Research map / Architectures and efficiency / Quantization
To Compress or Not? Pushing the Frontier of Lossless GenAI Model Weights Compression with Exponent Concentration
Zeyu Yang, Tianyi Zhang, Jianwen Xie, Chuan Li, Zhaozhuo Xu, Anshumali Shrivastava
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| losslessly compress FP8 weights with ECF8 (Huffman-coded exponents) | FP8 LLMs · Weight memory | memory reduction: 9.8% to 14.8% | Improved |
| losslessly compress FP8 weights with ECF8 | FP8 DiTs · Weight memory | memory reduction: 14.1% to 26.9% | Improved |
| use ECF8 memory savings to enlarge batch size under fixed memory | DeepSeek-R1-0528 · 640 GB budget | throughput: 1.55 -> 3.88 tokens/s (150.3%) | Improved |
| use ECF8 memory savings to enlarge batch size under fixed memory | DeepSeek-R1-0528 · 640 GB budget | per-request latency: 660.65 -> 263.95 s | Improved |
| use ECF8 under fixed memory | Llama-3.3-70B-Instruct-FP8-dynamic · 80 GB budget | throughput: 41.28 -> 45.96 tokens/s | Improved |
| use ECF8 under fixed memory | Qwen3-8B-FP8 · 12 GB budget | throughput: 208.80 -> 235.22 tokens/s | Improved |
| use ECF8 with DiffSynth VRAM management | FLUX.1-dev · DiffSynth GH200 | E2E latency: 24.29 -> 13.15 s | Improved |
| use ECF8 with DiffSynth VRAM management | Qwen-Image · DiffSynth GH200 | E2E latency: 111.14 -> 49.05 s | Improved |
| use ECF8 with DiffSynth VRAM management | Wan2.1-T2V-14B · DiffSynth GH200 | E2E latency: 476.21 -> 460.67 s | Improved |
| use ECF8 with DiffSynth VRAM management | Wan2.2-T2V-A14B · DiffSynth GH200 | E2E latency: 480.45 -> 461.41 s | Improved |
| ECF8 compression on generated outputs | Qwen-Image · Generated images, fixed seed | output fidelity: pixel-identical | No clear effect |
Closest papers by meaning
- 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11) (NeurIPS 2025 · similarity 0.89)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.86)
- MeLLoC: Lossless Compression with High-order Mechanism Learning (NeurIPS 2024 · similarity 0.85)
- Unified Scaling Laws for Compressed Representations (NeurIPS 2025 · similarity 0.85)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.85)
- SAFE-SVD: Sensitivity-Aware Fidelity-Enforcing SVD for Physics Foundation Models (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.
