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