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

 (twitter.com)

3 points | by Jr23_xd 6 hours ago ago

1 comments