Research map / Architectures and efficiency / Quantization
PrefixQuant: Eliminating Outliers by Prefixed Tokens for Large Language Models Quantization
Mengzhao Chen, Yi Liu, Jiahao Wang, Yi Bin, Wenqi Shao, Ping Luo
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 |
|---|---|---|---|
| add prefixed outlier tokens in KV cache (dynamic) (ablation) | Llama-3-8B W4A4KV4 with rotation and grid search · WikiText2 | perplexity: 11.70 -> 7.53 | Improved |
| add prefixed outlier tokens in KV cache (static) (ablation) | Llama-3-8B W4A4KV4 static quantization · WikiText2 | perplexity: 141.02 -> 7.93 | Improved |
| add block-wise fine-tuning (static) (ablation) | PrefixQuant-O2 Llama-3-8B W4A4KV4 · WikiText2 | perplexity: 7.93 -> 7.41 | Improved |
| add block-wise fine-tuning (dynamic) (ablation) | PrefixQuant-O1 Llama-3-8B W4A4KV4 · WikiText2 | perplexity: 7.53 -> 7.23 | Improved |
| add Hadamard rotation to RTN (ablation) | Llama-3-8B W4A4KV4 dynamic · WikiText2 | perplexity: 1282.34 -> 24.98 | Improved |
| add grid search initialization (ablation) | Llama-3-8B W4A4KV4 rotated · WikiText2 | perplexity: 24.98 -> 11.70 | Improved |
| use adaptively computed number of prefixed tokens instead of none (ablation) | PrefixQuant Llama-2-7B W4A4KV4 O2 · WikiText2 | perplexity: 17.63 -> 6.01 | Improved |
| use top-N high-frequency tokens instead of random prefix tokens (ablation) | PrefixQuant w/o FT Llama-2-7B · WikiText2 | perplexity: 66.51 -> 6.22 | Improved |
| use top-N tokens instead of only the highest-frequency token (ablation) | PrefixQuant w/o FT Llama-2-7B · WikiText2 | perplexity: 12.07 -> 6.22 | Improved |
| use Pile instead of C4 as fine-tuning dataset (ablation) | PrefixQuant-O2 Llama-3-8B W4A4KV4 · WikiText2 | perplexity: 7.60 -> 7.42 | Improved |
| fine-tune 20 epochs instead of none (ablation) | PrefixQuant-O2 Llama-3-8B W4A4KV4 · WikiText2 | perplexity: 7.93 -> 7.42 | Improved |
| increase fine-tuning tokens from 256x1024 to 1024x1024 (ablation) | PrefixQuant-O2 Llama-3-8B W4A4KV4 · WikiText2 | perplexity: 7.46 -> 7.41 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 5 of 5 shared results
- DuQuant — DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024) · better on 5 of 5 shared results
- SpinQuant — SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · similarity 0.91)
- Rethinking Channel Dimensions to Isolate Outliers for Low-bit Weight Quantization of Large Language Models (ICLR 2024 · similarity 0.91)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 2024 · similarity 0.90)
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.90)
- FlatQuant: Flatness Matters for LLM Quantization (ICML 2025 · similarity 0.90)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.90)
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.
