Struct SampleOpts 

pub struct SampleOpts {

    pub temperature: f32,
    pub top_p: f32,
    pub top_k: Option<u32>,
    pub repetition_penalty: f32,
    pub dynamic_temp: Option<(f32, f32)>,
    pub dynamic_temp_exponent: f32,
    pub typical_p: f32,
    pub top_n_sigma: f32,
    pub xtc_threshold: f32,
    pub xtc_prob: f32,
    pub dry_multiplier: f32,
    pub dry_base: f32,
    pub dry_allowed_length: usize,
    pub dry_max_ngram: usize,
    pub dry_sequence_breakers: Vec<u32>,
    pub mirostat: MirostatMode,
    pub mirostat_tau: f32,
    pub mirostat_eta: f32,
    pub mirostat_m: usize,
    pub frequency_penalty: f32,
    pub presence_penalty: f32,
    pub repetition_window: usize,
    pub min_keep: usize,
}

Sampling parameters. Greedy when temperature == 0 and no advanced sampler is enabled. All “advanced” knobs default to off / no-op so legacy callers see classic top-k/top-p/temperature behaviour.

into_chain() turns these flat fields into a SamplerChain that downstream backends can execute. Ordering follows llama.cpp’s canonical chain (penalties → temperature → top-k → typical → top-p → top-n-σ → xtc → mirostat).

Fields

temperature: f32

top_p: f32

top_k: Option<u32>

repetition_penalty: f32

dynamic_temp: Option<(f32, f32)>

Dynamic temperature [min, max] gated by softmax entropy. None ⇒ flat temperature only.

dynamic_temp_exponent: f32

Exponent used by crate::samplers::DynamicTemperature.

typical_p: f32

Locally-typical sampling (Meister et al. 2022). 1.0 ⇒ off.

top_n_sigma: f32

Top-n-σ cutoff (Hewitt et al. 2024). 0 ⇒ off.

xtc_threshold: f32

XTC: probability of dropping high-confidence top tokens.

xtc_prob: f32

dry_multiplier: f32

DRY repetition penalty knobs.

dry_base: f32

dry_allowed_length: usize

dry_max_ngram: usize

dry_sequence_breakers: Vec<u32>

mirostat: MirostatMode

Mirostat mode + parameters.

mirostat_tau: f32

mirostat_eta: f32

mirostat_m: usize

frequency_penalty: f32

Frequency / presence penalties (OpenAI-style).

presence_penalty: f32

repetition_window: usize

min_keep: usize

Minimum tokens kept by top-p / typical (avoid one-token nucleus).

Implementations

impl SampleOpts

pub fn greedy() -> Self

pub fn nucleus(temperature: f32, top_p: f32) -> Self

pub fn is_greedy(&self) -> bool

pub fn is_classic(&self) -> bool

True when only classic top-k/top-p/temperature are configured; backends can take a cheap fast path in this case (e.g. the existing sample_row CPU kernel) instead of building a chain.

pub fn into_chain(&self) -> SamplerChain

Build the SamplerChain corresponding to these options. The returned chain is ready to drive SamplerChain::sample against a logits row + history. Greedy decoding produces a chain with one Temperature{t:1e-6} step (which collapses to argmax after softmax) — callers that want true greedy can short-circuit via is_greedy() before building the chain.

Trait Implementations

impl Clone for SampleOpts

fn clone(&self) -> SampleOpts

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SampleOpts

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

Formats the value using the given formatter.

impl Default for SampleOpts

fn default() -> Self

Returns the “default value” for a type.