<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[improve_perfect的专栏]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/improve_perfect</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; improve_perfect]]></copyright><item><title><![CDATA[基于1层Llama草稿模型实现2.16推测解码加速]]></title><link>https://blog.youkuaiyun.com/improve_perfect/article/details/161642368</link><guid>https://blog.youkuaiyun.com/improve_perfect/article/details/161642368</guid><author>improve_perfect</author><pubDate>Wed, 03 Jun 2026 06:47:26 +0800</pubDate><description><![CDATA[在5分钟内实现了2.16x的推测解码加速比。]]></description><category></category></item><item><title><![CDATA[小米15 NPU部署Llama-3.2-1B-Instruct]]></title><link>https://blog.youkuaiyun.com/improve_perfect/article/details/149887091</link><guid>https://blog.youkuaiyun.com/improve_perfect/article/details/149887091</guid><author>improve_perfect</author><pubDate>Sun, 03 Aug 2025 22:34:23 +0800</pubDate><description><![CDATA[token-generation-time: 5191579 (微秒) ≈ 5.19 秒。time-to-first-token (首 Token 延迟): 531665 (微秒) ≈ 0.53 秒。token-generation-rate (Token 生成速率): 38.52 toks/sec。模型生成回复的速率（每秒生成的 Token 数）。模型处理输入提示的速率（每秒处理的 Token 数）。输入给模型的提示（Prompt）的 Token 数量。模型生成的回复的 Token 数量。]]></description><category></category></item><item><title><![CDATA[解决_ZN6google8protobuf8internal12kEmptyStringE未定义]]></title><link>https://blog.youkuaiyun.com/improve_perfect/article/details/89599639</link><guid>https://blog.youkuaiyun.com/improve_perfect/article/details/89599639</guid><author>improve_perfect</author><pubDate>Sat, 27 Apr 2019 11:52:20 +0800</pubDate><description><![CDATA[问题：

    运行nnie_mapper_11时的错，_ZN6google8protobuf8internal12kEmptyStringE未定义。



运行环境为：

    ubuntu 16.04.5，gcc-5.4.0，protobuf2.5



分析：

    通过nm查看libprotobuf.so.8.0.0中含有的符号为_ZN6google8p...]]></description><category></category></item><item><title><![CDATA[Tensorflow使用训练好的模型]]></title><link>https://blog.youkuaiyun.com/improve_perfect/article/details/79603730</link><guid>https://blog.youkuaiyun.com/improve_perfect/article/details/79603730</guid><author>improve_perfect</author><pubDate>Sun, 18 Mar 2018 20:15:59 +0800</pubDate><description><![CDATA[        深度学习大部分时间花费在利用海量的数据训练出一个精确的模型，并且对训练主机的性能有较高的要求。那么能不能直接使用别人训练好的模型，免去繁琐的训练过程以及超高配置的机器要求？答案是肯定的。        以识别下面这张图为例先上结果：    teddy, teddy bear (score = 0.99959)    toyshop (score = 0.00018)    ball...]]></description><category></category></item><item><title><![CDATA[ZC301摄像头，OK6410的开发板，linux3.0.1，内核Oops的解决方法]]></title><link>https://blog.youkuaiyun.com/improve_perfect/article/details/37729147</link><guid>https://blog.youkuaiyun.com/improve_perfect/article/details/37729147</guid><author>improve_perfect</author><pubDate>Sat, 12 Jul 2014 23:47:33 +0800</pubDate><description><![CDATA[在使用OK6





Unable to handle kernel NULL pointer dereference at virtual address 00000060   
pgd = c3d14000                                                                 
[00000060] *pgd=53d5]]></description><category></category></item></channel></rss>