ECHO: Early-layer Collaborative Hierarchical Orchestration with Bonus Logits in Speculative Decoding

While draft-model-free speculative decoding offers a promising path to efficient LLM inference, it is frequently constrained by stale draft candidates and the high computational cost of the verification. To address these challenges, we propose ECHO, a hierarchical dual-loop framework that exploits t