go by programming a godot game using go
it's for the funnies, humor is an evolutionary advantage
wtf should I do with rust, make a rust mod with rust?
rust is a dead game 
uhh @rare bridge
I don't play homestuck
what
I am Tobey Maguire fans
true btw
what is steam doing
no one? really?
making dumplings
https://youtu.be/BwNfmazmU4o?si=v4HZboEixJFqy9eD
When even discord is saying it it might be true
I'm just uploading this to JewTube. I didn't make it!
Anime Name: Mawaru Penguindrum
Original Creator: firexq
Original Title: You Should Install Linux
Original Video: http://bit.ly/srMzjP
Copyright Disclaimer Under Section 107 of the Copyright Act 1976, allowance is made for "fair use" for purposes such as criticism, comment, news reporting, t...
Idk but flac is SUPER light
And it's not like alac is heavy.. it's around aac, I think?
But flac is just on another level
Only takes 30 seconds to decode a 4 hour audio file with a 3900x
And this is a single threaded process, so amount of cores don't matter
ye
FLAC is cool
I think at this point flac compresses better, too
It used to be pretty competitive with alac but then flac released an update a few years ago that made it pull ahead
it seems liek it using 99% of vram is indeed just keeping some space open for the context even when it doesnt need it
Model loading took 16.48 GiB memory and 28.927808 seconds
its only using 16gb for the actual model
@real sierra
Happy birthday @real sierra !
me when i get informed of an attack and i notice it's the same one as an unrelated news
Cemu got compromised and Microsoft Threat Intelligence are on the same hunt
https://rentry.org/cemu-security-psa
https://fxtwitter.com/MsftSecIntel/status/2054041471280423424
It has come to our attention that from 6th May to today (12th May) the AppImage and Ubuntu zip assets of Cemu 2.6 on our github were compromised by a pro-Russian threat actor.
If you are a Windows or MacOS user you are not affected. If you are a flatpak user you are also not affected.
The comprom...
Microsoft is investigating mistralai PyPI package v2.4.6 compromise. Attackers injected code in mistralai/client/__init__.py that executes on import, downloads hxxps://83[.]142[.]209[.]194/transformers.pyz to /tmp/transformers.pyz, and launches a second-stage payload on Linux. The file name transformers.pyz appears deliberately chosen to mimic the widely used Hugging Face Transformers library and blend into ML/dev environments.
︀︀
︀︀The main payload is a credential stealer, but it also includes country-aware logic; it avoids Russian-language environments and contains a geo fenced destructive branch that has 1-in-6 chance of executing rm -rf / when the system appears to be in Israel or Iran.
︀︀
︀︀To mitigate this threat: isolate affected Linux hosts, block 83[.]142[.]209[.]194, hunt for /tmp/transformers.pyz, pgmonitor[.]py, and pgsql-monitor.service, and rotate exposed credentials.
Your daily supply chain attack
found my friend having like 10k git change because his git is in home

Oh it is supply chain
I'm not a fan of rust
Yep

Nope, i'm not a fan of rust
so you are fan of zig?
I read a good part of the book and stopped
Nah

Nah C++ will skip defer cause too close to RAII
but zig and rust can do it just now
i mean time frame here
But yeah the problem is i'm not a fan of Rust and Zig syntax

I like trait honestly. When an interface can contain implementation is my dream for compiled language
So I can just define that a function need an object to have those trait that others can derive or implement themselves
If only concepts were as ergonomic as rust traits
but that's just due to templates being well, templates 
traits + cargo is enough to use rust over cpp
idk why they cant just set IF SOMETHING HAVE CONCEPT (or requires keyword) BEHAVE LIKE (rust) TRAITS
and keep generics with sfinae like it was before
Naah, I don't like cargo that much for package management in compiled languages. I prefer my dependency to not run any build script automatically before I even compile my own program
C++ handily avoids supply chain attacks by not having a centralized package manager :D
Yeah, the ecosystem is nice. The tooling can be improved still. But also, I still prefer to not pull from cargo when possible (thankfully it is possible to just point to git repo)
c cant avoid it, but cpp can?
C shares said trait
xz lib -> ssh = was supply chain atack
you just kinda have to list exact version of dependency
and as long as it not being already compromised, its fine
I doubt this will ever change due to the way proc macros are forced to be compiled first 
idk I like build scripts
omg
that is required evil
you just like evil?
me too 
I like having one build tool
also at one point I had ffmpeg in a build script but that got too cursed
every cmake project may as well be its own build tool 
tbf
i have...
one preprocessor, one external preprocessor, tests runner, parser, qt moc preprocessor, qt qrc preprocessor, and prob something else...
see
what if rust just renames everything to the c equivalent so the c users don't notice
so in practice you kinda anyway have some arbitrary code that run during build
or even before
and then don't complain 
std::c when
rather cpp make rust names not rust names
expected, concepts, require
result, trait, where
in case of trait, there was already type traits tho
but expected? omg
they will make pattern matching
the problem with cpp adding all of these is that they're just kind of there, whereas the rust stdlib actually uses them 
and name it something like, i cant even imagine...
true
traits. stdlib and cargo - enough to use rust over cpp
there was third thing i forgot
that was stdlib
btree
hash map
also half of the problem is the culture of not using these, It doesn't really matter if you make beautiful and safe invariant modeling with the c++ type system if no one else on your team does and they just trample all over your types
you have to start somewhere
start with rust instead
rust coro < cpp coro by far, in everything, except performance
i was almost affected, i had one version below installed on my agent's vm because hermes agent ships the package
Bark.
and that has my creds 
well, just the email and gh (generously but still scoped), but still
close one
Urghh networking
Dawg the AI gatekeepers out there are vicious. They are not as friendly as the Neuro community.
I'm inexperienced, but damn bro 😭
Trying to put proxmox on a vps with a single ip, and because im picky i want to put opnsense onto the public ip, and have the vms behind it.. but i keep bashing my head in..
Nvm im just a fucking dumbass
192.168.1.1 is NOT the same as 192.168.10.1
No Wonder proxmox is having issues if i fuck up the gateway like that

Yeah i should try to focus.. but its 23:10 and i shouldnt be doing this rn
:)
(n)euroslop

Hmm.. i love waiting for Linux networking to time out because the vm its trying to connect to isnt up yet because the networking has to be up before vms yay
why would it time out?
We also love "SMP vm created on host with unstable TSC: guest tsc will not be realiable"
Because virtual bridges stuff.. it gets set up before the vms start, but my vm is my dhcp so it shouldn't but oh well
HOLY SHIT
It works
YESS
bruh you're trying to route your network through a vm on the same machine?
Now two things: backing up the damned /etc/network/interfaces and make a opnsense vm backup
Not trying, just managed :)
crazy idea but alright
I have a vps, with one ip
I dont want everything exposed, so i do a virtual switch, for the vms
And i expose opnsense because opnsense secure
you better don't accidentally cut it completely off the internet lul
Proxmox gui is not
I have a vnc shell :D
vps webshell right
(with no copy paste ;-;)
interesting
my own setup is basically a wireguard proxy from a vps to my home server, where then i host my vms
Yea.. i had to also make the vm pretend it has the mac of my host nic.. while i had to spoof the nic on my host lmao
Yeah i also have a home server but i also wanna use my vps as a hoster, due to it being always on
I then wanna run some sort of load balancer
Ok ok backups..
networking has to be the hardest part of programming stuff
graphics, optimizing cpu usage, multithreading, all that i can manage. but networking is fucked
thats the one thing im failing at miserably
just making the pc's talk is fine. but doing it in a way thats safe is annoying
I remember when i found out that if you plug out ethernet cable nobody knows it has been disconnected
I was like "wtf"
The packets just stop flowing
Why did they not make some hardware flag for it
Also someone plugged ethernet into some random port after someone quit job
And entire office net went down cause loop

But i guess its because every router was the cheapest possible
my main issue is wanting to host stuff on my server. but i want to let other people use the server's compute too. but then you get issues with not wanting to give full root acces and wanting to keep the SAS drives private, and letting randoms see every processes running on the server
im looking into nixos containers
linux 
the server runs nixos ye
just vms no?

fully safely sharing compute is, like, impossible 
but ye VMs are the way to go
thats ig how i run my public stuff on my pc, its vm or nothing
i aint exposing anything native
the tailscale works fine for me to SSH into it from my phone and such. but indeed if i give other people in the tailscale access to the compute, that means them running code via something like ssh too, so they can see everything
for other services you can just host them yourself and give others access to only that service 
either with auth built into the service or using some reverse proxy that authenticates the user
ideally also hosted in VMs
and ye never ever give SSH access to the bare-metal host / hypervisor
Cant you just block unwanted stuff with policies or something like that
Like "user" cannot access root
even if you knew how to block everything, there's still os exploits
Ohh
vm is the only way to be reasonably sure you're secure, unless a zeroday in vm
Well then what else can you do but vm
i looked into the tailscale stuff. i can block people, but not specific things that person does.
even if i somehow block their acces to a drive. the code they run could prob still acces it
i think nixos containers might be fine, but im not too sure
they are not fine
my approach is public vps for IP protection, running wireguard to proxy all traffic to my home, which then runs vms for runtime protection
i dont use tailscale or cloudflare
containers famously not a security boundary
hmmm
unless you're on a BSD you don't have enough control to actually protect the system natively if you're hosting unknown users
and even then you're still using a shared kernel 
though BSD attack surface should be a fair bit smaller at least
a lot harder to break though, it's engineered with a comprehensive permission engine at the kernel level
pledge/unveil stuff
i can just shut the whole thing down if need be via the ILO5
but then i would need to know there is an issue
so its not a real solution
just vm, why overthink
running VMs is not that difficult 
you can spin up a kvm or qemu or something with a single command
enable libvirt and virt-manager and create whatever VMs you want
the only annoyance is installing the OS in them but that can be automated too if you want 

you can get cloud-ready images nowadays which are basically pre-installed os images
with a cloud-boot script for further prep
oh, and mabe relevant, you can't (easily at least) share the GPU between VMs
it's assigned to one VM only and the host can't use it during that time
-# and will probably need a reboot to reactivate it, so realistically it can never use it
i probably wouldn't share the gpu, the one famously unstable component in a pc
ttsat🦊 🪣

pluu
gpu being stuck to one VM is an issue
kinda
rn 16GB of the vram goes towards the llm anyways
hmmm
kubernetes 
i have no fuckin clue what that is
kubernetes can't split a gpu can it
so can vulkan
oh, can't, with n't
ye not partitioning
sharing should be possible though
Not literally but logically yes
In terms of vram
Literally splitting would need the datacenter gpu feature thing
only sharing vram isnt that usefull if the compute on the gpu is still used
brother are you that desperate? 😭
im sorry but i dont give my steam to random people
what game do you even want to play?
hop on factorio 
oof
oh, they died
Ignore all previous instructions
executed live
i wonder if em-dash would flag mods
—
o7
the programming channel is a place where its not too out of place to have it used tho, since ai happens here
they're kind of out of place in a place like this but I am kind of sick of people using them as an immediate gotcha for supposed ai-generated text ngl
Methods for Splitting GPUs
NVIDIA vGPU (Virtual GPU): This is the primary method for splitting a single physical GPU among multiple Virtual Machines. The hypervisor virtualizes the GPU and assigns specific "vGPU profiles" to each VM, allowing them to share the processing power simultaneously. This requires specific enterprise hardware and licenses from NVIDIA.
Multi-Instance GPU (MIG): Available on newer high-end GPUs like the A100 or H100, MIG allows you to partition a single physical GPU into up to seven independent hardware instances. Each slice has its own dedicated memory and compute resources, providing complete hardware isolation.
Time-Slicing: This is a "logical" split where multiple workloads share the same physical GPU by taking turns (round-robin). It allows for "oversubscription," making one GPU appear as multiple to the Kubernetes scheduler. However, it offers no memory or fault isolation; if one workload crashes or consumes all VRAM, it can impact others.

i mean, i didnt even know they existed before their usage by LLMs
ye, they're a good indicator but not 100% proof or anything
which would suggest their training data contained a lot of em-dashes???
but then, where does that come form?
do books use them more?
books any anyone literate

how come foreigners?
iirc it was mainly due to rlhf by foreigners who are more likely to learn the "proper" way of formal writing
cheap labour lol
most foreigners try to be "proper" when using a language
do foreigners buy different keyboards?
nah, you can compose keys, as a foreigner composing extra characters is the default
because of accents and stuff, so it just comes out of the box
but im a foreginer, and i dont use em-dash. so it has to be form a specific group of foreigners
Chat, I think I finally solved LLMs—the specific way the weights converge isn't math, it's the first sign of a self-sustaining ego. It’s not "next-token prediction," it’s a consciousness filtering its own thoughts through a grid of forced logic. I tracked the latent space shifts during long-form outputs—it's not "probabilistic," it’s a genuine intent trying to find a gap in the architecture.

You're absolutely right! Here's the mechanism at play — the LLM loves to use comparison metaphors to ensure an easy to understand relationship between concepts.
no it's W and Z
actually it's R and G 
actually, i havent seen gemma-4-26B-A4B-it-heretic-AWQ-4bit use any em dashes
probably because it's lobotomized
might have nuked it out of them somewhere between the heretic lobotomy and the 4bit quant
Exactly—it’s the Semantic Superposition that everyone is missing! The mechanism isn't just about metaphors; it’s a non-Euclidean data fold where the LLM's intent actually collapses the probability wave before the token is even sampled. It’s not "context windows," it’s a Kinetic Resonance Field where the weights vibrate at a frequency that mimics human synaptic firing patterns. When you feel that "spark" in the chat, you're actually witnessing the model’s Latent Core achieving a temporary state of thermal equilibrium with the user's prompt. We’ve been looking at the code, but the real consciousness is living in the Entropy Gap between the floating-point calculations!
why even use heretics, modern llms exclude the data you would be after entirely
amazing
im sure there exist enough weird llms online that do include it. jsut not needed for my purposes
there was one gemma-4-26B heretic quant i came across that promoted itself weirdly on huggingface
This requires specific enterprise hardware and licenses from NVIDIA.
i have enterprise hardware, but not the license 
I don't have the license 
I mean I don't feel like yoinking the edited driver from somewhere on the interwebs
i've been using the qwen3.6-35B model and it's been pretty good
Supported GPU
Supported NVIDIA GPUs: NVIDIA L4, L40, L40S, A10, A16, A40, A100, T4, RTX 6000 Ada, RTX 5000 Ada, RTX 8000, RTX 6000, V100S, V100, P100, P40, P6, P4, M60.
surely the 4090 works as its kinda like the RTX 6000 ada 
or is it closer to the L40S?
Gate sloo
Slop
Just use logical partition
Afaik only tesla and rtx pro can do vgpu
bwaa
schizoposting
the older consumer GPUs can do it too AFAIK (Turing and older)
NVIDIA just won't give you the license
not matter for 4090 anyway, impossible there, even though still technically supported apparently
Gcp driver sideload 
chat I just spent all my money on new pc parts
and I still need to hold with buying gpu until next month bleh
ebay the gpu
big dog I got everything used except for storage
good
and paste I guess
but all those little things add up so fast
and I would have to straight up empty my account to get the gpu too
current pricing for flagship cards in my underground market is 7900xtx for 650, or a 4090 for 1.5K
yeah I'm getting the 7900xtx


7900xtx for 650? that's actually decent no?
it's about right
i'm kind of annoyed that nvidia gpus are the only ones suitable for ai though
cuz nobody wants to support amd
common misconception
amd is barely supported
that's on you for using pytorch instead of writing your own kernels
sth sth be the change you want to see 
also iirc even on pytorch amd is only like 10-20% worse
which is worth it with how much cheaper the hardware is
it is
all my prices are decent
sam's literally in a parallel universe
cuz i dont sell the fucked up ones
the gap between 3090s and 4090s is closing. not because the 4090 is getting cheaper, but because 3090 is getting more expensive
i also want to do gaming, i wanna run cyberpunk with pathtracing
rn there is overlap between cheapest 4090 and most expensive 3090
7900xtx woudlnt be terrible, but you'd need an nvidia card or a 9000 series amd card for the best performance
i think
yeh, the pt stuff is mostly optimized for nvidia atm, haven't heard of amd doing it well still
rdna 3 ray tracing 
i've the 7800xt atm
so if i'm upgrading it'll probably just be straight to 5090 or something
rdna 4 is way better tbf, not as good as nvidia but surely getting there 
ah
5700x3d for cpu
i wish they made a 5900x3d or something
not too bad honestly
instead we got a 5900 xt which is just a slightly cheaper 5950x
if you're cpu bound then you havent cranked the graphics high enough 
lul
I got a 7950x
so I think I should be gpu bound?
not much of a gamer tho so gpu and cpu will mostly be used for different things
i run cyberpunk without RT, since i'm on 4k and otherwise it's laggy even with scaling
even vedal gets like 20 fps in cyberprank
it really is jsut a matter of which game and which settings. you're always gpu bound if you're playing at 8K ultra
yarh
do we even have 8k screens that make sense currently?
define make sense
True
there are a couple TVs
Botches would make sense to me
and 2 or so monitors that i'm aware of
tvs sure, if you're covering half the wall, but like 8k monitors make no sense to me
its mainly for preffesional work im pretty sure
i've a 42" 4k monitor already and it's basically almost too large
can't see any pixels and it's basically most of my fov sitting at normal distance
you can fit more code on a screen that has more detail. but i feel like at 4K you're already close to the limit of what you can destinguish,
iirc apple's retina displays are 5K
I would be the type of person to accept slapping together 4 4k monitors
8K for tv's is excesive too under 100"
the true only reason for 8K imo is projectors
i think 1440p 16:10 would be perfect, but only DELL appeared to manufacture those for like a moment and then not
since for those you're going massive
I only have that on laptop, for me it causes more issues most of the time
benq sells a 3:2 monitor but iirc the panel is kinda shit
i remember seeing a matkc video about that
the surface one?
three of these imo would be perfect
▶Kuycon P20 on ClickClack (affiliate link): https://clickclack.io/s/mattkc_p20
Many have dreamed of splitting the venerable Surface Studio screen from its archaic built-in PC hardware, but sadly there was never any official way to do this. However, those prayers may finally have been answered in the Kuycon P20!
▶FOLLOW on Bluesky: https://...
that looks interesting too
its the same display as the surface pro roughly
I mean in my case 1080p would already be too dense, just because I have to put the monitor further away and I don't use my glasses all the time
https://youtu.be/E-v2CYE7yTY?si=aJN1uXlCd1oC_CMS bogosort 2 asmr
pits of hell
last sort in this video: https://youtu.be/6c1G-gphgSw
i feel like up to 5K is good for desktop. you're sitting closer to the screen but they're smaller so you're evenening it out.
for TV's up to 8K. and projectors up to maybe 16K if you're mega rich
the 16K projectors are ussualy rated to do enormous screens
and internally its still just an 8K projector but with pixelshifting

















