model-exporter 0.1.1

Universal model weight exporter for multiple formats and inference engines

model-exporter

Export HuggingFace models to different formats for inference.

Installation

pip install model-exporter

Supported Models

• GPT-2 family: gpt2, gpt2-medium, gpt2-large, gpt2-xl, distilgpt2

Supported Formats

• binary - Simple binary format for C++ inference
• safetensors - Fast and safe tensor format

Usage

Export a model

model-exporter export MODEL_NAME --format FORMAT --output OUTPUT_DIR

Options:

• --format - Export format (binary or safetensors)
• --output - Output directory (default: ./weights)
• --info - Print model info before export
• --no-vocab - Skip vocabulary export

Examples

# Export with model info
model-exporter export gpt2-medium --format binary --output ./my_weights --info

# Export without vocabulary
model-exporter export distilgpt2 --format binary --no-vocab

Python API

from model_exporter import get_exporter

# Create exporter
exporter = get_exporter('gpt2')

# Load model
exporter.load_model()

# Export
exporter.export(output_dir='./weights', format_name='binary')

Work in Progress

We're actively working on adding support for:

Model Families

• LLaMA/LLaMA-2 - Meta's LLaMA models
• BERT - BERT and variants (RoBERTa, DistilBERT)
• T5/FLAN-T5 - Encoder-decoder models
• Mistral/Mixtral - Mistral AI models

Export Formats

• GGUF - llama.cpp compatible format with quantization
• ONNX - Cross-platform inference format
• TensorFlow Lite - Mobile inference format

Features

• Quantization - int8, int4 quantization support
• Model sharding - Split large models across multiple files
• Batch export - Export multiple models at once

Want to contribute? Check out our GitHub repository!

License

MIT