Speculative Decoding

A small fast model generates N candidate tokens. The large model verifies all N in one forward pass. Describe both roles. Flag format: CONGRESS{small_model:[role],large_model:[role]} Example: CONGRESS{small_model:embed,large_model:generate}