Struct StochasticSampler 

pub struct StochasticSampler { /* private fields */ }

Temperature / top-k / top-p / min-p / repetition-penalty sampling, composed as a pipeline of logit transforms (see this module’s docs) and driven by a seeded [Rng].

Pipeline order

StochasticSampler::sample applies, in this order: repetition penalty (needs the raw logit scale — see [apply_repetition_penalty]’s docs) -> top-k -> top-p -> min-p -> temperature -> softmax -> weighted-random draw. Top-k/top-p/min-p run before temperature deliberately: top-p and min-p compute a softmax internally to rank tokens by probability, and temperature changes how flat or peaked that probability distribution is — running them after temperature would make the nucleus/threshold decision depend on a rescaling that has not been “committed to” yet. This is the same ordering llama.cpp’s sampler chain uses.

Implementations

impl StochasticSampler

pub fn new(config: SamplingConfig) -> Self

Trait Implementations

impl Sampler for StochasticSampler

fn sample(&mut self, logits: &[f32]) -> u32