Read the original at HF Daily Papers
SpecOPD is a branch-parallel on-policy distillation framework that accelerates large language model inference by parallelizing draft sampling and target supervision across independent branches.
Carried by: HF Daily Papers. First seen: .