This repository provides an efficient and accurate implementation of NF4 dequantization using Triton. It supports conversion to both FP16 and BF16 formats, and is designed to run on any modern GPU with minimal adjustments.

The implementation is written as a Jupyter Notebook, as it originated as a solution to the Unsloth Puzzle challenges. Despite its educational origin, the code is production-grade and intended to serve as a reference for NF4 dequantization using custom Triton kernels.

• Fast and correct NF4 to FP16/BF16 dequantization
• GPU-accelerated with Triton
• Portable across CUDA-compatible hardware
• Readable and modifiable code in notebook format

This code is free to use and modify for any purpose, provided that proper credit is given to the author.