autotiktokenizer 0.2.0

🧰 The AutoTokenizer that TikToken always needed -- Load any tokenizer with TikToken now! ✨

AutoTikTokenizer

🚀 Accelerate your HuggingFace tokenizers by converting them to TikToken format with AutoTikTokenizer - get TikToken's speed while keeping HuggingFace's flexibility.

Features • Installation • Examples • Supported Models • Benchmarks • Citation

Key Features

• 🚀 High Performance - Built on TikToken's efficient tokenization engine
• 🔄 HuggingFace Compatible - Seamless integration with the HuggingFace ecosystem
• 📦 Lightweight - Minimal dependencies, just TikToken and Huggingface-hub
• 🎯 Easy to Use - Simple, intuitive API that works out of the box
• 💻 Well Tested - Comprehensive test suite across supported models

Installation

Install autotiktokenizer from PyPI via the following command:

pip install autotiktokenizer

You can also install it from source, by the following command:

pip install git+https://github.com/bhavnicksm/autotiktokenizer

Examples

This section provides a basic usage example of the project. Follow these simple steps to get started quickly.

# step 1: Import the library
from autotiktokenizer import AutoTikTokenizer

# step 2: Load the tokenizer
tokenizer = AutoTikTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")

# step 3: Enjoy the Inferenece speed 🏎️
text = "Wow! I never thought I'd be able to use Llama on TikToken"
encodings = tokenizer.encode(text)

# (Optional) step 4: Decode the outputs
text = tokenizer.decode(encodings)

Supported Models

AutoTikTokenizer current supports the following models (and their variants) out of the box, with support for other models to be tested and added soon!

• GPT2
• GPT-J Family
• SmolLM Family: Smollm2-135M, Smollm2-350M, Smollm2-1.5B etc.
• LLaMa 3 Family: LLama-3.2-1B-Instruct, LLama-3.2-3B-Instruct, LLama-3.1-8B-Instruct etc.
• Deepseek Family: Deepseek-v2.5 etc
• Gemma2 Family: Gemma2-2b-It, Gemma2-9b-it etc
• Mistral Family: Mistral-7B-Instruct-v0.3 etc
• BERT Family: BERT, RoBERTa, MiniLM, TinyBERT, DeBERTa etc.

NOTE: Some models use the unigram tokenizers, which are not supported with TikToken and hence, 🧰 AutoTikTokenizer cannot convert the tokenizers for such models. Some models that use unigram tokenizers include T5, ALBERT, Marian and XLNet.

Benchmarks

Benchmarking results for tokenizing 1 billion tokens from fineweb-edu dataset using Llama 3.2 tokenizer on CPU (Google colab)

Configuration	Processing Type	AutoTikTokenizer	HuggingFace	Speed Ratio
Single Thread	Sequential	14:58 (898s)	40:43 (2443s)	2.72x faster
Batch x1	Batched	15:58 (958s)	10:30 (630s)	0.66x slower
Batch x4	Batched	8:00 (480s)	10:30 (630s)	1.31x faster
Batch x8	Batched	6:32 (392s)	10:30 (630s)	1.62x faster
4 Processes	Parallel	2:34 (154s)	8:59 (539s)	3.50x faster

The above table shows that AutoTikTokenizer's tokenizer (TikToken) is actually way faster than HuggingFace's Tokenizer by 1.6-3.5 times under fair comparison! While, it's not making the most optimal use of TikToken (yet), its still way faster than the stock solutions you might be getting otherwise.

Acknowledgement

Special thanks to HuggingFace and OpenAI for making their respective open-source libraries that make this work possible. I hope that they would continue to support the developer ecosystem for LLMs in the future!

If you found this repository useful, give it a ⭐️! Thank You :)

Citation

If you use autotiktokenizer in your research, please cite it as follows:

@misc{autotiktokenizer,
    author = {Bhavnick Minhas},
    title = {AutoTikTokenizer},
    year = {2024},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/bhavnicksm/autotiktokenizer}},
}