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

1. MEDUSA Multi-Head Training#

Target model: Llama 3.1 8B (frozen)
Add N extra lm_head'ler:
  head_1: predict token at t+1 (normal)
  head_2: predict token at t+2 (one step ahead)
  head_3: predict token at t+3
  head_4: predict token at t+4

Training: target model frozen, sadece head'leri train et
  - SFT data üzerinde (Llama-Instruct dataset)
  - Loss = sum(CE(head_i(h_t), token_t+i))

Inference:
  - Target forward 1 kez → 4 candidate token
  - Target verify: kabul ettiği token'lara devam et
  - Accept rate ~%60-75 (typical)
  - Speedup: ~2-3×