Research map / Architectures and efficiency / Quantization
NUPES: Non-Uniform Post-Training Quantization via Power Exponent Search
Edouard Yvinec, Arnaud Dapogny, Kévin Bailly
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 learned quantized weight values (NUPES learn W) to PowerQuant (ablation) | PowerQuant · COCO RetinaNet W4/A4 | mAP: 2.241 -> 32.692 | Improved |
| add learned per-layer exponent a on top of learned weights (ablation) | NUPES (learn W) · ImageNet EfficientNet B0 W4/A4 | accuracy: 18.132 -> 45.902 | Improved |
| add learned per-layer exponent a on top of learned weights (ablation) | NUPES (learn W) · ImageNet ViT h14 W4/A4 | accuracy: 87.190 -> 87.204 | No clear effect |
| add learned exponent only (no weight learning) to PowerQuant (ablation) | PowerQuant · ImageNet ResNet 50 W4/A4 | accuracy: 62.142 -> 62.468 | Improved |
| combine PowerQuant naively with AdaRound rounding learning (ablation) | PowerQuant · ImageNet ResNet 50 W4/A8 | accuracy: 74.892 -> 72.384 | Hurt |
| combine PowerQuant naively with AdaRound rounding learning (ablation) | PowerQuant · COCO RetinaNet W4/A8 | mAP: 21.562 -> 20.346 | Hurt |
| add numerical stability clipping in exponent gradient (ablation) | exponent learning · ImageNet ResNet 50 W4/A4 | accuracy: - -> 46,552 | Improved |
| add update of scale outside gradient descent (ablation) | exponent learning with balanced grads · COCO RetinaNet W4/A4 | mAP: 4.128 -> 22.424 | Improved |
| add balanced gradient averaging of weight and input contributions (ablation) | exponent learning with update scale · ImageNet ResNet 50 W4/A4 | accuracy: 61.997 -> 62.468 | Improved |
| use constant beta=20 scheduler for soft rounding instead of AdaRound or power schedulers (ablation) | NUPES · ImageNet and COCO | accuracy | Improved |
| add group-wise quantization (group 128) to PowerQuant at 3 bits | PowerQuant · OPT 13B common sense average W3/A16 | average accuracy: 46.448 -> 54.472 | Improved |
| use PowerQuant instead of DFQ for data-free W4/A16 quantization | DFQ · Dolly v2 3B common sense average | average accuracy: 47.921 (-4.182) -> 51.622 (-0.481) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PowerQuant — PowerQuant: Automorphism Search for Non-Uniform Quantization (ICLR 2023) · better on 8 of 8 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
- PowerQuant: Automorphism Search for Non-Uniform Quantization (ICLR 2023 · similarity 0.90)
- Don’t Waste Your Bits! Squeeze Activations and Gradients for Deep Neural Networks via TinyScript (ICML 2020 · similarity 0.89)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.88)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.87)
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.87)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · 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.
