Research map / Architectures and efficiency / Attention and transformers
Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing
Yelysei Bondarenko, Markus Nagel, Tijmen Blankevoort
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 clipped softmax with gamma<0 | BERT-base · BERT Wikipedia val | W8A8 perplexity: 1294 -> 4.52 | Improved |
| add gated attention | BERT-base · BERT Wikipedia val | W8A8 perplexity: 1294 -> 4.65 | Improved |
| add gated attention | OPT-125m · Wikipedia val | W8A8 perplexity: 21.18 -> 16.02 | Improved |
| add clipped softmax | OPT-125m · Wikipedia val | W8A8 perplexity: 21.18 -> 37.20 | Hurt |
| add clipped softmax | ViT-S/16 · ImageNet-1K val | W8A8 accuracy: 69.24 -> 79.77 | Improved |
| add gated attention | OPT-1.3B · Wikipedia val | W8A8 perplexity: 989.6 -> 29.95 | Improved |
| use gamma<0 in clipped softmax (clip at zero) (ablation) | BERT-base · BERT Wikipedia val | W8A8 perplexity: 1294 -> 4.55 | Improved |
| use zeta>1 only (clip at one) (ablation) | BERT-base · BERT Wikipedia val | W8A8 perplexity: 1294 -> 1469 | No clear effect |
| add zeta>1 on top of gamma<0 (ablation) | BERT-base · BERT Wikipedia val | W8A8 perplexity: 4.55 -> 4.56 | No clear effect |
| initialize gate bias for very low open probability (ablation) | BERT-6L and ViT · pretraining | FP and quantized performance | Hurt |
| add LayerNorm after patch embeddings (ablation) | ViT-S/16 vanilla · ImageNet-1K val | W8A8 accuracy: 69.24 -> 79.62 | Improved |
| add weight decay on LayerNorm gamma (ablation) | OPT-125m vanilla · Wikipedia val | W8A8 perplexity: 21.18 -> 39.46 | Mixed |
Closest papers by meaning
- From Attention to Activation: Unraveling the Enigmas of Large Language Models (ICLR 2025 · similarity 0.93)
- Outlier-Efficient Hopfield Layers for Large Transformer-Based Models (ICML 2024 · similarity 0.91)
- Systematic Outliers in Large Language Models (ICLR 2025 · similarity 0.89)
- Understanding and Minimising Outlier Features in Transformer Training (NeurIPS 2024 · similarity 0.88)
- SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration (ICLR 2025 · similarity 0.88)
- Rethinking Channel Dimensions to Isolate Outliers for Low-bit Weight Quantization of Large Language Models (ICLR 2024 · similarity 0.88)
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.
