TL;DR

Turn it on. Draft models are no longer needed; n-gram and EAGLE-style heads give 1.6–2.2× on chat-shaped traffic with no accuracy change.

What settled

The technique is now a flag in vLLM, SGLang and TGI. Default configurations are reasonable and the failure mode (slower on very short outputs) is documented.

What to check

Batch size. Speculation helps most at low concurrency; at high batch sizes the gain shrinks toward zero.