Research map / Architectures and efficiency / Quantization
SINQ: Sinkhorn-Normalized Quantization for Calibration-Free Low-Precision LLM Weights
Lorenz K Muller, Philippe Bich, Jiawei Zhuang, Ahmet Çelik, Luca Benfenati, Lukas Cavigelli
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 |
|---|---|---|---|
| replace HQQ/RTN single-scale with SINQ dual-scale Sinkhorn normalization (3-bit) | Qwen3-1.7B · WikiText2 (3-bit) | perplexity: 32.10 -> 22.39 | Improved |
| replace RTN with SINQ (4-bit) | Qwen3-32B · WikiText2 (4-bit) | perplexity: 8.92 -> 7.74 | Improved |
| replace Hadamard + RTN with SINQ (3-bit) | Qwen3-32B · WikiText2 (3-bit) | perplexity: 11.26 -> 8.79 | Improved |
| add SINQ normalization to AWQ (A-SINQ) at 3-bit vs GPTQ | Qwen3-14B · Avg flip rate (3-bit) | flip rate (%): 8.06 -> 7.56 | Improved |
| add SINQ to uncalibrated 3-bit quantization | Qwen3-14B · Avg flip rate (3-bit) | flip rate (%): 10.07 -> 7.73 | Improved |
| use NF4 levels in SINQ instead of uniform INT4 | SINQ · WikiText2 (4-bit) | perplexity: 8.76 -> 8.72 | Mixed |
| remove the shift (zero-point) from SINQ (ablation) | SINQ · Memory-perplexity Pareto, Qwen3 family | perplexity | Improved |
| vary auxiliary variable precision (int8 vs fp16) (ablation) | SINQ · Memory-perplexity Pareto, Qwen3 family | perplexity | No clear effect |
| use no-overhead SINQ (shared second scale absorbed in prior layer) instead of standard SINQ (ablation) | SINQ · WikiText2 (4-bit) | perplexity: 17.14 -> 17.63 | Hurt |
| apply no-overhead SINQ before GGUF Q3_K_S | Qwen3-0.6B · WikiText2 (GGUF) | perplexity: 35.59 -> 30.20 | Improved |
| add SINQ before AWQ scaling (ablation) | AWQ · Row-wise kurtosis, Qwen3-1.7B | kurtosis: 1.32x reduction | Improved |
| add the extra input-side scale vector at inference | SINQ with gemlite kernel · Batch size 1, D=1024 | latency overhead: 1.8% | No clear effect |
Closest papers by meaning
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.92)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 2024 · similarity 0.91)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.91)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.91)
- QERA: an Analytical Framework for Quantization Error Reconstruction (ICLR 2025 · similarity 0.91)
- OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models (AAAI 2024 · similarity 0.91)
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.
