maturin VS tokenizers

This story unfolds as a captivating journey where the agile Flounder, representing the Python programming language, navigates the vast seas of coding under the wise guidance of Sebastian, symbolizing Rust. Central to their adventure are three powerful tridents: cargo, PyO3, and maturin.

-- Maturin on GitHub

My new favorite way to write very fast libraries for Python is to just use Rust and Maturin:

https://github.com/PyO3/maturin

It basically automates everything for you. If you use it with Github actions, it will compile wheels for you on each release for every platform and python version you want, and even upload them to PyPi (pip) for you. Everything feels very modern and well thought out. People really care about good tooling in the Rust world.

Python first, you will be able to experiment quickly with the notebooks. Then maybe write (or rewrite) some modules in Rust that you can expose as python modules, with py03 and maturin. Feel free to publish useful packages on both crates.io and pypi.org, so you can contribute to Python and Rust ecosystems.

Now if you really want to use Rust, you can rewrite only the part that are slowing down your consumer. It's easy by using Py03 and maturin. Maybe also rayon to parallelize.

It's relatively easy to extend Python with project like Py03[0] and Maturin[1]. Polars[2] is the perfect example of that.

It's not easy to push coworkers/companies to use an unfamiliar language. Rust isn't fast to learn. You need very good arguments and a good usecase to make it works.

I doubt that learning Rust will help you more that learning more about the data engineers tools, so this isn't really "worth" your time.

[0] -- https://pyo3.rs/v0.18.3/

[1] -- https://github.com/PyO3/maturin

[2] -- https://www.pola.rs/

In this case, PyO3/maturin does all the setup and getting the module into Python. They also have docs going into a lot more depth on this.

Lastly if you're willing to introduce Rust, I'd consider a gradual approach using native libraries built in rust with PYO3. Check the maturin guide that helps you to streamline the build process of native libraries : https://github.com/PyO3/maturin . From there you could try to find hotspots in your python app and replace those with a native implementation.

Hugging Face seems to like Rust. They also wrote Tokenizers in Rust.

Your intuition is right. There are two ways (in increasing order of result performance) : 1. You can simply extend vocab file of the tokenizer and test the predictions 2. You can extend the vocab file and re-train your model on custom data which has these new tokens. Check the following issue on GitHub : https://github.com/huggingface/tokenizers/issues/247

SentencePiece -> implementation of some algorithms (there are several others, https://github.com/microsoft/BlingFire https://github.com/glample/fastBPE https://github.com/huggingface/tokenizers )

In sum I would like the idea to go with Rust as I more or less got to rewrite the whole thing anyway, but I am a bit skeptical if I will be able to interface with everything that might come up at some point. Or probably end up in a wrapper hell if I got to use more C++ libraries. On the other hand there are definitely a few Rust projects out there that might come in handy (for example https://github.com/huggingface/tokenizers). And the build process is pretty awful right now (CMake it is but with lots of hacks).

A full machine learning pipeline usually comprises far more than just the model, and this is the area where Rust may shine (the recent work by HuggingFace and their https://github.com/huggingface/tokenizers library is a good example)

As for other tokenizers, you can take a look at - Huggingface tokenizers library: https://github.com/huggingface/tokenizers - NLTK tokenize: https://www.nltk.org/api/nltk.tokenize.html - Polygot: https://pypi.org/project/polyglot/

Huggingface Tokenizers (https://github.com/huggingface/tokenizers), which are now used by default in their Transformers Python library, use pyO3 and became popular due to the pitch that it encoded text an order of magnitude faster with zero config changes.

It lives up to that claim. (I had issues with return object typing when going between Python/Rust at first but those are more consistent now)

huggingface/tokenizers (Rust): 💥Fast State-of-the-Art Tokenizers optimized for Research and Production