laya-core
The tensor-free core of laya-rs: everything on the CPU side of Laya typed decisions, with no candle dependency.
A gateway that only needs to decide where a request should go — or to build the prompt it would send — can depend on this crate without pulling in a GPU runtime.
What is in here
| Module | Purpose |
|---|---|
prompt |
Sequence construction: prefixes, [MASK] option markers, state budgeting |
question |
The Choice / Score / Noul question types and their parsing |
lang |
Script detection and the Latin-script language guess |
presets |
Upstream's built-in question presets |
calibrate |
Post-hoc temperature calibration and confidence |
email |
Quoted-history and signature stripping for email states |
tokenizer |
Thin wrapper over tokenizers, with the special-token ids Laya needs |
Parity
The tables in lang_tables.rs and preset_tables.rs are generated from the upstream
Python package rather than transcribed by hand, and the crate is tested against reference
fixtures produced by running that package. CI regenerates both and fails if they drift.
License
Apache-2.0. See NOTICE for
attribution: upstream Laya is by Convai Innovations, and the MLX port this is based on is
mizorewww/laya-mlx. Neither endorses this project.