[Official FAQ] Corrections and follow-up questions welcome.

Speculative decoding drafts tokens with a small model and verifies with the large one. When supported and well-matched, throughput can rise; when mismatched, it can stall or slow down.

Expectations:

• Not every backend/model pair supports it.
• Draft and target should be related; low acceptance rates kill gains.
• Ignore “2x for everyone” claims—measure on your stack.

Fix quant and GPU offload first; treat speculation as an advanced knob.