流水线并行Pipeline Parallelism
Pipeline Parallelism
解决什么问题What problem it solves
把模型按层切成几段，分别放在不同显卡上接力算。Cut the model into layer ranges and run them on different GPUs in relay.
这是什么What it is
vLLM 官方列出的分布式手段之一：张量、流水线、数据、专家、上下文并行。
含义是按层段把模型切到多张卡上接力计算。
One of vLLM's listed distributed techniques: tensor, pipeline, data, expert and context parallelism.
It lets a model larger than one GPU's memory run in pieces, at the cost of passing intermediate results.
为什么需要知道Why it matters
多卡怎么分工，官方就这几种口径。选错方式，卡再多也快不起来。These are the official ways to split work across cards. Pick wrong and extra cards do not help.