Struct GrammarCompiler 

pub struct GrammarCompiler { /* private fields */ }

A tokenizer-bound grammar compiler with an internal cache.

Each compiler instance is tied to one TokenizerInfo; use a separate compiler per vocabulary. Compilation results are cached (when cache_enabled) keyed by the request parameters.

Implementations

impl GrammarCompiler

pub fn new( tokenizer_info: TokenizerInfo, max_threads: usize, cache_enabled: bool, cache_limit_bytes: i64, ) -> Self

Construct a compiler bound to tokenizer_info.

• max_threads — bounds explicit top-k mask prewarming. Lazy matcher lookups still compute only the requested mask. Must be >= 1.
• cache_enabled — whether to cache compiled grammars.
• cache_limit_bytes — memory budget, ENFORCED by LRU eviction on both tiers. -1 means unlimited, which is now an explicit opt-in rather than the serve’s default: unbounded was the shape of the leak in issue #368. The budget splits as upstream does — ~1/3 to the rule-level cache, the remaining ~2/3 to this grammar-level one.

Port of the C++ GrammarCompiler constructor.

pub fn compile_grammar_from_ebnf( &self, ebnf: &str, root_rule_name: &str, ) -> Result<CompiledGrammar, CompileError>

Compile a grammar from an EBNF string. Port of GrammarCompiler::CompileGrammar(ebnf, root_rule_name).

pub fn compile_grammar(&self, grammar: GrammarData) -> CompiledGrammar

Compile an already-parsed GrammarData. Port of GrammarCompiler::CompileGrammar(const Grammar&) — keyed by the grammar’s printed EBNF, as the C++ does.

pub fn compile_builtin_json_grammar( &self, ) -> Result<CompiledGrammar, CompileError>

Compile the builtin “any JSON value” grammar. Port of GrammarCompiler::CompileBuiltinJSONGrammar.

pub fn compile_json_schema( &self, schema: &str, any_whitespace: bool, indent: Option<i32>, separators: Option<(String, String)>, strict_mode: bool, max_whitespace_cnt: Option<i32>, ) -> Result<CompiledGrammar, CompileError>

Compile a JSON schema. Port of GrammarCompiler::CompileJSONSchema.

pub fn compile_structural_tag( &self, structural_tag_json: &str, ) -> Result<CompiledGrammar, CompileError>

Compile a structural tag. Port of GrammarCompiler::CompileStructuralTag.

Delegates the structural-tag-JSON -> GrammarData conversion to the W5 structural_tag module (structural_tag_to_grammar).

pub fn clear_cache(&self)

Clear the internal compiled-grammar cache and the cross-grammar rule-level mask cache. Port of GrammarCompiler::ClearCache.

pub fn cache_size_bytes(&self) -> i64

Approximate bytes held by the cache — the grammar-level compiled grammars plus the cross-grammar rule-level mask cache. Port of GrammarCompiler::GetCacheSizeBytes.

pub fn cache_limit_bytes(&self) -> i64

The configured cache memory limit; -1 means unlimited. Port of GrammarCompiler::CacheLimitBytes.

pub fn tokenizer_info(&self) -> &TokenizerInfo

The tokenizer this compiler is bound to.

pub fn snapshot_identity(&self) -> SnapshotIdentity

The identity an on-disk mask snapshot must match to be usable with this compiler (#918).

pub fn load_mask_snapshot(&self, path: &Path) -> Result<usize, SnapshotError>

Seed the cross-grammar RuleLevelCache from a snapshot written by an earlier process (#918).

Returns the number of masks imported. 0 means a miss — no file, a snapshot from a different tokenizer or build, or a corrupt one — and the compiler simply computes as before. A compiler constructed with cache_enabled == false has no rule cache and always returns 0.

pub fn save_mask_snapshot( &self, path: &Path, max_entries: usize, ) -> Result<usize, SnapshotError>

Persist the cross-grammar RuleLevelCache so the NEXT process starts warm (#918). Returns the number of masks written; at most max_entries, most-recently-used first.

pub fn rule_cache(&self) -> Option<&RuleLevelCache>

The cross-grammar rule-level cache itself, when caching is on. Exposed for the #918 snapshot tests and for callers that want to share one cache across compilers.

pub fn rule_cache_len(&self) -> usize

Entries currently held by the cross-grammar rule-level cache — 0 when caching is disabled. Used by the #918 tests and by the serve path’s warm-up logging.