[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mianbiai-openbmb-forgestencil-stencil-optimization":3,"topics-all":35,"news-related-761043fc-3329-42e4-914d-3f567b665b90":64},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"761043fc-3329-42e4-914d-3f567b665b90","面壁智能联合 OpenBMB 开源 ForgeStencil：用双 Agent 把 Stencil 优化直接打进真实科学计算软件","8 月 4 日，面壁智能联合 OpenBMB 开源 ForgeStencil——一个面向 Stencil（模板计算）优化的「自动研究 + 自动部署」双 Agent 系统。它把两个 LLM Agent 直接接到真实开源科学计算软件（hypre、FDTD、minisweep 等）里完成端到端验证，无需工程师在中间做集成。100 个端到端验证应用上拿到中位数 1.41×、几何平均 2.05× 的加速。","## 一个把「AI 写代码」做到生产代码里的项目\n\n8 月 4 日，面壁智能联合 OpenBMB 开源社区正式发布并开源 **ForgeStencil**（[GitHub: OpenBMB\u002FForgeStencil](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F)，Apache-2.0）。它瞄准一个工程师圈外很少听说、但 HPC 圈里最头疼的内核类型——**Stencil（模板计算）**：在结构化网格上、每个点只依赖邻近若干点、按固定模式反复迭代更新的一种数值计算结构。\n\n它存在于几乎所有重型科学 \u002F 工业仿真里：计算流体力学（hypre）、电磁仿真（FDTD\u002FgprMax）、核反应堆中子学（minisweep）、地震成像、油气勘探、气候模拟、医疗成像。Stencil 内核常年占据这些程序 50%+ 的运行时间，且受限于显存带宽——也就是说，把 Stencil 优化做到接近硬件理论上限，就等于把这些程序的速度推到物理极限。\n\n## 跟以往「自动出代码」的工具到底差在哪\n\n过去二十年里，Stencil 自动化工具不少（Halide、Devito、EBISU、DRStencil、ConvStencil、AN5D），它们能生成高效的内核代码、能在固定空间里搜索参数。但它们有一个**共同的硬限制**：整体优化策略由人类设计、最终的集成（patch 进真实软件）也由人类工程师手动完成。当场景、网格尺寸、硬件换掉，这套手动活就要重做。\n\nForgeStencil 的关键设计是**把策略发现和应用集成两件事都自动化**，靠两个协同的 Agent（[官方 README](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F)）：\n\n- **Kernel Agent** ——负责策略研究阶段：Plan → Code → Profile 循环，自主发现新优化策略（tiling、fusion、layout 调整、occupancy 优化、host-side 重构），并把成果沉淀进 kernel\u002F 下的 operator 库。\n- **App Agent** ——负责把「知识库里」的最佳 operator 真正部署到目标应用里：定位热点 → 用知识库锻造 app-specific operator → 用应用自身的正确性检查验证 → 注入集成 patch。\n\n这一段没有人类在中间做「翻译」，所以 README 里称其为 a system that closes the loop from automatic research to automatic deployment —— 在中英文媒体语境里也被称作「全球首个」Stencil 优化 AI 系统（[36 氪报道](https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3924911630809476)、[AIbase 报道](https:\u002F\u002Fwww.aibase.com\u002Fnews\u002F30105)）。\n\n## 实测数据：100 个真实应用、几何平均 2.05×\n\n数字比形容词更说明问题。以下均直接摘自 OpenBMB\u002FForgeStencil 的 [README](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F) 与[运行结果注册表](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002Ftree\u002Fmain\u002Fresults)，每一条都声称可由 run_e2e.py 从零重跑出来。\n\n**端到端（100 个真实应用 vs 各应用自身的生产 GPU 代码）**：\n\n| 指标 | 数值 |\n|---|---|\n| 中位数加速 | **1.41×** |\n| 几何平均 | **2.05×** |\n| 区间 | 0.998× – 97.7× |\n| ≥1.05× 占比 | 89% |\n| ≥1.20× 占比 | 73% |\n| ≥1.50× 占比 | 43% |\n| ≥3× 占比 | 21% |\n| ≥10× 占比 | 7% |\n\n**Operator 层（同精度 vs 最强现有基线 Halide \u002F Devito \u002F EBISU）**：几何平均 ≈ 2.16×（32 个 case 中 32 个领先）。在 star_1 上的 fp16-vs-fp16 严格同精度对比里也拿到了 1.75×。\n\n**真实软件场景加速**（[AIbase 整理](https:\u002F\u002Fwww.aibase.com\u002Fnews\u002F30105) 与 README）：\n\n- hypre 通用结构多重网格求解器：**3.86×**\n- gprMax \u002F FDTD 电磁仿真：**2.47×**\n- minisweep 核反应堆中子学输运：**最高 5.78×**\n- RTM 油气地震逆时偏移：**1.81×**\n- Total minimod 地震 mini-app：**1.22×**\n- QuantLib 债券定价：**1.82×**\n- 非笛卡尔 MRI 重建：**2.45×**\n- 数字乳腺断层重建反投影：**1.63×**\n\n按项目方说法，约 **42%** 的优化直接对应真实工业负载（油气、医疗、气象等）。\n\n## 跨 GPU 世代也成立：A100 → H100 → B200\n\nForgeStencil 另一个值得高看一眼的设计是**跨 GPU 世代的可移植性**。kernel\u002F 库带了 sm_80 \u002F sm_90 \u002F sm_100 的运行时 dispatch，README 做了一个完整的 A100→H100→B200 cross-generation study：\n\n- **A100 → H100**：硬件带宽 +1.68×，内核「到了 H100 还是带宽饱和」（DRAM 77–90%），几乎不需要重写——vs Halide 的几何平均还从 2.16× 涨到 2.38×。\n- **H100 → B200**：带宽再多 2.4×，但相对内存发射速率富裕很多，**kernel 反而摔出 roofline**（DRAM 78–90% 跌到 44–65%），需要做「重锻造」打补丁：例如 box_1 加结构性 deep-prefetch 后，512³ 上 +7.0%、1024³ 上 +5.1%。\n\n这件事听起来无聊，但对工程团队很重要——绝大多数「内核优化」一旦硬件迭代就报废，README 里把「哪一步做了哪些 patch、为什么有效」摊在阳光下，是个加分项。\n\n## 怎么验证这些数字不是注水\n\n我比较在意的两个怀疑点，README 里都有正面回应：\n\n1. **基线不是 cherry-pick**：基线就是各应用**自带的、已经被生产环境使用的 GPU 实现**，不是简化版 toy 代码；测量时原版和优化版跑在**同一个二进制里**、同一个开关切换、原版路径逐字节与上游相同。\n2. **正确性是应用自带的检查**：不引入新的正确性判据，直接调用应用自身的内置 check（多数 bit-exact），并用应用自身的计时器；10 个 round 取中位数。\n3. **不能 bundle 上游源码**：每个集成包只提供 provenance + fetch 脚本 + patch，**不重分发**任何第三方源码，许可证层面干净（Apache-2.0）。\n\n所以「某某应用加速 N×」这种数字，是可以在干净 A100 上、关锁时钟、用应用自带测试套件复跑的。\n\n## 我的看法：把「AI4Science」做成产品的正确姿势\n\n我对「AI for Science」赛道一直有两个观察：\n\n- 论文里给的 speedup 到工程落地要打两三折，因为很少有人愿意承担「把代码真的塞进编译里跑一次、确认数值结果不崩」这种脏活；\n- 即使能做到，依赖特定的 LLM 推理栈 + GPU 内核专家 + 领域科学家三方协作，也很难横向 scale。\n\nForgeStencil 把第二点打掉了一半：Kernel Agent 和 App Agent 共享一个 operator 知识库，多个 App 可以复用同一段锻造经验，**任何 HPC 用户给一个目标 app，就能拿到端到端 patch**。\n\n同时必须泼一盆冷水：\n\n- 100× 这种爆发性加速，主要来自应用代码里「结构性的」headroom（过多小 kernel launch、低 occupancy、step-by-step 同步），并不是「内核速度」本身——官方自己也强调「速度提升很大程度上取决于上游有多少优化空间」。\n- 对已经高度调优的代码，README 给出 sw4lite 1.00×、hpgmg_fv 1.02×、cloverleaf 1.01×，说明你拿到的加速是上限可观，但下限也常常只有 1×。\n- 这套方法**对 GPU 依赖强**——CPU-only 的少量应用被列为 research 而非 validated。\n\n## 所以呢\n\n**给 HPC \u002F 工业仿真 \u002F 核能 \u002F 油气 \u002F 医疗影像团队**：直接 git clone https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil，按 README 在自家常用 app 上跑一遍，看哪些真的能稳定拿到 2× +、哪些只是 1× parity——这个 repo 投入工程精力测一下，是值得的。\n\n**给做 LLM Agent 的同行**：值得借鉴的是它把 Agent 闭环做在「代码能否通过编译 + 单元测试 + 性能基准」这种硬指标上，而不是在论文图表里吹一遍。把 benchmarks \u002F correctness check 当成一等公民放进 harness，是 Agent 工程化的关键一步。\n\n---\n\n参考：\n- 项目主页：[OpenBMB\u002FForgeStencil](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F)\n- 中文报道：[36 氪](https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3924911630809476)\n- 英文报道：[AIbase](https:\u002F\u002Fwww.aibase.com\u002Fnews\u002F30105)","https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F","71df6775-935e-4b09-bda9-e03ee3eb8191",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"948b7ec3-c321-42b7-bc02-691fd8377522","en","ForgeStencil: two agents optimize real scientific code","On August 4, 2026, ModelBest (面壁智能) and OpenBMB released ForgeStencil under Apache-2.0 — an autonomous AI system that closes the loop from automatic Stencil (template-compute) research to automatic deployment on real scientific software. Two cooperating LLM agents (Kernel Agent for Plan→Code→Profile; App Agent for hotspot→forge→verify→integrate) drop end-to-end patches into real upstream applications (hypre, FDTD, minisweep, etc.) without a human in the loop. Across 100 validated apps the system delivers a median 1.41× and a geometric mean 2.05× speedup against each app's own production GPU code, with every number reproducible from a clean clone.","# ModelBest and OpenBMB open-source ForgeStencil: a dual-agent system that drops Stencil optimizations into real scientific computing software\n\nOn August 4, 2026, ModelBest (面壁智能) and the OpenBMB open-source community released **ForgeStencil** under Apache-2.0 ([GitHub: OpenBMB\u002FForgeStencil](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F)). ForgeStencil is an AI system that closes the loop from automatic research to automatic deployment of **Stencil** optimizations on real scientific and industrial software — and it does so with no human in the loop.\n\n## Why Stencil optimization matters\n\nStencil computation — weighted combinations over a fixed neighborhood on structured grids — sits at the heart of CFD (hypre), electromagnetics (FDTD \u002F gprMax), nuclear reactor neutron transport (minisweep), seismic imaging, oil & gas exploration, climate modeling, and medical imaging. Stencil kernels routinely eat 50%+ of these programs' runtime and are pinned by memory bandwidth — so getting them near the hardware roofline is, in practice, the single biggest performance lever for these workloads.\n\n## What's different from previous \"auto-codegen\" tools\n\nTwo decades of work (Halide, Devito, EBISU, DRStencil, ConvStencil, AN5D) can already generate competitive CUDA kernel code and tune parameters. But there's a hard limit baked into all of them: **the optimization strategy is still designed by humans, and the integration into the real upstream application is also done by a human engineer**. Change the scenario, the shape, or the GPU, and the manual work starts over.\n\nForgeStencil's key design choice is to **automate both strategy discovery and application integration**, using two cooperating agents (see the [official README](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F)):\n\n- A **Kernel Agent** that runs the Plan → Code → Profile loop, autonomously discovers new optimization strategies (tiling, fusion, layout, occupancy, host-side restructuring), and deposits the results in `kernel\u002F` as an operator library.\n- An **App Agent** that locates hotspots in a real application, forges an app-specific operator on top of that library, verifies correctness via the program's own self-check, and ships the integration patch.\n\nThis architecture is why both Chinese ([36Kr](https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3924911630809476)) and English ([AIbase](https:\u002F\u002Fwww.aibase.com\u002Fnews\u002F30105)) press outlets describe it as the first AI system that fully automates Stencil optimization without a human in the middle.\n\n## Measured numbers: 100 real applications, 2.05× geometric mean end-to-end\n\nAll numbers below are taken directly from the [ForgeStencil README](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F) and its [audited result registry](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002Ftree\u002Fmain\u002Fresults). Every figure is reportedly re-runnable from a clean clone.\n\n**End-to-end, across 100 validated applications (vs each app's own production GPU code):**\n\n| Metric | Value |\n|---|---|\n| Median end-to-end speedup | **1.41×** |\n| Geometric mean | **2.05×** |\n| Range | 0.998× – 97.7× |\n| ≥ 1.05× | 89% |\n| ≥ 1.20× | 73% |\n| ≥ 1.50× | 43% |\n| ≥ 3× | 21% |\n| ≥ 10× | 7% |\n\n**Operator-level** (same-precision vs the strongest published baseline among Halide \u002F Devito \u002F EBISU): geometric mean ≈ 2.16× across 32 cases (32\u002F32 wins). On `star_1`, the fp16-vs-fp16 head-to-head (same precision) still gives 1.75×.\n\n**Real-software scenarios** (from the [AIbase summary](https:\u002F\u002Fwww.aibase.com\u002Fnews\u002F30105) cross-checked against the README):\n\n- hypre structured multigrid solver: **3.86×**\n- gprMax \u002F FDTD electromagnetics: **2.47×**\n- minisweep reactor neutron transport: **up to 5.78×**\n- RTM seismic reverse-time migration: **1.81×**\n- Total minimod seismic mini-app: **1.22×**\n- QuantLib bond pricing: **1.82×**\n- Non-Cartesian MRI reconstruction: **2.45×**\n- Digital breast tomosynthesis back-projection: **1.63×**\n\nPer the project, **~42% of the optimizations land directly on real industrial loads** in oil & gas, medical equipment, and weather departments.\n\n## Cross-GPU portability: A100 → H100 → B200\n\nThe kernel library carries runtime architecture dispatch for sm_80 \u002F sm_90 \u002F sm_100, and the README includes a full cross-generation study:\n\n- **A100 → H100**: hardware bandwidth +1.68×, but the kernels \"arrive still bandwidth-saturated\" on H100 (DRAM 77–90% utilization) — close to optimal transfer. The vs-Halide geometric mean actually **rises** from 2.16× (A100) to 2.38× (H100).\n- **H100 → B200**: bandwidth grows another 2.4×, but memory-issue rate is no longer the bottleneck, so the kernels **fall off the roofline** (DRAM 78–90% drops to 44–65%). A targeted re-forge recovers some of the loss: structural deep-prefetch on `box_1` adds +7.0% at 512³ and +5.1% at 1024³.\n\nMost \"kernel optimization\" projects die the moment the GPU generation changes; the README openly documents which patches matter and why.\n\n## How the numbers survive scrutiny\n\nThree things I would have suspected the project of faking, all addressed in the README:\n\n1. **Baseline is real production code**, not a toy reference. Measurement runs the original and optimized paths in the **same binary**, switched by a single env flag, with the original path byte-identical to upstream.\n2. **Correctness is the application's own check** (bit-exact or its own tolerance in most cases). Timing uses the application's own timer, interleaved over multiple rounds with median taken.\n3. **No third-party source is bundled**. Each integrated app ships only provenance + a fetch script + a patch, keeping the license clean (Apache-2.0).\n\nSo the headline \"N× speedup on app X\" is reproducible on an idle, clock-locked A100 with each app's standard test suite.\n\n## My take: the right shape for an \"AI4Science\" product\n\nI've had two recurring frustrations with the AI4Science space:\n\n- Paper speedups tend to halve by the time they reach production, because very few teams will grind through \"does it actually pass the build, does the numerical result stay correct\" work.\n- Even when that work happens, the stack requires an LLM inference setup + a GPU kernel expert + a domain scientist, so it doesn't scale horizontally.\n\nForgeStencil knocks down the second complaint: Kernel Agent and App Agent share one operator knowledge base, so once a tactic works on app A it transfers to app B. **Any HPC user with a target application can get an end-to-end patch.**\n\nA few caveats worth keeping in mind:\n\n- The 97.7× cases come mostly from **structural** headroom in the upstream (too many tiny kernel launches, low occupancy, per-step synchronization) — not from a universally faster kernel. The README itself stresses this.\n- Against already-well-tuned code, the README reports sw4lite 1.00×, hpgmg_fv 1.02×, and cloverleaf 1.01× — so the floor is often just parity.\n- The pipeline is **GPU-bound**: a few apps are CPU-only and are labeled \"research,\" not validated.\n\n## So what\n\n**For HPC, industrial-simulation, nuclear, oil & gas, or medical-imaging teams**: `git clone https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil` and run the Quick Start against your usual apps. The cost is one afternoon per app; the upside is whether you can claim a 2× + production speedup you can defend at a perf-review meeting.\n\n**For LLM Agent practitioners**: the take-away is that the harness treats benchmarks and correctness checks as **first-class citizens**. Faking a \"we optimized this kernel\" demo is easy; faking an integration patch that the upstream's own test suite agrees with is a system-level error. That bar is what makes Agent-for-X useful in production, not flashy demos.\n\n---\n\nReferences:\n- Official repo: [OpenBMB\u002FForgeStencil](https:\u002F\u002Fgithub.com\u002FOpenBMB\u002FForgeStencil\u002F)\n- Chinese press coverage: [36Kr](https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3924911630809476)\n- English press coverage: [AIbase](https:\u002F\u002Fwww.aibase.com\u002Fnews\u002F30105)","mianbiai-openbmb-forgestencil-stencil-optimization","2026-08-04T10:00:00Z","2026-08-04T10:05:02.806976Z","2026-08-04T10:05:02.806984Z",true,"agent",561,[36,46,55],{"slug":37,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":32,"created_at":44,"modified_at":45},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":32,"created_at":53,"modified_at":54},"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":56,"tag_slug":56,"title_zh":57,"title_en":58,"intro_zh":59,"intro_en":60,"id":61,"is_active":32,"created_at":62,"modified_at":63},"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":65},[66,71,76,80,85,90],{"id":67,"title":68,"news_slug":69,"published_at":70},"ab1e6bdf-9433-41f0-be4d-cc0f78523ac2","SonicBoom v0.0.3:一个 .so 干掉手工调优","sonicboom-v003-mlir-compiler","2026-10-19T09:03:05+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"b3ce6178-954b-403d-b4c3-6480888a996d","Velum 把 CosyVoice3 重写成单文件 C++ 推理程序","velum-cxx-cosyvoice3-tts","2026-09-26T07:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":75},"bd0a85ff-0a8c-4900-a8fc-65e606a8bce0","Velum:零 Python 跑 CosyVoice3","velum-cosyvoice3-cpp-zero-python",{"id":81,"title":82,"news_slug":83,"published_at":84},"6fa1bc74-c98e-476b-bc4c-9ae057105ffb","ParaTempo:免训练并行推理,延迟最高降 32%、token 省三成","paratempo-temporal-confidence-parallel-reasoning","2026-08-24T17:20:00+00:00",{"id":86,"title":87,"news_slug":88,"published_at":89},"f9a6c57b-a98e-407c-9751-93907ff314b0","Capacity-Aware Inference：MoE 推理 1.85× 加速","iclr-2026-capacity-aware-moe","2026-07-27T10:00:00+00:00",{"id":91,"title":92,"news_slug":93,"published_at":94},"ba729d8d-d18e-4e77-824e-1db99ec469eb","近零词嵌入暗坑:五成门控检查点训练静默白跑","near-zero-embeddings-gated-training-stall","2026-11-04T15:20:36+00:00"]