# Disaggregated Serving: Prefill/Decode Ayrımı — Mooncake + DistServe

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-disaggregated-serving-prefill-decode
> Updated: 2026-08-14T19:02:11.857Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part XV — Serving Engineering
**TLDR:** Modern LLM serving'in en yeni trend'i (2024-2026): prefill (input encoding) ve decode (token generation) farklı GPU'larda. Prefill compute-bound, decode memory-bound — ayrımı %30-50 throughput artırır. Mooncake (Kimi), DistServe (UCB) reçeteleri. RTX 4090 multi-GPU senaryosunda kavramsal.