Research map / Architectures and efficiency / Attention and transformers
SALE : Low-bit Estimation for Efficient Sparse Attention in Long-context LLM Prefilling
Xiaodong Ji, Hailin Zhang, Fangcheng Fu, Bin CUI
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 |
|---|---|---|---|
| use SALE sparse attention instead of FlashAttention2 | Llama-3.1-8B-Instruct · LongBench | average score: 48.77 -> 48.39 | No clear effect |
| use SALE sparse attention instead of FlashAttention2 | Llama-3.1-8B-Instruct · InfiniteBench | average score: 37.75 -> 37.11 | No clear effect |
| use SALE sparse attention instead of FlashAttention2 | Qwen3-4B-Instruct · InfiniteBench | average score: 31.85 -> 32.09 | Improved |
| use SALE sparse attention instead of FlashAttention2 | Llama-3.1-8B-Instruct · RULER | average score: 83.86 -> 84.67 | Improved |
| use SALE instead of FlashAttention2 | Llama-3.1-8B-Instruct · attention latency at 64K | speedup: 1.00× -> 3.36× | Improved |
| add per-head threshold calibration (ablation) | SALE · InfiniteBench (128K latency) | score | Improved |
| increase context length | SALE · single input speedup | speedup: 1.99× at 16K -> 3.88× at 128K | Improved |
| use SALE on short inputs (under about 8K tokens) | SALE · short-context prefill | latency | No clear effect |
| use one-pass selection+computation kernel instead of two-pass (ablation) | SALE · kernel design | workload | Hurt |
| use 4-bit instead of higher-bit quantization for estimation (ablation) | SALE · block selection | selection accuracy and speed | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FlexPrefill — FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.93)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.92)
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.92)
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.92)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.92)
- SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference (NeurIPS 2026 · similarity 0.91)
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.
