Research map / Architectures and efficiency / Quantization
I-ViT: Integer-only Quantization for Efficient Vision Transformer Inference
Zhikai Li, Qingyi Gu
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 integer-only INT8 quantization with I-ViT instead of FP32 | DeiT-S · ImageNet | top-1 accuracy: 79.85 -> 80.12 | Improved |
| use I-ViT INT8 instead of FP32 | Swin-S · ImageNet | top-1 accuracy: 83.20 -> 83.01 | Hurt |
| use I-ViT instead of I-BERT polynomial approximation for non-linear ops | Swin-S · ImageNet | top-1 accuracy: 81.86 -> 83.01 | Improved |
| replace ShiftGELU with polynomial GELU (ablation) | I-ViT DeiT-B · ImageNet | top-1 accuracy: 81.74 -> 80.88 | Hurt |
| replace Shiftmax with polynomial softmax (ablation) | I-ViT DeiT-B · ImageNet | top-1 accuracy: 81.74 -> 81.62 | Hurt |
| replace Shiftmax with LIS (Log-Int-Softmax) (ablation) | I-ViT DeiT-B · ImageNet | top-1 accuracy: 81.74 -> 81.66 | Hurt |
| replace I-LayerNorm with L1 LayerNorm (ablation) | I-ViT DeiT-B · ImageNet | top-1 accuracy: 81.74 -> 79.25 | Hurt |
| replace I-LayerNorm with L1 LayerNorm (ablation) | I-ViT DeiT-B · RTX 2080Ti latency | latency (ms): 7.93 -> 7.91 | No clear effect |
| use shifting-oriented Shiftmax/ShiftGELU instead of polynomial versions (ablation) | I-ViT DeiT-B · RTX 2080Ti latency | latency (ms): 8.10 -> 7.93 (ShiftGELU vs Poly) | Improved |
| integer-only inference instead of keeping non-linear ops in floating point | I-ViT vs FasterTransformer · RTX 2080Ti latency (DeiT-B) | speedup: x3.72 -> x4.11 | Improved |
Closest papers by meaning
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.89)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.88)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.88)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.87)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 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.
