Can you modify speculative decoding to work without a s...

Methods like Medusa eliminate the need for a separate draft model by adding multiple prediction heads to the target model. Each head predicts a token at a specific lookahead position based on the current hidden state. Because these heads do not see the full autoregressive history of the draft, they are less accurate than a standalone model. EAGLE improves upon this by feeding the target model’s features into a small, autoregressive draft module, which recovers most of the lost acceptance rate while maintaining low memory overhead.