tch-rs VS tokenizers

I wanted to do something like this but then I would miss on proper CUDA acceleration and lose performance compared to using torchlib.

I wrote a forgettable llama implementation for https://github.com/LaurentMazare/tch-rs (pytorch's torchlib rust binding).

I first encountered OCaml's PyTorch bindings because apparently they generate a C wrapper around PyTorch's C++ API, and Rust's PyTorch bindings use OCaml's C wrapper. See: https://github.com/LaurentMazare/tch-rs

You could try looking at the min-GPT example of tch-rs. I'd also strongly suggest watching Karpathy's video to understand what's going on.

If you pefer to see it in code there's a succint gpt implementation here https://github.com/LaurentMazare/tch-rs/blob/main/examples/m...

Rust is fully compatible w/ C bindings, so even Python libraries written in C can be easily set up to work in Rust (and have been). For example, see PyTorch Rust bindings, which actually works faster than in Python because all of the glue code around the C++ API is in Rust instead of Python.

Pytorch

I have looked into this a bit and found some crates which help in loading torchscript models called tch-rs

Can you also compare the performance with https://github.com/huggingface/tokenizers/? Would be helpful, since the benchmark in the tiktoken readme seems to be very outdated.

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