Acceptance profiles¶

How far speculative decoding advances a request per verification step. The registry holds measured acceptance lengths with their provenance; a workload, a scenario or the command line may state their own; and a run resolves exactly one of them and records where it came from.

The quantity¶

The canonical value is the mean acceptance length (AL): the tokens a request gains per verification step, counting the target's bonus token. For K speculative tokens it lies in [1, K + 1]. It is what InferenceX's golden curves publish and what the engines take (vLLM synthetic_acceptance_length, SGLang SGLANG_SIMULATE_ACC_LEN).

Two other quantities are derived from it, and each has one name here:

The acceptance rate is the accepted fraction of drafts, (AL − 1) / K. It is vLLM's logged "Avg Draft acceptance rate" and inference-sim's --speculative-acceptance-rate. The two agree only when every step drafts all K: with a drafter that verifies fewer on some steps, such as DSpark under enable_adaptive_verification, vLLM's logged rate (accepted over drafted tokens, vllm/v1/spec_decode/metrics.py:107-110) divides by the drafts actually proposed and diverges from (AL − 1) / K. The per-token probability p of the speculative-decoding papers is a different quantity and is never called a rate in this schema.
Per-position rates p₁..p_K follow vLLM's synthetic_acceptance_rates: p_i is the unconditional probability that the first i drafts are all accepted. They are non-increasing, each lies in [0, 1], and they sum to AL − 1.

No source in scope publishes per-position rates, so they are optional everywhere. A consumer that needs a distribution derives one with MinimumVarianceRates, which is vLLM v0.31.0's _acceptance_length_to_rates (vllm/config/speculative.py:557-564):

rates = [1.0] * floor(AL - 1) + [frac(AL - 1)] + [0.0] * rest, truncated to K

This is how vLLM, and the InferenceX runs that force a published AL, realise one, so a simulator using it reproduces those runs. Natural acceptance spreads more than this schedule does, so per-request tails derived from it are optimistic.

Acceptance is not the kernel's input. K and the method stay in the deployment's speculative block, because they change what a step costs; acceptance changes only how far requests advance.

Profile set¶

A registry file under acceptance/, named by its file. This one carries real values from InferenceX's golden SPEED-Bench curves:

kind: AcceptanceProfileSet
name: inferencex-golden-al
caveats:
  - Coding only. Every curve is SPEED-Bench Qualitative coding; other categories have none.
  - Measured on B300. Reuse on another part is recorded by the resolver.
  - DeepSeek-V4.1-Flash with DSpark at K 6-8 was refused by the image, so it has no value.
profiles:
  - model: deepseek-v4.1-flash
    speculative_method: dspark
    num_spec_tokens: 5
    thinking: "off"
    category: coding
    acceptance_length: 4.07
    method: measured
    provenance: &dsv41
      source: https://github.com/SemiAnalysisAI/InferenceX/blob/main/inferencex-e2e/infx/golden_al_distribution/dsv41flash_dspark.yaml
      run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/34493175056
      benchmark: SPEED-Bench Qualitative
      prompts: 80
      hardware: b300
      engine_image: vllm/vllm-openai:deepseekv41-flash-0909
      tp: 4
      sampling: {temperature: 1.0, max_tokens: 4096}
      settings:
        draft_sample_method: probabilistic
        rejection_sample_method: block
  - model: deepseek-v4.1-flash
    speculative_method: dspark
    num_spec_tokens: 5
    thinking: "on"
    category: coding
    acceptance_length: 3.51
    method: measured
    provenance: *dsv41
  - model: glm-5.3
    speculative_method: mtp
    num_spec_tokens: 3
    thinking: "on"
    category: coding
    acceptance_length: 2.99
    method: copied
    copied_from: glm-5.2-fp8/mtp/k3/thinking-on/coding
    rationale: >
      Provisional. GLM-5.3 shares GLM-5.2's base and layer topology, so GLM-5.2's curve is
      the closest measured reference; replace it with a run on GLM-5.3.
    provenance:
      source: https://github.com/SemiAnalysisAI/InferenceX/blob/main/inferencex-e2e/infx/golden_al_distribution/glm5.3_mtp.yaml
      run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28058352479
      hardware: b300
      sampling: {temperature: 1.0, top_p: 0.95, max_tokens: 4096}

Key	Type	Required	Meaning
`kind`	string	yes	Always `AcceptanceProfileSet`.
`name`	string	yes	The set's identity. By convention it equals the file name, which is what a scenario's `acceptance.profiles` uses.
`caveats`	list of strings	no	What the set does not cover and how far its values travel. No entry may be empty.
`profiles`	list	yes	At least one profile, below.

Profiles¶

A profile is keyed by model, speculative method, K, thinking mode and workload category. Two profiles with one key are an error.

Key	Type	Required	Meaning
`model`	string	yes	The catalog model name, not the source's spelling of it.
`speculative_method`	string	yes	The engine's method name, as a deployment's `engine.speculative.method` states it.
`num_spec_tokens`	integer	yes	K. At least 1.
`thinking`	string	yes	`on` or `off`. One of the thinking modes.
`category`	string	yes	The workload category, such as `coding`. Lower-case letters, digits, `_` and `-`.
`acceptance_length`	number	yes	AL, in [1, K + 1].
`per_position_rates`	list of numbers	no	K rates, each in [0, 1], non-increasing, summing to AL − 1 within 0.01.
`method`	string	yes	How the value was obtained. One of the methods except `not_charged`.
`provenance`	mapping	unless `method` is `assumed`	Where it was measured, below.
`rationale`	string	unless `method` is `measured`	The reasoning behind a value that was not measured.
`copied_from`	string	with method `copied`	The profile it was copied from, as `model/method/k<K>/thinking-<mode>/category`.

`provenance`¶

Key	Type	Required	Meaning
`source`	string	yes	The document the value was read from.
`run`	string	no	The job that measured it, such as a GitHub Actions run.
`benchmark`	string	no	The prompt set, such as `SPEED-Bench Qualitative`.
`prompts`	integer	no	How many prompts it covered. Not negative.
`hardware`	string	with method `measured`	The catalog chip it was measured on. A resolution on other hardware records this.
`engine_image`	string	with method `measured`	The engine image or release that ran.
`tp`	integer	no	The tensor-parallel width it ran at. Not negative.
`sampling`	mapping	with method `measured`	The decoding configuration, below.
`settings`	mapping	no	Other engine settings that bear on acceptance, by name, such as `rejection_sample_method`.

`provenance.sampling`¶

Key	Type	Required	Meaning
`temperature`	number	yes	Not negative. 0 is greedy decoding.
`top_p`	number	no	In (0, 1].
`top_k`	integer	no	At least −1; −1 and 0 disable it.
`max_tokens`	integer	no	Output cap. Not negative.

Inline values¶

A workload's acceptance, a scenario's acceptance.value, and the value a command line supplies all take this form. A value names the K it was stated for; at a deployment that speculates another K it is refused rather than rescaled.

Key	Type	Required	Meaning
`num_spec_tokens`	integer	yes	The K the value was stated for. At least 1.
`acceptance_length`	number	one of three	AL, in [1, K + 1].
`acceptance_rate`	number	one of three	(AL − 1) / K, in [0, 1]. Not with `acceptance_length`.
`per_position_rates`	list of numbers	one of three	K rates, as for a profile. Alone, they state AL as 1 + their sum; with either of the others, they must agree with it.
`source`	string	no	Where the value came from, recorded with the resolution.

Scenario pin¶

The value of a scenario's acceptance key.

acceptance:
  value:
    num_spec_tokens: 3
    acceptance_length: 2.8
    source: synthetic_acceptance_length in the scored run's server log

To use another profile set for the run, and to accept coding's curve for traffic that has none of its own:

acceptance:
  profiles: inferencex-golden-al
  fallback_category: coding

Key	Type	Required	Meaning
`profiles`	string	no	An acceptance-profile set under `acceptance/` in the registry, consulted in place of the default. Not with `value`.
`value`	mapping	no	An inline value, above.
`fallback_category`	string	no	The category to use when the workload's has no curve. The substitution is recorded. Not with `value`.

A pin must set at least one key.

Resolution¶

acceptance.Resolve takes the run's model, method and K, the registry sets, the workload's terms, the scenario's pin and the command line's value, and returns the acceptance length, the per-position rates, and which layer supplied them. The highest layer that states acceptance wins; layers are not merged.

Layer	Supplies acceptance by
`registry`	the profile matching (model, method, K, the workload's thinking mode and category) in the registry sets, the default
`workload`	the workload's own `acceptance` value
`scenario`	the pin's `value`, or the matching profile in the set its `profiles` names
`cli`	the command line's value, such as `--speculative-acceptance-rate` at the deployment's K

registry · workload · scenario · cli

Nothing falls back silently. A lookup is refused when

the workload states no thinking mode, or no category and no fallback was requested;
no profile matches and no fallback_category was requested, or the fallback has no profile either;
two registry sets both match;
the scenario names a set that was not supplied. The registry is not consulted in its place.

A value stated for a K other than the deployment's is refused at every layer. The resolution records the profile set and entry, the cited source, the category a fallback replaced, the hardware a profile was measured on when it differs from the run's, and whether the rates were derived.

Checks¶

A profile set is invalid if

kind is not AcceptanceProfileSet, name is empty, there are no profiles, or a caveat is empty;
two profiles share a key;
a profile has no model or speculative_method, thinking is not on or off, or category is malformed;
num_spec_tokens is less than 1, or acceptance_length is not finite or lies outside [1, K + 1];
per_position_rates has other than K entries, an entry outside [0, 1], rises, or sums more than 0.01 from AL − 1;
method is not recognized or is not_charged, or a field the method requires, per the table above, is missing;
a provenance has no source, a negative count, an empty setting, or sampling with no temperature, a negative one, a top_p outside (0, 1] or a top_k below −1.

An inline value is invalid if num_spec_tokens is less than 1, it states no acceptance, it states both acceptance_length and acceptance_rate, either lies outside its range, or its rates are invalid or disagree with it. A pin is invalid if it sets nothing, sets both profiles and value, sets fallback_category with value, or its value is invalid.

With a deployment in the bundle, every inline value (the scenario's pin, its trace's, its replay's, merged with its catalog entry when the bundle carries one, a catalog replay's, and the workload shape's) must name the K of each pool that speculates. When no pool speculates, a stated value draws a warning.

Source: spec/acceptance.