Matej Sirovatka
@m_sirovatka
Followers
1K
Following
872
Media
52
Statuses
584
It's that time of the year again and we're coming with another @GPU_MODE competition! This time in collaboration with @nvidia focused on NVFP4. Focused on NVFP4 and B200 GPUs (thanks to @sestercegroup ) we'll release 4 problems over the following 3 months: 1. NVFP4 Batched GEMV
7
10
191
nothing beats the feeling of seeing your model reach ~10% mfu on a full node, what pain awaits at me full scale? give me back Jax please
0
0
19
I hope the people behind Cursor composer will have their pillow cold from both sides
0
0
17
What's stopping you from waking up Sunday 4am to build? (please someone how do I fix jetlag š it's been a week already)
5
0
24
Hosted evals are live on our Environment Hub š«” Run your fav model, no setup needed. And share your benchmarks.
7
10
77
So much high signal in this blog, always bullish on Gauās posts š
0
2
28
If you'd like to win your own Dell Pro Max with GB300 we're launching a new kernel competition with @NVIDIAAI @sestercegroup @Dell to optimize NVF4 kernels on B200 2025 has seen a tremendous rise of pythonic kernel DSLs, we got on-prem hardware to have reliable ncu benchmarking
7
20
164
To register to be eligible for prices, follow this link https://t.co/Y8XHI1nI3c. As always, updates and discussions will be posted to https://t.co/5LyY8srz65. Good luck to everyone! š„
discord.com
An open source GPU programming community | 21509 members
1
1
22
As you're all accustomed to, this competition also comes with banger prices. The grand winner receives a GB300 Dell Pro Max from @Dell and winners per problem will receive DGX Sparks and RTX 50xx GPUs and last but not least, tickets for NVIDIA GTC, for the official award ceremony
3
2
28
you can just create environments
creating an environment with verifiers is as simple as writing a load_environment function and filling out a pyproject.toml, both of which are initialized for you when you do `prime env init` environments are packages, and can be used with prime-rl, skyrl, tinker, and more :)
1
7
65
Only in SF kinda story. I went to buy my matcha latte, listening to Clairo. Noticed a girl at this cafe and she seemed sad, I asked her whatās the problem, she said reward isnāt going up in her runs. I told her to try prime-rl - she starred the repo immediately.
Only in SF kinda story. Saw a beautiful girl at this cafe I was working from, well dressed, waiting for someone for over 25 mins & visibly distressed. Spoke to her, she was on a blind date and guy didn't show up (stupid guy). We spoke more. Told her what @mavehealth does,
9
5
283
Researchers figuring out a way to minimize mismatch while I'm happily rewriting the training loop to use CP + FlexAttention while the inference uses god knows what
FP16 can have a smaller training-inference gap compared to BFloat16, thus fits better for RL. Even the difference between RL algorithms vanishes once FP16 is adopted. Surprising!
4
3
76
Big tech doesn't want you to know this one simple trick - you can have 2 jobs at the same time, working the 2nd one while your "max-autotune-no-cudagraphs" flex attention is compiling
6
4
127
3 days later, a lot of multiprocessing and multithreading attempts later, weāre at linear scaling with number of inference nodes
Types in Python are a good idea until you open your profile trace and see this
2
0
30
Types in Python are a good idea until you open your profile trace and see this
14
2
107
We're scaling our Open-Source Environments Program As part of this, we're committing hundreds of thousands of $ in bounties and looking for partners who want to join our mission to accelerate open superintelligence Join us in building the global hub for environments and evals
12
59
428