Distributed Speculative Decoding (DSpec)

One request, N identical workers, one GPU each. Every worker drafts a different candidate continuation, verifies its own drafts with its target model, and the fleet commits the strand that reached furthest — every round, every worker advances at the pace of the best draft.
committed draft (per-worker) accepted by target bonus token (target sample) rejected forced (committed by a peer) ←/→ keys work too