README.md

metadata

title: Neural Network Quantizer
emoji: ⚡
colorFrom: indigo
colorTo: purple
sdk: docker
pinned: false
license: mit
app_port: 7860

Neural Network Weight Quantizer

Quantize neural network weights to lower precision formats (INT8, INT4, NF4) with interactive visualizations.

Features

• 🔢 Multi-bit quantization (4-bit, 8-bit)
• 📊 Interactive weight visualizations
• 🤗 HuggingFace model support (optional)
• ⚡ GPU acceleration (when available)
• 📈 Quantization error analysis
• 🔄 Method comparison (INT8 vs INT4 vs NF4)

Quick Start

1. Use the Quantizer tab to test on random weights
2. Compare different methods in the Analysis tab
3. Optionally load a HuggingFace model in the Models tab

API

The backend exposes a REST API at /api:

• GET /api/system/info - System capabilities
• POST /api/quantize/weights - Quantize custom weights
• POST /api/models/load - Load HuggingFace model
• POST /api/analysis/compare - Compare methods

🚀 Deployment

Hugging Face Spaces

This project is configured for Hugging Face Spaces using the Docker SDK.

1. Create a new Space on Hugging Face.
2. Select Docker as the SDK.
3. Push this repository to your Space:

   git remote add space https://huggingface.co/spaces/YOUR_USERNAME/YOUR_SPACE_NAME
   git push space main
   

Docker

Run locally with Docker:

docker build -t quantizer .
docker run -p 7860:7860 quantizer

Open http://localhost:7860.