Research map / Architectures and efficiency / Attention and transformers
Transformer Quality in Linear Time
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 FLASH instead of Transformer++ | C4 masked language modeling · C4, context 8192 | training latency: 252.1 -> 59.9 | Helped |
| Use FLASH instead of Transformer++ | Wiki-40B language modeling · Wiki-40B, context 8192 | PPLX: 15.254 -> 15.109 | Helped |
| Use FLASH mixed chunk attention instead of quadratic Transformer++ attention | Wiki-40B language modeling · Wiki-40B, context 8192 | training latency: 261.0 -> 62.5 | Helped |
| Use FLASH instead of Transformer++ | PG-19 language modeling · PG-19, context 8192 | PPLX: 43.26 -> 41.07 | Helped |
| Replace relu² attention with softmax (ablation) | GAU · LM and MLM | PPLX: 16.78 / 4.23 -> 17.04 / 4.31 | Hurt |
| Replace single-head attention with multi-head attention (ablation) | GAU · LM and MLM | PPLX: 16.78 / 4.23 -> 17.76 / 4.48 | Hurt |
| Add gating (ablation) | GAU · LM and MLM | PPLX: 17.45 / 4.58 -> 16.78 / 4.23 | Helped |
| Use head size 512 instead of 128 (ablation) | FLASH-Quad · TriviaQA validation, context 4096 | F1: 72.7 -> 74.8 | Helped |
| Use mixed chunk attention instead of full quadratic attention (ablation) | GAU / FLASH · C4 masked language modeling | PPLX increase: 0.0, 0.05, 0.06, 0.07, 0.07 | Mixed |
| Use mixed chunk attention instead of full quadratic attention (ablation) | Transformer++ / MC-TFM++ · C4 masked language modeling | PPLX increase: 0.36, 0.37, 0.49, 0.48, 0.43 | Hurt |
| Keep both local quadratic and global linear attention, rather than either component alone (ablation) | FLASH · Language modeling | quality | Helped |
| Use larger chunks, such as 512 instead of 256, when context exceeds 1024 (ablation) | FLASH · Long-context language modeling | quality | Helped |
Closest papers by meaning
- Viewing Transformers Through the Lens of Long Convolutions Layers (ICML 2024 · similarity 0.89)
- Understanding the Expressive Power and Mechanisms of Transformer for Sequence Modeling (NeurIPS 2024 · similarity 0.89)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.88)
- PolySketchFormer: Fast Transformers via Sketching Polynomial Kernels (ICML 2024 · similarity 0.88)
- Gated Linear Attention Transformers with Hardware-Efficient Training (ICML 2024 · similarity 0.88)
- Cached Transformers: Improving Transformers with Differentiable Memory Cachde (AAAI 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.