Benchmark metrics
Draft acceptance rate
Also known as acceptance rate, speculative acceptance rate, draft acceptance ratio
In plain English
Draft acceptance rate is the fraction of proposed tokens that pass verification, rather than the number emitted by each verification step.
Technical definition
Draft acceptance rate is the number of accepted draft tokens divided by the number of proposed draft tokens over a specified sample.
Relationship
draft acceptance rate = accepted draft tokens / proposed draft tokens
Engineering details
It is a dimensionless fraction, often reported as a percentage. Acceptance length instead counts tokens per verification step. With a fixed draft length and matching counting rules, the accepted-draft length equals draft length times acceptance rate. An engine may include a target-generated bonus token in its reported length, so that convention must be stated separately.
Why it matters
The same acceptance rate can correspond to different numbers of accepted tokens when draft lengths differ. Neither rate nor length alone determines speedup: draft generation, verification cost, batching, and rejected work also matter.
How to read it in InferenceX
AgentX uses externally calibrated acceptance behavior for synthetic replay. Compare the draft length, speculator, counting convention, and forced-versus-natural setting before treating two reported acceptance values as equivalent.
See the concept in real benchmarks
All articlesAgentX - 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