Draft page

This page is an outline. It describes what will be covered and is not yet complete technical documentation.

Prefill is compute-bound and decode is bandwidth-bound. Running both on the same replica means one phase’s requirements always compromise the other’s, and interference between them shows up directly in tail latency.

Disaggregation separates them, at the cost of moving KV cache between resources.

What you will learn

  • Why colocating prefill and decode causes interference.
  • How a disaggregated deployment routes a request through both stages.
  • What transferring KV cache between stages costs.
  • When the added complexity is justified.

This page is an outline. The subsections below are the planned structure; they are filled in as the handbook is written.

Two phases, two profiles

Draft

Interference in colocated serving

Draft

A disaggregated architecture

Draft

KV cache transfer

Draft

Scaling the stages independently

Draft

When to adopt it

Draft