Acceptance length
Also known as acceptance rate, draft acceptance
In plain English
Acceptance length is how many drafted tokens the full model actually approves per verification step, which is what decides whether speculation pays off.
Technical definition
Acceptance length is the average number of speculatively drafted tokens accepted by the target model in one verification pass.
Engineering details
Speculative decoding only saves time when drafts survive verification. An acceptance length near one means the draft and verify machinery ran for nothing, while a high value amortizes one expensive target-model step across several emitted tokens. The value depends on the speculator, the draft length, the model, and the content being generated.
Why it matters
Because acceptance depends on content, a benchmark can accidentally decide the result. Synthetic or anonymized text is out of distribution for a speculator trained on real language, so measured acceptance drifts away from what production would see, in either direction.
How to read it in InferenceX
AgentX replays anonymized traces filled with synthetic tokens, so it does not let acceptance emerge from that content. Runs instead apply a fixed acceptance length collected per model, speculator, draft length, and thinking mode on an external agentic coding dataset, which keeps the comparison vendor neutral.
Source material
See the concept in real benchmarks
AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?
$3 Million USD dataset open sourced, 1 Mil+ Context Length, Multiturn, Sub Agents 95%+ KVCache HitRate, GB300 NVL72, MI355X, B200
InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX
GB300 NVL72, MI355X, B200, H100, Disaggregated Serving, Wide Expert Parallelism, Large Mixture of Experts, SGLang, vLLM, TRTLLM
DeepSeekV4 1.6T Day 0 to Day 43 Performance Over Time — Huawei, GB300 NVL72, MI355X, B200
Day 0 Inference Performance, InferenceX, 100x performance improvement in 26 Days, Cost per Million Tokens, Huawei 950DT Inference Trace Analysis