RWKV Tokenizer

A fast RWKV Tokenizer written in Rust that supports the World Tokenizer used by the RWKV v5 and v6 models.

Installation

Install the rwkv-tokenizer python module:

$ pip install pyrwkv-tokenizer

Usage

>>> import pyrwkv_tokenizer
>>> tokenizer = pyrwkv_tokenizer.RWKVTokenizer()
>>> tokenizer.encode("Today is a beautiful day. 今天是美好的一天。")
[33520, 4600, 332, 59219, 21509, 47, 33, 10381, 11639, 13091, 15597, 11685, 14734, 10250, 11639, 10080]
>>> tokenizer.decode([33520, 4600, 332, 59219, 21509, 47, 33, 10381, 11639, 13091, 15597, 11685, 14734, 10250, 11639, 10080])
'Today is a beautiful day. 今天是美好的一天。'
>>> tokenizer.encode_batch(["Today is a beautiful day.", " 今天是美好的一天。"])
[[33520, 4600, 332, 59219, 21509, 47], [33, 10381, 11639, 13091, 15597, 11685, 14734, 10250, 11639, 10080]]

Performance and Validity Test

We compared the encoding results of the Rust RWKV Tokenizer and the original tokenizer using the English Wikipedia and Chinese poetries datasets. Both results are identical. The Rust RWKV Tokenizer also passes the original tokenizer's unit test. The following steps describe how to do the unit test:

$ pip install pytest pyrwkv-tokenizer
$ git clone https://github.com/cahya-wirawan/rwkv-tokenizer.git
$ cd rwkv-tokenizer
$ pytest

We did a performance comparison on the simple English Wikipedia dataset 20220301.simple* among following tokenizer:

• The original RWKV tokenizer (BlinkDL)
• Huggingface implementaion of RWKV tokenizer
• Huggingface LLama tokenizer
• Huggingface Mistral tokenizer
• Bert tokenizer
• OpenAI Tiktoken
• The Rust RWKV tokenizer

The comparison is done using this jupyter notebook in a M2 Mac mini. The Rust RWKV tokenizer is around 17x faster than the original tokenizer and 9.6x faster than OpenAI Tiktoken.

We updated the Rust RWKV world tokenizer to support batch encoding with multithreading. We ran the same comparison script from the Huggingface Tokenizers with the additional rwkv tokenizer. The result shows that the rwkv world tokenizer is significantly faster than the Tiktoken and Huggingface tokenizers in all numbers of threads and document sizes (on average, its speed is ten times faster).

*The simple English Wikipedia dataset can be downloaded as jsonl file from https://huggingface.co/datasets/cahya/simple-wikipedia/resolve/main/simple-wikipedia.jsonl?download=true

Tools using this tokenizer

We also created the json2bin application to convert datasets from JSONL format into binidx format, a data format used for training RWKV models. It uses multithreading to scale up the performance and can convert a dataset more than 70 times faster (around 360 MB/s) than the original json2binidx_tool written in Python.

Changelog

• Version 0.9.0
  • Added multithreading for the function encode_batch()
  • Added batch/multithreading comparison
• Version 0.3.0
  • Fixed the issue where some characters were not encoded correctly

This tokenizer is my very first Rust program, so it might still have many bugs and silly codes :-)