laghu-rs

The optional Rust core of laghu. laghu loads model2vec static embedding models with far less memory than model2vec and gives the same vectors. With laghu-rs installed, laghu runs normalization, pre-tokenization, the Unigram Viterbi search and mean pooling in Rust threads, over the same memory-mapped files. It also tokenizes BERT WordPiece models (a BertNormalizer, a BertPreTokenizer and a WordPiece model) in Rust, with a port of FastWordPiece from model2vec-rs, MIT, copyright (c) 2025 The Minish Lab (LICENSE-model2vec-rs).

pip install laghu-rs     # or: pip install 'laghu[rust]'

Nothing else changes: laghu finds laghu-rs and uses it for the models it supports. Output is identical, bit for bit, to laghu's pure-Python path and to model2vec 0.9.

500-character chunks on the 14 cores of an M3 Max, best of three runs of laghu.bench.speed, measured on 2026-10-24 at a load average of 6.5 to 9. potion-multilingual-128M (Unigram):

potion-base-8M (WordPiece):

The figure laghu-rs 0.1.0 gave for model2vec, 5,276 chunks/s, was measured with tokenizers on one core.

laghu falls back to pure Python when laghu-rs is not installed or LAGHU_RUST=0 is set. It keeps Python's normalization and pre-tokenization when the tokenizers crate inside laghu-rs is not the same version as the installed Python tokenizers, or when the two disagree on a set of probe texts. It tokenizes a WordPiece model with tokenizers under the same two conditions, and for any pipeline the port does not reproduce.

laghu-rs needs Python 3.11 or later. Building it from source needs Rust and a C compiler.