🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

grinding til 3am isn’t really a flex if you’re waking up at noon every day

nsys looks pretty cool actually, but information overload for a first-time user. Took me a bit to get good at Google's XProf too, so let's get started! QQ to my nsys expert followers: any specific pro-tips? Biggest bang-for-buck things/views to look at? Any good pytorch

Why on earth is the first thing I want to do always too bleeding edge. Every single time. Damnit, FML!

Heh nice

They have a cloud infrastructure?

guys I'm under observation now👀

torch.profiler.profiler.py Last commit: 1mo ago "add memory blablabla" torch.autograd.profiler.py Last commit: 2mo ago "Induce inductor blablabla" One of the two is legacy/deprecated, but you only learn that by looking at the docs of the other one, so if you land on the old one by

I like the Encoder-only Mask Transformer (EoMT): basically removing all the bells and whistles, and doing panoptic segmentation with an almost vanilla ViT. You're sliiiiightly worse for the same encoder size, but it's a lot simpler/faster and (likely) more scalable. I wish they

Oh wow, did you guys know that torch.compile can compile numpy code? And even run it on GPU? This is pretty neat for all kinds of "surrounding" code besides the model (like evals and fancy metrics) that I used to do with numba/numexpr (cuz CPU-XLA was pretty meh). Poll below

Did you know this?

The answer is that the name is weird, it's simply (almost) the whole flex_attention bwd compute, not just "zeros" as I thought the name implies. The way to find out would be looking at the call-stack, opening that generated file with the long name, and then go look at the actual

This might well be the start of a new big lab drama:

Cool work uses "visual anagrams": two images of different objects made out of the same image patches. Model must classify both correctly to score. Hence, higher scoring models use global geometry, lower use textures. SigLIP is GOAT of course, or I wouldn't repost this (jk)

Phew, reassuring that most of you didn't know either:

Interesting alternative to multi-token prediction, though the figure is a bit unintuitive. Instead of attaching a head for each +d'th prediction, pass a dummy input token for each extra prediction through the model. This is A LOT more expensive, e.g. doing 2-step prediction

Or, in other words, Gemini2.5 Pro succeeds at 30% of real world office tasks. That's pretty good, considering this is the worst Gemini Pro will ever be.