Research map / Architectures and efficiency / Attention and transformers
Sub-Linear Memory: How to Make Performers SLiM
Valerii Likhosherstov, Krzysztof Marcin Choromanski, Jared Quincy Davis, Xingyou Song, Adrian Weller
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 |
|---|---|---|---|
| reduce chunk size C from full to 64 (memory-efficient gradient) | Performer with memory-efficient gradient · Copying task L=512 | peak GPU memory (GB): 0.0449 -> 0.0374 | Improved |
| reduce chunk size C from full to 64 | Performer with memory-efficient gradient · Copying task L=512 | time per iteration (sec): 0.0474 -> 0.1228 | Hurt |
| reduce chunk size C from full to 256 | Performer with memory-efficient gradient · Penn Treebank L=1024 | peak GPU memory (GB): 0.300 -> 0.231 | Improved |
| reduce chunk size C from full to 256 | Performer with memory-efficient gradient · Penn Treebank L=1024 | time per iteration (sec): 0.1377 -> 0.3060 | Hurt |
| reduce chunk size C from full to 1366 | Performer with memory-efficient gradient · Enwik8 L=4096 | peak GPU memory (GB): 1.513 -> 0.909 | Improved |
| reduce chunk size C from full to 1366 | Performer with memory-efficient gradient · Enwik8 L=4096 | time per iteration (sec): 0.4598 -> 0.8654 | Hurt |
| train with memory-efficient algorithm instead of full gradient, or fine-tune with it | Performer with memory-efficient gradient · Copying, PTB, Enwik8 | learning curves | No clear effect |
| decrease C in gradient computation | Performer with memory-efficient gradient · random init models | relative gradient discrepancy | Hurt |
| use iterative instead of prefix-sum computation | Performer with memory-efficient gradient · Config II-IV | time and memory | Mixed |
Closest papers by meaning
- Understanding the Expressive Power and Mechanisms of Transformer for Sequence Modeling (NeurIPS 2024 · similarity 0.87)
- I/O Complexity of Attention, or How Optimal is FlashAttention? (ICML 2024 · similarity 0.86)
- Gated Linear Attention Transformers with Hardware-Efficient Training (ICML 2024 · similarity 0.85)
- Two Heads are Better than One: Simulating Large Transformers with Small Ones (NeurIPS 2025 · similarity 0.85)
- Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators? (ICLR 2024 · similarity 0.85)
- MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention (NeurIPS 2025 · similarity 0.85)
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.
