NeuralZip: Reusable Setup for Fast Lossless Compression
NeuralZip reuses a precomputed exponent setup to losslessly compress model checkpoints up to 21.33 times faster than baselines.
NeuralZip prepares the statistical structure of floating-point exponents once and reuses it for later lossless compression of model weights. It groups chunks with similar exponent distributions, shares Huffman codes, and packs recurring exponent tuples. After setup, checkpoint compression is 1.81–21.33× faster than baselines with exact bit-to-bit reconstruction, and the setup can transfer across compatible architectures as weights evolve. GPU experiments reduce active memory use by up to 27.5% while reproducing logits exactly.
- Shared Huffman codes and packed exponents add moderate compression
- Post-setup encoding is 1.81–21.33× faster with exact reconstruction
- Setup transfers across compatible architectures and evolving checkpoints
- GPU runs cut active memory by up to 27.5% with identical logits
From the original article
1 min read at arxiv.orgLossless compression can reduce the storage and movement of model weights without changing their floating-point values, but repeated statistical analysis and code construction add computational overhead. We study whether the statistical structure of exponents can be prepared once and reused.
By Martín Bravo · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.