← All questions
Inference & ServingMedium· Speculative Decoding· Medusa· EAGLE

What is the fundamental quality difference between Medu...

Medusa’s architecture relies on multiple heads that predict tokens based only on the target model’s hidden state, meaning each head is unaware of the tokens predicted by preceding heads. This lack of conditional dependency limits the quality of the draft. EAGLE addresses this by implementing an autoregressive draft process, allowing the draft module to account for the dependencies between tokens, which results in higher-quality proposals and better overall performance.

Code playground (Python)