Research map / Architectures and efficiency / Quantization
UniQL: Unified Quantization and Low-rank Compression for Adaptive Edge LLMs
Hung-Yueh Chiang, Chi-Chih Chang, Yu-Chen Lu, Chien-Yu Lin, Kai-Chiang Wu, Mohamed S. Abdelfattah, Diana Marculescu
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 QSVD (quantization-aware decomposition) after PTQ (ablation) | UniQL · Llama-3.1-8B 4-bit 25% pruning, no FT | average accuracy: 55.2% -> 65.0% | Improved |
| add QSVD after PTQ (ablation) | UniQL · Llama-3.1-8B 4-bit 25% pruning, with FT | average accuracy: 60.2% -> 67.7% | Improved |
| add QSVD after PTQ (ablation) | UniQL · Qwen-2.5-7B 4-bit 25% pruning, no FT | average accuracy: 56.1% -> 60.7% | Improved |
| add masked fine-tuning to 25% pruned model (ablation) | UniQL · Llama-3.1-8B FP16 25% pruning | average accuracy: 67.0% -> 69.6% | Improved |
| add masked fine-tuning to 25% pruned model (ablation) | UniQL · Qwen-2.5-7B FP16 25% pruning | average accuracy: 62.1% -> 65.8% | Improved |
| use alpaca instead of wikitext2 for masked fine-tuning (ablation) | UniQL · Llama-3.1-8B calibration ablation | average accuracy: 60.8% -> 65.5% | Improved |
| use alpaca instead of wikitext2 for weight sorting (ablation) | UniQL · Llama-3.1-8B calibration ablation | average accuracy: 65.5% -> 67.7% | Improved |
| use alpaca instead of wikitext2 for pruning ratio allocation (ablation) | UniQL · Llama-3.1-8B calibration ablation | average accuracy: 67.7% -> 68.6% | Improved |
| add fused RoPE kernel (ablation) | UniQL 4-bit, 0% pruning · Llama-3.1-8B A6000 | TPOT (ms): 9.9 -> 9.0 | Improved |
| add 35% pruning to 4-bit model | UniQL · Llama-3.1-8B A6000 | TPOT (ms): 9.0 -> 7.3 | Improved |
| use UniQL instead of MoDeGPT without fine-tuning | UniQL · Qwen-2.5-7B 15% pruning | average accuracy: 52.1% -> 69.1% | Improved |
| use UniQL instead of MoDeGPT without fine-tuning | UniQL · Llama-3.1-8B 15% pruning | average accuracy: 72.4% -> 70.5% | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 6 of 6 shared results
- MoDeGPT — MoDeGPT: Modular Decomposition for Large Language Model Compression (ICLR 2025) · better on 3 of 4 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
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.88)
- Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge (AAAI 2024 · similarity 0.87)
- PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression (NeurIPS 2024 · similarity 0.87)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs (ICML 2024 · similarity 0.87)
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.87)
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.
