LLM inference has two phases with different bottlenecks: LLM prefill runs once per request and is compute-intensive. Decoding repeats for every output token …
机构:Georgia Tech
来源:arXiv 2611.11761 | AI4Papers 论文推荐平台