Research map / Architectures and efficiency / Quantization
FrameQuant: Flexible Low-Bit Quantization for Transformers
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 FrameQuant, a post-training quantization method that quantizes transformer weights in a fusion frame (redundant tight frame) representation, enabling ~2-bit (and 2.x-bit via redundancy) weight quantization with small accuracy loss.
Problem
Post-training quantization of transformers degrades sharply at very low bit widths (around 2 bits), and existing methods only jump between integer bit widths, offering no fine-grained trade-off.
Approach
Compute a redundant fusion-frame representation of each layer's weight matrix (projected onto tight fusion frame subspaces with the previous layer's frame), clip at 2 sigma, and quantize the transformed weights with a modified GPTQ; redundancy factor r gives fractional bit widths (e.g., 2.2 bits at r=1.1) and frame robustness reduces quantization noise impact.
Key results
- On ImageNet, ViT-B at 2.2 bits (r=1.1) reaches 80.93% top-1 vs 85.10% full precision and 79.53% for GPTQ-based FrameQuant at 2 bits; QuIP with 2 bits gets 77.54%.
- Llama2-7B perplexity on WikiText2 drops to 8.48 at 2.2 bits vs 5.68 full precision and 26.02 for QuIP at 2 bits.
- On MNLI, BERT quantized by FrameQuant at 2.2 bits reaches 80.02% accuracy, above ZeroQuant at 3.66 bits (54.91%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase redundancy from r=1.0 to r=1.1 (ablation) | FrameQuant · ImageNet-1K val, ViT-S | Top-1 accuracy: 48.10 -> 61.51 | Helped |
| increase redundancy from r=1.0 to r=1.1 (ablation) | FrameQuant · ImageNet-1K val, ViT-B | Top-1 accuracy: 79.53 -> 80.93 | Helped |
| add TFF representation to GPTQ (ablation) | GPTQ · ImageNet-1K val, ViT-S | Top-1 accuracy: 0.4 -> 0.88 | Helped |
| add 2σ clipping to GPTQ + TFF (ablation) | FrameQuant · ImageNet-1K val, ViT-S | Top-1 accuracy: 0.88 -> 48.10 | Helped |
| apply 2σ clipping directly to GPTQ weights rather than to Fusion Frame representations (ablation) | GPTQ · WikiText2, Llama2-7B | Perplexity: 6.40e3 -> 9.45e3 | Hurt |
| increase redundancy from r=1.0 to r=1.1 (ablation) | FrameQuant · C4, Llama2-7B | Perplexity: 19.62 -> 11.23 | Helped |
| quantize activations to 8 bits rather than leave them at 32 bits (ablation) | FrameQuant W2, ViT-S · ImageNet-1K val | Top-1 accuracy: 48.17 -> 48.02 | No clear effect |
| quantize activations to 6 bits rather than leave them at 32 bits (ablation) | FrameQuant W2.2, ViT-S · ImageNet-1K val | Top-1 accuracy: 61.51 -> 58.35 | Hurt |
| use FrameQuant instead of GPTQ | Llama2-7B · inference | Inference speed (t/s): 1425.07 -> 974.20 | Hurt |
Surprising findings
- Clipping weights at 2 sigma directly (GPTQ with clipping) degrades Llama2 perplexity badly, whereas clipping in the fusion-frame representation helps.
- One-bit quantization with redundancy up to 1.8x was unsuccessful according to the authors.
Limitations
- Inference is slower than GPTQ due to additional frame transform operations.
- Gains are largest at very low bit widths; at higher bits (W6/A6, W4/A4) gains over RepQ-ViT are marginal.
- Redundancy increases effective bits (r=1.1 means 2.2 bits), so the 2.x bit advantage relies on fractional overhead.
Experimental setup
Datasets: ImageNet-1K WikiText2 C4 MNLI ARC BoolQ HellaSwag PIQA WinoGrande
Benchmarks: ImageNet-1K WikiText2 C4 MNLI ARC BoolQ HellaSwag PIQA WinoGrande COCO object detection and segmentation
Metrics: top-1 accuracy perplexity AP box/mask inference tokens/sec model size
Baselines: PTQ4ViT GPTQ QuIP ZeroQuant RepQ-ViT
Builds on: GPTQ QuIP PTQ4ViT fusion frame theory
Closest papers by meaning
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.90)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.88)
- Outlier-aware Slicing for Post-Training Quantization in Vision Transformer (ICML 2024 · similarity 0.88)
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.87)
- ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · similarity 0.87)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (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.