Одна інтеллектуальна одиниця

@JFPuget @kaggle @WalterReade 😢

.@kaggle @WalterReade how come the paper on Cryo ET competition does not even mention the names of the the authors of the models discussed in the paper? They are thanked anonymously at the end, that's all. Yet the paper disclose their findings. In other competitions, the winners were co authors. And in this case @kagglingdieter from the winning team did work with the host to make his model available on the host platform. Not citing kagglers in this context is profoundly unethical. I know you are the messengers, but please forward this to the full author list.

@wightmanr Not sure how data IO intensive runs you are going to have, but I went for RAID 0 configuration for a pair of NVMEs on my 4x3090 workstation (double capacity, double speed; reliability solved with prayers) Great of all sorts of 3d scans/video processing

I swore I'd never get another Seasonic PS after their previous Prime series completely shat itself with a restart issue running GPUs. I ended up throwing 3 of those away and replacing with EVGA SuperNovas (Super Flower) ... but those PSU are getting pretty dated with no major update... ATX 3.1's more aggressive hold time, etc reqs should address heavy GPU use so will give them another shot. The Noctua edition looks pretty neat.

I think it's time for a new 'AI' workstation. My 6 year old 2x3090 and 1x4090 (used to be 2x 3090) systems are getting rather long in the tooth. The NVIDIA Spark isn't going to cut it for train prototyping. I've got two recent M4 Apples but those are also paperweights for PyTorch training. The 'when is it going to arrive' GB300 workstations look like they'll probably end up $70-80K USD :/ So NVIDIA RTX Pro 6000 (Blackwell?) it is? Anyone built such a system recently? I said I wasn't going to do a custom build again, but none of the major PC makers with dB friendly workstations have a 9000 series Threadripper + RTX Pro 6000 option. I see some SuperMicro based options but last experience with those was LOOOUD. Hmm

Hi, I'm Ivan - a totally-not-suspicious aerospace prodigy. I just happen to have a charming Russian accent, a past internship at Sukhoi (no big deal), and a mild obsession with drone jet propulsion systems.

Trying an idle speculation is a challenge of its own. It’s an heart racing experience to have a good solution and not have submitted yet. If kaggle crashes today, I have nothing 😂

@ducha_aiki Any chance you have dynamic size of your inputs/targets? The way pin memory works is allocating a buffer on device memory where it would be copied. As long you have another size a new buffer must be allocated which over time leads to fragmentation of memory and even OOM

Answer: the bad guy(s?) is pin_memory=True + persistent_workers=True

Just secured 🥈2nd place in the BirdCLEF+ 2025 Kaggle competition with Fernando Gonçalves! 2,025 teams, tons of experiments, and a lot of fun. Huge thanks to the organizers and community! Write-up: kaggle.com/competitions/b… #kaggle #deeplearning #ecoacoustics #audio

@mobicham @jeremyphoward Great work! Does do_bench also call torch.cuda.synchronize at the end of iteration?

I just noticed that torch.nn.Embedding is a bit slow and a simple Triton kernel is 7-8x faster on the A100. Try it here: gist.github.com/mobicham/73962…

These children were killed by Russia – by its war. Boys and girls whose lives were cut short by Russian strikes and shelling, by hostilities that Ukraine never wanted for even a second – but from which we have been forced to defend ourselves for years already. It is on June 4 that we honor the memory of all such children. Today, we still do not know the exact number of those killed, as we have no access to the temporarily occupied territories. It is yet to be determined how many adults and children killed by Russia are buried in mass graves near Mariupol and other cities and villages destroyed by Russia. The world must neither forget nor forgive Russia for this terrible pain and aggression – this totally unprovoked and criminal Russian war against Ukraine. In memory of every child killed, in memory of every human life destroyed, there must be real global solidarity – from all across Europe, all across the the U.S., and all other parts of the world – so that Russia is held accountable for its actions and is forced to return the peace it stole from Ukraine and Ukrainians. @United24media

This is utterly disgusting. A russian game is available on steam where you can play as a russian war criminal to rape and kill innocent Ukrainians. If you have integrity- report this game into oblivion. For shame @Steam this is a disgrace.

@tunguz @trainxgb Is it socially acceptable to train xgboost in public?

Have you trained your XGBoost today? @trainxgb

@kagglingdieter @kaggle You thought, I thought, we all thought

I don’t understand where the perception comes from that @kaggle is about classical ML. For years now the competitions have been heavily deep learning and research focused. Or wait, maybe it’s because certain memefluencers propagate „xgboost is all you need for kaggle“ all day

As long your skill in certain domain is above “average” - probably don’t want to use cursor as it would generate worse code that you would otherwise write. But when if you don’t really shine in some areas (like with matplotlib) AI tools come to the rescue

Things where cursor doesn’t really do meaningful work: - suggesting new, out of the box ideas (just can’t) - finding a bug given symptoms and code (many FPs)

What I found @cursor_ai is good for - writing boilerplate code to visualize all sorts of plots. Really helpful at #kaggle, because let's admit - writing matplotlib calls is not fun yet necessary. Without proper model inspection you can't go to the top.

@ducha_aiki Because mean things are easier to do

Why I can write tensor.mean() and tensor.abs() in pytorch, but only array.mean() and not array.abs() in numpy?

Looking to give a try to #Cursor for a new Kaggle competition. @cursor_ai any promo-codes by chance for kaggle GMs and open-source contributors?

@francoisfleuret @__tinygrad__ We like it when we can buy it

@__tinygrad__ So we like the 5090 very much?

5090 memory bandwidth. Stated 1792 GB/s, tested @ 1956 GB/s. Confirmed not a scam.

Не дуже й хотілося якщо чесно :)

@francoisfleuret Memory aliasing :) lovely bugs

So ... I happened to subtract in-place from a tensor itself shifted. Do not do this kids. You mix this up with the GPUs doing their GPU things and you end up debugging for two days.

Russians, with US assistance, killed 11 civilians in Dobropillya. They hit a residential building.