Research map / Architectures and efficiency / Quantization
S-Quant: Rethinking Weight Quantization with Seed-Based Generation
Mingzi Wang, Lancheng Zou, Shuo Yin, Zhuolun He, Bei Yu
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 |
|---|---|---|---|
| use S-Quant instead of SeedLM at about 3 bits | SeedLM · WikiText-2 Llama 3 8B | perplexity: 10.1 -> 9.7 | Improved |
| use S-Quant instead of AWQ at 3 bits | AWQ · WikiText-2 Llama 2 7B | perplexity: 15.6 -> 6.5 | Improved |
| use S-Quant instead of SeedLM at about 4 bits | SeedLM · WikiText-2 Llama 3 70B | perplexity: 3.8 -> 3.6 | Improved |
| use S-Quant instead of SeedLM at about 4 bits | SeedLM · WikiText-2 Llama 2 7B | perplexity: 5.7 -> 5.7 | No clear effect |
| use S-Quant instead of SeedLM at 3 bits | SeedLM · zero-shot mean Llama 2 7B | mean accuracy: 64.48 -> 64.56 | No clear effect |
| use S-Quant instead of AWQ at 3 bits | AWQ · zero-shot mean Llama 2 7B | mean accuracy: 52.46 -> 64.56 | Improved |
| use S-Quant instead of SeedLM at about 4 bits | SeedLM · zero-shot mean Llama 3 70B | mean accuracy: 78.06 -> 78.11 | No clear effect |
| use S-Quant at 3.8 bit vs FP16 | FP16 baseline · zero-shot mean Llama 2 7B | mean accuracy: 68.74 -> 67.65 | Hurt |
| use adaptive basis count per block via explained energy ratio (ablation) | S-Quant · LLaMA 2-7B blocks | explained energy ratio | Improved |
| vary block size, seed length and explained-energy threshold (ablation) | S-Quant · WikiText-2 LLaMA2-7B | perplexity | Mixed |
| use S-Quant ASIC accelerator instead of uncompressed weights | uncompressed baseline · memory-bound LLM inference | latency: approximately 4x reduction in latency | Improved |
Closest papers by meaning
- LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models (ICLR 2024 · similarity 0.90)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.90)
- Treasures in Discarded Weights for LLM Quantization (AAAI 2025 · similarity 0.90)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.90)
- SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators (ICLR 2025 · similarity 0.90)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.89)
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.
