cudnn_cu12

Lord God, I come to You a sinner And I humbly repent for my sins I believe that Jesus is Lord I believe You raised Him from the dead I would ask that Jesus come into my life And be my Lord and Savior I receive Jesus to take control of my life And that I may live for Him from this day forward Thank you, Lord Jesus For saving me with Your precious blood In Jesus' name, amen

there is so much going on

@MiaAI_lab @NVIDIAAI how many do you have?

The amount of actual useful work I'm able to do on my @NVIDIAAI DGX Sparks is incredible and improving every day. These little machines are capable of serious work.

@datavorous_ thanks, i havent been able to crack under 10ms using agents only. will have to do it by hand

I tried to speedrun the QR_2 challenge of GPU MODE in 16 hours i stopped at 4.7ms, this was my first "proper" gpu mode contest. constraints being: no way to tokenmaxx, no gpu to profile against extensively. what I did to understand the math: > householder Reflection method to do QR decomposition - reading the problem statement multiple times - did a 3x3 matrix dry run by hand - consulted gemini for validation of my mental model - studied parallel prefix sums - drew out the tree structure on paper before writing a single line of code - thought of a simpler approach (blocked method) instead of directly jumping to the tree like implementation what worked: - blocked householder with GEMM trailing update - loading panels at ONCE into the SRAM - tuning block sizes to prevent spilling - tuning warp count for better occupancy - (unintended) pytorch fallback for n=4096 what did not work: - fp8/nvfp4 quantisation (precision violated) - Cholesky method (not allowed) - Mixed precision with iterative correction - TF32 global flag (failed tests lmao) - full panel resident in shared memory (spilled badly) takeaways: i didn't get a chance to do extensive kernel programming before, so I had to bet on my intuitions: - bytes moved is the real bottleneck, i have to reduce the round trips of data - asymptomatic wins can still lose in practise - if you're broke, tokenmaxxing is not an option. manual? doable, but time is a constraint. human x ai? hell yeah!

@blelbach thank u

This isn't one problem - it's a dozen or more! Specialized kernels for different shapes & kinds of matrices will win the day. Leveraging tensorcores & a mixture of numeric formats (FP64, FP32, TF32, FP16, FP8) is likely key to top answers. This may even end up memory bound.

@hotschmoe and you will try some kernel engineering?

@_proteuss_ Totally usable in my opinion. Mature W4A8/W8A8 should be much better imo to activate int8 fastpaths x.com/i/status/20685…

I almost bought a dgx spark at $2999, then I almost bought a rtx pro 5000 for $3400. Watched as prices continued to march. Not getting left behind, I decided to throw my hat in the ring with Intel, got 2xB70s for $1900 Has barely been 2 days and I can't believe how much fun this is. Excited to contribute to Intel optimizations and provide useful AI to my family and friends

@Dadahelper1 interested in how you make these!!

v2 is out on HF and the jump in agentic tool-use is huge — on tau2-bench telecom it goes from ~15% (base Gemma 4 12B) to ~55%: it actually diagnoses → fixes → verifies instead of bailing after step one. A lot of folks are hitting issues across different runtimes (llama.cpp, Ollama, LM Studio, koboldcpp…) — almost all of it is template / tool-format quirks, not the weights. I've written every fix up in the HF discussions, so check there first 🙏 And honestly? Google's models are way less convenient to work with than Qwen — the non-standard chat template + native tool-call format trip up half the ecosystem, while Qwen's standard ChatML just works everywhere. Shipped it anyway 💪 huggingface.co/yuxinlu1/gemma…

my version of alopex, a gradient free / correlative learning algorithm now beats adamW on a mlp learning mnist with zero probes and only forward passes. they said it couldn't be done

@andreisavin @oleksoleksoleks its not economical. i have vpc at $30 a month

@_proteuss_ @oleksoleksoleks I've had good success with vast.ai

You can just rent Blackwell hours at 90% discount of inference API prices. Model providers are playing you all like fiddles.

what are some examples of people that have made their own harness?

@0xSero thanks ill go thru it. but assume im dumb and if i want to make small version like this -what do i have to do?

@_proteuss_ Model cards are on huggingface.co/0xsero

GLM-5.2-REAP checks it's work to catch errors before serving me the game. I love this MF

i will try glm 5.2 again on a fresh repo for kernel competition but which harness, inference endpoint, ide should i use?

@Dadahelper1 @HuggingModels how was this fine-tuned?

Thank you @HuggingModels for the feature, this made my day 🙏 v1 has been incredible to see take off in the community. Good news: v2 is already cooking — dataset's basically done. It'll be stronger across the board. Releasing as soon as it's ready 🚀

@minu_who @hrishioa what harness , provider, ide are yall using?

glm 5.2 outperformed opus 4.8 on our undefeated (by humans and agents) take-home. it also moved fewer tokens and we've known it's cheaper, like 3 times cheaper. i'm kinda having a moment

@difficultyang i was using this for this and it couldn't get to a solution that was sub 2ms and i spent nearly $100 x.com/_proteuss_/sta…

My verdict on GLM-5.2 is that it passes the inflection point threshold. It isn't too cheap to meter but in interactive use you'll have trouble pushing to $100 spend a day, I think.

this algo loves ai slop articles only viewed by bots or something

im not in the weights

@ericzedd pc - i have some old gpus but no where near enough vram for large models

@_proteuss_ PC or Mac? 👀

ive hit usage limits on everything this month : codex, copilot, and claude. i really need to get a local coding model hmmm