Why LLM decode is memory-bound, not compute-bound(github.com)
4 pointsby harshuljain132 hours ago1 comment