Fusing a 27B ternary LLM's whole decode step into one CUDA kernel

(twitter.com) by Jr23_xd | Jul 16, 2026 | 0 comments on HN