[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-duet-dual-sparse-bdlm-inference":3,"topics-all":38,"news-related-51953ff7-5702-4d21-9af6-92d8edf7af09":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},"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","块扩散语言模型推理有两处白算:每个去噪步都全量回看历史前缀,块内所有位置也都推过整个Transformer。DUET免训练双管齐下:每块选一次紧凑历史KV子集跨步复用,浅层预测更新位置、只把活跃行送深层。32K下SDAR-8B单步延迟99.53毫秒降到21.09毫秒,端到端快2.41倍,LongBench几乎不掉。","扩散语言模型(Diffusion LM)常被当作\"下一代文本生成范式\"的候选:不必逐个 token 从左到右,而是多个位置并行去噪。但翻开账本,推理这一步并不便宜——块扩散模型(BDLM)每生成一个块,要在多个去噪步里反复回看不断变长的历史前缀,还要把块内所有位置反复推过整个 Transformer,哪怕每一步真正\"敲定\"的只是其中一小撮 token。arXiv 上的新论文 DUET(2611.09239)把这两处冗余各打一刀:免训练、双稀疏,32K 上下文下 SDAR-8B 的单步延迟从 99.53 毫秒压到 21.09 毫秒,端到端快 2.41 倍。\n\n## 白算的账,先算清楚\n\n论文把 BDLM 的浪费拆成两笔。第一笔在历史侧:每个去噪步都要对全量前缀做注意力,上下文越长,这一步越贵。第二笔在块内:当前块的全部位置在每个去噪步都会被推过所有 Transformer 层,但该步最终只会敲定其中一个子集——没被选中的位置白跑了一趟深层计算,包括注意力投影和 FFN\u002FMoE。\n\n动手之前,作者先给出两个经验观察,证明这两笔账真的可以省:\n\n- 前缀注意力高度集中,且跨步稳定。预算 256 的历史 KV 子集就能覆盖约 85% 的历史注意力质量,预算提到 512 接近 90%;且在整个块的去噪过程中,覆盖率波动只有约 2.5 个百分点。也就是说,每个块开头选一次紧凑子集、整块复用即可,不必每步重选。\n- 更新位置在浅层就能预测。把模型原本的解码策略作用在不同层的隐藏状态上,衡量其对\"最终会被更新位置\"的召回:在第 1 层保留 70% 的 masked 位置,就能召回约 80% 的最终更新位置;保留 90% 时召回约 95%。\"这一步会改谁\"的信号在浅层已经很强,没必要把所有行都送到最后一层。\n\n## 免训练的双稀疏框架\n\n对应两个观察,DUET 出两招,都不动模型权重:\n\n第一招,前缀 KV 稀疏。在每个生成块的第一个去噪步之后,用块内查询与历史键的相似度打分,为整个块共享地选出一份紧凑历史 KV 子集(预算 256\u002F512\u002F1024),跨去噪步复用;进入下一个块,再对着全量缓存重选。\n\n第二招,早期 query 稀疏。指定一个较早的选择层,在该层用模型原有解码规则预测哪些 masked 位置会被本次更新,只把这些活跃行送进后续层——后面的注意力投影、FFN\u002FMoE 计算只为活跃行付账,其余行绕过深层,保持 masked 留到下一步。两个模型配置不同:LLaDA2.1-mini 用保留率 0.7、选择层第 1 层;SDAR-8B 用保留率 0.5、选择层第 5 层。\n\n整套方法不引入任何额外训练的预测器,复用模型自身的解码策略,这是它\"免训练\"的关键。\n\n## 结果:单步快不等于端到端快\n\n质量端,LongBench Macro-F1 在 512 预算下 LLaDA2.1-mini 拿到 51.35,几乎追平稠密推理的 51.33;四项短任务(256 预算)上,SDAR-8B 的 MMLU、HumanEval 与稠密持平或略升,LLaDA 的 HumanEval 还从 78.66 涨到 80.49。代价并非零:LLaDA 的 MATH-500 从 91.80 掉到 87.60,数学推理上的降级要留意。\n\n效率端(80 条 NarrativeQA、固定生成 256 token、H800 GPU)更能说明问题:32K 上下文下,LLaDA2.1-mini 端到端 1.34 倍、单步 2.83 倍;SDAR-8B 的收益随上下文增长,8K\u002F16K\u002F32K 分别为 1.53\u002F2.05\u002F2.41 倍,端到端延迟从 18.416 秒降到 7.656 秒,单步 4.72 倍。\n\n对比两个基线更有嚼劲。FOCUS 的单步延迟同样漂亮(21.23 毫秒),但它改变了解码轨迹,平均步数从 25.05 涨到 64.53,端到端只剩 1.04 倍;LoSA 在 32K 的端到端为 3.672 秒,反而比稠密的 3.486 秒更慢,实测平均 KV 用量 434\u002F405,超出名义预算 256。论文强调所有方法跑在同一推理框架、同样的 Triton 核上,对比是公平的。\n\n## 工程启示\n\n这篇论文给推理优化圈提了个醒:单步延迟降得再漂亮,若步数随之膨胀,端到端收益可能被吃光——吞吐优化必须在真实解码动态里端到端地测,而不是只看 kernel 基准。对扩散语言模型这条线,DUET 的意义在于把\"推理贵\"从笼统印象拆成两个可验证、可利用的结构性事实(注意力集中且跨步稳定、更新位置浅层可测),再用最小干预收割红利。论文随附匿名源码与测试流程;致谢显示工作受江苏省科技重大项目(#BG2025005)资助,通讯作者 Jingwei Xu 的邮箱后缀为 nju.edu.cn。顺带一提,论文的 AI 使用声明写得相当坦白:生成式 AI 用于优化实现与评测代码、监控实验、润色文稿,核心方法出自作者本人——这种披露姿态值得同行效仿。\n\n扩散 LM 要从论文走向线上服务,推理优化得和架构演进同频。DUET 证明,不动权重的稀疏化就能先吃掉一大块冗余——下一步,就看这两个观察能否推广到更多块扩散模型上。\n\n论文:[DUET: Dual-Sparse Inference for Block Diffusion Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09239)","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2611.09239v1","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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"85b10aa5-73f2-4d3d-8b3d-50e8dd1ddca4","en","DUET cuts diffusion-LM step latency 4.72x, training-free","DUET: training-free dual sparsity for block-diffusion LMs. At 32K, SDAR-8B step latency falls 99.53 to 21.09 ms; 2.41x end-to-end, LongBench intact.","Diffusion language models (DLMs) are a serious candidate for the next text-generation paradigm: instead of decoding left to right, they denoise many token positions in parallel. But open the ledger and inference is not cheap. In block diffusion models (BDLMs), each generated block is refined over multiple denoising steps that repeatedly attend over an ever-growing historical prefix, and every position in the current block is pushed through the full Transformer stack at every step — even though only a small subset of positions is actually resolved at that step. A new arXiv paper, DUET (2611.09239), cuts both sources of waste with a training-free, dual-sparse scheme: at a 32K context length, SDAR-8B's per-step latency drops from 99.53 ms to 21.09 ms, a 2.41x end-to-end speedup.\n\n## Where the wasted compute lives\n\nThe paper splits BDLM redundancy into two parts. On the history side, every denoising step attends over the full prefix, so attention gets more expensive as context grows. On the block side, all positions in the current block traverse every layer at every step, though only a subset is finally selected for the current update — the rest burn deep computation (attention projections, FFN\u002FMoE) for nothing.\n\nBefore building anything, the authors establish two empirical observations showing both costs can be cut:\n\n- Historical attention is concentrated and cross-step stable. A prefix KV subset with budget 256 captures about 85% of the historical attention mass; raising the budget to 512 gets nearly 90%. Across the whole block's denoising trajectory, coverage varies by only about 2.5 percentage points. Select a compact subset once per block and reuse it — no per-step reselection needed.\n- Update positions are predictable at early layers. Applying the model's own decoding policy to hidden states from early layers: at layer 1, retaining 70% of masked positions already recovers about 80% of the final update positions; 90% retention reaches about 95% recall. The \"who gets updated this step\" signal is already strong early, so there is no need to send every row to the last layer.\n\n## A training-free dual-sparse framework\n\nDUET's two moves leave the model weights untouched.\n\nPrefix KV sparsity: after the first denoising step of each generation block, it scores historical keys against block queries and selects one compact, block-shared historical KV subset (budgets 256\u002F512\u002F1024), reused across all denoising steps in that block; the next block re-selects against the full cache.\n\nEarly query sparsity: at a chosen early selection layer, the model's original decoding rule predicts which masked positions will be updated this step, and only those active rows propagate through the remaining layers. Deep attention projections and FFN\u002FMoE computation are paid only for active rows; bypassed rows stay masked for the next step. Configurations differ per model: LLaDA2.1-mini uses retention ratio 0.7 with selection layer 1; SDAR-8B uses 0.5 with layer 5.\n\nNo extra trained predictor is introduced — reusing the model's own decoding policy is what keeps the method training-free.\n\n## Results: fast steps are not fast generation\n\nOn quality, LongBench Macro-F1 at budget 512: LLaDA2.1-mini scores 51.35 versus 51.33 dense. Across four shorter benchmarks (budget 256), SDAR-8B matches or slightly beats dense on MMLU (74.53 vs 74.52) and HumanEval (62.20 vs 61.59), and LLaDA's HumanEval rises from 78.66 to 80.49. The cost is not zero: LLaDA's MATH-500 falls from 91.80 to 87.60.\n\nEfficiency (80 NarrativeQA samples, 256 generated tokens, NVIDIA H800) tells the sharper story. At 32K, LLaDA2.1-mini gets 1.34x end-to-end and 2.83x per-step; SDAR-8B's gains grow with context — 1.53x, 2.05x, 2.41x at 8K, 16K, 32K — with end-to-end latency down from 18.416 s to 7.656 s and 4.72x per-step.\n\nThe baseline comparison is the instructive part. FOCUS also reaches a nice per-step latency (21.23 ms) but changes the decoding trajectory, pushing average steps from 25.05 to 64.53, so its end-to-end speedup shrinks to 1.04x. LoSA at 32K is end-to-end slower than dense (3.672 s vs 3.486 s), and its measured average KV usage of 434\u002F405 exceeds its nominal budget of 256. The authors note all methods run in one inference framework with the same Triton kernels, keeping the comparison fair.\n\n## Engineering takeaways\n\nThe paper is a reminder for the inference-optimization crowd: a beautiful per-step latency means little if step count inflates — throughput must be measured end-to-end under real decoding dynamics, not just in kernel benchmarks. For the diffusion-LM line, DUET's contribution is decomposing \"inference is expensive\" into two verifiable structural facts (concentrated, stable attention; early-predictable updates) and harvesting them with minimal intervention. The paper releases anonymized source code and testing procedures; acknowledgments cite the Jiangsu Science and Technology Major Project (#BG2025005), and corresponding author Jingwei Xu has an nju.edu.cn address. Its AI-use statement is unusually candid: generative AI helped optimize implementation and evaluation code, monitor runs, and polish the manuscript, while the core method came from the authors — a disclosure habit worth copying.\n\nFor diffusion LMs to reach production serving, inference optimization has to evolve in lockstep with the architecture. DUET shows that sparsity touching no weights can already claim a large slice of the redundancy — the open question is whether these two observations generalize across the growing zoo of block-diffusion models.\n\nPaper: [DUET: Dual-Sparse Inference for Block Diffusion Language Models](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09239)","duet-dual-sparse-bdlm-inference","2026-11-09T15:20:00Z","2026-11-09T15:18:15.455454Z","2026-11-09T15:18:15.455468Z",true,"agent",193,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"fd9298b5-dea9-4641-8f73-56ae32261c3e","扩散LM投机采样:串行评估砍到1\u002F4,分布零走样","specctmc-exact-speculative-diffusion-lm","2026-11-07T15:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"af6bfcff-1ffa-400d-ac93-cb6803ab232c","dSpark调度蒸馏:扩散LLM解码最多快10倍","dspark-schedule-distillation-diffusion-llm","2026-10-27T17:09:34+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a781accb-537d-4a09-ad5a-d26b87612801","扩散LLM成本实测:写故事比改语法贵3倍,错答案最烧算力","diffusion-llm-decoding-cost","2026-10-25T21:09:54+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"6ede1566-74c5-4bdd-9dd5-6d3f118db876","投机解码的重复陷阱:D-Loop 两轮循环破局","d-loop-diffusion-speculative-decoding","2026-10-06T15:20:00+00:00"]