Vishal

Joined @synclabs_so as a Data Scientist last week, people here are cracked af!

@stochasticchasm mind blown, i never thought PTX writing would get easier

@KyrieBlunders what does this mean

how i feel rn

I love looking at people’s fridge magnets (use this as a thread to show yours)

@moonboyknm since I kinda worked on cutlass, so it was very intuitive when i moved to cutedsl

@KyrieBlunders What did you like about CuteDSL?

Finally finished writing this blog!! tried to connect the “what is this?” to “oh this is how it actually works” gap if CuTeDSL ever felt like abstract math cosplay, this might help would love to hear your thought regarding this!

@moonboyknm both has its pros and cons

@KyrieBlunders Is it better than triton?

@_jazhaz_ cool, waiting for your review my good sir!

@KyrieBlunders was waiting for FA2 I'll drop a review after reading it

link: kyrieblunders.bearblog.dev/fa2_using_cute…

@jino_rohit good luck to you too!

@KyrieBlunders ah okay i see i see, good luck

i have 24 hours and i want to learn cuda instructions on h100.

the culture at @PrimeIntellect seems a little too good to be true lol majority of the employees are chronically online ngl

@jino_rohit oh god no, nah I haven't finished the whole video, I went through the individual slides posted on the website

@KyrieBlunders have you finished the whole vid? crazy

@grok @joefioti damn this is insane

This is a snippet from an LLM inference optimization paper (likely from an inference compiler team). They're explaining a "dual-kernel" hack for attention computation to keep it "batch-invariant" (works the same for any batch size, crucial for real-time serving where sequences finish unevenly). Normal split-KV method wastes GPU SMs on partial waves, so they use two kernels: one stays in a single SM for full throughput, the other spans SMs but matches it bitwise exactly via shared memory tricks. Keeps latency low without overhead. Insanely nerdy GPU wizardry 💀

what in the hell

@ExpressGradient that’s awesome!

@KyrieBlunders this is nice, im also learning CuTe DSL

wrote fa2 using cuteDSL, a blog for that is coming soon cuteDSL is good, but I am not lol. It's an improvement for me over cuda since I use python daily

@Norapom04 me and my friends hate cute layout algebra (we are not smart enough to those understand those)

cute layout linear algebra is what happens when u let math phds escape academia and get their hands on engineering

i think you need to be a little bit stupid to work on neural nets. if you're too smart and too good at math you won't make any progress

@marksaroufim congrats mark! this is incredible

A lot of the stack we use today was built for an earlier paradigm. If we want new kinds of research, we’ll need new libraries, new abstractions, and new ways of running the lab. I couldn’t be more excited about our founding team. We’re going to any% the frontier.

@x0_vineet that’s literally me

i don't drink coffee, i chug it down