Why is EAGLE faster than Medusa?
目录
Answer #
Medusa’s extra heads look only at the current hidden state and guess several future tokens in parallel. They never see which token was actually sampled in between, so they average over branches. The Chinese note puts that acceptance ceiling near 0.6.
EAGLE drafts the next feature, not the next token, and concatenates the token just sampled (a shifted-token input). The regression target becomes unique, acceptance goes up, and the first draft token in a round is accepted with probability 1. EAGLE-2 then grows the draft tree from context instead of a static shape. The same chapter notes that vLLM already integrates EAGLE, including an EAGLE-2-style dynamic tree.
What this page is not #
No new tokens/s table. The 0.6 figure and the “first token is always accepted” claim are the chapter’s reading of the papers.