跳到主要内容Skip to main content

预填充Prefill

Prefill

  • 硬件与部署Hardware & deployment
  • 第五站 · 跑起来Station 5 · Running It
解决什么问题What problem it solves

把你输入的那一整段先读完,再开始生成回答的那一步。The step where your whole input is read through before generation begins.

这是什么What it is

模型处理输入部分的阶段。输入越长,预填充越慢,首字延迟越大。

The stage where the model processes the input portion. Longer input means slower prefill and higher time to first token.

为什么需要知道Why it matters

解释了一个常见现象:塞了一大段资料后,模型「想」很久才开始回答。那是预填充在算。It explains a common observation: after you paste a long document the model "thinks" for a long time before answering — that is prefill computing.

出处Sources
  • NVIDIA AIPerf 指标参考NVIDIA AIPerf server metrics reference ↗
  • vLLM 指标设计文档vLLM metrics design ↗