Why it matters

EAGLE improves accept rate vs Medusa. Higher throughput for same architecture.

Advertisement

The architecture

Predict from features not tokens.

Tree draft.

Higher accept rate.

EAGLE speculativeFeatures draftnot tokensTree verifylarge treeHigher accept3x+ speedupLi et al. 2024; higher accept rate than Medusa; production frameworks add support
EAGLE speculative.
Advertisement

How it works end to end

Li et al. 2024.

Feature-based draft.

Larger tree.