Research map / Architectures and efficiency / Attention and transformers
FlashBlock: Attention Caching for Efficient Long-Context Block Diffusion
Zhuokun Chen, Jianfei Cai, Bohan Zhuang
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 |
|---|---|---|---|
| add block-external attention caching | Trado-8B-Thinking (block size 4) · 800k context, batch 128 | throughput (TPS): 312 -> 451 | Improved |
| add block-external attention caching | Trado-8B-Thinking (block size 8) · 800k context, batch 128 | throughput (TPS): 532 -> 674 | Improved |
| add block-external attention caching | Trado-8B-Thinking · GSM8K/MATH500/AIME | accuracy: 93.25 -> 93.12 (GSM8K, block 4) | No clear effect |
| add block-external attention caching | Trado-8B-Thinking (block size 8) · GSM8K | accuracy: 91.74 -> 90.22 | Hurt |
| add block-external attention caching | Trado-8B-Thinking · MBPP | pass@1: 25.60 -> 33.60 (block 4) | Improved |
| add block-external attention caching on top of SparseD (d=20%) | SparseD · GSM8K | accuracy: 34.72 -> 42.68 | Improved |
| add block-external attention caching on top of SparseD (d=20%) | SparseD · HumanEval | pass@1: 23.78 -> 33.54 | Improved |
| add block-external attention caching on top of SparseD (d=40%) | SparseD · GSM8K | accuracy: 84.61 -> 87.26 | Improved |
| add block-external attention caching on top of SparseD | SparseD · L1 distance to full attention, density 20% | L1 distance: 0.0028 -> 0.0008 | Improved |
| add reuse-aware distillation (ablation) | FlashBlock (tau=2) · MATH500 | accuracy: 80.40 -> 85.80 | Improved |
| add reuse-aware distillation (ablation) | FlashBlock (tau=4) · MATH500 | accuracy: 48.40 -> 66.20 | Improved |
| increase update threshold tau from 2 to 4 (ablation) | FlashBlock · MATH500 | accuracy: 85.80 -> 66.20 (with distillation) | Hurt |
Closest papers by meaning
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.91)
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · similarity 0.91)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.89)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.89)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · 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.
