比较有价值的点：论文提出了一种在线推理的、新的混部实现方式

1) At low loads, most of the GPU memory is allocated but not used, occupying the GPU memory and preventing it from being used by other services;

显存已分配未使用，其他服务也用不了 -> 论文提到了，他们用这个技术来做离线混部。

2) At high loads, due to the GPU memory allocation threshold set by the inference engine, up to 10%-20% of the GPU memory remains unused and idle. Hence, the current GPU memory management is inefficient;

由于推理过程存在不确定性，所以通常会预留一部分显存，导致 10%~20% 的浪费

3) The prefill and decode stages of the inference process have significantly different demands on GPU memory

prefill 和 decode 阶段的显存需求有显著差别（PD分离架构下，P节点不需要存储 kv cache，D节点需要）

整体架构