programming

[fast pagoda]

oh yeah rip

sama believes it's the best performing "open" model in the world tho dont worry

[uneven pulsar]

i don’t know why and i don’t want to know why but i ain’t touching gui again

[fast pagoda]

just generate some fligma and feed it to copilot dawg that's what the pros do these days according to them

[uneven pulsar]

you do have a point

[noble zodiac]

@faint sandal is there any validity to this HTTP/1.1 desync nonsense I've been annoyed with for weeks or is PR for more funding?

[faint sandal]

what

context hat

[noble zodiac]

Some def con talk about a HTTP smuggling vuln in some CDN provider but its worded as if HTTP/1.1 itself has a critical fault which I doubt

https://http1mustdie.com/

I hate this trend of researches clickbaiting and making a huge fuss about some relatively minor stuff

[rigid snow]

you need react in your life

[noble zodiac]

dont get me wrong, I get why they have to do it but I hate it still

[rigid snow]

(if you’re not doing immediate mode)

[faint sandal]

i'll be honest i don't know a lot about it

and i'm more inclined to believe portswigger and defcon + black hat USA reviewboard

the reviewboards for both are notoriously difficult to get past if your talk isn't of enough substance and strong evidence

i'd know because i've been rejected thrice

[noble zodiac]

I believe them that they got access to a major CDN but I think its something very specific to them and not an inherit problem of the protocol

[rough bloom]

after skimming their previous article on HTTP desyncs it looks like it's basically parser differentials except for HTTP and it's not always parsing

so maybe they found some issue with the HTTP spec that exacerbates this

[small anvil]

had a bit of trouble getting deserialization to work, apparently i was attempting to access buffers that were destroyed out of scope

[amber fractal]

OpenAI surely should have realized that their model quite literally sucks compared to any model released in the last year right?

[glass flower]

surely they know

[rough bloom]

wdym, the benchmarks look great!

[amber fractal]

(Ofc this is about GPT-oss)

[glass flower]

atleast the speed for the 20b model on ollama is fine now. (still takes wayyy too much vram)

[fast pagoda]

modified my fstab i'm sure this reboot will go flawlessly

every time i do this i fuck it up somehow

[sage crag]

we all know synthetic benchmarks are great at measuring real world performance of models

[fast pagoda]

the benchmarks cannot capture the vibez properly

[faint sandal]

[fast pagoda]

welcome

https://tenor.com/view/welcome-to-jurassic-park-we-made-it-dinosaur-arrived-gif-16076910

[rough bloom]

it automatically opens on startup now

[faint sandal]

time to switch to vscodium

[fast pagoda]

that'll definitely get rid of it

[uneven pulsar]

like i am trying to hash map a dict key inside of for loop of a dict key so the api appends the user password key

[fast pagoda]

it builds fast

[uneven pulsar]

i sometimes forget my own code

[faint sandal]

i'm falling asleep writing reports

[fast pagoda]

rand some ultra random series of commands on my laptop that i just tossed fedora 42 on

to get nvidia drivers working right

and rebooted

that shit worked

that's crazy

it was some guy saying it worked for him on fedora 39

yolo just worked

[jagged turtle]

why is it so squished

[fast pagoda]

made it as smal as could

neeeeever really thought about it but cronjob syntax isnt particularly exciting when it's describing a full time in like a clock syntax

like 50 4 6 8 3 was a few seconds ago and that's just

numbers

[sharp sierra]

ihate luau and luau hates me.

[cobalt sigil]

bro decided to support chromeOS but not win11 smh smh

all kins

[opaque sigil]

[rigid snow]

Compatible with all kins of motherboard Dos of CPU.

[opaque sigil]

I really like the QC pass sticker

[opaque wharf]

Wake me when it's F day

[sage crag]

too many day

[opaque wharf]

Only 4 more days no?

No, even less

[sage crag]

human dont sleep 4 day

[opaque wharf]

I am not human

[sage crag]

[opaque wharf]

I am an uncomputable function

[sage crag]

[south needle]

finally my 11 TB (10.5 Years) of audio data is done processing now I can finally start training a model with it :)

[desert plaza]

i thought you were a computed function

[olive sable]

Gooood morning

[opaque wharf]

I am a changed man

[desert plaza]

altered function

[olive sable]

Today the plan is to watch the ado 1st live bluray and play poker

[tender river]

[opaque wharf]

Meanwhile my battery:
Charge Full Design - 4351000
Charge Full - 4795000

[south needle]

battery saver

[sage crag]

the first number is 64% of the second number

[south needle]

sounds resonable for battery saver

[sage crag]

no it doesnt

[south needle]

mine is at 59% and its not charging anymore

[sage crag]

psychotic battery saver

there are diminishing returns for reducing it

well, each to their own

note that you have different status though

Full vs Not Charging

[opaque wharf]

60-80% is the sweet spot actually

[sage crag]

surely it also depends on whether your battery is li-ion or li-po

[south needle]

you dont want lipo in your laptop

[sage crag]

no one wants li-po

in general

but its cheap!

[south needle]

and lightweight, making it perfect for drone bombs

and drones in general

[opaque wharf]

60% for prolonged use with charger

[opaque wharf]

Both should be on the same range

What is wrong with my internet

Can you guys actually read my message?

[desert plaza]

yes

unless i can't see it

[opaque wharf]

Yeah, I sent that minutes ago

[desert plaza]

https://tenor.com/view/wlan-gif-12418525811703301139

[sage crag]

and more durable

typically bad for long term power storage, and also has a worse lifetime under repeated cycles

bwa

[tender river]

[tender river]

isnt this bad for drone bombs

[desert plaza]

how come linux brk is different from libc brk

[tender river]

i think the brk syscall is deprecated

libc probably implements it using mmap

[desert plaza]

i don't think so

brk gets called when i compile with libc

i see it in strace

[tender river]

okay brk is not in posix

[desert plaza]

[tender river]

https://cdn.discordapp.com/attachments/1060090228615958528/1399864129921093784/slep.gif

[desert plaza]

well, linux has a brk syscall which differ in behavior from the brk and sbrk on man page

compiling this with gcc uses brk at the beginning and end of program execution ```c
#include<stdlib.h>

int main(void) {
char* c = malloc(64);
free(c);
return 0;
}

i was wondering how i would allocate memory with just linux syscalls, no libc

[tender river]

you would typically use mmap

set fd to -1 and pass MAP_ANONYMOUS

[desert plaza]

[tender river]

the issue is of course that you have to allocate something divisible by page size

but its easier than requesting a heap and manually allocating memory from that pool

[desert plaza]

also, i wanna be able to see all allocated memory

[tender river]

mm maybe /proc/self/maps?

[desert plaza]

ugh, why does linux use text files for that

[tender river]

its linux

[faint sandal]

the anti debug I've seen on Linux is so funny because of that

I was like why is it looking for the word "racer" in /proc/self/status

[desert plaza]

i wish i could find reliable documentation about how linux works

[tender river]

https://www.kernel.org/doc/html/latest/filesystems/proc.html

[desert plaza]

[tender river]

also a lot of parts of linux are undocumented so just read the source code

[desert plaza]

i don't wannaa

[tender river]

you dont have to read linux's source code

reading other people's code that achieves the goals you want works just as well

[desert plaza]

i don't like reading source code, why do you think i'm writing in machine code

[tender river]

[desert plaza]

https://tenor.com/view/pepew-gif-46103354907324314

i wish linux docs were like intel docs

[opaque wharf]

I bet intel also has a lot of undocumented feature

[desert plaza]

well, at least the features i need are documented

[sage crag]

strictly setting fd to -1 isnt required if you pass MAP_ANONYMOUS

[tender river]

true! but its recommended

[sage crag]

yeh

[opaque wharf]

Honesty, the mov is cursed since it is turing complete

[desert plaza]

aight, but what is the value of MAP_ANONYMOUS

[sage crag]

in reality mov is a lot of different instructions in a trench coat

[tender river]

hehehe

thats where you git clone linux

good luck

[desert plaza]

[sage crag]

0x20

[tender river]

mhm

[sage crag]

here's how it was implemented in an older version of lily

$sys_mmap := 0x9
$prot_read := 0x1
$prot_write := 0x2
$map_private := 0x2
$map_anonymous := 0x20
$map_failed: ^u8 = @bit_cast(~0)
$alloc := fn(len: uint): ?^u8 {
    ptr: ^u8 = @syscall(sys_mmap, 0, len, prot_read | prot_write, map_private | map_anonymous, ~0, 0)
    $if config.optimise < .ReleaseFast {
        if ptr == map_failed return null
    }
    return ptr
}

[tender river]

what the hell is this i'm scared but i don't have the time to look into it

i found some msrs but this doesnt seem to be it

[faint sandal]

doing this in VMware suspends the machine

which yes, is a VMware thing than Intel but still

[desert plaza]

See Chapter 20,
“Input/Output,” in the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 1, for more infor-
mation on accessing I/O ports in the I/O address space.```

[faint sandal]

yuh it just writes to the io

[maiden geyser]

undocumented linux vs undocumented winapi

[sage crag]

at least you can read linux src code

i love port io

we use this all the time in osdev and its really fun and i love it

mhm

its really nice that x86_64 uses port io and both risc and arm use mmio

mhm

[desert plaza]

uh, what is this supposed to mean void addr[.length]

[sage crag]

context?

[desert plaza]

in ```c
void *mmap(void addr[.length], size_t length, int prot, int flags,
int fd, off_t offset);

on mmap man page

[sage crag]

modern c

[desert plaza]

?

[sage crag]

it implies to the compiler that the memory region at addr is at least length bytes

[desert plaza]

ah

[sage crag]

tricked you

im pretty sure its just mman

[desert plaza]

what

[sage crag]

its not actual c syntax

try compiling it

[desert plaza]

y'know, i really just need to know what number to put in ebx

[sage crag]

give me a minute

[desert plaza]

i can probably just put null

hm

[sage crag]

0 specifically yes

[opaque sigil]

i got excited seeing cuda 13 existing but cuda-gdb is still broken on wsl2

[sage crag]

@syscall(sys_mmap, 0, len, prot_read | prot_write, map_private | map_anonymous, ~0, 0)

[sage crag]

refer to A

tis correct for allocating memory

[desert plaza]

aight

[opaque wharf]

It is a sign from god Almighty to finally move to linux

[sage crag]

nix on hardware

[opaque sigil]

getting there

[sage crag]

1-phenyldodecane

[opaque sigil]

Also another glibc release, another patch needed for the cuda toolkit

[sage crag]

i miss static linking

[opaque wharf]

You wanted to have what happens now at rust?

[opaque sigil]

Honestly yes

[opaque wharf]

I think having something like windows .dll is nicer

[sage crag]

have you seen how egregiously rust fills its binaries with random stdlib functions that are never called

[opaque wharf]

I never bother to check

[sage crag]

all hblang programs have been statically linked and even the ones with hashmaps and formatters are smaller than glibc hello world

even though glibc hello world is dynamically linked

[sage crag]

this is the smallest binary i could make on demand

i rebuilt the standard library for this

# .cargo/config.toml
[build]
rustflags = ["-Zlocation-detail=none", "-Zfmt-debug=none"]

[unstable]
build-std=["std","panic_abort"]
build-std-features=["optimize_for_size","panic_immediate_abort"]

# Cargo.toml
[profile.release]
strip = true
opt-level = "z"
lto = true
codegen-units = 1
panic = "abort"

[tender river]

dynamic linking is very good for packagers

if theres a security issue in a dependency you can just swap it out for a patched version

without even recompiling anything

[sage crag]

yes

but its not great for portability

[tender river]

it requires an abstraction boundary and abstraction boundaries are all about indirection and late binding

and if you have generics you essentially have to make the dynamic linker bundle an entire compiler

or use boxed objects everywhere

not ideal

[sage crag]

its a bit crazy btw

[desert plaza]

when compiled with gcc, this program makes 28 syscalls btw ```c
int main(void) {
return 0;
}

[sage crag]

surely not

why

bwaa

[desert plaza]

sending screenshot

gimme a bec

[sage crag]

is this -Oz

[desert plaza]

no

[tender river]

it makes sense because it calls the dynamic linker

and the dynamic linker has to do a lot

[sage crag]

try statically linking

[tender river]

musl

[sage crag]

and with musl

[sage crag]

you can statically link glibc it just makes the binaries a megabyte large

[tender river]

[sage crag]

probably less with flto

[tender river]

i have not

[sage crag]

probably much less with the tuned for size options i use in the lily build script

[sage crag]

interesting

[sage crag]

works on my machine™

[sage crag]

does indeed read like dynamic linker

[tender river]

/lib64/ld-linux-x86-64.so.2

[sage crag]

i wonder why they use brk and mmap

brk is surely outdated

[tender river]

mmap is obvious, its for mapping the file into memory at the correct addresses

brk is less obvious but probably used internally for allocation

since the dynamic linker has to keep some intermediate data

[sage crag]

but they can use mmap for that

perhaps its a microoptimisation or for portability

[tender river]

its a program written in the stone age

[sage crag]

seems more likely

well, brk here does look like a faster call

[desert plaza]

brk is called at the beginning and end of the program

[sage crag]

though they are probably doing different things, so cant draw any conclusion

[sage crag]

what

beginning sure

what do they need brk for at the end

[tender river]

fini

[desert plaza]

wait nvm

[sage crag]

fini should just be a handoff to libc fini surely

[desert plaza]

i'm thinking of a different program

[sage crag]

ideally they would have loaded that at the beginning of the program

[tender river]

this is the entire file

[sage crag]

who is we and why do they know that the __curbrk is not exported from ld.so

[tender river]

looks like glibc falls back to mmap

[sage crag]

why do they bother with brk then

[noble zodiac]

40 years of technical debt does things to a codebase

[sage crag]

ill bet that it was written without mmap

[sage crag]

then someone made a bug and a patch got submitted to fall back on mmap

but they didnt replace the whole thing to avoid breaking userspace

[maiden geyser]

i can read some book in swahili, doesn't mean i will understand it

[tender river]

they are committing war crimes in there

[sage crag]

it makes me feel less bad about my code

[tender river]

its legacy but its the kind of legacy code that somehow works and is potentially better than alternatives just heaps more complicated

[desert plaza]

[noble zodiac]

You underestimate how much of very elemental mechanisms are just guess work. It should behave in a certain way but someone somewhere in the last 50 years did it differently and you have to drag that shit along

[desert plaza]

i don't think i know how to use nixos

[sage crag]

just seems to be my machine has a funny quirk

rather than yours

[tender river]

one reason brk might be better is it requests a fixed page thats not likely to be clobbered by the program itself in the address range in elf header

[tender river]

nix-shell -p gcc glibc.static
allegedly

[sage crag]

great binary

i get that modern compilers add NOPs for alignment purposes, but it peeves me

1e82:       0f 1f 00                nopl   (%rax)
    1e85:       66 66 2e 0f 1f 84 00    data16 cs nopw 0x0(%rax,%rax,1)

:ReallyMad:

[tender river]

90 90 90 90 90 90 90 90

[sage crag]

bwaleaf

[tender river]

today i'll be installing windows on a jet engine

[sage crag]

why it get larger when i build for musl

bwa

[tender river]

musl itself is smaller but individual functions may be larger

[sage crag]

i dont trust it

bwa

30040 bytes

evil

2424:       48 89 05 1d 4f 00 00    mov    %rax,0x4f1d(%rip)        # 7348 <_ZN3std3sys3pal4unix14stac
k_overflow3imp13MAIN_ALTSTACK17hbfc09eed90bace91E.0>

:mhm:

mangled functions

[tender river]

itanium

ther should be a command line flag to demangle it maybe

[sage crag]

i saw a movups in there

[tender river]

if not you can use https://cxxfilt.com/

[sage crag]

im not wasting my time analysing a rust binary

it looks like sadness

i just saw movzwl

and movups and movaps

cmove

[tender river]

the compiler paid for the entire instruction set, it's going to use the entire instruction set

[sage crag]

nopw

if you didnt have enough fun with nop, here is nop for word

excuse me

[noble zodiac]

nop is the most relatable instruction

[desert plaza]

90    # xchg eax, eax
90    # xchg eax, eax
90    # xchg eax, eax
90    # xchg eax, eax
90    # xchg eax, eax
90    # xchg eax, eax
90    # xchg eax, eax
90    # xchg eax, eax

[sage crag]

2c4b:       f0 0f b1 0d c5 46 00    lock cmpxchg %ecx,0x46c5(%rip)        # 7318 <_ZN3std3sys3pal4unix14stack_overflow11thread_info4LOCK17hd28eda68fd74036cE>

why are we doing atomics and threads in hello world

stop it

threadinfo

bwa

2cfa:       f0 48 0f b1 0d 0d 46    lock cmpxchg %rcx,0x460d(%rip)        # 7310 <_ZN3std6thread8ThreadId3new7COUNTER17ha657e37bd8dd76c1E>

[tender river]

but what if the hello world stack overflows

[sage crag]

its ok

the binary is too large and too filled with ud2 for anything to happen

i give hblang's codegen too much flak

zig made the right choice abandoning clang

i dont think this is clang's fault though to be fair

probably rust's llvm IR generation

[tender river]

i think long prologues and epilogues are usually various exception/unwind handlers?

plus a bit of boring initialization like parsing the env vars and command line

[sage crag]

why are we locking stdout so much i only printed hello wlrd

[tender river]

in case of rust it also ensures you can println from multiple threads which requires atomics

[sage crag]

i cant really blame rust for handling multithreading behaviour by default in println

but also bwa

[sage crag]

this definitely still more complex than it needs to be

[tender river]

agreed its too branch heavy

[sage crag]

i literally turned off all panic handlers and formatters

[tender river]

it should be a simple lock + loop + unlock

[sage crag]

just lock once and write the string you doofus compiler

[tender river]

can you try manually calling std::io::stdout().write_all("a") or whatever it is

[sage crag]

5s

[noble zodiac]

I assume they have some good reasons for it

[tender river]

i forgot but also print impls in all languages buffer input

[sage crag]

looks like loop unrolling gone wild

[tender river]

to avoid too may syscalls

so its buffering in addition to locking in addition to inlining i guess

[desert plaza]

ah, only 14 syscalls now

great

[sage crag]

use std::io::Write;

fn main() {
    _ = std::io::stdout().write_all(b"Hello, World!");
}

[tender river]

reasonable implementation of hello world

https://cdn.discordapp.com/emojis/1126222727422955601.gif?size=16&quality=lossless

[sage crag]

insert each character as a leaf of a btree

pop each leaf

genius

binary went from 19000 to 28000

why

i should just write a #![no_std] #![no_main] binary and define _start myself

i guess its unsurprising the println macro is going to generate smaller code than writing directly to stdout

rust's io has a lot of machinery

[tender river]

use std::io::Write;

pub fn main() {
    std::mem::forget(std::io::stdout().write_all(b"Hello, World!"));
}

this is better

[sage crag]

why in tf is it doing btree stuff though

[sage crag]

yes it went from 27840 to 27760

nice

looks like we are still btreeing it up

[tender river]

this is actually part of rust's memory safety guarantees

your software will be more secure if reverse engineers will cry looking at the assembly

[simple goblet]

hey if you guys dont mind me asking what is this mostly for?

[simple goblet]

i know that but what are guys talking abour rn?

[sage crag]

binary warcrimes

[tender river]

about low level code and compilers

[simple goblet]

oh ok

[simple goblet]

so i guest this how you say hello world in rust?

[sage crag]

om

no

[tender river]

you can write it in an easier way, we're just trying to minimize binary size

[sage crag]

println!("Hello, World!");

[desert plaza]

guys, this executable is 734880 bytes how to make smaller

[sage crag]

musl-gcc

[rigid snow]

what is !

[simple goblet]

why were you doing that code then?

[noble zodiac]

its a macro

[sage crag]

macro

specifically macro rules

[rigid snow]

why does printing need a macro

[sage crag]

println!("{}", 1)

arbitrary formatting args

[tender river]

implementing a type safe printf in the language itself requires dependent types

[rigid snow]

you guys are not real istg

[tender river]

idris does it good job idris

[sage crag]

im not real im konii

[noble zodiac]

format macros are a nightmare

[tender river]

i am definite and real

[sage crag]

me when im gumi vocaloid and i sing ECHO by Crusher

[rigid snow]

you are llms

all of you

[sage crag]

im small osu model

som

[simple goblet]

we are all just ai

[opaque wharf]

Man I need to sleep

[rigid snow]

how does an osu model have language capabilities emerge

[opaque wharf]

See you guys in 4-6 hour

[sage crag]

you mean you dont want format!("{time:.1?}"); in your codebase

[sage crag]

ermermgent properties

[tender river]

➡️ ⬅️

[noble zodiac]

no, I mean that I dont want to implement it

[sage crag]

no one wants to implement macros

macros are vile to write

[tender river]

and to use

[sage crag]

fun to use except when you need to debug their generated code

[noble zodiac]

(except in lisp)

but lisp macros are just better

[sage crag]

lisp macros are just lisp

[rigid snow]

what is a macro

[sage crag]

its when you press a button in excel and it deletes your spreadsheet

[tender river]

macros have similar problems to optimizers, they erase some info during translation

[sage crag]

[rigid snow]

that’s stupid, that’s what taking a hammer to a hard drive is for

[tender river]

which makes error reporting hard

[sage crag]

bubles

[simple goblet]

what is even gogin on rn

[sage crag]

if you arent following then you need to go do some leetcode

[rigid snow]

these bubles have a stupid mirrored hdri in them

immersion ruined literally unplayable

[sage crag]

what is a reflectiono

[simple goblet]

why do you post bubbles

[desert plaza]

nice, i can now return 0 with a 16KB statically linked executable

[sage crag]

why do you post why do oyu post bubbles

mhm

[sage crag]

you can get it smaller

give me a second

[desert plaza]

[desert plaza]

how come exit_group isn't showing up in strace with -c option though

[sage crag]

try
musl-gcc -Oz -static -o test.o test.c
clang -fuse-ld="gold" test.o test -static -Wl,--gc-sections,--build-id=none,--as-needed,--icf=all,-N -s
objcopy --strip-section-headers test

yeah

this will work first try

surely

[desert plaza]

surely

[tender river]

ereader update: suspend didnt work but i decided i'm fine with just turning it off when i dont use it

someone will magically decide to fix it some day

[sage crag]

the good ol "i cant fix it so i give up and wait"

[tender river]

well, it technically did work, but it stops receiving touch input

[desert plaza]

/nix/store/z34dss3rj1rnp1g820r64w0na7452jy1-binutils-2.41/bin/ld.gold: internal error in set_offset, at output.cc:4734
clang: error: linker command failed with exit code 1 (use -v to see invocation)

[sage crag]

erm

switch gold for lld

[desert plaza]

clang: error: invalid linker name in argument '-fuse-ld=lld'

[sage crag]

er

you dont have lld installed

ok just remove -fuse-ld=lld entirely

[sage crag]

its using ld

so icf=all doesnt exist

nice

gotta remove that one too

[desert plaza]

hm

mhm yeah

[sage crag]

nice

linkers

i love linkers

[desert plaza]

i love linkers

[sage crag]

thats why i have this all in the thingy

[hoary lion]

morn

[sage crag]

ing

[hoary lion]

[desert plaza]

it's crazy to me that a c program which only exits with status code 0 is at least 16KB

[tender river]

4 pages

[desert plaza]

4 pages

what

[sage crag]

memory

[desert plaza]

ok

[sage crag]

only if you use c runtime

also i had it at a much more modest 4kb when i did the testing myself

[tender river]

you can pass -nostdlib

but then you have to write a _start that calls the syscalls

[sage crag]

this is good advertisement for hblang

hblang return 0 is about 120 bytes

[tender river]

in my experience gold produced slightly fatter binaries because it put all segments on separate 4096 byte chunks

[desert plaza]

damn, that's a lot

[sage crag]

that's true when it respects alignment

[sage crag]

that's because you write handrolled machine code you doofus

[desert plaza]

yuh, it's best

[sage crag]

its silly

[desert plaza]

nuh uh

[tender river]

when doing advent of code i've had the most fun with prolog

[sage crag]

aoc in hblang

[tender river]

too easy

[sage crag]

probably

[tender river]

i want to learn scala it has cool concepts

but i dont have a use for apps running on jvm

maybe i should write erf in scala

[olive mango]

[tender river]

[olive mango]

what are we talking about:D?

[sage crag]

my

pc cant resolve any domains

i wonder if updating systemd took down resolvd

[tender river]

resolved

i dont use it i like using my router's dns as is

[sage crag]

do you use pihole

i dont

[tender river]

i use unbound

[sage crag]

i see

i use nextdns because im lazy

[tender river]

with a custom module for storing russian domains' ips in my routing table so they bypass the vpn

[sage crag]

[tender river]

i wish i could just scan client hellos and do it on a connection basis but i dont wanna write kernel modules

so this is the easiest way (even if cdns sometimes break it)

i dont do any dns filtering though

since all of my devices run nixos theres no ads to block

[tender river]

behold https://github.com/chayleaf/dotfiles/blob/master/system/hosts/router/default.nix

by the way you can run unbound on any device, doesnt have to be your router

at some point i ran it on my server instead

[desert plaza]

terminal keeps crashing, need to get off gnome

[tender river]

you dont have to

you can just install foot

and use it instead of gnome-terminal

[desert plaza]

oh

yeah

[tender river]

[desert plaza]

i could do that

[sage crag]

ghostty if youre trendy

[rigid snow]

caught

[tender river]

i use kitty

[desert plaza]

foot has been installed

[tender river]

huge

i might switch

[desert plaza]

now i just need hand

[rigid snow]

and you already have head

[cosmic sphinx]

@sick owl unsloth should be ok now, ima test what ig atm

[sick owl]

Yeah I've been testing it today and its done a bit better

They also fixed the recommended parameters

[desert plaza]

no, i have cpu

[rigid snow]

i was referring to head the binary that you have installed already

[cosmic sphinx]

where do they post them in github

[desert plaza]

oh

i do have that

how did you know

[cosmic sphinx]

it still noms my RAM so much damn

[rigid snow]

unix or somethign

[sick owl]

Its in the params file on the huggingface repo

[cosmic sphinx]

yuck

they released more quants so maybe I should try those instead

[sick owl]

It seems to me that something's still buggered because I just saw it make a typo

[rigid snow]

and part of posix iirc

[tender river]

oh ghostty uses gtk... i'm not sure whether i dislike the python dependency of kitty or the gtk dependency of ghostty more

[rigid snow]

gimp toolkit

[cosmic sphinx]

have you tested out the other quants

[tender river]

gnu's not unix image manipulation program toolkit

[sick owl]

I'm going to try the BF16 one to see if there's a difference

[rigid snow]

ghostty does not depend on gtk on macos, skill ecosystem issue

[tender river]

how does it work then

what

[sick owl]

Doesn't seem like there's much of a difference on first impressions

[rigid snow]

Ghostty is a cross-platform terminal emulator but we don't aim for a least-common-denominator experience. There is a large, shared core written in Zig but we do a lot of platform-native things:

• The macOS app is a true SwiftUI-based application with all the things you would expect such as real windowing, menu bars, a settings GUI, etc.
• macOS uses a true Metal renderer with CoreText for font discovery.
• The Linux app is built with GTK.

[cosmic sphinx]

i'll try running the Q8, shouldnt be that big of a diff

[desert plaza]

where is config

[tender river]

home-manager programs.foot

[rigid snow]

it's a fucking xcode project in their repo that depends on libghostty

[desert plaza]

home-manager?

[tender river]

https://github.com/nix-community/home-manager/

its ~/.config/foot/foot.ini, probably

[rough bloom]

did they always have this logo

[desert plaza]

it does not exist

[tender river]

because you didnt create it

https://codeberg.org/dnkl/foot/src/branch/master/foot.ini

[rigid snow]

they also appear to have an ios target???

from an issue about the ios build crashing:

iOS isn't an important target for us but I like to keep it working so that libghostty remains iOS compatible.

absolute psychos

[cosmic sphinx]

@sick owl is there a way to change the reasoning effort on the unsloth gpt-oss like on the og

[sick owl]

I presume you can just set the system prompt to Reasoning:high

[tender river]

okay since i dislike python and gtk dependencies equally kitty wins because of the name case closed

i might revisit it in the future

[rigid snow]

them supporting apple platforms and ios in particular is a big downside, correct choice

[cosmic sphinx]

[sick owl]

Do you have flash attention enabled?

[cosmic sphinx]

no

[sick owl]

That'll be why

Hang on lemme get you my launch command

[cosmic sphinx]

Experimental

[sick owl]

llama-server -hf unsloth/gpt-oss-20b-GGUF:F16 -c 131072 -fa --temp 1.00 --min-p 0.00 --top-k 0 --top-p 1.0 --repeat-last-n 0 --repeat-penalty 1.00 -ngl 99 --jinja --mlock --reasoning-format none

[sick owl]

Trust me, you're gonna want flash attention

Obviously you may want to edit context since you have less vram to play with

[cosmic sphinx]

its already at the reccomended 16k

going any lower makes this model absolutely useless for me

[sick owl]

You want more than 16k for these models

[cosmic sphinx]

[sick owl]

I'd say go for 65536 at least

It'll fit, trust me

[cosmic sphinx]

it'll fit, but seeing 98% on my 'memory' aint gonna make 🥶

even the q8 which should ideally fit on my gpu alone with the full memory, somehow still takes up RAM

[sick owl]

You shouldn't be using Q8 for this model

Use F16

[cosmic sphinx]

too large

[sick owl]

Also bear in mind flash attention massively reduces the size of the context cache

If you're running without it you're not seeing the true memory usage of the model under ideal circumstances

[warped narwhal]

Is there even enough memory left to store the response?

[sick owl]

You have 16GB VRAM

Its not too large

[cosmic sphinx]

[sick owl]

I thought you said you had a 4080?

[cosmic sphinx]

yes

[sick owl]

That's a 16GB VRAM card

[cosmic sphinx]

and im using 23/24 layers

[warped narwhal]

There is a 12gb version iirc

[sick owl]

You can offload every layer to it in F16/MXFP4

And at the very least fit a minimum of 32768 tokens

[desert plaza]

[cosmic sphinx]

how is it already full

the gpu mem

[sick owl]

Like I said, enable flash attention with -fa

[nocturne olive]

No, that became the 4070Ti

[cosmic sphinx]

i toggled it on

[sick owl]

And your context is at?

[nocturne olive]

What size of model even is this?

[tender river]

try XF86ScrollUp/XF86ScrollDown maybe?

[sick owl]

Its GPT OSS 20B

[nocturne olive]

What quant?

[sick owl]

But something is clearly wrong because I can fit that model in 18 gigs max of vram at F16 131075 tokens of context

[cosmic sphinx]

ok im not clocking whats wrong here

[sick owl]

He's running Q8 instead of the F16/MXFP4 version for some reason

[nocturne olive]

FP16 of a 20B model requires 40GB for weights

[cosmic sphinx]

thats a 4090

[nocturne olive]

Then it's not true FP16

[sick owl]

Its the model sizes I'm pointing out

[sick owl]

No its not, its mixed precision FP4

But the output and embedding layers are FP16

GPT OSS is weird

[nocturne olive]

Explains why there's some weirdness with the model size

But either way, I always go Q4 when trying a model

[cosmic sphinx]

its not a FP16 infact

[sick owl]

Yeah even at F16 the MoE layers are still FP4 here anyway

So I wouldn't quantize further personally

Its already plenty small and fast

[sick owl]

Like I said, not pure FP16

[desert plaza]

those are invalid too, weird part is i can ctrl+scroll to change font size when i'm not in nvim, but it gives me that error

[nocturne olive]

Weird models indeed
Well, luckily I have no reason to use them

[tender river]

well just leave it commented out

[desert plaza]

well, how do i space text away from window border

[desert plaza]

i'm feeling a lot of jank with this

[cosmic sphinx]

tldr 4080 sucks for local models

[sick owl]

Like I said with flash attention this should easily fit and run pretty damn fast

[tender river]

thats generally not something that terminals do

[desert plaza]

[sick owl]

I used to have a 4080 super with the exact same amount of vram so I'm more than familiar

[desert plaza]

text to close to border, need to fix

gnome console was better in that way

[sick owl]

Try just running it with this edited launch command @cosmic sphinx

llama-server -hf unsloth/gpt-oss-20b-GGUF:F16 -c 32768 -fa --temp 1.00 --min-p 0.00 --top-k 0 --top-p 1.0 --repeat-last-n 0 --repeat-penalty 1.00 -ngl 99 --jinja --mlock --reasoning-format none

[cosmic sphinx]

in lm studio?

[sick owl]

In llama.cpp

[cosmic sphinx]

[sick owl]

You can just unzip the latest github release then run the command from a batch file or command prompt window in the same folder

https://github.com/ggml-org/llama.cpp

Download the non cudarts CUDA 12 labelled zip file from releases

[tender river]

idk i just use

      main = {
        font = "Noto Sans Mono:size=16,Noto Sans Mono CJK JP:size=16,Symbols Nerd Font Mono:size=16";
        dpi-aware = false;

and it looks fine to me

but i only use foot on my phone

i use kitty on my laptop

kitty also has a very small margin

[desert plaza]

so, hypothetically

[tender river]

if you like wasted space you can try sway with gaps

[desert plaza]

could i write a terminal emulator in my programming style

[tender river]

you could, but not in asm

modern software is complicated

theres a reason we built a tower of abstractions on top of everything

[cosmic sphinx]

ion think this would be safe to run

[sick owl]

The absolute worst thing that could happen is you get an out of memory error and the program crashes

[cosmic sphinx]

or I run into another BSOD

[sick owl]

Or that if you're especially unlucky but that's not the end of the world

[desert plaza]

those pesky abstractions

surely i must be able to create a window in machine code

[sick owl]

Just remove mlock from the command if you're worried

[tender river]

heh

[desert plaza]

is that a challenge

[tender river]

no its a statement of fact

[desert plaza]

hmm

[sick owl]

WAUT

[tender river]

you could do a simple x terminal probably

[sick owl]

WAIT

WAIT

OH MY GOD

[tender river]

definitely not wayland

[sick owl]

@cosmic sphinx DO YOU HAVE A LAPTOP 4080

ITS ALL MAKING SENSE NOW

[cosmic sphinx]

no

[sick owl]

Oh nvm lol

[opaque sigil]

writing a terminal in assembly does sound like pure horror i can't lie

[sick owl]

I was gonna say the laptop one has 12gb vram

[cosmic sphinx]

blud it literally

[sick owl]

But yeah you're safe to run it, remove mlock if you're worried about the program or your PC crashing

That way if it doesn't fit it'll just spill into swap and you can close the program

[cosmic sphinx]

wouldnt I need to download the model AGAIN

ion have bf16 instaled

[desert plaza]

not assembly, machine code

[cosmic sphinx]

and it'd be easier to run speed experiments with the q8 that ig rn

[desert plaza]

writing hex representation of machine code and converting to binary

with xxd

[opaque sigil]

so you wanna be a human assembler?

[desert plaza]

i'm currently experimenting with mmap syscall in machine code

so i already kinda am

[sick owl]

Yes, but you'll probably get faster download speeds anyway

And at least this version won't exhibit weird behaviour to the same degree

llama.cpp is open source and much more reliable than lmstudio in my experience

And the BF16/F16 ggufs are gonna be way less buggered than the weird Q8 version you're currently on

[desert plaza]

time to put "terminal emulator hand-written in x86 machine language" onto my todo list

[sick owl]

Seems like the F16 on the huggingface link points to bf16 in the repo so I guess that's what you're getting

Shouldn't matter either way though

[desert plaza]

oh, and i'm also trying to make a "compiler" in machine code, to simplify the process of writing machine code

hence the mmap experimentation

need memory for compiler

[sick owl]

This is how much GPU memory that setup uses at 128k context @cosmic sphinx

Drop that down to 32k and I guarantee it'll comfortably fit

[cosmic sphinx]

can probably run

[sick owl]

Nothing ventured nothing gained

[cosmic sphinx]

even q8 might be overweight

[sick owl]

Start at 16384 like you were before if you're worried

I guarantee you won't run out of memory at F16 with 16k context using that launch command

llama-server -hf unsloth/gpt-oss-20b-GGUF:F16 -c 16384 -fa --temp 1.00 --min-p 0.00 --top-k 0 --top-p 1.0 --repeat-last-n 0 --repeat-penalty 1.00 -ngl 99 --jinja --mlock --reasoning-format none

Here you go, a 16k context version

[cosmic sphinx]

where is it gonna download the model

[sick owl]

C:\Users\%username%\AppData\Local\llama.cpp

[cosmic sphinx]

ion need that on C drive

[sick owl]

You can download the gguf file manually in that case

Just replace -hf unsloth blah blah blah with -m pathtofile.gguf

I'm running it that way

I recommend Bf16 if you download it manually

[cosmic sphinx]

gotta add llama-server to PATH holdon

[nocturne olive]

Or you could, run it in the terminal from the folder

[sick owl]

You can just save the command as a batch file in your llama cpp directory and skip that if you're just testing

Or run it as a command from command prompt with the folder mounted

Wot superbox said

cd C:\pathtollama.cpp

Then run it

Then once the models loaded you can access the interface by going to https://localhost:8080

[cosmic sphinx]

hold on lemme run into 220 more problems rq

[olive sable]

hello

i am here

[opaque sigil]

hello here

[rough bloom]

I was too slow

[sick owl]

Are you still running lmstudio?

[olive sable]

hello warm slice of bread

[sick owl]

If so close it, I think it also hosts on that port if I remember correctly

[cosmic sphinx]

pretty sure I shut it down just now and tried again

[sick owl]

You can use these commands to find the ID of the process using the port and kill it

netstat -aon | findstr ":8080"

taskkill /PID PROGRAMPROCESSIDHERE /F

[olive sable]

the viewing of the ado bluray with my friends has finished, so now i shall code

[sick owl]

The process ID will be a numeric string like this

[tender river]

first time i've used (physical) windows in years

15 minutes and it bsods

[sick owl]

Classic

[opaque sigil]

what are you doing

[tender river]

it crashed in kbdclass.sys and i was cleaning the keyboard of the laptop that hasnt been booted in years so i assume it couldnt handle the repeated keypresses

[opaque sigil]

damn

[uneven pulsar]

hi guys

[sick owl]

Hi I'm guys

[desert plaza]

physical windows

[olive sable]

do you need to install the hblang repo to your working directory or is just on the C drive fine?

[tender river]

nonvirtual idk what else to call it

[opaque sigil]

does hblang even run under windows

[tender river]

it doesnt

[olive sable]

what

[uneven pulsar]

no

[olive sable]

but you have a windows install guide

[uneven pulsar]

Windows so limiting

[opaque sigil]

wsl time for you

[tender river]

where

[olive sable]

wtf is this sabout then?>

[tender river]

the compiler itself runs on windows

since its just zig

but it cant target windows

[olive sable]

but you cant code it on windows?

oh my days

[tender river]

hi vani

[patent walrus]

(what's the point of including windows instructions atp)

[tender river]

it can still target hbvm

so its still potentially useful

[olive sable]

actually the git clone doesnt even work lmoa

[cosmic sphinx]

ok, i decided to run a first test on the downloaded q8 model with only 4k context

first when its loading up it also fills up my memory to the brim, but when its finished it actually clears up and fucking works

[tender river]

ntfs issue

[rough bloom]

Windows file name constraints

[cosmic sphinx]

i'll test it at 16k context next, if that also runs fine then i'll probs step up to a bigger quant

[opaque sigil]

that file name is crazy

[tender river]

autogenerated from md headers

[opaque sigil]

ah

[tender river]

committing it allows seeing how compiled code changes between compiler commits

though its not necessarily good for repo size

[opaque sigil]

unrelated question

[olive sable]

can i be bothered to use wsl? cuz idk

[noble zodiac]

bakushin

[rough bloom]

yes

[opaque sigil]

does nix copy the entire folder that the system flake lives at into the store

[cosmic sphinx]

:HOLY:

[tender river]

nix copy doesnt

but any nix command that uses flakes copies the flake dir into the local store

[opaque sigil]

even when there's a folder that is not referenced at all?

[desert plaza]

[olive sable]

can wsl use the windows zig or do i need to install a linux version?

[opaque sigil]

technically yes but install the linux version

[olive sable]

mf

[opaque sigil]

also do yourself a favour and try not to access windows files inside wsl

[rough bloom]

if it's a local folder without a Git repo then yeah, it'll always copy everything
if it has a repo then I think it should only copy the files added to the repo? but not 100% sure

[tender river]

reading time

[opaque sigil]

asking cause i have two folders in my dotfiles repo, one for the actual dotfiles and one for the nix config

thought about moving the flake files into the root to avoid having to specify the directory but if it's going to copy the other files then uhh yeah no

[rough bloom]

if they're large then yeah don't do that
especially if you're not using a CoW filesystem

[cosmic sphinx]

honestly, why didnt I think of directly downloading the files myself

lmstudio is such a netbloat for this

[opaque sigil]

only 135MB

[olive sable]

[rough bloom]

sir your Git repo is in OneDrive again

[olive sable]

as usual

[opaque sigil]

[olive sable]

is that the issue here?

[opaque sigil]

no

[olive sable]

well then im staying on onedrive

[rough bloom]

putting it on a WSL path certainly wouldn't hurt

[opaque sigil]

the issue is that your zig thinks you're using a linux file system

moving it into one such file system would probably fix it

[olive sable]

hmm

aight ok

[rough bloom]

sharing the compiler through OneDrive will suck anyway

you can put your project files on there but maybe not the compiler

[cosmic sphinx]

@sick owl so, unlike the Q8, the F16 completely gobbles up my memory at 16k context

[opaque sigil]

rule #1 of wsl2, do NOT access files on the windows host if you can avoid it

[olive sable]

i dont know how to open vscode in the wsl directory tho

[opaque sigil]

you can just open the one from windows with the wsl path

it'll open in windows and connect to wsl

[cosmic sphinx]

im not even sure why it uses up all the RAM when the VRAM isnt even full

and the speed is meh

[nocturne olive]

Windows likes to prevent entirely filling the VRAM

[cosmic sphinx]

yeah, lets instead fill up the RAM so it fucking crashes itself 👍

[nocturne olive]

Meanwhile me on Linux with a dual GPU setup with a total of 36GB of VRAM

-# And I don't even do LLMs

[rough bloom]

VS Code has an extension for WSL support

just need that and it should be fine

if it's somehow not fine and you're on Win 11 you should also be able to install VS Code inside of WSL but that's a bit inconvenient

[olive sable]

it was already installed

im in the ubuntu folfder now and still get the same issue

[rough bloom]

oh, I think I know what it is

did you clone with submodules

[olive sable]

i did exactly this

[rough bloom]

you forgor the submodules

do this
git submodule update --init --recursive

[olive sable]

idk what the submodules are

sure

its cloning shit

[cosmic sphinx]

if gpt-oss was this disappointing, ion even wanna talk about GPT-5..

[rough bloom]

there will probably be an error but we will fix it

[olive sable]

yep

[rough bloom]

that

[rigid snow]

"live5tream" they prolly think they're cool for that

[tender river]

how subtle

[rough bloom]

the file .gitmodules has a line like [submodule "vendored-tests/lily/"]
turn that into [submodule "vendored-tests/lily"] (remove the trailing slash)
then run the submodule command again

[olive sable]

it seems to have worked

no "succes" message or anythin, just didnt error

[opaque sigil]

gpt-5 at least makes them money so i'd hope they care a little more

[olive sable]

its probably fine

[tender river]

success messages are generally not a thing

no output and return code 0 = success

[olive sable]

thats sad

I need more positive messages in my terminals

[rough bloom]

you can add a success message to your shells prompt if you wish

[tender river]

makes scripts easier to write when you dont have to pipe every single command into dev/null

[rough bloom]

but it'll almost always say success so it just wastes space

[uneven pulsar]

the swarm will follow you

[tender river]

i like my prompt

obviously its much wider normally

[rough bloom]

I still have the default fish prompt
it has been okay

[olive sable]

nah same error still

maybe the vscode marketplace guide is out of date

lemme use the ablecorp site

[cosmic sphinx]

"this one will send hype through the roof, trust"

[rough bloom]

if it failed on line 36 again then it still can't find a submodule

[sick owl]

Its because they're on ollama

Like I said it runs like arse at the moment compared to cpp due to borked memory management

Or thats been my experience anyway

[cosmic sphinx]

honestly, cpp gave me a good experience only on q8

f16 has been a flop

[olive sable]

nah i still get the same error

@sage crag i cant get the compiler to build

[sage crag]

its the compiler dev's fault

[olive sable]

ah

[sage crag]

he hasnt fixed a bug with the submodules yet

[rough bloom]

we patched that up already

[olive sable]

no we didnt

[rough bloom]

yes we did

[olive sable]

it didnt change anything

[sage crag]

this frickin forward slash

and then you git submodule update --init --recursive

[olive sable]

hmm

oh its doing stuff now

i guess i didnt do it properly last time?

[sage crag]

are you doing this on wsl/linux?

[olive sable]

wsl ubuntu yes

[sage crag]

i dont think the compiler has been tested on windows

[olive sable]

i couldnt get it to install on windows anyways. probably due to the same error tho

well ive compiled the compiler now, how do you use the compiler?

[sage crag]

well, typically you wouldnt just use it freestanding because that's a bit of a headache

[tender river]

just pass the source file to the compiler and specify the --target flag (you'll probably want x86_64-linux)

then link it

[sage crag]

having the compiler installed on your machine is only there for the extension

[tender river]

lily's build script does it all for you

[sage crag]

yeh

[tender river]

but note that it bundles its own version of the compiler in out

also you'll still have to set some env vars to build

[olive sable]

uhh

[tender river]

you can just call the compiler manually instead and link with clang or ld or whatever

[olive sable]

and how do you pass stuff the the compiler?

what is the command?

[sage crag]

the command is hbc, but you likely dont have it in your PATH

[olive sable]

i think i might have it in path, i followed the hblang vscode exstension tutorial

[rough bloom]

thank you Common Lisp cookbook, very cool
(this is the whole explanation for that function)

[tender river]

what other info could you possibly need

[opaque wharf]

Then it will be trial by fire

[sage crag]

type hbc --help then

[olive sable]

seems to be in path yep

[sage crag]

cool

there you go then

[olive sable]

also is zig build test having 19/21 normal?

[sage crag]

yes, test errors if it doesnt match the existing test

doesnt matter

[olive sable]

sure

[sage crag]

its a bit of a hassle to compile for x86 without using lily

and you wont have stdlib features

but if you want to, here's how

[olive sable]

oh so it doesnt compile to x86 normally?

[sage crag]

it does, it's just not automatic

[olive sable]

what is the standart format then?

[sage crag]

well, if you just run hbc main.hb > out you will get an ableos executable

// main.hb
main := fn(): i32 {
  return 0
}
``````bash
hbc main.hb --optimizations release --target x86_64-linux > main.o
ld main.o -o main

here's the simplest way of doing it

and probably what you want

[opaque wharf]

Bother the compiler dev some more

[sage crag]

that's intended behaviour

[olive sable]

im sadly not running on ableos

[sage crag]

refer to A

[olive sable]

[opaque wharf]

Tell me when you have TCP/IP stack on lily. I wanna try to make something

[tender river]

will it work without a start handler

[sage crag]

yes, if you dont use -nostdlib in the ld

it will just link it to libc

[tender river]

right

okay wrong

[olive sable]

is the last main supposed to be an exe?

[tender river]

binary, exe is a windows thing

[sage crag]

an executable, not an exe

[tender river]

mhm

[sage crag]

to be pedantic

[olive sable]

uh

sure

how do i run it?

[sage crag]

./main

[tender river]

you have to link with clang or gcc rather than ld

otherwise it wont link to libc

[sage crag]

ah

makes sense

[tender river]

or just zig cc

[sage crag]

zig cc is the easiest way

[tender river]

zig cc -o main main.o

[olive sable]

its dumping its core

[sage crag]

use zig cc instead of ld

[tender river]

.

[olive sable]

oh i thought that was jsut for using c

[patent walrus]

whens the first game integration in hblang

[sage crag]

first game written in hblang or first game that uses hblang for scripting

both exist

[tender river]

i did implement json de/serialization, but also no thanks

[tender river]

nuro

[sage crag]

ah

hell no

websockets

[tender river]

tls

[patent walrus]

you're so close you just have to implement websockets!

[tender river]

[patent walrus]

[tender river]

actually not even that hard since i can use syscalls for tcp

and since i dont have to support wss