Hacker News new | ask | show | jobs
Fusing a 27B ternary LLM's whole decode step into one CUDA kernel (twitter.com)
3 points by Jr23_xd 15 days ago