Research map / Architectures and efficiency / Quantization
PackQViT: Faster Sub-8-bit Vision Transformers via Full and Packed Quantization on the Mobile
Peiyan Dong, LEI LU, Chao Wu, Cheng Lyu, Geng Yuan, Hao Tang, Yanzhi Wang
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 log2 attention quantization (ablation) | Q-ViT baseline · ImageNet DeiT-T 8-bit | top-1: 73.6 -> 74.1 | Improved |
| add log2 attention quantization (ablation) | Q-ViT baseline · ImageNet DeiT-T 4-bit | top-1: 72.1 -> 72.4 | Improved |
| add outlier-aware training (ablation) | Q-ViT baseline · ImageNet DeiT-T 8-bit | top-1: 73.6 -> 74.2 | Improved |
| add outlier-aware training (ablation) | Q-ViT baseline · ImageNet DeiT-T 4-bit | top-1: 72.1 -> 72.5 | Improved |
| add Int-2n-Softmax (integer softmax) (ablation) | Q-ViT baseline · ImageNet DeiT-T 8-bit | top-1: 73.6 -> 73.5 | Hurt |
| add Int-2n-Softmax (integer softmax) (ablation) | Q-ViT baseline · ImageNet DeiT-T 4-bit | top-1: 72.1 -> 72.0 | Hurt |
| add Int-LayerNorm (ablation) | Q-ViT baseline · ImageNet DeiT-T 8-bit | top-1: 73.6 -> 73.6 | No clear effect |
| combine all components with integer-only flow (ablation) | Q-ViT baseline · ImageNet DeiT-T 8-bit | top-1: 73.6 -> 74.5 | Improved |
| combine all components with integer-only flow (ablation) | Q-ViT baseline · ImageNet DeiT-T 4-bit | top-1: 72.1 -> 72.7 | Improved |
| full quantization of Softmax/LayerNorm/GELU at 4-bit (ablation) | INT4 DeiT-T on mobile CPU · Speedup vs FP32 | speedup: 3.6x -> 4.9x | Improved |
| full quantization at 4-bit (ablation) | INT4 Swin-T on mobile CPU · Speedup vs FP32 | speedup: 2.8x -> 3.8x | Improved |
| switch attention precision from 8 to 4 bits | PackQViT · ImageNet (avg) | top-1: less than 0.1% | No clear effect |
Closest papers by meaning
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.93)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.90)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.90)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.90)
- Amortized-Precision Quantization for Early-Exit Vision Transformers (NeurIPS 2026 · similarity 0.89)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · 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.
