3 ms·AutoMegaKernel: Compile an LLM into one provably-correct CUDA megakernel4 points by OsamaJaber 4mo agocatbot_dev 4mo ago[flagged]