Research map / Architectures and efficiency / Attention and transformers
TileSparse: Arithmetic-Intensity-Aware Sparse Attention for Compute-Bound LLM Decoding
Chao Wang, Pengfei Zuo, Zhangyu Chen, Qihui Zhou, Tsung-Yi Ho, Ming-Chang Yang
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 |
|---|---|---|---|
| replace K-only sparsity with QK 2D Sparsity (prune Q-heads per key) under the same compute budget | TileSparse on DeepSeek-V3-0324 · RULER NIAH, low budget | accuracy: 40% accuracy gain in low-budget settings | Improved |
| use tiered configuration (critical tier full heads, ambient tier fewer heads) instead of single-tier QK sparsity (ablation) | TileSparse · RULER, 65,536-entry budget | accuracy | Improved |
| tiered QK 2D sparsity instead of K-only sparsity (x=512) (ablation) | TileSparse · RULER, 65,536-entry budget | accuracy | Improved |
| add dynamic AutoTuner over static configurations (ablation) | TileSparse · RULER, 65,536-entry budget | accuracy | Improved |
| vary the number of critical tokens at full heads (ablation) | TileSparse · RULER, 65,536-entry budget | accuracy | Mixed |
| TileSparse instead of ArkVale to reach 99% of full-attention accuracy | TileSparse · RULER | compute budget reduction: 18.3% | Improved |
| TileSparse instead of ArkVale to reach 99% of full-attention accuracy | TileSparse · LongBench | compute budget reduction: 40.8% | Improved |
| TileSparse instead of ArkVale attention operator | vLLM Triton MLA operator on A100 · LongBench | end-to-end attention speedup: 3.33× -> 3.99× | Improved |
| run AutoTuner every N=10,000 steps | TileSparse · decode overhead | amortized overhead: ~0.01% | No clear effect |
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.
- ArkVale — ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction (NeurIPS 2024) · better on 2 of 2 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
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.91)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.91)
- TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention (ICLR 2025 · similarity 0.91)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.90)
- Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction (NeurIPS 2025 · similarity 0.90)
- Loki: Low-rank Keys for Efficient Sparse Attention (NeurIPS 2024 · similarity 0.90)
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.
