https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Mon, 15 Jun 2026 10:00:00 GMT https://www.hansking.cn/blog/2026/AI-xi-tong-cheng-ben-he-suan-TokenGPU-cun-chu-yu-ren-gong-cheng-ben https://www.hansking.cn/blog/2026/AI-xi-tong-cheng-ben-he-suan-TokenGPU-cun-chu-yu-ren-gong-cheng-ben [{"value":"AI 系统成本核算：Token、GPU、存储与人工成本","url":"#ai-系统成本核算tokengpu存储与人工成本","depth":1},{"value":"一、先算一次“成功任务”的成本","url":"#一先算一次成功任务的成本","depth":2},{"value":"二、Token 成本不只取决于用户说了多少话","url":"#二token-成本不只取决于用户说了多少话","depth":2},{"value":"三、Embedding 和索引是容易被漏算的长期成本","url":"#三embedding-和索引是容易被漏算的长期成本","depth":2},{"value":"四、自建 GPU 不能只除以运行小时","url":"#四自建-gpu-不能只除以运行小时","depth":2},{"value":"五、缓存是成本优化，也是数据边界","url":"#五缓存是成本优化也是数据边界","depth":2},{"value":"六、把人工成本放进模型里","url":"#六把人工成本放进模型里","depth":2},{"value":"七、成本归因要能定位到功能和租户","url":"#七成本归因要能定位到功能和租户","depth":2},{"value":"八、优化顺序：先消除浪费，再换模型","url":"#八优化顺序先消除浪费再换模型","depth":2},{"value":"总结：便宜不是目标，单位价值才是","url":"#总结便宜不是目标单位价值才是","depth":2}] Mon, 15 Jun 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM成本优化GPU工程实践 https://www.hansking.cn/blog/2026/ContinuousBatching-yu-fu-wu-bing-fa https://www.hansking.cn/blog/2026/ContinuousBatching-yu-fu-wu-bing-fa [{"value":"Continuous Batching 与服务并发","url":"#continuous-batching-与服务并发","depth":1},{"value":"一、静态 Batch 为什么浪费","url":"#一静态-batch-为什么浪费","depth":2},{"value":"二、Continuous Batching 如何动态加入请求","url":"#二continuous-batching-如何动态加入请求","depth":2},{"value":"三、Prefill 和 Decode 是两种不同工作","url":"#三prefill-和-decode-是两种不同工作","depth":2},{"value":"四、KV Cache 是并发的硬约束","url":"#四kv-cache-是并发的硬约束","depth":2},{"value":"五、调度目标不只有吞吐","url":"#五调度目标不只有吞吐","depth":2},{"value":"六、队列和背压要尽早生效","url":"#六队列和背压要尽早生效","depth":2},{"value":"七、不同请求不要互相拖累","url":"#七不同请求不要互相拖累","depth":2},{"value":"八、流式输出不能掩盖调度问题","url":"#八流式输出不能掩盖调度问题","depth":2},{"value":"九、压测要模拟真实长度分布","url":"#九压测要模拟真实长度分布","depth":2},{"value":"总结：并发优化是在安排等待","url":"#总结并发优化是在安排等待","depth":2}] Sun, 12 Oct 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMContinuous BatchingGPU并发优化 https://www.hansking.cn/blog/2026/FlashAttention-zhu-yi-li-ji-suan-de-IO-you-hua https://www.hansking.cn/blog/2026/FlashAttention-zhu-yi-li-ji-suan-de-IO-you-hua [{"value":"FlashAttention：注意力计算的 IO 优化","url":"#flashattention注意力计算的-io-优化","depth":1},{"value":"一、标准注意力为什么占内存","url":"#一标准注意力为什么占内存","depth":2},{"value":"二、GPU 有很多层存储","url":"#二gpu-有很多层存储","depth":2},{"value":"三、分块计算怎么保持正确结果","url":"#三分块计算怎么保持正确结果","depth":2},{"value":"四、因果 Mask 和变长序列","url":"#四因果-mask-和变长序列","depth":2},{"value":"五、它优化的是 IO，不是把复杂度变成线性","url":"#五它优化的是-io不是把复杂度变成线性","depth":2},{"value":"六、为什么实际速度可能没有论文那么快","url":"#六为什么实际速度可能没有论文那么快","depth":2},{"value":"七、使用时先确认版本和回退路径","url":"#七使用时先确认版本和回退路径","depth":2},{"value":"八、和量化、KV Cache 一起看","url":"#八和量化kv-cache-一起看","depth":2},{"value":"九、从工程角度如何判断值得使用","url":"#九从工程角度如何判断值得使用","depth":2},{"value":"总结：性能优化常常是少搬数据","url":"#总结性能优化常常是少搬数据","depth":2}] Sun, 09 Nov 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMFlashAttentionGPU性能优化 https://www.hansking.cn/blog/2026/KVCacheLLM-tui-li-wei-shen-me-hui-bei-xian-cun-ka-zhu https://www.hansking.cn/blog/2026/KVCacheLLM-tui-li-wei-shen-me-hui-bei-xian-cun-ka-zhu [{"value":"KV Cache：LLM 推理为什么会被显存卡住","url":"#kv-cachellm-推理为什么会被显存卡住","depth":1},{"value":"一、先理解自回归生成","url":"#一先理解自回归生成","depth":2},{"value":"二、KV Cache 为什么会占很多显存","url":"#二kv-cache-为什么会占很多显存","depth":2},{"value":"三、Prefill 和 Decode 对 Cache 的使用不同","url":"#三prefill-和-decode-对-cache-的使用不同","depth":2},{"value":"四、为什么连续批处理需要更聪明的 Cache","url":"#四为什么连续批处理需要更聪明的-cache","depth":2},{"value":"五、前缀缓存可以复用，但必须确认边界","url":"#五前缀缓存可以复用但必须确认边界","depth":2},{"value":"六、Cache 量化和压缩要用质量换算","url":"#六cache-量化和压缩要用质量换算","depth":2},{"value":"七、Cache 管理要考虑取消、抢占和恢复","url":"#七cache-管理要考虑取消抢占和恢复","depth":2},{"value":"八、上下文窗口不是无限记忆","url":"#八上下文窗口不是无限记忆","depth":2},{"value":"九、如何压测 KV Cache","url":"#九如何压测-kv-cache","depth":2},{"value":"总结：缓存是速度和空间的交换","url":"#总结缓存是速度和空间的交换","depth":2}] Sun, 28 Sep 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMKV CacheGPU推理优化