
Currently JAX & Pallas @GoogleDeepMind 🇨🇭 (prev XLA @Google 🇨🇭🇧🇻, PL PhD @UiB 🇧🇻) | ⚪⚫ Go player, 🇫🇷🥈'25 | Prev CTF @FlatNetworkOrg. Opinions my own
Joined September 2010
- Tweets1.6K
- Following1.1K
- Followers1.1K
- Likes4.6K
Benjamin Chetioui retweeted
Want to improve GPU compute/comms overlap? We just published a new short tutorial for you!
A few small changes to the Pallas:MGPU matmul kernel is all it takes to turn it into an all-gather collective matmul that overlaps NVLINK comms with local compute: docs.jax.dev/en/latest/palla…
Showcasing one really cool outcome of a small MGPU hackathon we had a few weeks back!
Curious how to write SOTA performance Blackwell matmul kernels using MGPU? We just published a short step-by-step tutorial: docs.jax.dev/en/latest/palla…
At each step, we show exactly what (small) changes are necessary to refine the kernel and the final kernel is just under 150 lines.
Benjamin Chetioui retweeted
Replying to @jeremyphoward
We have some Pallas tutorials (docs.jax.dev/en/latest/palla…) and in particular a reference guide for Mosaic GPU (docs.jax.dev/en/latest/palla…). It is of course far from being comprehensive, but we're working on it. If you have questions you should also just reach out and we'll help :)
Benjamin Chetioui retweeted
Making LLMs run efficiently can feel scary, but scaling isn’t magic, it’s math! We wanted to demystify the “systems view” of LLMs and wrote a little textbook called “How To Scale Your Model” which we’re releasing today. 1/n
... And now the thesis is online (bora.uib.no/bora-xmlui/handl…)!
And the defense is scheduled next Wednesday (uib.no/ii/174068/disputas-be…) 😱
It'll in principle be streamed—feel free to DM if you're interested in attending :)
... And now the thesis is online (bora.uib.no/bora-xmlui/handl…)!
And the defense is scheduled next Wednesday (uib.no/ii/174068/disputas-be…) 😱
It'll in principle be streamed—feel free to DM if you're interested in attending :)
Genuinely excited about this one, can't wait to find out where Mosaic GPU can take H100 performance!
Many of you are excited about H100 attention, so it’s a good time to show you Mosaic GPU: a Python DSL for H100s.
The attention example matches FA3 performance, while being only ~200 lines of Python: github.com/google/jax/blob/m…
It's easy to install too! Latest JAX packages have it.
Benjamin Chetioui retweeted
Finally got around to writing a guide for matrix multiplication on TPUs using Pallas. Check it out!
jax.readthedocs.io/en/latest…
Benjamin Chetioui retweeted
As much as I'm personally enjoying the AI ride at the moment, the attacks we are seeing are worryingly familiar web app attacks and one wonders if anyone in the AI world is aware of how you do web app security?
Case in point: arxiv.org/abs/2403.06634
A very good paper firstly
Benjamin Chetioui retweeted
Tired of “LLM hacking” hype with no code? Here’s a breath of fresh air. arxiv.org/abs/2402.11814
1. Challenges: open source ✅
2. Solution framework: open source ✅
If you’re interested in hackbots in offsec and you’re craving something you can RUN, you gotta read this
Benjamin Chetioui retweeted
High quality steganographers
The Teaser 2024 is over! Congratulations to 🥇 @FlatNetworkOrg, 🥈 @justCatTheFish, 🥉 @hxpctf and everyone else!
We hope to see you guys at SwissTech Convention Center in Lausanne.
Thanks everyone for participating Insomni'hack Teaser 2024, #INSO24, #CTF
Benjamin Chetioui retweeted
Writing is an absolutely critical skill… and one that can be improved dramatically. Doing so requires not just repetitions, but deliberate efforts to improve. Some tips here: docs.google.com/document/d/1…
This quoted post is unavailable.
Benjamin Chetioui retweeted
Not announcing anything publicly yet, but please get in touch with me if you're interested in funded PhD opportunities in applied formal methods. I will make a proper announcement once everything is confirmed, but it would be good to make connections now.
Benjamin Chetioui retweeted
"You go into your toolchest and you select, I don't know, that sad excuse for a Scala effect system, for instance, because you're trying to tell the world that you take yourself too seriously to care about monad tutorials or abstract category-theory. [...]
And you're also blithely unaware that in 2013, Andrej Bauer demonstrated a collection of effects in Eff, which helped pioneer the idea of using in-language handlers, using Pitts’ theory of minimal invariant relations to prove its core calculus adequate to the task. And then I think it was HOPE, which for years continued to push for non-MTL-based systems of effects. And then handlers started showing up in the effect systems of eight different designers. And then it filtered down, through Ocaml, into the .NET languages and finally over into some tragic open source clearance bin.
It's sort of comical how you think that you've made a choice that exempts you from category theory, when in fact you are using the effect system that was selected for you."
1 year and 2 months after the end of my PhD contract and I wrote a thesis abstract today. Maybe I'll defend my PhD one day, after all :)
Benjamin Chetioui retweeted
There's a huge opportunity to use ML to replace hand-written heuristics in compilers.
Want to try? Check out this @kaggle
contest organized by Ashley Chow, Bryan Perozzi, HCL-Jevster, inversion, Mangpo Phothilimthana, and Sami Abu-El-Haija!
kaggle.com/competitions/pred…
Benjamin Chetioui retweeted
Replying to @tonofcrates @alpha_convert
You appear to be advocating a new:
[ ] dependent [ ] refinement [ ] "functional"
type system. Your type system will not work. Here is why it will not work.
You appear to believe that:
[ ] All code has a precise specification
[ ] SAT solving can be fast because P = NP


















