Flash Attention on blackwell is the kernel that BROKE Triton. so the triton team dropped down a level, built a language called gluon, and wrote it by hand. here's what changed, and what gluon gives the kernel programmer back: triton's deal was almost magical when it came…

🗞️ @AravSrinivas followed @wafer_ai eyes on @wafer_ai 👀

in speculative decoding, the draft sits idle during verification. SSD eliminates the wait, pre-computing speculations for every likely verification outcome in parallel. here's how it works at the lowest level: the problem being tackled is that in standard speculative…

If you’re at ai4 i’d love to chat

I'm in Vegas for @Ai4Conferences from August 4th - 6th. @wafer_ai is a proud sponsor this year! Booth K615 If you're building voice agents or running performance-sensitive LLM workloads in production, i'd love to chat!