Clover: Regressive Lightweight Speculative Decoding with Sequential Knowledge

arXiv:2405.00263v1 Announce Type: cross
Abstract: Large language models (LLMs) suffer from low efficiency as the mismatch between the requirement of auto-regressive decoding and the design of most contemporary GPUs. Specifically, billions to trillions of parameters must be loaded to the GPU cache through its limited memory bandwidth for computation, but only a small batch of tokens is actually computed. Consequently, the GPU spends most of its time on memory transfer instead of computation. Recently, parallel decoding, a type of speculative …

abstract arxiv auto bandwidth cache computation cs.ai cs.cl cs.lg decoding design efficiency gpu gpus knowledge language language models large language large language models llms low memory parameters through type