# Speculative Decoding FT: Draft Model + EAGLE-2 + MEDUSA Head Training

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-speculative-decoding-ft-eagle-medusa
> Updated: 2026-08-22T19:16:33.850Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part XIII — Custom Kernels & Performance Surgery
**TLDR:** Speculative decoding'in FT versiyonu: draft model'i target ile pair'le, kabul oranını maksimize et. EAGLE-2 head training (Li et al. 2024, +%94 throughput), MEDUSA multi-head training, target model frozen tutarak ek head'ler eğitme. RTX 4090 + Llama 8B target + MEDUSA 4-head ~2-3 saat training.