Benjamin Chetioui

1.6K posts

Benjamin Chetioui

Benjamin Chetioui

@_SIben_

Currently JAX & Pallas @GoogleDeepMind 🇨🇭 (prev XLA @Google 🇨🇭🇧🇻, PL PhD @UiB 🇧🇻) | ⚪⚫ Go player, 🇫🇷🥈'25 | Prev CTF @FlatNetworkOrg. Opinions my own

Katılım Eylül 2010
1.1K Takip Edilen1.1K Takipçiler
Benjamin Chetioui
Benjamin Chetioui@_SIben_·
@norpadon If Pallas/Mosaic GPU doesn't have what you need, plgpu.inline_mgpu is your escape hatch. Here is an example of how to use it to insert inline PTX #L81-L114" target="_blank" rel="nofollow noopener">github.com/openxla/tokama…, or arbitrary MLIR code #L166-L211" target="_blank" rel="nofollow noopener">github.com/openxla/tokama…. Please let us know if useful abstractions are missing!
English
0
0
5
191
Artur Chakhvadze
Artur Chakhvadze@norpadon·
Fellow JAX users, what do you use for GPU kernel programming when pallas model doesn't give you enough control? CuTe/CUTLASS? TileLang? CuTile? Afaik only CUTLASS has good jax interopability, which is annoying
English
7
2
66
7.8K
Benjamin Chetioui retweetledi
Adam Paszke
Adam Paszke@apaszke·
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/pall…
English
8
46
302
33.4K
Benjamin Chetioui retweetledi
Jacob Austin
Jacob Austin@jacobaustin132·
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
Jacob Austin tweet media
English
25
389
1.9K
466.3K
Benjamin Chetioui
Benjamin Chetioui@_SIben_·
@_xjdr @IlyasHairline @finbarrtimbers FWIW, we have the ability to use Triton to lower a lot of our code in XLA:GPU (and in fact, do it)! I'm curious exactly what are the performance cliffs that you're facing, and whether we can help with that!
English
0
0
3
101
xjdr
xjdr@_xjdr·
@IlyasHairline @finbarrtimbers jax has pallas which is amazing but now you are getting into a whole different beast. I just need something that works for research and dev to start. Optimizations come at prod deployment time
English
2
0
13
777
xjdr
xjdr@_xjdr·
the performance of jax + GPU is abysmal compared with an equivalent pytorch implementation. This is sad but unsurprising. jax might be relegated to TPU only (for me) for a little while longer. that said, there is nothing that touches jax + TPU for large scale perf (405B bby)
English
22
7
379
33.9K
Benjamin Chetioui
Benjamin Chetioui@_SIben_·
Genuinely excited about this one, can't wait to find out where Mosaic GPU can take H100 performance!
Adam Paszke@apaszke

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: #L146-L354" target="_blank" rel="nofollow noopener">github.com/google/jax/blo… It's easy to install too! Latest JAX packages have it.

English
0
0
1
277
Benjamin Chetioui
Benjamin Chetioui@_SIben_·
@PatrickKidger Likewise just moved to Zürich! I followed some of your work while at Google, would love to meet up :)
English
1
0
2
62
Patrick Kidger
Patrick Kidger@PatrickKidger·
Okay, so what does that mean now? 🇨🇭 If any of you are in Zurich then LET ME KNOW! Let's be friends :D 🧪 We're starting to think about publishing / academic collaborations / open-sourcing / etc. Keep your eyes peeled on this front. 3/
English
4
0
12
1.4K
Patrick Kidger
Patrick Kidger@PatrickKidger·
✨Life update - I have joined Cradle.bio in Zurich!✨ We're a startup/scaleup on ML for protein design. (And about half ex-Googlers, haha!) The team here are some of the best at this of anyone in the world. My official job title is "machine learning wizard"! 1/
English
19
9
312
30.3K
Benjamin Chetioui retweetledi
Daniel Cuthbert
Daniel Cuthbert@dcuthbert·
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
English
3
7
27
3.8K
Benjamin Chetioui retweetledi
shane
shane@shncldwll·
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
English
0
112
230
58.7K
Ólafur Páll Geirsson
Ólafur Páll Geirsson@olafurpg·
My wife gave birth to two identical girls yesterday and I’m now a lucky father of three under 2yo. Wish me luck! And forgive me if I tweet more family stuff for some time.
English
32
1
153
4.7K
Mathis Hammel-Brylinski
Mathis Hammel-Brylinski@MathisHammel·
En ce moment je redécouvre le bonheur d'aller à la bibliothèque pour bosser : - c'est gratuit 7j/7 - aucune distraction - mais surtout REGARDEZ CET ENVIRONNEMENT DE TRAVAIL omg
Mathis Hammel-Brylinski tweet media
Français
43
38
882
124.3K
Soumith Chintala
Soumith Chintala@soumithchintala·
Knowing all the development history, I'd say -- JAX+XLA will be better on TPUs and slightly better on bf16 compatibility. PyTorch will be better on NVIDIA and AMD GPUs, server-class and desktop-class CPUs, fp16 compat, vastly better on dynamic-shaped workloads. They'll probably be neutral on other hardware. Those are the sweet-spots they optimized for historically, and each are evolving towards the other direction -- but the evolution isn't complete for either software stacks, so you'd find the gaps.
English
2
2
21
10.1K
Alex Laterre
Alex Laterre@AlexLaterre·
Enjoyed this paper from @sarahookr team 👏 Key point -- JAX significantly outperforms PyTorch in hardware portability, which suffers 44% GPU to TPU failure rate. This echoes my first-hand experience and confirms our choice to adopt JAX early on, and its ongoing benefits 💪
Cohere Labs@Cohere_Labs

How portable are popular ML software frameworks? 🚚 Our recent cross-institutional collaboration reveals how costly straying from a narrow set of hardware-software combinations can be. 📜 arxiv.org/pdf/2309.07181…

English
4
8
70
23.4K
Odin
Odin@oas004·
🎉 Incredibly honored to receive the Google Open Source Peer Bonus Award for my contributions to the Accompanist library this year! Thank you so much @bentrengrove for the nomination, and @GoogleOSS for the award! 😊🥳 #opensource #android
Odin tweet media
Norway 🇳🇴 English
6
0
10
1.3K

Keşfet

@norpadon @jeremyphoward @_xjdr @IlyasHairline @finbarrtimbers @Sh0ckFR @PatrickKidger @elonmusk