Why it matters

Medusa avoids draft model deployment. Growing adoption.

Advertisement

The architecture

Target model + K medusa heads.

Each head predicts token at position +N.

Verify accept prefix.

Medusa structureBase modelpredicts token 1Medusa headspredict 2, 3, 4Tree verifykeep valid prefixTrained on target model + heads; no separate draft to deploy
Medusa flow.
Advertisement

How it works end to end

Heads: N MLPs on final hidden state.

Training: fine-tune heads on model outputs.

Tree attention: multiple candidates.