ohh lmk if you're open to prs (500 toks / sec) for fused nvfp4 moe kernels we made a while ago

this is why im doing the ting

dude yess! specifically the L2 stuff i tried similar stuff too wow (well specifically prefetching into L2 haha)

this alone is worth, so sick, got me coming out of vacation, time to master cute -> ARC-AGI-3

i did WarpGBM, PackBoost, and most recently specialize in making llms work on lesser compute with more speed:

more about the co-design here (with illustrations):

ill keep updating the co-design(s) for dense nvfp4 moe here: maintained by the homie sol these sessions are equally important i think (for scaling up esp) - more images in the doc xCalibrr

aight im making it public during development so ya'll can follow along! xCalibrr (xCaliber - name wip)

coming soon, our goal is (prolly), ARC-AGI-3 and to have fun 😁 thinking about gpumode competitions too