Research map / Architectures and efficiency / Quantization

GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers

NeurIPS 2025 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GPLQ, a two-stage ViT quantization method that first trains only activation quantizers (FP32 weights, 1 epoch, PCA feature-mimicking loss) and then applies PTQ (RepQ-ViT + QwT) to weights, reaching near-FP32 W4A4 generalization at a fraction of QAT cost.

Problem

Low-bit (4-bit) quantization of Vision Transformers: PTQ suffers large accuracy drops, while QAT is extremely expensive, unstable, and generalizes poorly to downstream tasks, and lacks open-source code.

Approach

Sequential "activation-first, weights-later" quantization: activation-only QAT with FP32 weights for one epoch, guided by a PCA-subspace feature-mimicking loss to stay in the FP32 optimization basin, followed by weight-only PTQ with activations frozen at 4 bits.

Key results

• GPLQ W4A4 Swin-T reaches 79.5% ImageNet top-1 and 70.77% average FGVC accuracy, versus 81.2% and 70.79% for FP32, while RepQ-ViT W4A4 gets 73.0% and 62.79%.
• GPLQ Act-QAT takes 35 minutes on Swin-T versus roughly 12 days for Q-Var/OFQ QAT, with 6.6 GB peak memory versus 7.2 GB for FP32 training.
• On COCO with Mask R-CNN, GPLQ W4A4 Swin-T (3x) achieves 0.401 AP box versus 0.361 for RepQ-ViT and 0.460 for FP32.

Measured effects reported in the paper

Surprising findings

• QAT models with the highest ImageNet accuracy (Q-Var 82.4%) show internal FP32 accuracy dropping to 68.9%, suggesting they leave the original optimization basin and generalize worse downstream.
• Quantizing activations to 4 bits hurts accuracy more than quantizing weights to 4 bits under PTQ.

Limitations

• Comparison with QAT methods is limited because few QAT methods are open-source and they were only run on small models.
• Second-stage performance depends on existing PTQ methods and inherits their limitations.
• Evaluation focuses on 4-bit quantization; lower bit widths are only briefly explored in the appendix.

Experimental setup

Datasets: ImageNet-1k COCO 2017 Aircraft Food101 Flowers102 Pets Cars

Benchmarks: ImageNet-1k COCO 2017 Aircraft Food101 Flowers102 Pets Cars

Metrics: Top-1 accuracy AP box AP mask average downstream accuracy training time peak memory

Baselines: OFQ Quantization-Variation RepQ-ViT QwT PTQ4ViT APQ-ViT DeiT-B

Builds on: TCS LSQ RepQ-ViT QwT Mask R-CNN

Closest papers by meaning

• AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.92)
• ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · similarity 0.91)
• GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.91)
• APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.90)
• RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.90)
• VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.90)