<br />
<b>Notice</b>:  Function _load_textdomain_just_in_time was called <strong>incorrectly</strong>. Translation loading for the <code>betterdocs</code> domain was triggered too early. This is usually an indicator for some code in the plugin or theme running too early. Translations should be loaded at the <code>init</code> action or later. Please see <a href="https://developer.wordpress.org/advanced-administration/debug/debug-wordpress/">Debugging in WordPress</a> for more information. (This message was added in version 6.7.0.) in <b>/data/user/htdocs/wp-includes/functions.php</b> on line <b>6114</b><br />
<br />
<b>Notice</b>:  Function _load_textdomain_just_in_time was called <strong>incorrectly</strong>. Translation loading for the <code>jnews-view-counter</code> domain was triggered too early. This is usually an indicator for some code in the plugin or theme running too early. Translations should be loaded at the <code>init</code> action or later. Please see <a href="https://developer.wordpress.org/advanced-administration/debug/debug-wordpress/">Debugging in WordPress</a> for more information. (This message was added in version 6.7.0.) in <b>/data/user/htdocs/wp-includes/functions.php</b> on line <b>6114</b><br />
<br />
<b>Notice</b>:  Function _load_textdomain_just_in_time was called <strong>incorrectly</strong>. Translation loading for the <code>wp-statistics</code> domain was triggered too early. This is usually an indicator for some code in the plugin or theme running too early. Translations should be loaded at the <code>init</code> action or later. Please see <a href="https://developer.wordpress.org/advanced-administration/debug/debug-wordpress/">Debugging in WordPress</a> for more information. (This message was added in version 6.7.0.) in <b>/data/user/htdocs/wp-includes/functions.php</b> on line <b>6114</b><br />
<br />
<b>Notice</b>:  Function _load_textdomain_just_in_time was called <strong>incorrectly</strong>. Translation loading for the <code>wpdiscuz</code> domain was triggered too early. This is usually an indicator for some code in the plugin or theme running too early. Translations should be loaded at the <code>init</code> action or later. Please see <a href="https://developer.wordpress.org/advanced-administration/debug/debug-wordpress/">Debugging in WordPress</a> for more information. (This message was added in version 6.7.0.) in <b>/data/user/htdocs/wp-includes/functions.php</b> on line <b>6114</b><br />
<br />
<b>Notice</b>:  函数 _load_textdomain_just_in_time 的调用方法<strong>不正确</strong>。 <code>jnews</code> 域的翻译加载触发过早。这通常表示插件或主题中的某些代码运行过早。翻译应在 <code>init</code> 操作或之后加载。 请查阅<a href="https://developer.wordpress.org/advanced-administration/debug/debug-wordpress/">调试 WordPress</a>来获取更多信息。 （这个消息是在 6.7.0 版本添加的。） in <b>/data/user/htdocs/wp-includes/functions.php</b> on line <b>6114</b><br />
<br />
<b>Notice</b>:  函数 _load_textdomain_just_in_time 的调用方法<strong>不正确</strong>。 <code>jnews-like</code> 域的翻译加载触发过早。这通常表示插件或主题中的某些代码运行过早。翻译应在 <code>init</code> 操作或之后加载。 请查阅<a href="https://developer.wordpress.org/advanced-administration/debug/debug-wordpress/">调试 WordPress</a>来获取更多信息。 （这个消息是在 6.7.0 版本添加的。） in <b>/data/user/htdocs/wp-includes/functions.php</b> on line <b>6114</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
<br />
<b>Warning</b>:  Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in <b>/data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php</b> on line <b>1893</b><br />
{"id":30519,"date":"2024-05-13T12:55:58","date_gmt":"2024-05-13T04:55:58","guid":{"rendered":"https:\/\/linguaresources.com\/?p=30519"},"modified":"2024-07-18T21:36:24","modified_gmt":"2024-07-18T13:36:24","slug":"30519","status":"publish","type":"post","link":"https:\/\/linguaresources.com\/?p=30519","title":{"rendered":"Phi-3 \u548c\u9ad8\u6027\u80fd iPhone LLM \u7684\u5f00\u7aef"},"content":{"rendered":"<p>\u8bfb\u8fc7\u6211\u4ee5\u524d\u4f5c\u54c1\u7684\u8bfb\u8005\u53ef\u80fd\u8fd8\u8bb0\u5f97\u6211\u66fe\u62a5\u9053\u8fc7 &#8220;\u6559\u79d1\u4e66\u5c31\u662f\u4f60\u6240\u9700\u8981\u7684\u4e00\u5207&#8221;\uff0c\u8fd9\u662f\u5fae\u8f6f\u516c\u53f8\u53d1\u8868\u7684\u4e00\u7bc7\u8bba\u6587\uff0c\u5c55\u793a\u4e86\u9ad8\u8d28\u91cf\u6570\u636e\u5982\u4f55\u5bf9\u6a21\u578b\u6027\u80fd\u4ea7\u751f\u5de8\u5927\u5f71\u54cd\u3002\u8be5\u8bba\u6587\u7684\u7814\u7a76\u7ed3\u679c\u76f4\u63a5\u9a73\u65a5\u4e86 &#8220;\u6a21\u578b\u5fc5\u987b\u975e\u5e38\u5e9e\u5927\u624d\u80fd\u80dc\u4efb &#8220;\u7684\u89c2\u70b9\u3002\u8fd9\u7bc7\u8bba\u6587\u80cc\u540e\u7684\u7814\u7a76\u4eba\u5458\u7ee7\u7eed\u4ed6\u4eec\u7684\u5de5\u4f5c\uff0c\u5e76\u53d1\u8868\u4e86\u4e00\u7bc7\u8ba9\u6211\u611f\u5230\u65e0\u6bd4\u5174\u594b\u7684\u8bba\u6587\u3002<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u6807\u9898\u4e5f\u8bb8\u89e3\u91ca\u4e86\u6700\u5927\u7684\u53d1\u73b0\uff1a &#8220;Phi-3\u6280\u672f\u62a5\u544a\uff1a \u624b\u673a\u672c\u5730\u7684\u9ad8\u80fd\u529b\u8bed\u8a00\u6a21\u578b&#8221;\u3002<\/p>\n<p>\u8ba9\u6211\u4eec\u6df1\u5165\u4e86\u89e3\u4e00\u4e0b\u4f5c\u8005\u5bf9 Phi-2 \u6a21\u578b\u505a\u4e86\u54ea\u4e9b\u6539\u52a8\uff0c\u4ed6\u4eec\u662f\u5982\u4f55\u8bad\u7ec3\u5b83\u7684\uff0c\u4ee5\u53ca\u5b83\u5728 iPhone \u4e0a\u662f\u5982\u4f55\u5de5\u4f5c\u7684\u3002<\/p>\n<p><strong><b>\u5173\u952e\u672f\u8bed<\/b><\/strong><\/p>\n<p>\u5728\u6df1\u5165\u4e86\u89e3\u67b6\u6784\u4e4b\u524d\uff0c\u6709\u51e0\u4e2a\u5173\u952e\u6982\u5ff5\u9700\u8981\u4e86\u89e3\u3002\u5982\u679c\u4f60\u5df2\u7ecf\u77e5\u9053\u8fd9\u4e9b\u6982\u5ff5\uff0c\u8bf7\u76f4\u63a5\u8df3\u5230\u4e0b\u4e00\u8282\u3002<\/p>\n<p>\u6a21\u578b<strong><b>parameters<\/b><\/strong>\u662f\u6307\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5230\u7684\u6743\u91cd\u548c\u504f\u7f6e\u7684\u6570\u91cf\u3002\u5982\u679c\u6709 10 \u4ebf\u4e2a\u53c2\u6570\uff0c\u90a3\u4e48\u5c31\u6709 10 \u4ebf\u4e2a\u6743\u91cd\u548c\u504f\u7f6e\u51b3\u5b9a\u4e86\u6a21\u578b\u7684\u6027\u80fd\u3002\u53c2\u6570\u8d8a\u591a\uff0c\u795e\u7ecf\u7f51\u7edc\u5c31\u8d8a\u590d\u6742\u3002<strong><b>head<\/b><\/strong>\u6307\u7684\u662f\u53d8\u5f62\u5668\u4e2d\u81ea\u6211\u5173\u6ce8\u673a\u5236\u6240\u62e5\u6709\u7684\u952e\u3001\u503c\u548c\u67e5\u8be2\u5411\u91cf\u7684\u6570\u91cf\u3002<strong><b>Layers<\/b><\/strong>\u662f\u6307 Transformer \u795e\u7ecf\u7f51\u7edc\u4e2d\u5b58\u5728\u7684\u795e\u7ecf\u6bb5\u7684\u6570\u91cf\uff0c\u9690\u85cf\u7ef4\u5ea6\u662f\u6307\u5178\u578b\u9690\u85cf\u5c42\u4e2d\u795e\u7ecf\u5143\u7684\u6570\u91cf\u3002<\/p>\n<p><strong><b>Tokenizer <\/b><\/strong>\u662f\u5c06\u8f93\u5165\u6587\u672c\u8f6c\u6362\u4e3a\u5d4c\u5165\u5f0f\u6587\u672c\u7684\u8f6f\u4ef6\uff0c\u8f6c\u6362\u5668\u968f\u540e\u5c06\u5bf9\u5176\u8fdb\u884c\u5904\u7406\u3002<strong><b>V<\/b><\/strong><strong><b>ocabulary size<\/b><\/strong>\u6307\u7684\u662f\u6a21\u578b\u6240\u8bad\u7ec3\u7684\u552f\u4e00\u6807\u8bb0\u7684\u6570\u91cf\u3002\u8f6c\u6362\u5668\u7684<strong><b>block<\/b><\/strong><strong><b>\u00a0<\/b><\/strong><strong><b>structure<\/b><\/strong>\u662f\u6307\u6211\u4eec\u4e3a\u7279\u5b9a\u6a21\u578b\u9009\u62e9\u7684\u5c42\u3001\u5934\u3001\u6fc0\u6d3b\u51fd\u6570\u3001\u6807\u8bb0\u5316\u5668\u548c\u5c42\u5f52\u4e00\u5316\u7684\u7ec4\u5408\u3002<\/p>\n<p>\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff08GQA\uff09\u662f\u6211\u4eec\u4f18\u5316\u591a\u5934\u6ce8\u610f\u529b\u4ee5\u51cf\u5c11\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u4e2d\u8ba1\u7b97\u5f00\u9500\u7684\u4e00\u79cd\u65b9\u6cd5\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0cGQA \u91c7\u7528\u7684\u662f\u4e2d\u5eb8\u4e4b\u9053&#8211;\u6211\u4eec\u4e0d\u662f\u5c06 1 \u4e2a\u503c\u548c 1 \u4e2a\u5173\u952e\u5b57\u4e0e 1 \u4e2a\u67e5\u8be2\u914d\u5bf9\uff0c\u800c\u662f\u91c7\u7528 1:1:M \u7684\u65b9\u6cd5\uff0c\u5373\u591a\u4e2a\u67e5\u8be2\u5c0f\u4e8e\u6574\u4e2a\u67e5\u8be2\u4f53\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\u662f\uff0c\u65e2\u80fd\u4ece\u591a\u67e5\u8be2\u6ce8\u610f\uff08MQA\uff09\u4e2d\u83b7\u5f97\u8bad\u7ec3\u6210\u672c\u4f18\u52bf\uff0c\u53c8\u80fd\u6700\u5927\u9650\u5ea6\u5730\u964d\u4f4e\u968f\u4e4b\u800c\u6765\u7684\u6027\u80fd\u4e0b\u964d\u3002<\/p>\n<p><strong><b>phi<\/b><\/strong><strong><b>\u67b6\u6784<\/b><\/strong><\/p>\n<p>\u6211\u4eec\u5148\u6765\u770b\u770b\u8fd9\u4e2a\u6a21\u578b\u80cc\u540e\u7684\u67b6\u6784\u3002\u7814\u7a76\u4eba\u5458\u53d1\u5e03\u4e86 3 \u79cd\u4e0d\u540c\u7684\u89e3\u7801\u5668\u6a21\u578b\uff1aphi-3-mini\u3001phi-3-small \u548c phi-3-medium\uff0c\u6bcf\u79cd\u6a21\u578b\u90fd\u6709\u4e0d\u540c\u7684\u8d85\u53c2\u6570\u3002<\/p>\n<ul>\n<li><i><\/i><em><i>phi-3-mini<\/i><\/em><\/li>\n<\/ul>\n<p>&#8211; 38 \u4ebf\u4e2a\u53c2\u6570<\/p>\n<p>&#8211; 32 \u4e2a\u5934<\/p>\n<p>&#8211; 32 \u5c42<\/p>\n<p>&#8211; 3072 \u4e2a\u9690\u85cf\u7ef4\u5ea6<\/p>\n<p>&#8211; 4k \u6807\u8bb0\u9ed8\u8ba4\u4e0a\u4e0b\u6587\u957f\u5ea6<\/p>\n<p>&#8211; 32064 \u8bcd\u6c47\u91cf<\/p>\n<p>&#8211; \u6743\u91cd\u5b58\u50a8\u4e3a bfloat16<\/p>\n<p>&#8211; \u5728 3.3 \u4e07\u4ebf\u4e2a\u4ee4\u724c\u4e0a\u8bad\u7ec3<\/p>\n<ul>\n<li><i><\/i><em><i>phi-3-small<\/i><\/em><\/li>\n<\/ul>\n<p>&#8211; 70 \u4ebf\u4e2a\u53c2\u6570<\/p>\n<p>&#8211; 32 \u4e2a\u5934<\/p>\n<p>&#8211; 32 \u5c42<\/p>\n<p>&#8211; 4096 \u4e2a\u9690\u85cf\u7ef4\u5ea6<\/p>\n<p>&#8211; 8k \u6807\u8bb0\u9ed8\u8ba4\u4e0a\u4e0b\u6587\u957f\u5ea6<\/p>\n<p>&#8211; 100352 \u8bcd\u6c47\u91cf<\/p>\n<p>&#8211; \u6743\u91cd\u5b58\u50a8\u4e3a bfloat16<\/p>\n<p>&#8211; \u5728 4.8 \u4e07\u4ebf\u4e2a\u4ee4\u724c\u4e0a\u8bad\u7ec3<\/p>\n<ul>\n<li><i><\/i><i><\/i><em><i>phi-3-medium<\/i><\/em><\/li>\n<\/ul>\n<p>&#8211; 140 \u4ebf\u4e2a\u53c2\u6570<\/p>\n<p>&#8211; 40 \u4e2a\u5934<\/p>\n<p>&#8211; 40 \u5c42<\/p>\n<p>&#8211; 3072 \u4e2a\u9690\u85cf\u7ef4\u5ea6<\/p>\n<p>&#8211; \u5728 4.8 \u4e07\u4ebf\u4e2a\u4ee3\u5e01\u4e0a\u8fdb\u884c\u8bad\u7ec3<\/p>\n<p>\u5173\u4e8e\u5176\u4e2d\u7684\u4e00\u4e9b\u5dee\u5f02\uff0cphi-3-mini\u6a21\u578b\u662f\u4f7f\u7528\u5178\u578b\u7684\u591a\u5934\u6ce8\u610f\u529b\u8fdb\u884c\u8bad\u7ec3\u7684\u3002\u867d\u7136\u8bba\u6587\u4e2d\u6ca1\u6709\u63d0\u53ca\uff0c\u4f46\u6211\u8ba4\u4e3a\uff0c\u7531\u4e8e\u8be5\u6a21\u578b\u7684\u89c4\u6a21\u5927\u7ea6\u662f\u5176\u4ed6\u4e24\u4e2a\u6a21\u578b\u7684\u4e00\u534a\uff0c\u56e0\u6b64\u4e0e\u591a\u5934\u76f8\u5173\u7684\u8bad\u7ec3\u6210\u672c\u5e76\u4e0d\u9ad8\u3002\u5f53\u7136\uff0c\u5f53\u4ed6\u4eec\u6269\u5927 phi-3-small \u7684\u89c4\u6a21\u65f6\uff0c\u4ed6\u4eec\u91c7\u7528\u4e86\u5206\u7ec4\u67e5\u8be2\u5173\u6ce8\uff0c4 \u4e2a\u67e5\u8be2\u4e0e 1 \u4e2a\u5173\u952e\u5b57\u76f8\u8fde\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u5c3d\u53ef\u80fd\u4f7f phi-3-mini \u7684\u533a\u5757\u7ed3\u6784\u63a5\u8fd1 LLaMa-2 \u7ed3\u6784\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\u662f\u4e3a\u4e86\u8ba9\u5f00\u6e90\u793e\u533a\u80fd\u591f\u7ee7\u7eed\u5229\u7528 Phi-3 \u5bf9 LLaMa-2 \u8fdb\u884c\u7814\u7a76\u3002\u8fd9\u5bf9\u4e8e\u8fdb\u4e00\u6b65\u4e86\u89e3\u8be5\u5757\u7ed3\u6784\u7684\u5a01\u529b\u662f\u6709\u610f\u4e49\u7684\u3002<\/p>\n<p>\u7136\u800c\uff0cPhi-3-small\u5e76\u6ca1\u6709\u4f7f\u7528LLaMa\u7684\u533a\u5757\u7ed3\u6784\uff0c\u800c\u662f\u9009\u62e9\u4f7f\u7528 tiktoken \u6807\u8bb0\u5668\uff0c\u5e76\u4ea4\u66ff\u4f7f\u7528\u5bc6\u96c6\u6ce8\u610f\u5c42\u548c\u65b0\u7684\u533a\u5757\u89e3\u6790\u6ce8\u610f\u5c42\u3002\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u4e3a\u8fd9\u4e9b\u6a21\u578b\u7684\u8bad\u7ec3\u6570\u636e\u96c6\u6dfb\u52a0\u4e86 10% \u7684\u591a\u8bed\u8a00\u6570\u636e\u3002<\/p>\n<p><strong><b>\u8bad\u7ec3\u548c\u6570\u636e\u4f18\u5316\u7ec4\u5408<\/b><\/strong><\/p>\n<p>\u4e0e Phi-2 \u7c7b\u4f3c\uff0c\u7814\u7a76\u4eba\u5458\u5728\u9ad8\u8d28\u91cf\u6570\u636e\u65b9\u9762\u6295\u5165\u4e86\u5927\u91cf\u7cbe\u529b\u3002\u5728\u751f\u6210\u7528\u4e8e\u8bad\u7ec3\u6a21\u578b\u7684\u6570\u636e\u65f6\uff0c\u4ed6\u4eec\u4f7f\u7528\u4e86\u4e0e\u4e4b\u524d\u7c7b\u4f3c\u7684 &#8220;\u6559\u80b2\u4ef7\u503c &#8220;\u8303\u5f0f\uff0c\u5e76\u9009\u62e9\u4f7f\u7528\u6bd4\u4e0a\u6b21\u591a\u5f97\u591a\u7684\u6570\u636e\u3002\u4ed6\u4eec\u5206\u4e24\u4e2a\u9636\u6bb5\u521b\u5efa\u6570\u636e\u3002<\/p>\n<p>\u7b2c\u4e00\u9636\u6bb5\u662f\u5bfb\u627e\u4ed6\u4eec\u8ba4\u4e3a\u5bf9\u7528\u6237\u5177\u6709\u9ad8 &#8220;\u6559\u80b2\u4ef7\u503c &#8220;\u7684\u7f51\u7edc\u6570\u636e\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\u662f\u4e3a\u6a21\u578b\u63d0\u4f9b\u5e38\u8bc6\u3002\u7136\u540e\uff0c\u7b2c\u4e8c\u9636\u6bb5\u4ece\u7b2c\u4e00\u9636\u6bb5\u7684\u6570\u636e\u4e2d\u63d0\u53d6\u4e00\u4e2a\u5b50\u96c6\uff0c\u751f\u6210\u80fd\u6559\u4f1a\u6a21\u578b\u5982\u4f55\u8fdb\u884c\u903b\u8f91\u63a8\u7406\u6216\u83b7\u5f97\u7279\u5b9a\u6280\u80fd\u7684\u6570\u636e\u3002<\/p>\n<p>\u8fd9\u91cc\u7684\u6311\u6218\u5728\u4e8e\u5982\u4f55\u786e\u4fdd\u6bcf\u4e2a\u8bed\u6599\u5e93\u7684\u6570\u636e\u7ec4\u5408\u4e0e\u6240\u8bad\u7ec3\u6a21\u578b\u7684\u89c4\u6a21\u76f8\u9002\u5e94\uff08\u5373 phi-3-small \u4e0e phi-3-mini\uff09\u3002\u8fd9\u5c31\u662f &#8220;\u6570\u636e\u6700\u4f18 &#8220;\u673a\u5236\u80cc\u540e\u7684\u7406\u5ff5\uff0c\u5373\u63d0\u4f9b\u7ed9 LLM \u8fdb\u884c\u8bad\u7ec3\u7684\u6570\u636e\u80fd\u8ba9\u5b83\u53d1\u6325\u5176\u5757\u7ed3\u6784\u7684\u6700\u4f73\u80fd\u529b\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5982\u679c\u60a8\u8ba4\u4e3a\u6570\u636e\u662f\u8bad\u7ec3\u4f18\u79c0 LLM \u7684\u5173\u952e\u56e0\u7d20\uff0c\u90a3\u4e48\u627e\u5230\u901a\u8fc7\u6570\u636e\u5411\u6a21\u578b\u5c55\u793a\u7684\u6b63\u786e\u6280\u80fd\u7ec4\u5408\u4e0e\u627e\u5230\u4f18\u79c0\u6570\u636e\u540c\u6837\u5173\u952e\u3002\u7814\u7a76\u4eba\u5458\u5f3a\u8c03\uff0c\u4ed6\u4eec\u5e0c\u671b\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u5f3a\u4e8e\u77e5\u8bc6\u80fd\u529b\uff0c\u56e0\u6b64\u4ed6\u4eec\u4ece\u7b2c\u4e8c\u9636\u6bb5\u8bed\u6599\u5e93\u4e2d\u9009\u62e9\u4e86\u6bd4\u7b2c\u4e00\u9636\u6bb5\u8bed\u6599\u5e93\u66f4\u591a\u7684\u6570\u636e\u3002<\/p>\n<p>\u6709\u8da3\u7684\u662f\uff0c\u5f53\u4ed6\u4eec\u4f7f\u7528\u4e0e\u8bad\u7ec3 phi-3-small \u65f6\u5927\u81f4\u76f8\u540c\u7684\u6570\u636e\u6df7\u5408\u7269\u8bad\u7ec3 phi-3-medium \u65f6\uff0c\u4ed6\u4eec\u6ce8\u610f\u5230\u4ece 7B \u53c2\u6570\u5230 14B \u53c2\u6570\u7684\u6539\u8fdb\u8fdc\u6bd4\u4ece 3.8B \u5230 7B \u53c2\u6570\u7684\u6539\u8fdb\u6709\u9650\u3002\u4f5c\u8005\u6000\u7591\u8fd9\u4e0d\u662f\u533a\u5757\u7ed3\u6784\u7684\u9650\u5236\uff0c\u800c\u662f\u4ed6\u4eec\u7528\u6765\u8bad\u7ec3 phi-3-medium \u7684\u6570\u636e\u6df7\u5408\u7269\u7684\u9650\u5236\u3002<\/p>\n<p><strong><b>\u540e\u671f\u8bad\u7ec3<\/b><\/strong><\/p>\n<p>\u7814\u7a76\u5c0f\u7ec4\u4f7f\u7528\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u6765\u6539\u8fdb\u8bad\u7ec3\u540e\u7684\u6a21\u578b\u3002\u6709\u5174\u8da3\u6df1\u5165\u4e86\u89e3 DPO \u7684\u8bfb\u8005\u53ef\u4ee5\u70b9\u51fb\u8fd9\u91cc\u67e5\u770b\u6211\u7684\u535a\u6587\u3002\u76d1\u7763\u5fae\u8c03\u662f\u4e00\u79cd\u8fc1\u79fb\u5b66\u4e60\uff0c\u6211\u4eec\u4f7f\u7528\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u6765\u63d0\u9ad8 LLM \u5728\u8be5\u6570\u636e\u96c6\u4e0a\u7684\u80fd\u529b\u3002\u4f5c\u8005\u4f7f\u7528 SFT \u63d0\u9ad8\u4e86\u6a21\u578b\u5728\u6570\u5b66\u3001\u7f16\u7801\u3001\u63a8\u7406\u548c\u5b89\u5168\u7b49\u4e0d\u540c\u9886\u57df\u7684\u80fd\u529b\u3002\u7136\u540e\uff0c\u4ed6\u4eec\u4f7f\u7528 DPO \u8fdb\u884c\u804a\u5929\u4f18\u5316\uff0c\u5f15\u5bfc\u804a\u5929\u4ece\u4ed6\u4eec\u5e0c\u671b\u907f\u514d\u7684\u56de\u7b54\u8f6c\u5411\u7406\u60f3\u7684\u56de\u7b54\u3002<\/p>\n<p>\u5728\u8fd9\u4e00\u9636\u6bb5\uff0c\u4f5c\u8005\u5c06 phi-3-mini \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4ece 4k \u5b57\u8282\u6269\u5c55\u5230 128k \u5b57\u8282\u3002\u4ed6\u4eec\u91c7\u7528\u7684\u65b9\u6cd5\u79f0\u4e3a &#8220;\u957f\u7ef3&#8221;\uff08Long Rope\uff09\u3002\u4f5c\u8005\u58f0\u79f0\uff0c\u4e24\u79cd\u8bed\u5883\u7c7b\u578b\u7684\u6027\u80fd\u662f\u4e00\u81f4\u7684\uff0c\u8fd9\u5728\u8bed\u5883\u957f\u5ea6\u5927\u5e45\u589e\u52a0\u7684\u60c5\u51b5\u4e0b\u662f\u5f88\u91cd\u8981\u7684\u3002\u5982\u679c\u6709\u8db3\u591f\u7684\u5174\u8da3\uff0c\u6211\u5c06\u5c31\u8be5\u8bba\u6587\u4e2d\u7684\u7814\u7a76\u7ed3\u679c\u5355\u72ec\u64b0\u5199\u4e00\u7bc7\u535a\u6587\u3002<\/p>\n<p><strong><b>\u624b\u673a\u4f7f\u7528\u7684\u91cf\u5316<\/b><\/strong><\/p>\n<p>\u5c3d\u7ba1\u8fd9\u4e9b\u6a21\u578b\u5f88\u5c0f\uff0c\u4f46\u8981\u8ba9\u8fd9\u4e9b\u6a21\u578b\u5728\u624b\u673a\u4e0a\u8fd0\u884c\uff0c\u4ecd\u9700\u8981\u8fdb\u4e00\u6b65\u6700\u5c0f\u5316\u3002LLM \u7684\u6743\u91cd\u901a\u5e38\u5b58\u50a8\u4e3a\u6d6e\u70b9\u6570\uff1b\u4f8b\u5982\uff0cPhi-3 \u7684\u539f\u59cb\u6743\u91cd\u4e3a\u00a0bfloat16\uff0c\u8fd9\u610f\u5473\u7740\u6bcf\u4e2a\u6743\u91cd\u5360\u7528\u00a016 \u4f4d\u5185\u5b58\u3002\u867d\u7136\u00a016 \u4f4d\u770b\u8d77\u6765\u5fae\u4e0d\u8db3\u9053\uff0c\u4f46\u5982\u679c\u8003\u8651\u5230\u6a21\u578b\u4e2d\u5927\u7ea6\u6709\u00a010\u2079\u00a0\u4e2a\u53c2\u6570\uff0c\u4f60\u5c31\u4f1a\u610f\u8bc6\u5230\u6bcf\u4e2a\u989d\u5916\u4f4d\u7684\u589e\u52a0\u901f\u5ea6\u6709\u591a\u5feb\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u4f5c\u8005\u5c06\u6743\u91cd\u4ece 16 \u4f4d\u538b\u7f29\u5230 4 \u4f4d\u3002\u5176\u57fa\u672c\u601d\u60f3\u662f\u51cf\u5c11\u5b58\u50a8\u6bcf\u4e2a\u6570\u5b57\u6240\u9700\u7684\u6bd4\u7279\u6570\u3002\u4e3e\u4e2a\u6982\u5ff5\u6027\u7684\u4f8b\u5b50\uff0c\u6570\u5b57 2.71828 \u53ef\u4ee5\u538b\u7f29\u4e3a 2.72\u3002\u867d\u7136\u8fd9\u662f\u4e00\u79cd\u6709\u635f\u64cd\u4f5c\uff0c\u4f46\u4ecd\u80fd\u6355\u6349\u5230\u5927\u90e8\u5206\u4fe1\u606f\uff0c\u540c\u65f6\u5927\u5927\u51cf\u5c11\u4e86\u5b58\u50a8\u7a7a\u95f4\u3002<\/p>\n<p>\u4f5c\u8005\u5728\u914d\u5907 A16 \u82af\u7247\u7684 iPhone \u4e0a\u8fd0\u884c\u4e86\u91cf\u5316\u7a0b\u5e8f\uff0c\u53d1\u73b0\u5b83\u6bcf\u79d2\u6700\u591a\u80fd\u751f\u6210 12 \u4e2a\u6587\u5b57\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u8fd0\u884c LLaMa-2 \u91cf\u5316 4 \u4f4d\u7684 M1 MacBook \u7684\u901f\u5ea6\u5927\u7ea6\u4e3a\u6bcf\u79d2 107 \u4e2a\u6807\u8bb0\u3002\u6211\u89c1\u8fc7\u7684\u6700\u5feb\u4ee4\u724c\u751f\u6210\u901f\u5ea6\uff08Groq\uff09\u4e3a\u6bcf\u79d2\u751f\u6210 853.35 \u4e2a\u4ee4\u724c\u3002\u8003\u8651\u5230\u8fd9\u4ec5\u4ec5\u662f\u4e2a\u5f00\u59cb\uff0c\u6211\u4eec\u80fd\u5728 iPhone \u4e0a\u770b\u5230\u7528\u8fd9\u79cd\u6a21\u578b\u751f\u6210\u6807\u8bb0\u7684\u901f\u5ea6\u4e4b\u5feb\u4ee4\u4eba\u60ca\u53f9\u3002\u770b\u6765\u63a8\u7406\u901f\u5ea6\u53ea\u4f1a\u8d8a\u6765\u8d8a\u5feb\u3002<\/p>\n<p><strong><b>\u5c06 Phi-3 \u4e0e\u641c\u7d22\u914d\u5bf9<\/b><\/strong><\/p>\n<p>\u5c0f\u578b\u6a21\u578b\u7684\u4e00\u4e2a\u5c40\u9650\u662f\uff0c\u5b83\u5728\u7f51\u7edc\u4e2d\u53ef\u4ee5\u5b58\u50a8\u4fe1\u606f\u7684\u5730\u65b9\u8f83\u5c11\u3002\u56e0\u6b64\uff0c\u5728\u9700\u8981\u5e7f\u6cdb\u77e5\u8bc6\u7684\u4efb\u52a1\u4e2d\uff0cPhi-3 \u7684\u8868\u73b0\u4e0d\u5982 LLaMa-2 \u7b49\u6a21\u578b\u3002<\/p>\n<p>\u4f5c\u8005\u8ba4\u4e3a\uff0c\u5982\u679c\u5c06 Phi-3 \u4e0e\u641c\u7d22\u5f15\u64ce\u642d\u914d\u4f7f\u7528\uff0c\u8be5\u6a21\u578b\u7684\u80fd\u529b\u5c06\u5f97\u5230\u663e\u8457\u63d0\u9ad8\u3002\u5982\u679c\u771f\u662f\u8fd9\u6837\uff0c\u6211\u8ba4\u4e3a\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u5f88\u53ef\u80fd\u4f1a\u7ee7\u7eed\u5b58\u5728\uff0c\u5e76\u6210\u4e3a\u5e2e\u52a9\u5c0f\u578b\u6a21\u578b\u4e0e\u5927\u578b\u6a21\u578b\u4e00\u6837\u51fa\u8272\u7684\u5173\u952e\u90e8\u5206\u3002<\/p>\n<p><strong><b>\u7ed3\u8bba<\/b><\/strong><\/p>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u770b\u5230\u4e86\u9ad8\u6027\u80fd\u5c0f\u578b\u6a21\u578b\u7684\u96cf\u5f62\u3002\u867d\u7136\u8fd9\u4e9b\u6a21\u578b\u7684\u8bad\u7ec3\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4ecd\u7136\u4f9d\u8d56\u4e8e\u9ad8\u6027\u80fd\u7684\u786c\u4ef6\uff0c\u4f46\u5bf9\u5b83\u4eec\u8fdb\u884c\u63a8\u7406\u6b63\u65e5\u76ca\u6c11\u4e3b\u5316\u3002\u8fd9\u5e26\u6765\u4e86\u4e00\u4e9b\u6709\u8da3\u7684\u73b0\u8c61\u3002<\/p>\n<p>\u9996\u5148\uff0c\u53ef\u4ee5\u5728\u672c\u5730\u8fd0\u884c\u7684\u6a21\u578b\u51e0\u4e4e\u53ef\u4ee5\u5b8c\u5168\u4fdd\u5bc6\uff0c\u7528\u6237\u53ef\u4ee5\u5411\u8fd9\u4e9b LLM \u63d0\u4f9b\u4ed6\u4eec\u53ef\u80fd\u4e0d\u613f\u610f\u901a\u8fc7\u4e92\u8054\u7f51\u53d1\u9001\u7684\u6570\u636e\u3002\u8fd9\u4e3a\u66f4\u591a\u7684\u4f7f\u7528\u6848\u4f8b\u6253\u5f00\u4e86\u5927\u95e8\u3002<\/p>\n<p>\u5176\u6b21\uff0c\u8fd9\u4e9b\u6a21\u578b\u5c06\u63a8\u52a8\u79fb\u52a8\u786c\u4ef6\u7684\u6027\u80fd\u8fdb\u4e00\u6b65\u63d0\u9ad8\u3002\u56e0\u6b64\uff0c\u6211\u5e0c\u671b\u5728\u9ad8\u7aef\u667a\u80fd\u624b\u673a\u4e0a\u770b\u5230\u66f4\u591a\u7684\u7247\u4e0a\u7cfb\u7edf\uff08SoC\uff09\uff0c\u5c24\u5176\u662f CPU \u548c GPU \u5171\u4eab\u5185\u5b58\u7684\u7247\u4e0a\u7cfb\u7edf\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\u3002\u6b64\u5916\uff0c\u4e3a\u8fd9\u4e9b\u786c\u4ef6\u63d0\u4f9b\u9ad8\u8d28\u91cf\u7684\u63a5\u53e3\u4e5f\u81f3\u5173\u91cd\u8981\u3002\u4efb\u4f55\u65b0\u8fdb\u5165\u6d88\u8d39\u786c\u4ef6\u9886\u57df\u7684\u786c\u4ef6\u90fd\u53ef\u80fd\u9700\u8981\u50cf\u82f9\u679c\u7845\u7684 MLX \u8fd9\u6837\u7684\u5e93\u3002<\/p>\n<p>\u7b2c\u4e09\uff0c\u672c\u6587\u8868\u660e\uff0c\u5728 LLM \u4e2d\uff0c\u9ad8\u8d28\u91cf\u6570\u636e\u5728\u5f88\u591a\u65b9\u9762\u90fd\u80fd\u8d85\u8d8a\u66f4\u591a\u7684\u7f51\u7edc\u590d\u6742\u6027\uff0c\u56e0\u6b64\uff0c\u4e0d\u4ec5\u8981\u627e\u5230\u800c\u4e14\u8981\u751f\u6210\u9ad8\u8d28\u91cf\u6570\u636e\u7684\u7ade\u4e89\u53ea\u4f1a\u8d8a\u6765\u8d8a\u6fc0\u70c8\u3002<\/p>\n<p>\u8fd9\u662f\u4e00\u4e2a\u4ee4\u4eba\u5174\u594b\u7684\u65f6\u4ee3\u3002<\/p>\n<p><a href=\"https:\/\/towardsdatascience.com\/phi-3-and-the-beginning-of-highly-performant-iphone-models-d413d8ea0714?gi=2fc5e133e481\"><u>https:\/\/towardsdatascience.com\/phi-3-and-the-beginning-of-highly-performant-iphone-models-d413d8ea0714?gi=2fc5e133e481<\/u><\/a><\/p>\n<p>\uff08\u673a\u5668\u7ffb\u8bd1\uff0c\u8f7b\u5ea6\u8bd1\u540e\u7f16\u8f91\uff0c\u4ec5\u4f9b\u53c2\u8003\uff09<\/p>\n<p>\u7f16\u8f91\uff1a\u5218\u6167<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bfb\u8fc7\u6211\u4ee5\u524d\u4f5c\u54c1\u7684\u8bfb\u8005\u53ef\u80fd\u8fd8\u8bb0\u5f97\u6211\u66fe\u62a5\u9053\u8fc7 &#8220;\u6559\u79d1\u4e66\u5c31\u662f\u4f60\u6240\u9700\u8981\u7684\u4e00\u5207&#8221;\uff0c\u8fd9\u662f\u5fae\u8f6f\u516c\u53f8\u53d1 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":30520,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[391],"tags":[],"class_list":["post-30519","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-391"],"_links":{"self":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/30519","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=30519"}],"version-history":[{"count":2,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/30519\/revisions"}],"predecessor-version":[{"id":30522,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/30519\/revisions\/30522"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/media\/30520"}],"wp:attachment":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=30519"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=30519"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=30519"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}