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, 28 Sep 2025 10:00:00 GMT 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 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