Struct GrammarSampler 

pub struct GrammarSampler { /* private fields */ }

A grammar being applied to one generation, over one vocabulary.

Implementations

impl GrammarSampler

pub fn new( grammar: Grammar, n_vocab: usize, piece_of: impl Fn(usize) -> Vec<u8>, is_eog: impl Fn(usize) -> bool, ) -> Self

Snapshot n_vocab token pieces and their end-of-generation flags, and start applying grammar over them.

pub fn vocab_size(&self) -> usize

The vocabulary size this was built over.

pub fn grammar(&self) -> &Grammar

The live grammar, for a caller that wants to ask it something.

pub fn mask_logits( &self, logits: &mut [f32], ) -> Result<MaskOutcome, ConstraintError>

Set every logit the grammar forbids to -inf, in place.

llama_grammar_apply_impl. A logit already at -inf is left alone and never offered to the grammar: it is forbidden whatever the grammar thinks, and the walk in reject_candidates is linear in the candidates it is given.

pub fn accept(&mut self, token: usize) -> Result<(), ConstraintError>

Advance the grammar over a sampled token.

llama_grammar_accept_impl. An end-of-generation token does not consume characters – it asserts the parse is finished – so it goes to Grammar::accept_eog, which refuses if it is not.

The order of the two tests is upstream’s and matters: the trigger check comes FIRST, so while a lazy grammar is awaiting, even an end-of-generation token is buffered rather than asserted against a parse that has not started.

pub fn allows_eog(&self) -> bool

Whether a parse is complete, so generation may end here.

True throughout an untriggered lazy grammar: it has not been applied, so it has no say in when generation ends.

pub fn is_awaiting_trigger(&self) -> bool

Whether this is a lazy grammar that has not switched on yet.

A caller deciding whether it needs full vocabulary logits must NOT read this as “unconstrained”: the trigger can fire on any token, so the grammar needs a real logit vector from the first one. It is for reporting and for tests.