Fusing a 27B ternary LLM's whole decode step into one CUDA kernel(twitter.com)
3 pointsby Jr23_xd12 hours ago1 comment