blt-tokenizer 0.2.0

High-performance byte-level tokenizer with BPE support

BLT Python Bindings

High-performance Python bindings for BLT (Byte-Level Tokenizer), providing fast byte-level tokenization with BPE support.

🚀 Features

• High Performance: Rust-powered tokenization with multi-threading support
• BPE Support: Byte-Pair Encoding with custom merge rules
• Memory Efficient: Memory-mapped I/O and configurable memory usage
• Easy to Use: Simple Python API with comprehensive error handling
• Cross-Platform: Works on Linux, macOS, and Windows

📦 Installation

pip install blt-tokenizer

Development Installation

# Clone the repository
git clone https://github.com/jtrefon/blt.git
cd blt/blt_python

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install maturin and build
# On Linux
pip install maturin[patchelf]

# On macOS or Windows
pip install maturin

# Build and install for development
maturin develop

# Install development dependencies
pip install -e .[dev]

🔧 Usage

Basic File Tokenization

import blt

# Create a tokenizer
tokenizer = blt.ByteTokenizer()

# Tokenize a file
tokenizer.tokenize_file("input.txt", "output.bin")

BPE Tokenization

import blt

# Load BPE merges from file
merges = blt.load_bpe_merges("merges.txt")

# Or define merges manually
merges = {
    (97, 98): 256,   # 'a' + 'b' -> token 256
    (99, 100): 257,  # 'c' + 'd' -> token 257
    (101, 102): 258, # 'e' + 'f' -> token 258
}

# Create tokenizer with BPE merges
tokenizer = blt.ByteTokenizer(merges=merges)
tokenizer.tokenize_file("input.txt", "output.bin")

Advanced Configuration

import blt

# Create tokenizer with custom settings
tokenizer = blt.ByteTokenizer(
    content_type="Text",    # "Text" or "Bin"
    threads=4,              # Number of processing threads
    chunk_size="1MB",       # Chunk size for processing
    memory_cap=50           # Memory usage cap (0-100%)
)

tokenizer.tokenize_file("large_file.txt", "output.bin")

📖 API Reference

ByteTokenizer

Main tokenizer class for byte-level tokenization.

Constructor

ByteTokenizer(
    merges=None,        # Dict[Tuple[int, int], int] - BPE merge rules
    content_type=None,  # str - "Text" or "Bin"
    threads=None,       # int - Number of threads
    chunk_size=None,    # str - Chunk size (e.g., "1MB")
    memory_cap=None     # int - Memory cap percentage (0-100)
)

Methods

• tokenize_file(input_path, output_path): Tokenize a file and save results
  • input_path (str): Path to input file
  • output_path (str): Path to output file
  • Raises: RuntimeError, IOError

Utility Functions

• load_bpe_merges(path): Load BPE merges from file

  • Returns: Dict[Tuple[int, int], int]
  • Raises: IOError, ValueError

• version(): Get library version

  • Returns: str

🧪 Testing

# Run tests
python -m pytest tests/

# Run with coverage
python -m pytest tests/ --cov=blt

# Run benchmarks
python -m pytest tests/ -k benchmark

⚡ Performance

The Python bindings maintain the same high performance as the CLI version:

File Size	Processing Time	Throughput
1MB	~0.4ms	~2.5 GB/s
10MB	~3.8ms	~2.6 GB/s
100MB	~38ms	~2.6 GB/s

🔍 Examples

See the examples/ directory for comprehensive usage examples:

• basic_usage.py - Basic tokenization examples
• bpe_example.py - BPE tokenization with custom merges
• performance_test.py - Performance benchmarking

🐛 Troubleshooting

Common Issues

ImportError: No module named 'blt'

# Ensure package is installed
pip install blt-tokenizer

# Or for development
maturin develop

RuntimeError during tokenization

# Check file permissions and paths
# Reduce memory usage if needed
tokenizer = blt.ByteTokenizer(memory_cap=50)

Performance issues

# Increase thread count and chunk size
tokenizer = blt.ByteTokenizer(threads=8, chunk_size="16MB")

📄 License

This project is licensed under the Apache License, Version 2.0. See LICENSE for details.

🤝 Contributing

Contributions are welcome! Please see the main CONTRIBUTING.md for guidelines.

🔗 Links

• Main Repository
• Documentation
• Issue Tracker
• PyPI Package