Struct SamplingConfig 

pub struct SamplingConfig {
    pub temperature: f32,
    pub top_k: Option<usize>,
    pub top_p: Option<f32>,
    pub min_p: Option<f32>,
    pub repeat_penalty: f32,
    pub repeat_last_n: usize,
    pub seed: u64,
}

Configuration for StochasticSampler’s logit-transform pipeline.

Every threshold is Option-or-plain depending on whether “disabled” is a meaningful state: top_k/top_p/min_p default to None (that stage of the pipeline is skipped entirely), while temperature and repeat_penalty always run (a temperature of 1.0 and a repeat penalty of 1.0 are each that stage’s identity value, so leaving them at their defaults is equivalent to skipping them, without needing a second Option layer).

Fields

temperature: f32

Divides every logit before the final softmax. 1.0 (the default) leaves the distribution’s shape unchanged; values below 1.0 sharpen it toward the top-scoring tokens, values above 1.0 flatten it toward uniform. <= 0.0 is treated as “skip stochastic sampling entirely and fall back to greedy_argmax” — see StochasticSampler::sample’s docs for why that is an exact equivalence, not an approximation.

top_k: Option<usize>

Keep only the k highest-scoring tokens before sampling. None (the default) skips this stage.

top_p: Option<f32>

Nucleus sampling: keep the smallest prefix of tokens (sorted by probability, descending) whose cumulative probability is the first to exceed p. None (the default) skips this stage.

min_p: Option<f32>

Keep only tokens whose probability is at least min_p times the single most likely token’s probability. None (the default) skips this stage.

repeat_penalty: f32

Divides a previously-generated token’s logit by this value if it was positive, or multiplies it if negative — the standard (Keskar et al., 2019, CTRL) repetition penalty. 1.0 (the default) is the identity: no penalty.

repeat_last_n: usize

How many of the most recently generated tokens StochasticSampler remembers for Self::repeat_penalty’s history window.

seed: u64

Seeds StochasticSampler’s internal PRNG — see this module’s docs for why this is mandatory rather than defaulting to system entropy.

Trait Implementations

impl Clone for SamplingConfig

fn clone(&self) -> SamplingConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SamplingConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for SamplingConfig

fn default() -> Self

Returns the “default value” for a type.