DRelay Uses Global Draft Context to Repair Speculative Decoding Errors
October 2, 2026
DRelay improves parallel speculative decoding for LLMs by using a global reader to extract predictive information across an entire draft block. A causal selector uses this global context to perform prefix-aware repairs of candidate selections, preventing single early errors from discarding valid subsequent tokens.
HOW THIS AFFECTS YOU
●
builderImplementing this could increase the effective throughput of speculative decoding by reducing rejected candidate blocks.
●
researcherThis introduces a new method for mitigating the prefix-length limitations of parallel drafting.