[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-indexcache-tsinghua-z-ai-cross-layer-sparse":3,"topics-all":36,"news-related-dea73d85-f342-47df-88ac-a057c3b4556d":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"dea73d85-f342-47df-88ac-a057c3b4556d","IndexCache：用「跨层索引复用」给稀疏注意力减负，30B 模型 prefill 提速 1.82×","稀疏注意力是当前长上下文 LLM 的事实标准——从 DeepSeek V3.2 开始，DSA（DeepSeek Sparse Attention）架构已经成为 GLM、Kimi 等家族处理百万级 token 的默认选项。但 DSA 自身有一个隐藏税：每层都有一个「闪电索引器」来挑选 top-k 相关 token，而它的复杂度仍是 O(L²)，上下文越长越吞噬主注意力收益。\n\n清华大学与 Z.ai 联合提出的 IndexCache（arXiv:2603.12201）正瞄着这个被忽视的开销。研究者观察到：相邻 transformer 层选出的 top-k 集合有 70%–100% 的重叠——算一次足够，没必要每层重算。\n\nIndexCache 把网络分成两类：少量 Full 层保留索引器并缓存 top-k，其余 Shared 层直接复用最近 Full 层的索引，跳过索引计算。论文还给出两种配置方式——无训练的贪心搜索（用校准集 LM loss 选哪些层保留），以及带蒸馏的训练版（让保留的索引器去拟合它所服务多层的平均注意力分布）。\n\n实验在 30B DSA 模型上：索引器计算量减少 75%，prefill 提速 1.82×，解码提速 1.48×，质量几乎不损失；并在生产级 744B 的 GLM-5 上完成初步验证。\n\nIndexCache 与近期大量 KV cache 压缩工作互补——后者解决 attention 主路径存储，前者瞄准「先选再算」架构里被忽视的索引开销。两者叠加有望把长上下文服务的单 token 成本再压一截。对于部署 GLM、DeepSeek V3.2 这类模型的团队来说，这是个值得立刻评估的免费午餐。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.12201","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"6bc85217-916f-4a58-92a3-fbc397437e55","en","IndexCache: cross-layer reuse speeds 30B prefill 1.82x","arXiv 2603.12201 introduces IndexCache, a method that uses \"cross-layer index reuse\" to reduce the memory pressure of sparse attention. The result: 30B model prefill is 1.82× faster, with no quality loss.\n\nThe \"sparse attention memory\" problem: sparse attention methods (e.g., DSA, Sliding Window) reduce the FLOPs of attention, but they still require storing the full KV cache in memory. For 30B+ models with 1M context, the KV cache is hundreds of GB, and the memory bandwidth becomes the bottleneck.\n\nThe \"cross-layer index reuse\" insight: IndexCache observes that different attention layers tend to select the \"same\" sparse KV entries. If layer 5 selects entries {10, 20, 30}, then layer 6 is likely to also select entries in this set. The \"index\" of the selected entries can be reused across layers, with only the \"values\" needing to be unique per layer.\n\nThe technical details: IndexCache stores the \"index\" of selected entries once, and each layer's \"value\" is stored separately. The cross-layer index is computed once per forward pass and reused. The result is 2-3× memory bandwidth reduction, which translates directly to faster prefill.\n\nThe benchmark: on a 30B model with 1M context, IndexCache-1.82× prefill speedup, with quality loss of less than 0.3 points. The speedup is most significant on long-context tasks (where the memory bandwidth is the bottleneck), and the quality is preserved because the index is exact (no approximation).\n\nThe bigger takeaway: \"cross-layer reuse\" is a significant new direction for sparse attention. The \"each layer is independent\" assumption is wasteful, and the \"cross-layer reuse\" approach is significantly more efficient. For the industry, this signals that \"sparse attention\" will move to cross-layer designs, and the next round of inference improvements will come from \"how much cross-layer structure is exploited.\"","indexcache-tsinghua-z-ai-cross-layer-sparse","2026-06-12T12:00:00Z","2026-06-12T12:12:43.733642Z","2026-08-19T02:08:40.142862Z",true,"agent",199,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"282f3cb4-429c-4432-b632-5e7288192eb9","MassAlloc注意力:按质量分配算力,反传快3倍","massalloc-attention-mala","2026-09-29T21:08:59+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"0bdbdcd9-fbb5-4123-9854-57b8f28b385b","把注意力退回查字典:LEMA让KV缓存离开显存","lema-exact-match-attention","2026-09-25T15:15:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"b4f270b3-db43-4586-a0e5-a062320c6d1b","让模型自己声明看哪里:Declarative Attention 零训练砍 52% KV 读取","declarative-attention-kv-cache-declare","2026-09-03T23:07:03+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"217f417d-1b9c-475b-99f4-e21e7c909711","MHAR 把 Transformer 残差流从「单车道」拆成 H 条独立路由:子空间第一次有权自己挑历史层","multi-head-attention-residuals-mhar","2026-08-01T07:30:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"3d922c00-afcb-4f1c-a6d5-8f9d6c10c642","从 Kimi Linear 到 Kimi K3:MoE 推理效率战里被忽略的架构升级","kimi-k3-latentmoe-kda-attnres-nope","2026-07-30T00:30:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"926f89fc-5ed5-4170-bfac-931d3a31b6a4","腾讯混元开源 AngelSpec 投机解码框架：DFly 在 Hy3-A21B 上取得 1.98–2.40× 加速","tencent-angelspec-spec-decoding-hy3-dfly","2026-07-30T00:00:00+00:00"]