Struct PenaltyWindow 

pub struct PenaltyWindow<'a> { /* private fields */ }

The tokens the penalties may see: a prompt and the tokens generated after it, in that order.

Borrowed rather than owned because this is built once per sampled token on every decode loop in the workspace; an owning window would clone the whole sequence per token.

The two halves are kept separate rather than concatenated because a decode loop already holds them separately, and concatenating would mean an allocation per token for a value only ever read back as “the last N of the two”.

Implementations

impl<'a> PenaltyWindow<'a>

pub fn new(prompt: &'a [usize], generated: &'a [usize]) -> Self

The window over prompt followed by generated.

prompt is the tokens the model was fed before generation started, and it belongs in the window: see the module docs for the llama.cpp lines that put it there.

pub fn len(&self) -> usize

Total tokens in the sequence, before penalty_last_n truncates it.

pub fn is_empty(&self) -> bool

pub fn recent(&self, last_n: usize) -> impl Iterator<Item = usize> + '_

The most recent last_n tokens of prompt ++ generated, oldest first.

This is llama.cpp’s ring buffer expressed as a view: the buffer holds at most penalty_last_n entries and the oldest is dropped on every accept (src/llama-sampler.cpp:2707-2716), so its contents are exactly the tail of the accepted sequence.

Order does not matter to any caller – only the multiset does – but it is the natural one anyway, and a test reads it.

Trait Implementations

impl<'a> Clone for PenaltyWindow<'a>

fn clone(&self) -> PenaltyWindow<'a>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<'a> Copy for PenaltyWindow<'a>

impl<'a> Debug for PenaltyWindow<'a>

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

Formats the value using the given formatter.