The article describes three parallel drafting algorithms—P-EAGLE, DFlash, and DSpark—that improve speculative decoding for large language models by generating multiple draft tokens in a single step rather than sequentially. These approaches decouple drafting latency from the number of tokens speculated, enabling larger speculator models and simplified parameter tuning compared to autoregressive methods like EAGLE-3. The algorithms achieve this by conditioning on the verifier model’s hidden states through different architectural paths, with each implementing distinct training optimizations to manage computational costs.
