@Simon_Vt I think the flahsinfer repo is Apache 2. Pls consult with your lawyer.
English
Alex Zhurkevich
339 posts
Alex Zhurkevich
@cudagdb
Select difficulty 👁️🗨️
United States شامل ہوئے Aralık 2017
540 فالونگ578 فالوورز
@cudagdb @cudagdb license wise: can we use this code as we please or do we have to take care of some licensing issue
github.com/aleozlx/flashi…
English
Trtllmgen kernels are now open. Fastest prefill and decode kernels for our target workloads. We wrote these to win InferenceX, MLPerf, other benchmarks. Powering some of today’s top served models. Dive in, learn, use them, or level up your own. Enjoy.
github.com/flashinfer-ai/…
English
Give Meta 5 years of life, they'll call you a pokemon.
Soumith Chintala@soumithchintala
@marksaroufim You build communities that last, and that’s rare. GPU Mode, PyTorch… they worked because you showed up with honesty and zero ego. This makes you a rare pokemon. Thanks for your service. Excited for what’s next. I'm sure you’re definitely not done leveling up.
English
openai reseacher running in place towards anthropic every day
poetic
gabriel@gabriel1
realized the door im running in front of every day of is anthropics office let's see attendance on a saturday
English
@MaziyarPanahi @vllm_project It's been there for a while. However, I don't think it's used to full potential
English
I can’t believe we built this in the late 2025’s and then just never built anything like it again
English
@Leik0w0 Trtllmgen predates cute dsl. We prefer writing CUDA+PTX kernels, though Cutlass and Cute are also excellent. It’s just a matter of style - our code even uses some Cutlass components like pipelines.
English
+4,906,004
😵💫
Alex Zhurkevich@cudagdb
Trtllmgen kernels are now open. Fastest prefill and decode kernels for our target workloads. We wrote these to win InferenceX, MLPerf, other benchmarks. Powering some of today’s top served models. Dive in, learn, use them, or level up your own. Enjoy. github.com/flashinfer-ai/…
cash me outside picojoule for picojoule how bout dah
stochasm@stochasticchasm
impressive, very nice. now let's compare a 31b dense to a 31b active 670b total instead. flop for flop
English
deepmind more like deep in the mines haha amirite
sophie@netcapgirl
google shutting down a deepmind hedge fund
English
the size of a MoE is determined not by how high the parameter count is, but by how low the mfu is
Zach Mueller@TheZachMueller
What do we call a "large scale MoE" nowadays? 100B+ 500B+?
English
@stuart_sul May I ask what is special about your novel variant of NVFP4?
English
Happy to share this technical report! Building MXFP8/NVFP4 training kernels for Composer 2 with ThunderKittens/ParallelKittens was a lot of fun. We share some details in the report, including our novel variant of NVFP4:
Cursor@cursor_ai
We're releasing a technical report describing how Composer 2 was trained.
English
Do not forget to enable CUTLASS_ENABLE_PERFORMACE or you won't get any
Aaron Gokaslan@SkyLi0n
If you do this, don’t forget to set CUTLASS_LIBRARY_KERNELS to all True too. Enjoy the dreadful compile times though!
English
Pay attention to this one. Anne and folks are cooking smth special 👩🏻🍳
Anne Ouyang@anneouyang
Excited to share @Standard_Kernel's seed round and some reflections on what we’ve learned about kernel generation and what we believe is next. Grateful to our amazing team, supporters, and the broader community pushing this space forward.
English
Alex Zhurkevich ری ٹویٹ کیا
Checkout blog.vllm.ai/2026/02/01/gpt….
Powered by trtllmgen. Get yours in flahsinfer.
English