What is Hackers' Pub?

Hackers' Pub is a place for software engineers to share their knowledge and experience with each other. It's also an ActivityPub-enabled social network, so you can follow your favorite hackers in the fediverse and get their latest posts in your feed.

0
0
0
0
0
0

Meine gern getragene Jeansjacke hat langsam aber sicher jede Farbe verloren. Man könnte sie wegwerfen oder man färbt sie nochmal neu.
Nach dem Färben fiel mir auf, dass die Stöße ganz abgenutzt waren und es einige sehr dünne Stellen gab. Meine Chance, visible mending auszuprobieren! 😁
Es wird ein Langzeitprojekt, aber bis hierhin mag ich es. Und ja, bunt und kindisch. Aber ich brauch das gerade. 🫠

Handgearbeitete Stickereidetails auf einer Jeansjacke. Zu sehen sind 5 Pilze in Gras. Drei Fliegenpilze, ein Steinpilz und ein Austernseitling.Handgearbeitetes visible mending an den Stößen der knopfleiste einer Jeansjacke. Die Flickarbeit ist beidseitig in Regenbogenfarben.
0
0
0

70% Size, 100% Accuracy: Lossless LLM Compression via Dynamic-Length Float

Link: arxiv.org/abs/2504.11651
Discussion: news.ycombinator.com/item?id=4

arXiv logo

70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float

Large Language Models (LLMs) have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM size by 30% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) decomposition of memory-intensive lookup tables (LUTs) into compact LUTs that fit in GPU SRAM, (ii) a two-phase kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on recent models, including Llama-3.1, Qwen-2.5, and Gemma-3, validates our hypothesis that DFloat11 achieves around 30% model size reduction while preserving bit-for-bit exact outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 1.9-38.8x higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.3-13.17x longer context lengths than uncompressed models. Notably, our method enables lossless inference of Llama-3.1-405B, an 810GB model, on a single node equipped with 8x80GB GPUs. Our code and models are available at https://github.com/LeanModels/DFloat11.

arxiv.org · arXiv.org

0

おちごとしようぬ :saba:

a97647158c (upstream/main) Add REST API for featuring and unfeaturing a hashtag (#34489)
49b6a49c76 Change "Pin on profile" to "Feature on profile" for posts in web UI (#34492)
d4944a2467 Fix incorrect redirect in response to unauthenticated API requests in limited federation mode (#34549)
91db45b197 Change account search to be more forgiving of spaces (#34455)
7a70d95435 Add warning for Elasticsearch index analyzers mismatch (#34515)
1326c8cb1d chore(deps): update dependency rqrcode to v3 (#34541)
199acce481 Fix sign-up e-mail confirmation page reloading on error or redirect (#34548)
b1b949f16c New Crowdin Translations (automated) (#34546)

0

“Evaluating The Effects Of Anycast On DNS Root Nameservers”: ripe.net/publications/docs/rip

some interesting statistics for (among other things) reliability of stateful (TCP) connections on anycast nodes, specifically in the context of DNS.

it is from 2006, so i wonder how numbers have changed in the mean time.

0
0
0
0

Setting up the first part of the second network cabinet at the new 40k sq/ft warehouse today. (Just the essentials to get a few desktops up, and setup outgoing network connections for them; did the cameras previously).

0
0
0
0

Join us for an exciting talk on using directly in R to boost data workflows!

Alexandra Posekany from TU Wien will cover the essential background of modern LLMs, then take us on a deep dive into practical integration techniques in R with leading models, and demonstrate concrete use cases.

R users of all skill levels and identities are welcome!

🗓️ When: Monday, 28 Apr 2025, 18:00 - 19:30
📍 Where: TU Wien, Campus Freihaus
🔗 RSVP: meetup.com/rladies-vienna/even

0

"On tourne vert" met à votre disposition une gamme complète d’outils et de ressources pour rendre vos tournages, studios d’animation ou d’effets visuels plus durables.

Venez découvrir des guides, un répertoire de fournisseur, l’accréditation, le calculateur carbone adapté aux productions, des études et bien plus encore pour vous aider à intégrer des pratiques écoresponsables dans vos activités quotidiennes.

ontournevert.com/fr/

Logo On Tourne Vert
0
0

The Rust Foundation believes that the Rust community is stronger when more voices are heard.

If you're thinking about submitting a talk to @rustconf 2025, taking place in Seattle & online this September, this is your moment! 🦀🧡

Speaking application closes on April 29: sessionize.com/rustconf-2025

0
0
0
0

Still messing about with and have Picom installed and setup for transparency in the terminal so far. I've also been tinkering with Xterm and Alacritty terminals and for all the later is good I keep returning to Xterm as it seems to do everything that I want. I have resisted the urge to return to days gone by when running Xorg where I used Kitty. I have for now peeled myself away from said machine to go have a soak in the bath with a bath bomb.
0

nuintari's rules of networking 0x62:

When you move infrastructure to the cloud, you remove the necessity of repairing much of it if and when it fails.

That's good!

You also lose the ability to repair most of it when it fails.

That's bad.

You also lose an enormous amount of visibility into said failures.

...

That's bad.

0
0
0
1
0
0
0
0
0
0
0
0
0
0
0

I am Jeremy Soller.

I work at @system76System76 :popos: :ubuntu: as Principal Engineer where I maintain our Linux distribution @pop_os_officialPop!_OS :popos:, port @coreboot@mastodon.technology and open source embedded controller firmware to our laptops, work on the new desktop environment @COSMIC_desktop, and more!

I also am the creator and BDFL of a microkernel operating system primarily written in @rustRust Language called @redoxRedox OS. I am damn near crazy about and use it anywhere and everywhere!

Follow if you are interested in these things 🙌 🦀

0
0
0

How to achieve Ultimate developer skills:

Step 1. Humble yourself by starting with the raw, unforgiving power of C.

Step 2. Attempt to purify your soul and memory safety by rewriting it all in Rust.

Step 3. Ascend to a higher plane of functional enlightenment (or utter confusion) by rewriting it again in Haskell.

Step 4. Congratulations! You are now Ultimate developer. Your code is perfect and immutable, and absolutely no one else on your team will understand a single line of it.

0
1
0

Setting up the first part of the second network cabinet at the new 40k sq/ft warehouse today. (Just the essentials to get a few desktops up, and setup outgoing network connections for them; did the cameras previously).

0

This is a _very_ cool project that just landed on @crowdsupply. Had I not been laid off last week (hire me! ) I'd definitely order one.

"Using its high-quality sensors, Air Lab measures CO2, temperature, relative humidity, air pollutants (VOC, NOx), and atmospheric pressure. The device can record measurements for multiple days, which you can analyze directly on-device—no computer or smartphone required.”

crowdsupply.com/networked-arti

0
0
0
0

🎧 Just dropped a new episode of the Lorenzo's Music Podcast featuring @benpateBen Pate 🤘🏻 discussing Bandwagon, an online platform for musicians! We talk about how it connects to the Fediverse and its connection to the online station Indie Beat.

lorenzosmusic.com/2025/04/ben-

0
0
0
1
1
0
0
0
0

@xezpeletaXabi yes, using a local LLM works with datasette-extract already

I'd suggest using Ollama to run a model like Gemma 3 or Mistral Small 3.1 (though others could work well too) - with the llm-ollama plugin installed datasette-extract should work fine

I don't know of any local models that handle PDFs in Ollama yet but plenty of them can handle images

0

if someone disagrees with your* (correct, of course) take, it does not mean they: - are evil - desire the full resulting conclusions of their beliefs - have fully expressed all their thinking - have disqualified themselves from the conversation *does not apply to me or my opinions

0

Did you know that @forgejo doesn’t just replace GitHub for hosting — it can even run GitHub Actions (née Microsoft® Azure® Pipelines) right out of the box?

The first @organicmaps workflows were up and running with minimal tweaks — no major rewrites needed. Thanks to @zloidemonVeniamin Gvozdikov for making it smooth!

Need a solid home for your project? @forgejo has you covered.
Find out why it's time to : sfconservancy.org/GiveUpGitHub

0
0
0
1