Struct Sampler 

pub struct Sampler { /* private fields */ }

Stateful sampler: owns the RNG + the recent-token history (for penalties).

Implementations

impl Sampler

pub fn new(cfg: SamplerConfig) -> Self

pub fn is_greedy(&self) -> bool

pub fn is_spec_sampling(&self) -> bool

Sampled spec in its FASTEST regime: pure temperature, no truncation filters, no penalties. Filters and penalties are also distribution-exact under the rejection verify (spec.rs applies both symmetrically to draft q and target p), so they remain spec-ELIGIBLE — see spec_eligible in memra-server’s worker, the authoritative predicate. What they cost is the in-graph draft chain: the captured sampled draft samples from the RAW softmax and can hold neither per-row filter stats nor a varying penalty history, so spec.rs engages graph_s only in this pure-temp regime (pure_temp) and otherwise falls back to the eager draft chain. This predicate names that regime; it is NOT an eligibility test.

pub fn top_k(&self) -> usize

pub fn penalty_last_n(&self) -> usize

pub fn penalty_repeat(&self) -> f32

pub fn penalty_freq(&self) -> f32

pub fn penalty_present(&self) -> f32

pub fn top_p(&self) -> f32

pub fn min_p(&self) -> f32

pub fn temperature(&self) -> f32

pub fn seed(&self) -> u64

pub fn accept(&mut self, token: u32)

Record an emitted token so subsequent penalties see it.

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

Sample the next token id from raw logits [n_vocab]. Does NOT mutate logits in place beyond a local copy. Returns the chosen token id. (Caller should accept() it afterwards.)