Research map / Architectures and efficiency / Quantization
Float8@2bits: Entropy Coding Enables Data-Free Model Compression
Patrick Putzky, Martin Genzel, Mattes Mollenhauer, Sebastian Schulze, Thomas Wollmann, Stefan Dietzel
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EntQuant instead of HQQ at 2 bits | LLaMA-2 70B · C4 | perplexity: 32.24 -> 6.47 | Improved |
| use EntQuant instead of HQQ at 3 bits | LLaMA-2 70B · C4 | perplexity: 6.02 -> 5.74 | Improved |
| use EntQuant (data-free) instead of calibration-based GPTQ at 2 bits | LLaMA-2 70B · zero-shot average | accuracy: 34.4 -> 68.6 | Improved |
| use EntQuant instead of QuIP# at 2 bits | LLaMA-2 70B · zero-shot average | accuracy: 70.9 -> 68.6 | Hurt |
| compress below 2 bits (1.7 bits) (ablation) | EntQuant · LLaMA-3.1 8B C4 | perplexity: 9.74 (3-bit) -> 135.24 (1.7-bit) | Hurt |
| use block-wise decompression instead of layer-wise decoding (ablation) | EntQuant inference · inference speed | speed: approximately 50% speed-up | Improved |
| exclude layers with super weights (ablation) | EntQuant Int8 · LLaMA-2 7B C4 | perplexity | Improved |
| exclude layers with super weights (ablation) | EntQuant Float8 · LLaMA-2 7B C4 | perplexity | Improved |
| exclude layers with super weights (ablation) | HQQ at 2 bits · LLaMA-2 7B C4 | perplexity | No clear effect |
| use Int8 base format instead of Float8 (ablation) | EntQuant · LLaMA-2 7B C4 | perplexity | No clear effect |
| add dynamic 8-bit activation quantization (W8A8) (ablation) | EntQuant 3-bit · LLaMA-2 70B C4 | perplexity: 5.74 -> 7.35 | Hurt |
| add dynamic 8-bit activation quantization (W8A8) (ablation) | EntQuant 3-bit · LLaMA-2 7B C4 | perplexity: 7.55 -> 7.76 | Hurt |
Closest papers by meaning
- MeLLoC: Lossless Compression with High-order Mechanism Learning (NeurIPS 2024 · similarity 0.86)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.86)
- Minimum Entropy Coupling with Bottleneck (NeurIPS 2024 · similarity 0.85)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 2025 · similarity 0.85)
- BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments (ICLR 2025 · similarity 0.85)
- Fix the Structural Bottleneck: Context Compression via Explicit Information Transmission (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
