Compiling LLMs into a MegaKernel: A path to low-latency inference

Compiling LLMs into a MegaKernel: A path to low-latency inference. 76 comments, 314 points on Hacker News.