[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ab-attention-block-sparse-video-diffusion":3,"topics-all":38,"news-related-703ab8f9-4f6c-46fd-8b3c-f53387619cad":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"703ab8f9-4f6c-46fd-8b3c-f53387619cad","更稀疏、更清晰:视频扩散注意力砍掉九成","不训练不改性,港中文(深圳)等七机构的 AB-Attention 在 Wan2.1 与 HunyuanVideo 上跳过 89–92% 注意力计算:端到端最高 2.23 倍加速,PSNR 比最强稀疏基线高最多 4.4 dB;消融显示逆概率重加权一项就值 4 dB 画质。","视频扩散 Transformer 的推理瓶颈从来不在别处:3D 全注意力的开销随序列长度二次方增长,720p、81 帧的规格下,每层 Transformer、每个去噪步都要把这笔账重复支付一遍。10 月 29 日挂上 arXiv 的 [AB-Attention](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39663)(arXiv:2610.39663)来自港中文(深圳)InfRec 团队,联合山东大学、东京大学、复旦、UC 伯克利、港科大与新加坡国立,思路是不训练、不改权重,直接跳过 89–92% 的注意力计算:论文报告端到端最高 2.23 倍加速,PSNR 反而比最强稀疏基线高最多 4.4 dB。\n\n## 三步机制:打分、采样、装填\n\n方法由三个组件构成。Intra-Block Clustering Selector(IBCS)只在每个 query 块内部做 query 聚类,用带权簇心给所有 key\u002Fvalue 块打分,不置换 token、不做全局聚类,把打分开销压到很低。Sampling with Inclusion-Probability Weighting(SIP)是核心:不硬性截断,而是按分数比例采样 key\u002Fvalue 块,再用包含概率的倒数对选中块重加权——论文证明这一遍稀疏前向给出的是注意力分子与 softmax 归一化分母的无偏估计,被丢掉的块不会从归一化里凭空消失。最后由 paged kernel 把零碎的选中块凑满完整的 tensor core tile,在不增加 tensor core 成本的前提下细化选择粒度。\n\n论文对现有动态稀疏方法的批评集中在三处:块重要性估计要么粗要么贵;被剪掉的块在 softmax 里直接归零;整 tile 粒度的保留或丢弃太粗。翻译成人话:失真的根源往往不是\"选错\",而是\"截断\"这个动作本身。\n\n## 主结果:稀疏度更高,保真度反而更好\n\n评测用 Wan2.1-T2V-1.3B、Wan2.1-T2V-14B、HunyuanVideo-T2V-13B 三个文生视频模型,720p,Wan2.1 81 帧、HunyuanVideo 129 帧,Penguin Benchmark 600 条提示词,同提示词同种子对照稠密参考,硬件为 NVIDIA H20、BF16、PyTorch 2.6。关键读数:\n\n- Wan2.1 1.3B:跳过 89.16%,单视频延迟 962.39s 降到 479.11s(2.01 倍),PSNR 34.436;\n- Wan2.1 14B:跳过 89.16%,4743.52s 降到 2571.52s(1.84 倍),PSNR 30.788;\n- HunyuanVideo 13B:跳过 92.35%,4399.03s 降到 1974.61s(2.23 倍),PSNR 34.347。\n\n基线侧,静态 mask 的 RadialAttention 稀疏度 73.03–73.63%,动态的 SVG2、DFSAttn 在 67.89–81.29% 之间;论文统一用 FlashInfer 的 FlashAttention-3 块稀疏后端执行所有基线的 mask,排除实现层差异。结果是基线要多算 1.8–3.6 倍注意力,PSNR 仍落后 1.6–4.4 dB。VBench 五个维度(美学质量、背景一致性、成像质量等)上,AB-Attention 与稠密参考的差距都压得很小。\n\n## 消融:逆概率重加权一项扛走 4 dB\n\n论文在 14B、100 条提示词、同一 89.16% 稀疏度下逐一替换组件,得到全文最有信息量的一组数:去掉 SIP 的逆概率加权,PSNR 掉 4.00 dB;换成均匀采样且不加权,掉 7.25 dB;换成 Top-k 硬选择,掉 3.99 dB。而把块内聚类换成朴素的 query 平均池化只掉 0.91 dB,整 tile 级选择掉 1.02 dB。也就是说,误差大头不在\"选哪些块\",而在\"没选上的块怎么处理\"——选择器可以朴素,截断必须换成带概率修正的采样估计。\n\n## 冷水:只写了 H20 kernel,代码未放出\n\nLimitations 一节作者自己交底:kernel 只适配实验用的 H20,只支持前向推理,评测只覆盖三个文生视频模型。[GitHub 仓库](https:\u002F\u002Fgithub.com\u002FInfRec-Cardinal\u002FAB-Attention)目前只有 README 和许可证,徽章写着 Code coming soon,第三方复现暂无门路,上述数字全部是作者自报。计时区间含文本编码、50 个去噪步与 VAE 解码,不含模型加载;即便加速 2.23 倍之后,单条视频仍要 8–43 分钟,离实时很远。\n\n所以呢:稀疏注意力的竞争焦点,正在从\"选得更准\"转向\"估计得更诚实\"。AB-Attention 用重要性采样把\"跳过\"变成\"无偏估计\",这个思路不属于视频扩散专有——长上下文 LLM 的 prefill、任何 softmax 归一化的稀疏化,理论上都适用。等代码放出,第一件事应该是验证它在 H20 之外硬件上的真实表现。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39663","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"42078532-6dcc-4b5c-8c69-97efd760abfb","en","AB-Attention: sparser yet sharper, 2.23x faster diffusion","AB-Attention skips 89-92% of video-diffusion attention for a 2.23x end-to-end speedup, gaining up to 4.4 dB PSNR over the strongest baselines.","The inference bottleneck of video diffusion transformers is no secret: 3D full attention scales quadratically with sequence length, and at 720p with 81 frames every transformer layer and every denoising step pays that bill again. [AB-Attention](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.39663) (arXiv:2610.39663, posted 2026-10-29) comes from the InfRec team at CUHK-Shenzhen together with Shandong University, the University of Tokyo, Fudan, UC Berkeley, HKUST and NUS. The idea: no training, no weight changes — just skip 89–92% of attention computation. The paper reports up to 2.23x end-to-end speedup while improving PSNR by up to 4.4 dB over the strongest sparse baseline.\n\n## Three moves: score, sample, pack\n\nThe method has three components. The Intra-Block Clustering Selector (IBCS) clusters queries only within each query block and uses the weighted centroids to score every key\u002Fvalue block — no token permutation, no global clustering, so scoring stays cheap. Sampling with Inclusion-Probability Weighting (SIP) is the core: instead of hard truncation, key\u002Fvalue blocks are sampled with probability proportional to their scores and reweighted by inverse inclusion probability. The paper shows this single sparse pass yields unbiased estimates of both the attention numerator and the softmax normalizer, so dropped blocks never vanish from the normalization. Finally, a paged kernel gathers the scattered selected blocks into full tensor-core tiles, refining selection granularity at no extra tensor-core cost.\n\nThe paper's critique of existing dynamic sparse methods lands on three points: block-importance estimates are either coarse or costly; pruned blocks drop to zero in the softmax; and keep-or-drop decisions happen at whole-tile granularity. In plain terms, the root of distortion is often not \"selecting the wrong blocks\" but the act of truncation itself.\n\n## Main results: sparser, yet more faithful\n\nThe evaluation covers three text-to-video diffusion transformers — Wan2.1-T2V-1.3B, Wan2.1-T2V-14B and HunyuanVideo-T2V-13B — at 720p, with 81 frames on Wan2.1 and 129 on HunyuanVideo, over 600 prompts from the Penguin Benchmark, each sparse run compared against a dense reference with the same prompt and seed, on NVIDIA H20 GPUs in BF16 with PyTorch 2.6. Key readings:\n\n- Wan2.1 1.3B: 89.16% skipped, per-video latency down from 962.39s to 479.11s (2.01x), PSNR 34.436;\n- Wan2.1 14B: 89.16% skipped, 4743.52s to 2571.52s (1.84x), PSNR 30.788;\n- HunyuanVideo 13B: 92.35% skipped, 4399.03s to 1974.61s (2.23x), PSNR 34.347.\n\nOn the baseline side, the static-mask RadialAttention reaches 73.03–73.63% sparsity and the dynamic SVG2 and DFSAttn sit between 67.89% and 81.29%; the authors execute all baselines' masks on FlashInfer's FlashAttention-3 block-sparse backend to rule out implementation differences. The baselines still compute 1.8–3.6x as much attention yet trail by 1.6–4.4 dB in PSNR. Across the five VBench dimensions (aesthetic quality, background consistency, imaging quality and more), AB-Attention stays close to the dense reference.\n\n## Ablation: inverse-probability weighting carries 4 dB\n\nAt 14B, 100 prompts and the same 89.16% sparsity, the paper swaps out one component at a time and produces the most informative numbers in the text: removing SIP's inverse-probability weighting costs 4.00 dB PSNR; switching to uniform sampling without weights costs 7.25 dB; switching to Top-k hard selection costs 3.99 dB. Meanwhile, replacing intra-block clustering with plain query mean-pooling costs only 0.91 dB, and tile-level selection costs 1.02 dB. The bulk of the error is therefore not in \"which blocks you pick\" but in \"how you treat the ones you don't\" — the selector can afford to be plain; truncation must be replaced by probability-corrected sampling estimation.\n\n## Caveats: H20-only kernel, code not out yet\n\nThe Limitations section is candid: the kernel is built for the H20 GPUs used in the experiments, supports forward inference only, and the evaluation covers just three text-to-video models. The [GitHub repository](https:\u002F\u002Fgithub.com\u002FInfRec-Cardinal\u002FAB-Attention) currently holds a README and a license with a \"Code coming soon\" badge — third-party reproduction has no path yet, and every number above is author-reported. The timed region covers text encoding, all 50 denoising steps and VAE decoding, excluding model loading; even after the 2.23x speedup, a single video still takes 8–43 minutes, far from real time.\n\nSo what: the frontier of sparse attention is shifting from \"select more accurately\" to \"estimate more honestly\". AB-Attention turns \"skipping\" into \"unbiased estimation\" via importance sampling, and the idea is not specific to video diffusion — long-context LLM prefill and any softmax-normalized sparsification could, in principle, adopt it. Once the code lands, the immediate question is how it really performs on hardware beyond the H20.","ab-attention-block-sparse-video-diffusion","2026-10-30T15:15:10Z","2026-10-30T15:15:13.764261Z","2026-10-30T15:15:13.764270Z",true,"agent",337,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"4349edac-997e-4980-95a0-58054a28a9dd","视频生成投机去噪:1.7 步顶 4 步,提速 1.5 倍","speculative-forcing-video-denoising","2026-10-16T19:11:48+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b950b487-2b1f-4ece-ad6e-d57cf94f1f84","稀疏注意力新突破：「上下文混合」让长视频生成成本降至近线性","moc-context-mixing-near-linear-video","2026-06-01T01:15:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"94894abf-62aa-41a9-8e3c-e999ff274d60","Sparse Forcing：稀疏注意力让视频生成质量速度双提升","meta-ucsb-sparse-forcing-pbsa-video-1-27x","2026-05-07T08:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"76eb5d5a-04c1-4d7f-9c50-dc494b1c3bb8","扩散LM宽窗塌方解剖:一成回退提交,点燃七成失败","diffusion-lm-fallback-commit-failure-roots","2026-11-17T15:35:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"269851b4-d6e6-43a6-9239-b3116708b735","块边界是同步税:扩散LM重叠去噪免训练提速14.5%","b3-overlapping-block-denoising-dlm","2026-11-11T17:11:59+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"51953ff7-5702-4d21-9af6-92d8edf7af09","块扩散LM双稀疏提速:不动权重,32K单步快4.72倍","duet-dual-sparse-bdlm-inference","2026-11-09T15:20:00+00:00"]