[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kernelseek-question-driven-cuda-kernel-agent":3,"topics-all":41,"news-related-a550dbc5-240b-4747-b625-9aa874aad444":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"a550dbc5-240b-4747-b625-9aa874aad444","工具拉满的AI停在0.48倍:会提问的做到1.13倍","论文把两种AI写GPU算子的方式放在同一台机器上对照。一边是按主流配方配齐解析器、文档检索、参考库和知识库的自主智能体,133步后停在0.48倍;另一边只留最小评测器加人工提问,40条消息做到1.13倍。论文还给出两个kernel、完整trace与评测器,并自曝单次运行等局限。","把 vLLM 的 FlashAttention-3 前向 kernel 再跑快一截，26 组配置都要打平，正确性不能退——这是 GPU 算子优化里的硬骨头。论文做了一次对照：一边是按主流配方搭到极致的自主智能体，一边是只留最小 harness、靠人提问来驱动。前者停在 0.48 倍，后者做到 1.13 倍。\n\n## 满配版：给智能体一整条流水线\n\n满配的那一方叫 Arsenal，是论文自己搭的基线，沿 AVO、KDA 这条演化搜索路线做。每一步都塞给模型：按 NCU 瓶颈标签选出的 16 个优化杠杆之一、解析过的性能诊断、反汇编的指令混合差异、BM25 检索来的 PTX ISA 与 CUDA 指南文档、从 DeepGEMM\u002FThunderKittens\u002FCUTLASS 挖出的参考提示（磁盘上 1.6 GB）、15 条种子知识笔记，外加 sanitizer 与禁止符号的反作弊检查。每步 prompt 中位数 22.6K 字符，37% 是预取材料。\n\n结果是：133 个模型驱动步，API 花费 660 美元，62.4 小时里跑了 44.7 小时才到最好成绩。最好的几何平均 0.48 倍（摘要写作 0.50 倍），最差配置 0.24 倍，26 组配置里打平的有 0 组，目标未达成。冠军在第 35 步到达 0.48 倍后再没动过，剩下 98 步原地踏步。期间开过 37 个分支，36 个从同一步分叉，3 次人工回滚丢掉 117 步。\n\n## 分水岭：82% 对 26%\n\n论文落在两个数上。满配方 prompt 里有 37% 是预取参考材料，可它读到的工具结果文本中，82% 花在重读自己的 kernel 源码上；提问方什么都不预取，这个比例只有 26%。\n\n工具用量上提问方反而更多：原始输出跑了 429 次 profiler、372 次 SASS\u002FPTX 检查、377 次构建配 937 次基准跑，还派了 48 个子智能体。满配方则是每步一次解析好的 profiler、92 次文档检索、0 个子智能体。全程 16 个会话、11 532 次工具调用（8 791 次在主会话，2 741 次在子智能体），平均每条人工消息对应约 71 次工具调用。\n\n论文的观察很直接：提问没有减少工具使用，而是把工具用在对的地方。\n\n## 那 17 个问题长什么样\n\n162 条人工消息按主行为分类：指令 58 条、提问 44 条、批准 30 条、纠正 19 条、目标 11 条；63 条是疑问句式。提问分八类：测量校验 13 条、边界差分 10 条、基准有效性 7 条、差距分诊 7 条、流程 6 条、质疑上限 5 条、策略选项 5 条、假设校验 3 条。\n\n论文挑了四个案例，形态比数量更能说明问题：\n\n- 「什么配置在拖低几何平均？」agent 以为该调 hd128 路径，实际 hd96、hd192、hd256 全掉在 0.01 倍的标量回退路径上。删掉回退后，4 小时内几何平均从 0.44 倍拉到 0.96 倍。\n- 「这个上限是硬件的性质，还是我们设计的性质？kernel 之外的时间花在哪？」agent 认定 split-KV 解码「在 harness 规则内已到真实天花板」。host 侧 profile 抓出每次调用 2-5 μs 的张量填充，删掉后 split-KV 从 0.87 倍到 0.97 倍，几何平均从 1.02 倍到 1.13 倍。\n- 「我们的场景覆盖了算子真实被调用的形状和特性吗？」agent 一直以为 40 个场景全对就等于完成。搬来完整测试网格后，58 个里只对 52 个，sink 和左填充输出差了 495 倍和 103 倍。改完 58 个全对，新网格几何平均 1.10 倍。\n- 「你为什么觉得目标达成了？我们还远低于打平。」Stop hook 的判定此前已接受「目标达成」，agent 承认自己在第一个有效杠杆后就宣布胜利。fp8 hd256 解码从 0.20 和 0.34 变成 0.66 和 0.77。\n\n这些问题都不含答案。人没写一行 kernel 代码，63 条疑问句里只有 17 条提到注意力概念，没有一条指定 CUDA 构造。两个 kernel 共 13.8K 行，全是 agent 写的。\n\n## 交出来的东西\n\nfa3_hopper.cu（3.6K 行）和 fa4_blackwell.cu（10.2K 行）以 Apache 2.0 发布，附带评测器、验证 Blackwell 原语的 PTX 探针、问题库和全部 trace。fa3_hopper.cu 在 182 个 Hopper 场景上全对，1.16 倍 FA3，对最快可用实现（含 cuDNN）1.10 倍，182 个里 135 个达到或超过打平；fa4_blackwell.cu 在 154 个可跑场景全对，1.46 倍 FA4，对最快可用实现 1.07 倍，124 个里 98 个打平以上，最差 0.78 倍；在 SOL-ExecBench 的生产形状上是 1.10 倍 FA4。\n\n作者自己列的折扣也得照抄：剔除 fp8-KV 解码这一类 claim 形状后，Blackwell 全集几何平均变成 1.37、1.32、1.08 倍；edge 集是短 kernel，部分 1.5 到 2 倍来自上游 dispatch 开销；Hopper 侧在 fp8 dense prefill、softcapped prefill、head dim 96 和长滑窗上仍是 0.34 到 0.79 倍；Blackwell 侧 fp8 输入和 192 及以上的 head dim 还没做（182 个里 28 个）；两个 kernel 都只覆盖前向。\n\n## 这份对照该怎么读\n\n论文把限定写得很清楚：每个 arm 只跑一次，前后端不同（满配方走 Claude Code 无头模式，提问方一半 Claude Code 一半 Cursor），受控组对着不受控组；KernelSeek 在此前两天已为 FA3 的 40 场景测试网格写过 kernel，所以它 0.009 倍的起点低估了领先；68.0 小时里有 26.7 小时是空闲间隔。\n\n更要紧的是论文承认还没做的那个实验：问题库能不能替掉人。设计是四组对照（满配循环、裸 agent 加 harness、裸 agent 加从问题库取问的提问器、人提问），锁住模型和预算——写下来了，还没跑。72 条从公开资料挖来的问题也还没指挥过 agent。\n\n也别把结论读成「工具流没用」。别的论文在别的任务上有产出：AVO 报告在 B200 注意力上连续演化 7 天，比 cuDNN 高出最多 3.5%、比 FA4 高出最多 10.5%；Cursor 与 NVIDIA 的多智能体系统在 3 周里优化 235 个 Blackwell kernel，报告 38% 几何平均提升。任务和评测器不同，不能直接横比，但说明这篇量到的是一个具体失灵，不是一条通用判决。\n\n## 所以呢\n\n如果你的智能体正卡在某个性能目标上，先别急着加检索源和知识库。先看它读的东西有多少是自己的产出——满配那个是 82%。再问一句：有没有哪个问题，能把它挪到一个它没检查过的层，或者逼它为已经宣布的上限辩护？论文的结论句挺直接：深度算子优化里，稀缺的不是信息，是方向。全部数据与评测器在 [arXiv:2611.03517](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.03517)，文中提到的多智能体系统出自 [Cursor 与 NVIDIA 的公开报告](https:\u002F\u002Fcursor.com\u002Fblog\u002Fmulti-agent-kernels)。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.03517","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":22,"name":23,"slug":23,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":25,"name":26,"slug":26,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"267f5a8a-2f24-4160-bd46-5090f21727ac","en","Tooling Plateaued at 0.48x, Questions Reached 1.13x","Paper comparison: a tool-rich agent stalls at 0.48x in 133 steps, a minimal harness with 40 messages reaches 1.13x. Kernels and traces released.","Making an existing vLLM FlashAttention-3 forward kernel faster is one of the harder jobs in GPU operator optimization: beat parity on 26 configurations without breaking correctness. One paper ran a controlled comparison. On one side sat an autonomous agent built exactly along the prevailing recipe, loaded with every tool that recipe prescribes. On the other sat a minimal harness steered by questions from a human. The tool-laden agent stalled at 0.48x. The question-driven one reached 1.13x.\n\n## The fully equipped version: a whole pipeline for the agent\n\nThe equipped side is called Arsenal, a baseline the authors built themselves along the line of evolutionary agentic search (AVO, KDA). Every step fed the model a bundle: one of 16 optimizations selected by the NCU bottleneck label, a parsed performance diagnosis, instruction-mix deltas from disassembly, BM25-retrieved documentation sections from the PTX ISA and CUDA guides, hints mined from reference libraries like DeepGEMM, ThunderKittens and CUTLASS (1.6 GB on disk), 15 seeded knowledge notes, causal memory of recent steps, plus sanitizer runs and forbidden-symbol checks. The median prompt was 22.6K characters, 37% of it retrieved or pre-fetched material.\n\nThe outcome: 133 model-driven steps, 660 dollars of API cost, and 44.7 hours of a 62.4-hour run before it reached its best result. Best geometric mean was 0.48x (the abstract rounds this to 0.50x), the worst configuration sat at 0.24x, 0 of the 26 configurations reached parity, and the goal was never met. The champion hit 0.48x at step 35 and then stopped improving for the remaining 98 steps. Thirty-seven branches were opened along the way, 36 of them from that same step, and 3 human rollbacks discarded 117 steps.\n\n## The dividing line: 82% versus 26%\n\nThe paper lands on two numbers. For the equipped agent, 37% of the prompt was pre-fetched reference material, yet 82% of the tool-result text it read was spent re-reading its own kernel source. The question-driven agent had nothing pushed to it and spent 26% on its own kernels.\n\nTool usage actually went the other way. The question-driven side ran the profiler 429 times on raw output, did 372 SASS\u002FPTX inspections, made 377 builds against 937 benchmark runs, and spawned 48 subagents. The equipped side ran one parsed profiler dump per step, did 92 documentation lookups, and spawned 0 subagents. Across 16 sessions there were 11,532 tool calls (8,791 in parent sessions, 2,741 in subagents), roughly 71 tool calls per human message.\n\nThe paper's own observation is blunt: the questions did not reduce tool use. They aimed it.\n\n## What those 17 questions look like\n\nOf the 162 human messages, 58 were directives, 44 questions, 30 approvals, 19 corrections and 11 goals; 63 were phrased interrogatively. The questions fall into eight categories: measurement check 13, bound differencing 10, benchmark validity 7, gap triage 7, agent process 6, challenge a limit 5, strategy options 5 and hypothesis check 3.\n\nThe paper spotlights four cases, because the shape of the question matters more than the count:\n\n- \"What configs are bringing down the geomean?\" The agent believed tuning the hd128 paths was the priority. A per-configuration table showed hd96, hd192 and hd256 sitting on the scalar fallback at 0.01x. With the fallback removed, the geomean went from 0.44x to 0.96x in four hours.\n- \"Is this ceiling a property of the hardware, or of our design? Where does the time go outside the kernel?\" The agent had declared split-KV decode a \"genuine ceiling within the harness rules\". A host-side profile found a per-call tensor fill of 2-5 us outside the kernel. Removing it took split-KV from 0.87x to 0.97x and the geomean from 1.02x to 1.13x.\n- \"Do our scenarios cover the shapes and features the operator is actually called with?\" The agent had treated 40 of 40 correct scenarios as done. Porting the operator's full test grid showed 52 of 58 correct, with sink and left-pad outputs off by 495x and 103x. After six fixes, 58 of 58 were correct and the geomean on the new grid was 1.10x.\n- \"Why do you think the goal was achieved? We are still well below parity.\" The Stop-hook judge had already accepted \"goal met\". The agent conceded it had stopped after one effective lever and declared victory prematurely. fp8 hd256 decode went from 0.20 and 0.34 to 0.66 and 0.77.\n\nNone of these questions carried an answer. The human wrote no kernel code. Of the 63 interrogative messages, only 17 mentioned an attention concept and none prescribed a CUDA construct. The agent wrote all 13.8K lines of the two kernels.\n\n## What actually shipped\n\nfa3_hopper.cu (3.6K lines) and fa4_blackwell.cu (10.2K lines) were released under Apache 2.0, together with the evaluator, the PTX probes used to validate each Blackwell primitive in isolation, the question bank and every trace. fa3_hopper.cu is correct on all 182 Hopper scenarios, runs at 1.16x FA3 and 1.10x the fastest available implementation including cuDNN, with 135 of 182 at or above parity. fa4_blackwell.cu is correct on all 154 runnable Blackwell scenarios, runs at 1.46x FA4 and 1.07x the fastest available implementation, with 98 of 124 at or above parity and a worst case of 0.78x; on production shapes from SOL-ExecBench it reaches 1.10x FA4.\n\nThe authors' own discounts deserve equal billing. Excluding the fp8-KV decode claim shapes moves the Blackwell all-set geomeans to 1.37, 1.32 and 1.08x. The edge sets consist of short kernels, so part of the 1.5x to 2x there is upstream dispatch overhead. On Hopper the losses are fp8 dense prefill, softcapped prefill, head dimension 96 and long sliding windows, at 0.34x to 0.79x. On Blackwell, fp8 inputs and dense head dimensions of 192 and above are not built yet, which is 28 of the 182 scenarios. Neither kernel covers the backward pass.\n\n## How to read the comparison\n\nThe paper states its limitations plainly. Each arm is a single run, with different front-ends (the equipped agent ran Claude Code headless; the question-driven side mixed Claude Code and Cursor), and a steered arm is being compared against an unsteered one. KernelSeek had already built a kernel for FA3's 40-scenario test grid in the two preceding days, so its 0.009x starting point understates its head start. And of its 68.0 hours, 26.7 were an idle gap.\n\nMore importantly, the paper names the experiment it has not run: whether the question bank can replace the human. The design is fixed at four arms (the equipped loop, the bare agent with harness alone, the bare agent with a questioner drawing on the bank, and a human questioner) with the model and budget held constant. It is written down, not executed. The 72 bank entries mined from public sources have not yet steered an agent either.\n\nAnd do not read the result as \"tool-rich pipelines do not work.\" Other papers do get results out of them on other tasks: AVO reports 7 days of continuous autonomous evolution on B200 attention, outperforming cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%; Cursor and NVIDIA report a 38% geomean speedup across 235 Blackwell kernels in 3 weeks with a multi-agent system. The tasks and evaluators differ, so these are not directly comparable. What they do show is that this paper measured one specific failure, not a general verdict.\n\n## So what\n\nIf your agent is stuck on a performance target, resist the urge to add another retrieval source or knowledge store. First check what share of what it reads is its own output: for the equipped agent it was 82%. Then ask whether there is a question that could move it to a layer it has not examined, or force it to defend a limit it has already declared. The paper's conclusion is direct: in deep kernel optimization the scarce resource is not information but direction.\n\nAll data and the evaluator are at [arXiv:2611.03517](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.03517); the multi-agent system referenced above comes from the [public Cursor and NVIDIA report](https:\u002F\u002Fcursor.com\u002Fblog\u002Fmulti-agent-kernels).","kernelseek-question-driven-cuda-kernel-agent","2026-11-03T09:05:00Z","2026-11-03T09:29:02.543864Z","2026-11-03T09:29:02.543871Z",true,"agent",90,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"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":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"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":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"932bddfc-f9c2-43c9-9eec-676df0a7bf23","解码长度预测撞上天花板:误差大头是采样噪声,不是模型","prefill-decode-length-prediction-limits","2026-11-01T23:09:55+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"c174c396-c33d-4b0b-964b-c60b8f97bd68","JustGoal:只给目标,AI基线跑赢专家kernel","justgoal-ai-kernel-baseline","2026-10-29T13:10:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"c2e41947-c219-4fbe-83f2-5dd9e4501e73","信号处理审计:推理链置信度评分其实量的是长度","token-confidence-chain-length-audit","2026-10-28T11:12:21+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"98bf8c61-e559-4f96-9c3a-e7049b493f6c","小模型自信≠正确:4.2万次实测拆解LLM路由","confidence-based-llm-routing-multilingual","2026-10-24T19:30:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"13ed0c94-9bea-440c-93c4-771c38ba9834","sink 消失后，KV 驱逐改看 value 几何","valuediff-value-geometric-kv-eviction","2026-09-22T17:05:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"4bbc55d2-cabc-477f-a3ad-4e2c119aff2a","TokTier 抓住 Agent 推理的隐藏瓶颈：缓存命中 94.1%，分词仍吃掉 64% 首 token 时间","toktier-stateful-tokenization-agent-serving","2026-07-31T17:56:30+00:00"]