Research map / Architectures and efficiency / Attention and transformers
FSA: An Alternative Efficient Implementation of Native Sparse Attention Kernel
Ran Yan, Youhe Jiang, Zhuoming Chen, Haohui Mai, Beidi Chen, Binhang Yuan
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 |
|---|---|---|---|
| invert kernel loop order (loop KV blocks outer, queries inner) | NSA kernel · small GQA groups (g<8) | kernel latency: up to 3.5x, avg 1.6x lower latency (abstract) | Improved |
| use FSA instead of NSA kernel | end-to-end training · Llama3-8B, Qwen3-14B, Qwen2.5-32B | training latency: avg 1.09x, up to 1.25x | Improved |
| use FSA instead of NSA kernel | inference prefill · LLMs, 32K-64K | prefill latency: avg 1.11x, up to 1.36x | Improved |
| use FSA instead of NSA kernel | decoding · LLM inference | decoding latency | No clear effect |
| use FSA at GQA group size 8 | kernel · H20 prefill 128K, GQA=8 | latency: 0.45 vs 0.45 s | No clear effect |
| disable inner-loop optimization (ablation) | FSA kernel · kernel benchmarks | performance: avg 11.9% drop, up to 18.9% | Hurt |
| disable early return optimization (ablation) | FSA kernel · kernel benchmarks | performance: avg 18.2% drop, up to 25.2% | Hurt |
| separate reduction kernel and online softmax kernel | FSA · kernel design | memory access and FLOPs: FSA lower than NSA for g<=8 | Improved |
| larger GQA group | FSA vs NSA kernel speedup · kernel latency | speedup: gap widens for g in {1,2}; gone at g=8 | Mixed |
Closest papers by meaning
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.90)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.90)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.90)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.89)
- Block Sparse Flash Attention (NeurIPS 2026 · similarity 0.89)
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.
