programming

[tender river]

[sage crag]

it not demonstration

[tender river]

true

[trim valve]

time to

convert laptop to flakes

[tender river]

is it your second machine

if so you can use same flake as first one

[trim valve]

first machine is not on flakes

😔

[tender river]

[trim valve]

too lazy to move it

or something

[tender river]

you can do both at the same time

[trim valve]

true ‼️

[tender river]

if have ssh access it super easy to deploy remotely too

https://nixos.asia/en/configuration-as-flake but you dont need me to send you this probably

[trim valve]

they're both on a tailnet so yeah

I saw you post that earlier

its like the first actual guide that didn't miss the important stuff

[tender river]

[trim valve]

😭

[tender river]

it still misses tons of QoL features

but good enough

[trim valve]

[tender river]

nixos-rebuild takes --target-host argument (like --target-host root@1.2.3.4)

[trim valve]

is there a good guide on managing multiple devices

[tender river]

thats it really

[trim valve]

ooh

god this laptop has such shitty palm rejection

[tender river]

well you have to add #<hostname> after flake path

like this nixos-rebuild switch --build-host root@10.0.0.2 --target-host root@10.0.0.2 --flake ~/code/nix/dotfiles2#server

i have my own nixos-rebuild because official one didnt work very well
https://github.com/chayleaf/dotfiles/blob/master/nixos-rebuild.py

[trim valve]

[rough bloom]

there's a few different tools with different philosophies on how to handle deployments, Nixops being the "official" tool, but they all end up effectively running the command leaf just posted in one way or another
not needed if you have just a few hosts

[trim valve]

its basically just my laptop and my pc

I might eventually add a server to the mix but that will probably be a different thing anyways

[rough bloom]

you can share code between different hosts / NixOS configurations however you want since they're in one flake, it's completely independent of how you deploy it
I have a set of NixOS modules that I import in multiple configurations

[tender river]

its very short build on host i want -> copy from build host to target host -> deploy on target host (which depends on which action i passed, some dont deploy at all, some run particular commands on target host after copying the derivation output from build to target)

[sage crag]

[tender river]

[rough bloom]

exponential

[trim valve]

my only gripe with this is the flow of using my laptop kinda doesn't fit it

my pc lives as being on most of the time

my laptop is more of a "on when I need it and not much else"

[tender river]

can at least check that the config builds

[trim valve]

true

[tender river]

then its nixos-rebuild build i think

[trim valve]

maybe I just need to ssh -c "nixos-rebuild switch --target-host laptop" pc

and pray I have fast internet

[rough bloom]

either that or just sync the flake files and build locally

[tender river]

ye git

[trim valve]

laptop weak

I'd rather build on pc

ik nix has something for that

[tender river]

building is virtually nothing

unless you use patches

[trim valve]

true i guess

[tender river]

most of nix updates is just evaluate new config + download the files

reason i suggest building on target host is

dont have to download and then transfer

can just download

[trim valve]

yee fair

[sage crag]

10 year of ssd lifespan

[tender river]

good thing ssd cheap

[sage crag]

in 2026

[tender river]

spoiler

[rough bloom]

can use --use-substitutes to help with this
but ye just build locally, way simpler
can always use PC as remote builder if needed but that seems annoying to set up

[trim valve]

I actually already tried it

[quick condor]

Im glad wear isnt a concern on my 1.92Tb ssds... the endurance is 1 DWPD for 5 Years (3.5PB)

[trim valve]

main issue was

I wasn't using flakes for my config

so it wasn't doing anything with flakes

[rough bloom]

[trim valve]

because I think it only does it for flakes or something?

not sure

ig I also just don't do enough stuff where I don't have anything that actually gets built

[tender river]

flakes do make managing multiple configs in one codebase easier because no channels

otherwise they shouldnt really have anything special

i think

[rough bloom]

ye I think so too

but never tried without flakes

[trim valve]

[rough bloom]

syncing channels between hosts seems bad

[trim valve]

command looks different

[tender river]

you can just use ifd and no channels

[sage crag]

spspsps

[trim valve]

also need to figure out vscode remote server

and stop myself from forking sunshine

[tender river]

sunshine

rainbows

unfortunate name

here

nothing but

industrial wasteland

[trim valve]

these darn linux users

and their "I wanna do stuff properly"

lemme just copy the windows version 1:1

its near useless because of linux shenanigans

• doesn't work if not signed in, because it runs as a user service instead of a system service or something
• can't wake up the screens, and if the screens aren't on then it can't share screens

second problem will be solved when I move to virtual displays (kinda)

but the first one is seemingly a linux + wayland moment

also for whatever reason the hardware encoding looks terrible on linux but not windows

[young plover]

🤔 Figure out what? I've used it through SSH to an embedded device and it just worked, albeit slowly.

[trim valve]

it has issues running on nixos hosts

[young plover]

Hmm, must be expecting different paths I guess

[sage crag]

pspspspsp

[trim valve]

https://nixos.wiki/wiki/Visual_Studio_Code#Any_client_to_NixOS_host

wants flakes and stuff

[sage crag]

brbbebebqmwme

bwrr

https://tenor.com/view/bunny-rabbit-pats-pet-bunny-pats-gif-19861222

[tender river]

[sage crag]

[trim valve]

chat I am suddenly questioning my choice of laptop

surely this cpu has hardware accel in the igpu

[sage crag]

https://tenor.com/view/bunny-rabbit-pats-pet-bunny-pats-gif-19861222

[tender river]

bnuuuy

[hard delta]

my gut feelings tell me you're missing some angy proprietary driver somewhere that goes "there it is."

[trim valve]

probably

[sage crag]

i think discord broke the cdn

[tender river]

work fine here

[trim valve]

actually nixos question

[sage crag]

not for mee

[trim valve]

how does multi host hardware-configuration work

[sage crag]

o

[sage crag]

o

flakd

flake

[rough bloom]

ye your route to CDN broken

[tender river]

each host use its own

[sage crag]

keep hardware specific stuff in

one file

[tender river]

me not use generated hardware config anyway

write my own

[sage crag]

ye mee too

[trim valve]

do those go into version control

[tender river]

yes

[sage crag]

ye make

branch

[tender river]

flake wont pick up anything not in git repo anyway

[trim valve]

ooh wait

[sage crag]

or folder

[trim valve]

I see in your repo

[sage crag]

i dont have more than 1 system so my hardware conf is entirely

in the thingy ye

[tender river]

system/hardware
have hardware specific config

system/devices
have device specific config

system/hosts
have device/hardware agnostic config

[trim valve]

[tender river]

anyway just make something that seems simple to you

[rough bloom]

I have a nixos-modules/ directory with individual files containing anything shared and individual hosts/<hostname>/ directories for the host configs (+ some metadata)

doesn't distinguish between devices or hardware or whatever other shared config other than being named differently

but ye just pick something simple and expand from there

[tender river]

ye me have system/modules too

stuff in there included for every host

[sage crag]

https://cdn.discordapp.com/attachments/1067638175478071307/1452069761512112159/copy_CFBD3ED9-4A88-4341-B0AF-2450399B6EE7.gif

wrr

[silent cloak]

got my 3D printer finally all working

[tender river]

can it print dirt

[trim valve]

grr

what do you mean computer

[quick condor]

Well because I love overly complex systems, I am now using 5 total models for my ai (Main instruct, reasoning, whisper, embedding and 8b enhanced embedding)

Wait, while probably unnecessary, Is there any reason for Zen to pass (Zen VTT?) -> Zen text ->Zen TTS-> Neuro VTT -> zen text -> Neuro input, Couldnt they just do (Zen VTT?) -> Zen text -> Neuro input?

its in (?) because idk if they are just using text into TTS or voice

[fathom crescent]

Hello there, i'm learning abt front and back development, anyone got pages to learn javascript, node.js and any sql databases like postgre ?¿

[quick condor]

IMO the best way to learn is to pick a simple project and learn how to make it using the thing you want to learn

[fathom crescent]

and off topic anyone knows a good keyboard ? with the > sign ? i'm a spanish speaker and i'm struggling with those sign in javascript bc my keyboard doesn't have it lol

[sage crag]

off topic

you should see the topics that normally go on in here

[sage crag]

in most english layout keyboards, the . key is the same as the > key

[fathom crescent]

yeah sorry abt that

[sage crag]

[fathom crescent]

yeah but as i said i'm a spanish speaker and english and spanish layouts are way different

[rough bloom]

ideally something that motivates you
if you really don't care then just make a TODO list app or something, it's a classic

[sage crag]

a todo list is just a regular list with a todo header

[rough bloom]

exactly

very simple

[fathom crescent]

i can't get used to english layout so the simple solution is to get a keyboard that has those signs

[sage crag]

wrr

[fathom crescent]

i also read that but i'm struggling in learning the basics

[fierce light]

right alt + . or , <>

if u have the euro layout with you language's specific memecharacters then right alt has all the gimmick symbols

[tender river]

i dont know of any keyboards with extra physical keys for > specifically

ye use alt gr sound about right

can also just learn english layout its not that hard me know 4 layouts

[sage crag]

On-Target Earnings

[fierce light]

i can't type in slavshit on english kb 😭

[stray dragon]

hi konii

[sage crag]

hi t

[stray dragon]

hi leafchay

[sage crag]

hi t

[fierce light]

i need all the ů and others

[stray dragon]

hi konii

hi konii

[sage crag]

hi t

[stray dragon]

hi konii

[sage crag]

hi t

hi t

[stray dragon]

hi konii

[sage crag]

hi t

[stray dragon]

hi konii

[sage crag]

true

ctrl+:

on my IME

clipboard

[obsidian mantle]

true?????????

[sage crag]

very useful

i didnt know this

[fathom crescent]

for me to use that i have to win + space to change the language of my windows 😭

[sage crag]

[sage crag]

ohh its japanese

お手

"paw"

[fathom crescent]

yee there some keyboards with the <> signs right to left shift

[sage crag]

@tender river i thought it was acronym

evilDentge

[fathom crescent]

i had one of those but i changed to a mechanical keyboard and it doesn't have those key signs

peposadge

[sage crag]

https://tenor.com/view/my-jolly-reaction-jolly-my-honest-reaction-gif-5490890317762380451

[obsidian mantle]

is it still jolly out there or it ends on 26th

[sage crag]

vedalPuppet

christmas ends at new year

i'll keep using this gif though

mainly because it annoys people

[tender river]

ye i was about to send but was busy finishing curry

[sage crag]

i typed out

you could say that i said it faster than you could finish curry
and then i realised it made no sense as a joke

me sense of humour

autismu

[tender river]

you could indeed say that

[sage crag]

i think its supposed to be a play on

"if X, Y faster than you can say Z"

but

i failed the concept of the joke

evilDentge

its probably funny in some language

so im fine

[tender river]

o right forgot about those

me laptop have that key

[sage crag]

why they call it laptop when top on desk in on the not the lap

[trim valve]

I enjoy laptop on lap

[sage crag]

you have no lap

[trim valve]

[sage crag]

no lap

no legs

[tender river]

they burned away

from

gaming laptop

on lap

[trim valve]

[sage crag]

why doesnt discord allow us to customise spacing on emojis

ALERT would be so much better

and

neuroNom could be generic

neuroNom<M>

[tender river]

template<emoji E> emoji neuroNom

[sage crag]

i not speak psycho

[tender river]

coming in 2026

for

nitro users

on boosted server

[olive sable]

hi

i have mellow cakes

i passed mito, cloud and chay in the server ranking

[nocturne olive]

[sage crag]

ohh what me ranking

i rank 41

[olive sable]

since we're doing #cooking

i ownder how hard it would be to make mellowcakes form scratch

[sage crag]

113k

yay

wrr

[tender river]

where repo

me run

[olive sable]

https://github.com/samvanmaele/vulkan

[silent cloak]

Vulkan my beloved

[tender river]

me have 27k llvmpipe 40k igpu 10k dgpu

[olive sable]

10K dgpu

strange

what cpu and gpu?

[tender river]

AMD Ryzen 9 5900HX with Radeon Graphics

AMD Radeon RX 6700M (RADV NAVI22)

[olive sable]

maybe move around the window a bit

idk why but ive had strange stuff on dgpu

letting it rebuild the swapchain might help

[sage crag]

im upgrading to kernel 6.18

get ready for 130k

[tender river]

na not help

[olive sable]

not even with resizing the window?

[tender river]

nope

[olive sable]

hmmm

idk

strange

[tender river]

ok it does jump up to 14k in fullscreen

with the caveat that

[olive sable]

[tender river]

i dont actually see it

oh here managed to get 12k while seeing it

[hard delta]

hey prog chat would you cash in on a new pc now while ram is expensive (but there's still supply) or would you wait for those prices to seep into gpu etc as well 😟

[olive sable]

depends on how badly you need a pc rn

most stuff you should jsut get 2ndhand and hunt ofr deals

[hard delta]

i have an okay laptop (ryzen 7 4800, rtx 3060, 32gb ddr4)
but i've been wanting to get a PC around last summer, just had to postpone because of irl shit happening
i really don't feel like waiting another 2 years only for the flat overall price to rise to 4x <current> then fall back to 1.75x <current>
but i also don't feel like i can just yank $3k at new stuff
honestly i'm mostly just curious about your predictions
haven't checked used market yet

target is some AAA gaming at 1440 and some dicking around with training not too big stuff (backend dev with medium determination to catch up with the mostly sailed AI hype boat)
also i really just wanna get away from mobile hardware i've spent too much % of my life on it

[olive sable]

get a 2ndhand 3090, an older cpu that supports ddr4, and some cheap-ish but still 2 sticks ddr4 ram

even if prices are high you would still want dual channel ram

[hard delta]

also i'm somewhat averse to trading used goods because i generally dislike potentially having to nag people selling faulty stuff etc

[olive sable]

used 3090s are just generally the best deal

[umbral thorn]

not big training?

like wot

becuz maybe 3080Ti better deal 4 value

[olive sable]

from what i can find in my area 3080ti are same price as 3090

only 3080 non ti is cheaper

by like a full 200 bucks

3080 - 300euro
3080ti and 3090 - 500euro

[hard delta]

so is 40-50 series little added value overall?
architectural changes and such
i assume new fancy features are mostly geared towards gaming and aren't that useful for training

[olive sable]

most of the new features afaik are ai upscaling and pathtracing

dlss and stuff

[stray dragon]

faster inference with 8 bit and 4 bit quantization respectively

4bit is lossy anyways

[rough bloom]

the newer cards are also way faster for training, at least comparing the 3090 and the 4090
presumably the lower tier cards are about the same

but also way more expensive

[olive sable]

im not up-to-date on ai training cards

[fierce light]

30xx also starts getting deprecated in some snakeoils

[olive sable]

but 4090 are more than double the price of 3090 used

so not worth it

[hard delta]

there's also the fact that a used 3090 might have been running for 3-5 years now

[olive sable]

unless you can find a good deal you're often getting better value with 3090s

[stray dragon]

yeah maybe if you wanted something purely for model inference where response speed is a concern, 4090 or 5090 would be worth

for training though? 3090

[rough bloom]

ye exactly
something like 1.5x - 2x the performance but also double the price
except you still get the same amount of VRAM

[stray dragon]

you won't be training in 8bit precision probably

[fierce light]

buying into 30xx now is insane

[hard delta]

5090 4090 is probably out the budget window tbh

[olive sable]

not really

[stray dragon]

no

[fierce light]

if you have them, keep them

but don't buy more

[nocturne olive]

Nah inference 3090 is within 10% of 4090

[hard delta]

probably floating around an 5070ti - 4080 - 3090 line
obviously the 3090 wins with the vram, but there's the nasty used surprise, architectural changes, etc factors

[stray dragon]

depends on quantization, no?

[fierce light]

inference isn't everything

[nocturne olive]

No it's all bandwidth

Inference is bandwidth-bottlenecked, not compute

[rough bloom]

is there a better relatively cheap card with good VRAM per dollar yet?

[olive sable]

its all about pricing.
sure 4090 is better, but the 5090 is beterer.
if the 3090 is a way better deal tho then get a 3090

[stray dragon]

i see

[nocturne olive]

I'd buy more 3090s, VRAM + NVLink is good

[rough bloom]

(from NVIDIA, AMD and Intel have decent offerings but they're not exactly comparable)

[olive sable]

ye

[fierce light]

you are buying a retro vehicle before a bubble burst that will launch server grade cards onto used market

[olive sable]

"retro"

lmao

2020 is retro now

[nocturne olive]

3090 is plenty relevant

[fierce light]

not for long

[nocturne olive]

I got one recently and it's been so amazing for NeuroSynth

[olive sable]

you gotta be trolling

[nocturne olive]

And what's gonna magically replace it?

[fierce light]

30XX already has deprecations >right now< and it's only get worse

[olive sable]

and what ewxaclty is deprecating it?

[fierce light]

since you'll get all the H100s

[nocturne olive]

Nah H100s aren't gonna become affordable

[fierce light]

or even lesser cards

[rough bloom]

when will that happen though? can you reliably plan around it?

[fierce light]

the A series i think

[nocturne olive]

A100? Not gonna become that affordable either

[fierce light]

AI is gonna collapse due to vibecoding and inbred models

[stray dragon]

those need a custom motherboard solution, no?

[fierce light]

cloud will die

there are PCIe version

i am almost 98% certain many of the datacenter cards can run pcie

[olive sable]

to keep with your analogy of cars. you're basicly saying "why buy a used chevy for a great price when this porsche might maybe potentially get a sale in 7 years"

[fierce light]

the HBM or whatever is another class

HBX?

uh

[nocturne olive]

SXM?

[fierce light]

i think there are multiple

[olive sable]

HNM is high bandwith memory

thats ram

[nocturne olive]

HBM is VRAM yeah

[rough bloom]

not really, just PCIe and SXM for NVIDIA cards AFAIK

AMD and Intel use OAM

[nocturne olive]

But do you really think A100, even the 40GB version, will magically become the price of a 3090?

[fierce light]

amd and intel do not matter

[rough bloom]

indeed

[fierce light]

i am not talking about price of a 3090

i am talking general price drop

[nocturne olive]

Sure if that happened I would buy it, I'd like the VRAM, but that's just not gonna happen

[fierce light]

that will net better it/s

per money

[nocturne olive]

A small price drop won't affect me much

[fierce light]

small

[olive sable]

so you're basicly saying the 3090 is irrelevant because more expensive cards exist?

[nocturne olive]

False, the A100 has similar compute as 3090, just with more VRAM

[fierce light]

2008 will look like a joke

[olive sable]

you're missing the point tho

[hard delta]

all that said, i'm still just buying a PC for home that i'd be using for gaming too, not some server on the side purely for training
might look out for a good 3090 if it comes along but otherwise eh, probably should stick to new enough
i'm just worried that any used 4090s out there have been melted to bits by overuse too xD

[fierce light]

no i am not

the point is still you will be able to get overall better price to performance ratio

3090 will only be relevant for people who already own them

but buying them now?

nah

[olive sable]

2008 is quite litteraly 18 years ago. ofcourse those gpus arent on par with current gpus.
but the 3090 is 5 years old, and still doing great
the best price/performance ratio i know of in this space

[nocturne olive]

I would get an A100 if it dropped to a reasonable price, but if it's not gonna drop by a massive amount and stay in the thousands I'd much rather have a modded 48GB 4090

[fierce light]

l'd rather splurge on the 5070TiS

[olive sable]

thats L take

[nocturne olive]

More compute and VRAM compared to the A100 40GB and 3090

[fierce light]

wakarimasen lol

[nocturne olive]

Ew so few VRAM

[fierce light]

also the ability to load better models

[nocturne olive]

Useless card

[fierce light]

like 3090 these days?

[stray dragon]

"load better models" less vram loads worse models

[nocturne olive]

Yeah

[fierce light]

there's 50gig models left and right

[nocturne olive]

3090 is still super relevant

[stray dragon]

3090 will continue to stay relevant

[fierce light]

and if you run goofs of big models thats all 70+

[nocturne olive]

24GB can load 30B param models easily

[olive sable]

3090- 500 euro - 24gb vram
5070ti - 900 euro - 16 gb vram

i could get nearly 2 3090s for that price and have triple the vram

[nocturne olive]

Two of them can load 70B param models

[rough bloom]

small issue
this is historically true all the time
cards only really get cheaper in the long term, but you gotta buy at some point
who cares if it's cheaper in the far future, I can't buy at those future prices right now

[stray dragon]

gpu prices right now aren't super high, no reason to hold out

[rough bloom]

the bubble will pop and cards will get cheaper, but nobody really knows when exactly that will happen
so can't plan around it

[hard delta]

that's the source of the dilemma btw - i expect prices to stop dropping or even raise thanks to the ram shortage

which is why it's a should i buy now situation

[olive sable]

if the 3090 cant fit it the 5070ti DEF cant fit it. nevermind that 3090 has nvlink and the 5070ti doesnt

[fierce light]

the ram memes are temporaray

[nocturne olive]

Yeah I'd take the 3090s any day
Basically right now the only thing I'd take over a 3090 is a 48GB 4090D

[fierce light]

china will 100% reroute capacity to steal teh market

[rough bloom]

yeah, obviously it'll get cheaper again eventually
the question is when

[nocturne olive]

The 3090 is and will stay relevant as long as there's no better 24GB+ good compute card on the market

And in a similar price range too

The 3090 is such good price-VRAM-compute balance

[olive sable]

the high ram prices are 100% here till at least 2027 and im willing to bet 50 bucks on it

asus might be planning on producing ram now, but that production def wont be anytime soon

[rough bloom]

I don't think there's a good answer to this, so might as well just purchase the current best options as required instead of waiting for something that may only happen in 5-10 years

[mighty thorn]

They also aren’t producing new DRAM (the thing we need more of)

[nocturne olive]

24GB of VRAM with speeds almost as good as the 4090, compute equivalent to an A100 or 4070 super, and prices as low as 500€, you're not gonna replace that with a couple thousand euro enterprise cards

[hard delta]

i saw some headlines saying that was a hoax

[nocturne olive]

Not to forget that a 3090 can render, unlike an A100

Well, as in render games

[mighty thorn]

They are just also now purchasing said DRAM to further spread them thinner and mark them up for more money

[nocturne olive]

Output display

All that stuff

[fierce light]

you can have one card for inference and for gaems

and with every card running on 70% powerlimit it won't even be such a killer

[nocturne olive]

Sure, but the best option for that inference card will be a 3090 for the average person

[olive sable]

the only way the 3090 will lose relevance is when the 4090 becomes a better deal compared to the 3090. which wont happen for a while since its still more than double and it would need to drop to 1.5x the price of the 3090

[mighty thorn]

(never)

[fierce light]

the average person should secure as much as iron aspossible

[mighty thorn]

Also chat

[olive sable]

in 7 years at most it should

[nocturne olive]

The average person doesn't have 4K to spend on an A100 even if it was to drop in price

My budget is highly limited and thus the 3090 is basically the only option

[stray dragon]

big heatsink

[olive sable]

iron?
why iron?

is this another conspiracy?

[hard delta]

iron means hardware

[nocturne olive]

Yeah like the 4090 is only really better for training and gaming

[olive sable]

iron is like, the most abundant metal on the whole entire planet

[nocturne olive]

It doesn't have more VRAM and the VRAM is barely even faster

[rough bloom]

give up on computers
trade iron instead

[olive sable]

[hard delta]

train iron

become a blacksmith who hand crafts trained AI

by hammering iron

until the weights are nice and balanced

[nocturne olive]

For me, it's either get more 3090s, or if I magically find money, 4090D with 48GB mod

[olive sable]

for me

i dont really need more gpu

[nocturne olive]

Well you aren't developing an ML vocal synth

[olive sable]

im not developing ai stuff

yet

[nocturne olive]

By the way if you magically find a really insane deal for 4090D 48GB somehow let me know because I'd like one of those

[olive sable]

48gb model will never reach the belgian used market

[nocturne olive]

Unfortunate

Guess if I want one it's 2400€ or something from China not including shipping and stuff

I don't have that kinda money though soo 3090s it will be for now

[true hemlock]

the fp8 and fp4 don't even work lol

its just fp16, its faster because its a scaled block that gets rescaled back during computation just to be more bandwidth efficient

[true hemlock]

scales with bandwidth and depending on your configuration for batch, compute aswell mostly

[rough bloom]

ye, bandwidth is about the same but you have way more compute

[quick condor]

only so much though

[true hemlock]

false, A100 has similar compute as 4090 (tensor)

[nocturne olive]

Oh? Then where did I hear the 3090 equivalence from?

[stray dragon]

damn so it's just bullshit marketing then

[true hemlock]

A100 can render, just barely can output because lack of rops

[stray dragon]

thanks nshittia

[nocturne olive]

Scamvidia

[rough bloom]

nah they do work
bandwidth and cache efficiency matter quite a lot

[true hemlock]

always have been. actual fp8 fp4 inference is so losey i just considered them obsolete

quantization works better

but even Q4 is still computed with fp16 afaik

[rough bloom]

but indeed not native fp4 or fp8 computation AFAIK
pretty sure that tensor core accumulators are (mostly) fp32 anyway

[true hemlock]

yeah which is why i said quantization and scaled blocks

[quick condor]

I usually am using exl3 for runtime

[olive sable]

idk if it counts as advertising so i wont link it. but i saw a video if rin penrose playing a programming game and it was painful

" no you need to use double = for if statements"

[hard delta]

so like

if (thing = (double) 5) {...}

[nocturne olive]

And then apparently on RVC there's even a noticeable difference in quality when using FP32

[olive sable]

needs to be python syntax btw

[nocturne olive]

This is according to Wispers who was accidentally using FP32 and couldn't figure out why his models always sounded better than ones from CJ

[quick condor]

I mean Ive been working with a groop testing some things that has found even keeping lm head as f16/bf16 vs f8 is a major change in quality

[umbral thorn]

bf16 better

[quick condor]

depends in some edge cases

[silent cloak]

bf16 my beloved

[quick condor]

Like I said, There are situations for both Generally though bf16 is better

[nocturne olive]

I actually wonder how BF16 vs FP16 vs FP32 goes in RVC, because currently FP32 is apparently noticeably better than FP16

[young plover]

That would make sense because 9 bits of precision is pretty terrible for audio.

[mighty thorn]

She does programming?

Insane lore drop

[olive sable]

mo

or she does

but really badly

he*

rin is a prince or something iirc

[amber fractal]

British moment

[olive sable]

https://tenor.com/view/azumanga-daioh-azumanga-osaka-ayumu-kusaga-ayumu-gif-20393239

goodnight

[hard delta]

looked a bit at what model sizes i could train on what card, and oh boy

i'm starting to think i should just give up on the more relevant ones, play around with training a smaller one to get the idea, maybe use bigger ones for inference at most, and otherwise wait out

idk if i'd be producing real, pays-off kind of value anyway, smh

[nocturne olive]

Yeah it takes insane hardware to train any kind of usable LLM
Vocal synths and such are way easier

[hard delta]

plus if i get idk a 5070ti and similar, i have a nice dev PC for 6-8 years
if in that time i ever feel the bottleneck from profit, i can just bump it up because it's for profit

[patent shard]

yeah, people train instrumental and vocal removal models with 8gb cards... those are easy
but can't say the same for all things...

[quick condor]

It depends Quanting and training require pretty high training requirements beyond 12B

[long hazel]

Ok, what the heck is going on with my Raspberry Pi? I don’t code by the way so apologies if I don’t understand something simple going on

I’m getting a line it keeps freezing on saying “End Kernel Panic - Not syncing: VFS: Unable to mount root fs on unknown block (179,7)”

[amber fractal]

[amber fractal]

Can't say much as I don't know much on the pis but good luck
May need to check in a few hours as a lot of chatters are sleeping

[long hazel]

Alrighty

I’m tempted to get a new card reader and try reformatting it again

[amber fractal]

Do that, some form of drive corruption did happen

[visual pawn]

how the hell is a rasp pi $100 now

[long hazel]

I think it depends on what one you’re buying

I have the 2

[mighty thorn]

Just copy my strat

Finetune gemma3-270m on your data locally

Then, if it shows promise, either train LoRA for a bigger model or scale up with cloud compute

[tender river]

why is there a koniifer on my screen

[sage crag]

sorry ill go

[tender river]

come back

[sage crag]

[tender river]

[sage crag]

bzz bzzz bz bzzz

[rigid snow]

[sage crag]

[covert bane]

anyone who knows go, can you help?

warning: `neuro-desktop` (bin "neuro-desktop") generated 3 warnings (run `cargo fix --bin "neuro-desktop"` to apply 1 suggestion)
    Finished `release` profile [optimized] target(s) in 1.90s
     Running `target\release\neuro-desktop.exe`
Neuro Desktop Application Starting...
Calling Go function...
Name: World
error: process didn't exit successfully: `target\release\neuro-desktop.exe` (exit code: 0xc000013a, STATUS_CONTROL_C_EXIT)

I'm stuck on why rust wouldn't call the go function, rn I'm just trying to find a way to get this to work before I implement the actual logic

[sage crag]

@tender river

[tender river]

meringue аскорбинка

[sage crag]

[rigid snow]

wha

t

[tender river]

[sage crag]

acid

[rigid snow]

rust go interop

cursed

[sage crag]

just use one or other

[rigid snow]

yee

[sage crag]

why use both

[rigid snow]

exactly

[covert bane]

idk

fun

[rigid snow]

why it cursed

[covert bane]

use case was that, the neuro integration code is in go. Rust just bundles it all into a standalone file, and python provides the actions that the integration uses

#1428980640895733820

[sage crag]

but

[tender river]

i assume you followed https://amirmalik.net/2023/02/15/embedding-go-in-rust ?

[sage crag]

there are rust bindings

[covert bane]

yes

[rigid snow]

go weird language why function to convert cstring to native called C.GoString

bad name

[sage crag]

[tender river]

not a solution to your problem but

https://docs.rs/neuro-sama

just felt like self advertising

[sage crag]

self promo

[covert bane]

I tried that

[sage crag]

[covert bane]

Rust was too complicated for me. So I just decided, let's use Go.

and make rust the parent

[rigid snow]

lmao?

[sage crag]

[tender river]

why use rust in the first place then

[rigid snow]

using crate too complicated so let me do interop instead

[sage crag]

rust too complicated lets do more complicated

[rigid snow]

yuh make sense

[sage crag]

[rigid snow]

yeyeye

[covert bane]

for some math algrothims, for the mouse control abstraction that neuro desktop will provide

[sage crag]

[rigid snow]

[tender river]

can do in go

[covert bane]

neuro just says what she wants to do, another AI (custom built one) will figure out the mouse movement and keyboard controls

[sage crag]

can do in python if you want

[covert bane]

another reason was... I just wanted to use multiple languages, for absolutely no reason, other than making neuro-desktop a multi-language project

[sage crag]

[covert bane]

cause I was bored while thinking of the new project structure, when I started refractoring neuro desktop

[covert bane]

well, would anyone know what is wrong?

rust was hanging so I ctrl+c it

idk why its doing that

[sage crag]

[rigid snow]

using multiple languages NEVER a good idea btw. interop just sucks period

[covert bane]

first time hearing that, didnt know interop sucked

oh! new idea! why not just call them as binaries instead of linking! :D

[sage crag]

[covert bane]

wdyt?

[sage crag]

[rigid snow]

like if your api is 3 functions it's fine
anything more than that and suddenly you have 5k loc of type marshalling

and other abstractions

[tender river]

protobuf grpc microservices apache cassandra cloud native

[sage crag]

protobuf

[covert bane]

I'll try this again, I guess. Any example implementation for it?

[sage crag]

use docs

[covert bane]

cause I don't wanna go thru the doc, to find the full example impl

[stray dragon]

hi mintcandy

[sage crag]

t

[stray dragon]

hi konii

[rigid snow]

hi t

[sage crag]

hi t

[stray dragon]

hi konii

[sage crag]

[tender river]

ye https://github.com/chayleaf/rust-neuro-sama-game-api/tree/master/test-game

[stray dragon]

wait not again

[sage crag]

i learned

from

last time

[covert bane]

ty

[tender river]

still read docs though there isnt that much stuff

[rigid snow]

i have rust+dart+java codebase

[tender river]

[rigid snow]

[tender river]

is it an android app

[sage crag]

no android uses only

kotlin

[rigid snow]

flutter not only android

[tender river]

true

android superbox

[sage crag]

[tender river]

java only android

famously doesnt run on

any other device

[rigid snow]

[sage crag]

upgraded from android sandbox

to android superbox

now phone is brick

[tender river]

ye only kotlin run

android require

kernel

[sage crag]

[tender river]

[sage crag]

rewrite kernel in kotlin

gc in kernel

ee

[tender river]

you can keep the kernel minimal by

making a jvm cpu

builtin gc

not my problem

[rigid snow]

still have gc pauses

i would think

[tender river]

dumb cpu vendor not use concurrent gc

[rigid snow]

me

garbage collection cotroller

gcc

[sage crag]

[tender river]

wasnt there a

[sage crag]

dentge

[tender river]

gcc java

https://gcc.gnu.org/wiki/GCJ

As of GCC 7, the GCC Java frontend and associated libjava runtime library have been removed from GCC

[opaque sigil]

Gone and forgotten

[sage crag]

Gone and forgotten

[tender river]

GCJ compiles Java source code to Java virtual machine (JVM) bytecode or to machine code for a number of CPU architectures

[sage crag]

[opaque sigil]

What's that other jvm called again

[tender river]

graal

[opaque sigil]

Yeah that

[rigid snow]

[sage crag]

that other jvm

[tender river]

the good version of graalvm is proprietary

afaik

[sage crag]

juice vm

[sage crag]

graalvm ee

proprietary require oracle license

[rigid snow]

https://cdn.discordapp.com/attachments/1072697081443131476/1409679696441901086/EvilLightbulbJuice.gif

juice vm

[sage crag]

[tender river]

[covert bane]

I will

||guess everything, and try it until it works||

[tender river]

[covert bane]

:D

[sage crag]

wrr wrr wrr wrr

[tender river]

https://cdn.discordapp.com/emojis/1455273822625333375.webp?size=128

[sage crag]

[tender river]

[sage crag]

[tender river]

that me

[sage crag]

me send chatgpt 20mb birt

ddos

[tender river]

@samaltman

tos

not use chatgpt for

bird

[sage crag]

20mb birt

[tender river]

[sage crag]

slop eee

[tender river]

it constantly finds new ways to be weird

[sage crag]

yee

semi serious

close enough

[tender river]

anecdotal but

korean yuri tends to be more drama heavy

[sage crag]

enub

[tender river]

not have statistics just gut feeling from what i read

[sage crag]

enub

[tender river]

theres less cute stuff and more fucked up dynamics

[sage crag]

did you turn into chatgpt

[tender river]

me always chatgpt

[sage crag]

[tender river]

leaf invented in

2022

before that not possible due to technical limitations

[sage crag]

[tender river]

new technology please be patient

[rigid snow]

cheese sommelier

[tender river]

other thing me notice is almost all manhwa in general fall into one of two categories

[sage crag]

man or hwa

[rigid snow]

fried egg with cheese taste like milk

[sage crag]

you

odd

[tender river]

shounen (edgy game mechanics where you become super rich/influential due to luck)
shoujo (otome game mechanics, sometimes edgy, where you become some prince's wife)

like genuinely almost all manhwa falls into one of those

there are some good works in both categories, but otome game ones are more readable on average because the edginess of shounen manhwa is just unbearable sometimes

[sage crag]

wrrr

[tender river]

the less dramatic korean yuri tends to be about OLs

[rigid snow]

[rpogram,ign

[sage crag]

[rigid snow]

i'm an awful typist

[sage crag]

[rigid snow]

i tried learning touch typing for two month every day

0 progress i still come back to my cursed way

[sage crag]

2 months per day

effective time management

[rigid snow]

every day for two months

[sage crag]

[tender river]

why touch type if can

not

[sage crag]

[rigid snow]

because slow otherwise

i'm throttled in speed by how i type

[tender river]

[sage crag]

me not use correct technique still 140wpm

[rigid snow]

konii

[tender river]

me around 100wpm not remember

[rigid snow]

i'm 35-40wpm

i'm saying

awful

[sage crag]

[rigid snow]

i mean awful

[sage crag]

if you think about it, anything over 60wpm is pretty high

thats a word every second

[tender river]

second long time

[sage crag]

300 wpm

5 word per second

[tender river]

[rigid snow]

friend requests by people with no mutual servers and friends

[tender river]

i dont accept

[sage crag]

[rigid snow]

why so many where are you coming from

[sage crag]

secret

[rigid snow]

tpyo filtered

[sage crag]

[rigid snow]

sadv tpyo

[sage crag]

[tender river]

ignore for 3 years

[rigid snow]

leaked

https://cdn.discordapp.com/attachments/1072697081443131476/1355617829939974144/ezgif-5267616c5eb024.gif

[sage crag]

how lamp

[tender river]

think quack got it

[sage crag]

[tender river]

yee

[sage crag]

long dog

[rigid snow]

[sage crag]

https://tenor.com/view/borzoi-snoopa-lord-foog-dog-funny-gif-20125052

[tender river]

https://tenor.com/view/pinkemon-gif-8810031288398944013

[rigid snow]

[sage crag]

https://tenor.com/view/be-happy-dog-cute-gif-15281532823757571186

[tender river]

https://tenor.com/view/borzoi-borzoi-dog-dog-dog-borzoi-furry-gif-118235196246393650

[sage crag]

ant eater dog

[tender river]

they

slightly different

[sage crag]

https://tenor.com/view/borzoi-dog-water-hydrate-hydrated-gif-2063020374237287172

long nose

ant eater

[tender river]

slop

[sage crag]

crow

[tender river]

anteater nose twice as long as borzoi apparently

[sage crag]

nn

nnn

no

look

same

[tender river]

quiz time

https://tenor.com/view/tamandua-anteater-not-a-peta-gif-26465583

borzoi or anteater

[rigid snow]

[fierce light]

I HATE FRUTIGER AERO

[rigid snow]

whatever you say

[tender river]

me eat sour meringue with sweet honey cupcake taste great

[rigid snow]

meringue сакорбинка

аск

tpyo russia

n

[tender river]

ye just without the

ascorbic acid

[zenith scarab]

have any of you guys worked on a relay pc before

[rigid snow]

do you mean relay computer

[zenith scarab]

i kinda wanna redo my old project a bit better

[zenith scarab]

yes

[rigid snow]

sounds cool

[zenith scarab]

my initial design had a braindead design flaw which ended my motivation immediately

i had done it all on one PCB so it became obscenely large in design

[young plover]

have Google Pixel 10
it might as well not have a GPU

[rigid snow]

[zenith scarab]

wanted to know if anybody had experience with card edge connectors and their holding power

assuming this is a technology channel and not just a it channel

😅

[tender river]

me reveal answer after poll end

[tender river]

its neither

[rigid snow]

there is a correct answer??>?

[tender river]

ye

[zenith scarab]

ah okay

[tender river]

this channel only for baking

but people go off topic often

[zenith scarab]

i wanted something akin to PCIE connectors but wasn't quite sure how much weight they can hold down so didn't pick them (yet)

the design i had in mind was a one that had each section of the pc on its own little PCB that I then connected into the main board

it would make the design larger but whatever

making it actually modular is something i can think about in version 3

[young plover]

Well I don't have a support bracket for my 5090 Ventus which is... heavy
gonna look that weight up

1.9 kg 4.1 lbs

[zenith scarab]

it'd be 8 bit with the instruction set of:

• ADD/SUB
• NAND
• NOR
• XOR
• NOT
• MOVE
• JUMP

so as to keep it simple

[zenith scarab]

that is a fair bit

[white flower]

Actually I've been wondering about this too recently, just thinking about how much a GPU PCIE connector could take. I guess one could look up the mechanical moduli of FR-4, and then you could use that to determine how much max bending force (given a max allowed bending moment) it can handle for your geometry

[young plover]

It's a block of radiator almost the size of my forearm lol

[zenith scarab]

true that. I had initially considered laying it flat or making a little support extension onto the design of each PCB that blocks rotation pulling outwards

[young plover]

Of course it is screwed to the case at the back, dunno if you were thinking of a free-standing connector

[zenith scarab]

sideways bending like you'd see in a gpu was more worrying to me so i had considered putting the connector vertical instead

when the whole thing is laying flat on the ground, it should be able to handle sideways movements fairly easily

it's only really when standing it up that i have to think about it

[patent shard]

That's light
My 4090 is 2.4kg

[young plover]

💀 4 slots?

[patent shard]

Yeah

[young plover]

4090s were massively overbuilt expecting 5090 power draw.

And even the 5090 doesn't need that.

[patent shard]

Yeah
I also undervolt it, idk why I got such a big one
I mean, I do like low noise.. but it's undervolted, soooo

It's also apparently jammed into the PCIe slot
So that'll be fun to remove whenever I need to

[zenith scarab]

i know PCIe slots are not meant to deliver power but wow i did not expect it to be so hard to find parts with documented current ratings

[young plover]

Is it not 6A@12V?

[fiery fossil]

anyone else seen this? https://github.com/Tencent-Hunyuan/HY-Motion-1.0

[rigid snow]

that timing

weights dropped today woa

[fiery fossil]

no kidding lol

[covert bane]

So I ended up using Go again, and using IPC to communicate with Rust

:D

[tender river]

look like answer self explanatory

everyone answer correctly

[fierce light]

man all the 3d models are so tiny

trellis is also very small

[tender river]

3d models

[zenith scarab]

that might not be enough depending on the size of the circuits i go with

or i'll have to go with relays that arent as loud

the relays i had in mind take 12V@~100mA

i could also just connect multiple pins together

i wanted bigger relays for their satisfying clicking noise, granted they are also more expensive and take more current

[tender river]

you make dubstep

[rigid snow]

WHAT

GETOUR

GET OUT

???

you'rer fdreaking me out

[tender river]

[rigid snow]

you're unironically in my walls

are you listening

[tender river]

ye

[rigid snow]

no itr's just

i turned on on eof my song

to listen back

on speaker

you send that in 30s

you're actually scaring me

yeah no

fuck tythis

[tender river]

me scary

https://media.discordapp.net/stickers/1432138279679033354.png?size=160&name=YouShouldBeFrightened

eclair time

[jagged turtle]

chayleaf in mlntcandy's house now

[tender river]

will be yours soon

[jagged turtle]

what

[tender river]

[jagged turtle]

I don't see anyone outside right now

[tender river]

why would you see me

[jagged turtle]

[tender river]

eclair cream done

eclair next

i ran out of vanilla and almost ran out of sugar but i dont need any more of it luckily