https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Sun, 09 Nov 2025 10:00:00 GMT 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/On-deviceAI-duan-ce-tui-li-de-xing-neng-yu-yin-si-quan-heng https://www.hansking.cn/blog/2026/On-deviceAI-duan-ce-tui-li-de-xing-neng-yu-yin-si-quan-heng [{"value":"On-device AI：端侧推理的性能与隐私权衡","url":"#on-device-ai端侧推理的性能与隐私权衡","depth":1},{"value":"一、先决定哪些任务值得放到设备端","url":"#一先决定哪些任务值得放到设备端","depth":2},{"value":"二、先做内存预算，再谈模型大小","url":"#二先做内存预算再谈模型大小","depth":2},{"value":"三、量化不是免费午餐","url":"#三量化不是免费午餐","depth":2},{"value":"四、推理优化的第一目标是减少等待","url":"#四推理优化的第一目标是减少等待","depth":2},{"value":"五、端侧隐私强，但不是天然安全","url":"#五端侧隐私强但不是天然安全","depth":2},{"value":"六、模型文件也需要保护","url":"#六模型文件也需要保护","depth":2},{"value":"七、端云路由需要可解释","url":"#七端云路由需要可解释","depth":2},{"value":"八、我建议用一张设备矩阵收尾","url":"#八我建议用一张设备矩阵收尾","depth":2},{"value":"总结：端侧 AI 的价值来自克制","url":"#总结端侧-ai-的价值来自克制","depth":2}] Wed, 22 Apr 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMOn-device AI端侧推理性能优化 https://www.hansking.cn/blog/2026/SpeculativeDecoding-yong-xiao-mo-xing-jia-su-da-mo-xing https://www.hansking.cn/blog/2026/SpeculativeDecoding-yong-xiao-mo-xing-jia-su-da-mo-xing [{"value":"Speculative Decoding：用小模型加速大模型","url":"#speculative-decoding用小模型加速大模型","depth":1},{"value":"一、为什么自回归生成慢","url":"#一为什么自回归生成慢","depth":2},{"value":"二、基本流程是什么","url":"#二基本流程是什么","depth":2},{"value":"三、小模型不需要和大模型一样聪明","url":"#三小模型不需要和大模型一样聪明","depth":2},{"value":"四、速度收益取决于接受率","url":"#四速度收益取决于接受率","depth":2},{"value":"五、为什么结果可以保持目标模型分布","url":"#五为什么结果可以保持目标模型分布","depth":2},{"value":"六、KV Cache 管理决定能不能跑得快","url":"#六kv-cache-管理决定能不能跑得快","depth":2},{"value":"七、不同任务的收益不同","url":"#七不同任务的收益不同","depth":2},{"value":"八、工程上常见的失败原因","url":"#八工程上常见的失败原因","depth":2},{"value":"1. 草稿模型太慢","url":"#1-草稿模型太慢","depth":3},{"value":"2. Tokenizer 不兼容","url":"#2-tokenizer-不兼容","depth":3},{"value":"3. 采样参数不一致","url":"#3-采样参数不一致","depth":3},{"value":"4. 并发下显存不足","url":"#4-并发下显存不足","depth":3},{"value":"5. 只测平均值","url":"#5-只测平均值","depth":3},{"value":"九、上线前怎么评估","url":"#九上线前怎么评估","depth":2},{"value":"总结：让小模型猜，让大模型负责","url":"#总结让小模型猜让大模型负责","depth":2}] Sun, 26 Oct 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMSpeculative Decoding推理优化性能优化 https://www.hansking.cn/blog/2026/Test-timeCompute-yong-geng-duo-tui-li-huan-geng-gao-zheng-que-l https://www.hansking.cn/blog/2026/Test-timeCompute-yong-geng-duo-tui-li-huan-geng-gao-zheng-que-l [{"value":"Test-time Compute：用更多推理换更高正确率","url":"#test-time-compute用更多推理换更高正确率","depth":1},{"value":"一、训练时计算和推理时计算有什么不同","url":"#一训练时计算和推理时计算有什么不同","depth":2},{"value":"二、最简单的方式：Best-of-N","url":"#二最简单的方式best-of-n","depth":2},{"value":"三、自洽性不是简单投票","url":"#三自洽性不是简单投票","depth":2},{"value":"四、搜索与分支：把答案当成一个小型搜索问题","url":"#四搜索与分支把答案当成一个小型搜索问题","depth":2},{"value":"五、动态预算比固定长思考更合理","url":"#五动态预算比固定长思考更合理","depth":2},{"value":"六、推理时计算的成本怎么算","url":"#六推理时计算的成本怎么算","depth":2},{"value":"七、质量收益必须通过对照实验证明","url":"#七质量收益必须通过对照实验证明","depth":2},{"value":"八、什么时候不值得使用更多计算","url":"#八什么时候不值得使用更多计算","depth":2},{"value":"九、生产环境必须设置护栏","url":"#九生产环境必须设置护栏","depth":2},{"value":"十、我的总结：把计算当成预算，而不是信仰","url":"#十我的总结把计算当成预算而不是信仰","depth":2}] Sun, 02 Feb 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型Test-time Compute性能优化