async-std VS tokenizers

But maybe check out the discussion here https://github.com/async-rs/async-std/pull/631 or something (the blog post was linked on the end of it)

Have a look at these: https://github.com/async-rs/async-std/tree/main/examples

In fact, there are a lot of crates in Rust where in other programming languages, it would be included in the standard library. Examples are regex, random number generators, additional iterator methods, macros for other collections, num traits, loggers, HTTP libraries, error handling, async runtimes, serialization and deserialization, date and time, and many more.

Two major projects (non std lib but extremely commonly used) stand out in the area of async programming: Async std and Tokio - no doubt familiar to anyone that has turned an eye towards Rust for a second too long. Async architecture in general is likely very familiar to JavaScript programmers but in Rust there are some extra considerations (like ownership of the data that is thrown into an async function). Tokio is fast becoming a heavily supported and road tested async framework, with a thread scheduling runtime "baked in" that has learned from the history of Go, Erlang, and Java thread schedulers.

Ah... https://github.com/tokio-rs/tokio and https://github.com/async-rs/async-std ?

async-std: Basically a Tokio alternative with a few different design decisions.

Go's solution is for the scheduler to notice after a while when a goroutine has blocked execution and to shift goroutines waiting their turn to another thread. async-std pondered a similar approach with tasks, but it proved controversial and was never merged.

This indicates curl, zlib, openssl, and libnghttp2 as well as a bunch of WASM-related things are being dynamically linked into my executable. To resolve this, I looked at the build features exposed by surf and found that it selects the "curl_client" feature by default, which can be turned off and replaced with "h1-client-rustls" which uses an HTTP client backed by rustls and async-std and no dynamically linked libraries. Enabling this build feature removed all -sys dependencies from androidx-release-watcher, allowing me to build static executables of it.

And also, the actual PR never got merged.

Async in rust needs a runtime (aka executor) to run. You can maybe get a better description from the rust docs. As an example, Tokio attempts to provide an interface for a developer that is minimal change to the more common blocking code. So you'd end up putting #[tokio::main] above your main function to spin up the executor and most of the rest of the code is similar to a non-async version with a few sprinkles of .await, which you can see in the hello world for tokio. In contrast, async-std provides a more hands-on/low-level approach. If you are unlucky enough to have libraries that choose different stacks to work on, you'll possibly (probably?) have to handle both.

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