alice-llm 2.0.0

Pure Rust LLM inference engine — GGUF v3 parser, K-quants, hybrid CPU/GPU DeltaNet+Attention, wgpu compute shaders, speculative decoding, OpenAI-compatible HTTP server.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

//! Grammar-constrained decoding (Phase X.8 Guided Generation).
//!
//! Opt-in via the `grammar` feature. Provides the building blocks for
//! constraining LLM sampling to follow a formal grammar so that the model can
//! only emit strings recognised by the target language (JSON schema, an
//! embedded DSL such as ALICE-LOL, tool-call payloads, etc.).
//!
//! Currently implemented:
//!
//! - [`gbnf`] (B-1) — llama.cpp-compatible GBNF parser (subset).
//! - [`fsm`] (B-2) — `Grammar` → finite-state machine driving constrained
//!   decoding.
//! - Sampling integration (B-3) — [`crate::sampling::mask_logits_by_grammar`]
//!   (naïve, per-token probe; reference implementation).
//! - [`trie`] (B-10) — [`TokenTrie`] sharing FSM work across tokens with a
//!   common prefix; [`crate::sampling::mask_logits_by_grammar_trie`] is the
//!   production mask and what `generate_grammar` uses.

pub mod fsm;
pub mod gbnf;
pub mod trie;

pub use fsm::{CharSet, Fsm, FsmError, DEFAULT_MAX_DEPTH};
pub use gbnf::{parse_gbnf, Alternative, CharClass, GbnfError, Grammar, Symbol};
pub use trie::TokenTrie;