Research map / Architectures and efficiency / Attention and transformers
Faster Than Flash: Exploiting Attention Sparsity for Efficient Long-Context Decoding
Zhigeng Liu, Zhiyuan Ning, Ruixiao Li, Xiaoran Liu, Yuerong Song, Min Zhang, Ziwei He, Xipeng Qiu
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 top-k with top-delta selection (FP16 scan, split) (ablation) | FFD · Ablation, 16k prefill + 128 decode, RTX 4090 RULER | RULER AVG: 86.82 -> 90.39 | Improved |
| replace top-k with top-delta selection (FP16 scan, split) (ablation) | FFD · Ablation, 16k prefill + 128 decode, RTX 4090 | throughput (tok/s): 17.43 -> 21.64 | Improved |
| add 2-bit scanning under top-k selection (ablation) | FFD · Ablation, 16k prefill + 128 decode, RTX 4090 | throughput (tok/s): 17.43 -> 14.80 | Hurt |
| add 2-bit scanning under top-delta selection (ablation) | FFD · Ablation, 16k prefill + 128 decode, RTX 4090 | throughput (tok/s): 21.64 -> 24.65 | Improved |
| add fused kernel and CUDA-graph execution under top-k (ablation) | FFD · Ablation, 16k prefill + 128 decode, RTX 4090 | throughput (tok/s): 14.80 -> 21.21 | Improved |
| add fused kernel and CUDA-graph execution under top-delta (ablation) | FFD · Ablation, 16k prefill + 128 decode, RTX 4090 | throughput (tok/s): 24.65 -> 53.96 | Improved |
| use top-delta instead of top-k at matched keep ratio (ablation) | FFD · Per-head selection latency | latency (ms/head): 0.0090 -> 0.0044 | Improved |
| use top-delta instead of top-p at matched keep ratio (ablation) | FFD · Per-head selection latency | latency (ms/head): 0.1115 -> 0.0044 | Improved |
| use top-delta instead of fixed-budget top-k on CWE (split execution) (ablation) | FFD · RULER 32k CWE | score: 24.6 -> 68.2 | Improved |
| use 2-bit + 8-bit residual instead of 8+8 quantization (ablation) | FFD · Decode kernel latency, RTX 4090 16k | speedup vs FlashAttn: 2.82 -> 10.65 | Improved |
| raise delta from 5 to 7 (ablation) | FFD · RULER 32k | average score: 87.1 -> 89.4 | Improved |
| raise delta from 5 to 7 (ablation) | FFD · LongBench | average score: 25.86 -> 26.35 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Quest (0.5) — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 7 of 7 shared results
- Twilight — Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning (NeurIPS 2025) · better on 3 of 4 shared results
- FlashAttention-2 — FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022) · better on 1 of 1 shared result
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 0.89)
- TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention (ICLR 2025 · similarity 0.89)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.89)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.88)
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.88)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · 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.
