Struct DraftDist 

pub struct DraftDist { /* private fields */ }

The distribution one drafted position was sampled from, as its complete support: (token id, probability) pairs summing to 1.

Sparse rather than a dense vocabulary-length vector because the distributions drafters actually produce are sparse: a prompt-lookup drafter’s support is a single token, and a real drafter’s is a top-k, not 150k floats per drafted position per step.

Contract

The support must be the distribution the draft token was actually sampled from. Losslessness does not require the drafter to be good, or even sane – the rejection rule corrects any q – but it does require q to be honest. A drafter that truncates its own softmax to a top-k before sampling must report the truncated, renormalised distribution, not the full softmax it started from.

Implementations

impl DraftDist

pub fn deterministic(token: usize) -> Self

A drafter that is certain: all probability on one token. This is what a lookup-table drafter honestly reports – it did not sample, it asserted.

pub fn from_dense(probs: &[f32]) -> Self

The nonzero entries of a dense probability vector.

pub fn from_support(support: Vec<(usize, f32)>) -> Self

Builds from explicit (token, probability) pairs.

pub fn support(&self) -> &[(usize, f32)]

pub fn prob(&self, token: usize) -> f32

q(token), or 0.0 for a token outside the support.

Trait Implementations

impl Clone for DraftDist

fn clone(&self) -> DraftDist

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DraftDist

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

Formats the value using the given formatter.

impl Default for DraftDist

fn default() -> DraftDist

Returns the “default value” for a type.

impl PartialEq for DraftDist

fn eq(&self, other: &DraftDist) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for DraftDist