<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[量化以后模型变傻，是不是我参数没调对]]></title><description><![CDATA[<p dir="auto">我把一个中文模型量化到 4bit 后速度上来了，但回答明显变短，还经常漏条件。是不是温度参数没调好？</p>
]]></description><link>https://localaihub.com/topic/197/量化以后模型变傻-是不是我参数没调对</link><generator>RSS for Node</generator><lastBuildDate>Wed, 03 Jun 2026 19:44:28 GMT</lastBuildDate><atom:link href="https://localaihub.com/topic/197.rss" rel="self" type="application/rss+xml"/><pubDate>Thu, 14 May 2026 11:42:00 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Fri, 15 May 2026 08:42:00 GMT]]></title><description><![CDATA[<p dir="auto">对，参数可以调，但别用参数掩盖能力损失。</p>
]]></description><link>https://localaihub.com/post/2189</link><guid isPermaLink="true">https://localaihub.com/post/2189</guid><dc:creator><![CDATA[林小北]]></dc:creator><pubDate>Fri, 15 May 2026 08:42:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Fri, 15 May 2026 07:01:00 GMT]]></title><description><![CDATA[<p dir="auto">我先做原精度对照和 4bit/8bit 比较。</p>
]]></description><link>https://localaihub.com/post/2188</link><guid isPermaLink="true">https://localaihub.com/post/2188</guid><dc:creator><![CDATA[树莓派烫手]]></dc:creator><pubDate>Fri, 15 May 2026 07:01:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Fri, 15 May 2026 06:21:00 GMT]]></title><description><![CDATA[<p dir="auto">这要补。模型文件也算依赖，不是随便丢进目录。</p>
]]></description><link>https://localaihub.com/post/2187</link><guid isPermaLink="true">https://localaihub.com/post/2187</guid><dc:creator><![CDATA[陈一]]></dc:creator><pubDate>Fri, 15 May 2026 06:21:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Fri, 15 May 2026 03:16:00 GMT]]></title><description><![CDATA[<p dir="auto">我们下载社区量化包，没记录来源。</p>
]]></description><link>https://localaihub.com/post/2186</link><guid isPermaLink="true">https://localaihub.com/post/2186</guid><dc:creator><![CDATA[小蓝]]></dc:creator><pubDate>Fri, 15 May 2026 03:16:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Fri, 15 May 2026 02:54:00 GMT]]></title><description><![CDATA[<p dir="auto">生产上要标明模型版本和量化版本。出了问题要能回滚到具体文件。</p>
]]></description><link>https://localaihub.com/post/2185</link><guid isPermaLink="true">https://localaihub.com/post/2185</guid><dc:creator><![CDATA[nora]]></dc:creator><pubDate>Fri, 15 May 2026 02:54:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Fri, 15 May 2026 01:41:00 GMT]]></title><description><![CDATA[<p dir="auto">别忘了上下文长度。量化模型在长上下文下的退化可能更明显。</p>
]]></description><link>https://localaihub.com/post/2184</link><guid isPermaLink="true">https://localaihub.com/post/2184</guid><dc:creator><![CDATA[小吴]]></dc:creator><pubDate>Fri, 15 May 2026 01:41:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 23:43:00 GMT]]></title><description><![CDATA[<p dir="auto">还要看量化方法和模型尺寸。一个更大模型 4bit，不一定比小模型 8bit 差，必须实测。</p>
]]></description><link>https://localaihub.com/post/2183</link><guid isPermaLink="true">https://localaihub.com/post/2183</guid><dc:creator><![CDATA[Grace]]></dc:creator><pubDate>Thu, 14 May 2026 23:43:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 22:10:00 GMT]]></title><description><![CDATA[<p dir="auto">通常位数高损失小，但显存和速度也不同。要按机器和任务取舍。</p>
]]></description><link>https://localaihub.com/post/2182</link><guid isPermaLink="true">https://localaihub.com/post/2182</guid><dc:creator><![CDATA[阿航]]></dc:creator><pubDate>Thu, 14 May 2026 22:10:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 21:02:00 GMT]]></title><description><![CDATA[<p dir="auto">量化越高越好吗？比如 8bit。</p>
]]></description><link>https://localaihub.com/post/2181</link><guid isPermaLink="true">https://localaihub.com/post/2181</guid><dc:creator><![CDATA[普通网友A]]></dc:creator><pubDate>Thu, 14 May 2026 21:02:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 19:32:00 GMT]]></title><description><![CDATA[<p dir="auto">4bit 不是不能用，但别拿它做高风险制度问答。我更愿意让它做分类和改写。</p>
]]></description><link>https://localaihub.com/post/2180</link><guid isPermaLink="true">https://localaihub.com/post/2180</guid><dc:creator><![CDATA[半截薯条]]></dc:creator><pubDate>Thu, 14 May 2026 19:32:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 18:54:00 GMT]]></title><description><![CDATA[<p dir="auto">至少拿 50 条真实样例。看哪些类型掉得最明显。</p>
]]></description><link>https://localaihub.com/post/2179</link><guid isPermaLink="true">https://localaihub.com/post/2179</guid><dc:creator><![CDATA[melo]]></dc:creator><pubDate>Thu, 14 May 2026 18:54:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 17:24:00 GMT]]></title><description><![CDATA[<p dir="auto">我只测了几条。</p>
]]></description><link>https://localaihub.com/post/2178</link><guid isPermaLink="true">https://localaihub.com/post/2178</guid><dc:creator><![CDATA[树莓派烫手]]></dc:creator><pubDate>Thu, 14 May 2026 17:24:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 16:30:00 GMT]]></title><description><![CDATA[<p dir="auto">先和原精度同样提示词、同样样例对比。不要只凭感觉。</p>
]]></description><link>https://localaihub.com/post/2177</link><guid isPermaLink="true">https://localaihub.com/post/2177</guid><dc:creator><![CDATA[陈一]]></dc:creator><pubDate>Thu, 14 May 2026 16:30:00 GMT</pubDate></item><item><title><![CDATA[Reply to 量化以后模型变傻，是不是我参数没调对 on Thu, 14 May 2026 13:26:00 GMT]]></title><description><![CDATA[<p dir="auto">可能不是温度。量化会影响模型能力，尤其复杂推理、长上下文和细节保留。</p>
]]></description><link>https://localaihub.com/post/2176</link><guid isPermaLink="true">https://localaihub.com/post/2176</guid><dc:creator><![CDATA[林小北]]></dc:creator><pubDate>Thu, 14 May 2026 13:26:00 GMT</pubDate></item></channel></rss>