Vladimir Vlejd Macko (@vlejd) profile picture

I like taking things from 0 to 1. From nothing to something. Sprinkled with ML when necessary.

Joined March 2011
Photo by Vladimir Vlejd Macko (@vlejd)Photo by Vladimir Vlejd Macko (@vlejd)Photo by Vladimir Vlejd Macko (@vlejd)Photo by Vladimir Vlejd Macko (@vlejd)Photo by Vladimir Vlejd Macko (@vlejd)Photo by Vladimir Vlejd Macko (@vlejd)Photo by Vladimir Vlejd Macko (@vlejd)
Unstructured weight #sparsity made practical. 50% unstructured weight sparsity was considered too low for real GPU speed up without specific hardware support (like @cerebras). With @bozavlado we built MACKO-SpMV - a new matrix format + SpMV kernel to change that. 🧵
4
7
1
30
6,399
if you are at ICML2026, you should definitely check this free tour store. I was suspicious, but it is actually legit! #icml2026 . They are sad that not many people stop by. Still free spots for tours tonight and tomorrow!
1
3
150
Just came to #ICML2026 to present a spotlight paper on how we made the fastest general sparse matrix vector multiplication GPU kernel. github.com/vlejd/macko_spmv Come talk: Thu, Jul 9, 2026 • 5:00 PM – 6:45 PM KST HALL A #2713
1
6
10
593
Flying to #ICML2026 to present a spotlight paper on how we made the fastest general sparse matrix vector multiplication GPU kernel. arxiv.org/pdf/2511.13061 Come talk: Thu, Jul 9, 2026 • 5:00 PM – 6:45 PM KST HALL A #2713
1
1
7
645
Vladimir Vlejd Macko retweeted
Now, accepted to ICML as a spotlight paper. I am super proud of my student @vlejd.
Unstructured weight #sparsity made practical. 50% unstructured weight sparsity was considered too low for real GPU speed up without specific hardware support (like @cerebras). With @bozavlado we built MACKO-SpMV - a new matrix format + SpMV kernel to change that. 🧵
1
4
50
4,093
Vladimir Vlejd Macko retweeted
opus 4.5 is really good at GPU programming, but somehow it’s even better at GPU programming jokes (h/t @Si_Boehm)
20
45
9
533
86,114
And yes, it translates to real LLM inference speed ups.
1
2
165
🛠️ Next step: server GPUs. If you know how to implement a minimal CUDA matvec on H100 that hits ≥95% of cuBLAS 👉 My DMs are open.
1
147
It is funny how little correct information is there about how to properly benchmark a CUDA kernel. Most papers are wrong, eval libraries are hard to inspect and even this could have a problem because it may include the kernel launch depending on clear_cache implementation
btw, I think BackendBench just uses triton's do_bench function, which uses a very similar timing mechanism to the one exploited here and wouldn't be robust to the same side-stream shenanigans
116
Vladimir Vlejd Macko retweeted
holy shit they found a power series solution to ALL polynomial equations!! (bypassing Galois which says you can’t solve them in radicals)
30
94
24
1,111
169,345
I do model compression and optimization. It is essential to have access to different GPUs and that would be impossible without @vast_ai . Happy to finally meet you guys at #ICML2025 . And thanks a lot for the Nintendo Switch!
3
7
1,054
Just visited with Vast.ai at ICML 2025. Without them my research would be almost impossible. They make gpu kernel development much more accessible. @vast_ai
1
3
197
Vladimir Vlejd Macko retweeted
We are introducing Quartet, a fully FP4-native training method for Large Language Models, achieving optimal accuracy-efficiency trade-offs on NVIDIA Blackwell GPUs! Quartet can be used to train billion-scale models in FP4 faster than FP8 or FP16, at matching accuracy. [1/4]
20
75
8
389
65,107
Before #ICLR2025 I was nervous about the size of our poster. Now I am tempted to think outside of the box and push some boundaries. Cool papers btw.
2
108