Notice: Function _load_textdomain_just_in_time was called incorrectly. Translation loading for the betterdocs 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 init action or later. Please see Debugging in WordPress for more information. (This message was added in version 6.7.0.) in /data/user/htdocs/wp-includes/functions.php on line 6114

Notice: Function _load_textdomain_just_in_time was called incorrectly. Translation loading for the jnews-view-counter 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 init action or later. Please see Debugging in WordPress for more information. (This message was added in version 6.7.0.) in /data/user/htdocs/wp-includes/functions.php on line 6114

Notice: Function _load_textdomain_just_in_time was called incorrectly. Translation loading for the wp-statistics 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 init action or later. Please see Debugging in WordPress for more information. (This message was added in version 6.7.0.) in /data/user/htdocs/wp-includes/functions.php on line 6114

Notice: Function _load_textdomain_just_in_time was called incorrectly. Translation loading for the wpdiscuz 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 init action or later. Please see Debugging in WordPress for more information. (This message was added in version 6.7.0.) in /data/user/htdocs/wp-includes/functions.php on line 6114

Notice: 函数 _load_textdomain_just_in_time 的调用方法不正确。 jnews 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 init 操作或之后加载。请查阅调试 WordPress来获取更多信息。（这个消息是在 6.7.0 版本添加的。） in /data/user/htdocs/wp-includes/functions.php on line 6114

Notice: 函数 _load_textdomain_just_in_time 的调用方法不正确。 jnews-like 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 init 操作或之后加载。请查阅调试 WordPress来获取更多信息。（这个消息是在 6.7.0 版本添加的。） in /data/user/htdocs/wp-includes/functions.php on line 6114

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893

Warning: Cannot modify header information - headers already sent by (output started at /data/user/htdocs/wp-includes/functions.php:6114) in /data/user/htdocs/wp-includes/rest-api/class-wp-rest-server.php on line 1893
{"id":32399,"date":"2024-08-20T08:46:38","date_gmt":"2024-08-20T00:46:38","guid":{"rendered":"https:\/\/linguaresources.com\/?p=32399"},"modified":"2024-08-20T08:46:38","modified_gmt":"2024-08-20T00:46:38","slug":"%e8%bf%84%e4%bb%8a%e4%b8%ba%e6%ad%a2%e6%9c%80%e5%bc%ba%e5%a4%a7%e7%9a%84%e5%bc%80%e6%ba%90-llm%ef%bc%9a-meta-llama-3-1-405b","status":"publish","type":"post","link":"https:\/\/linguaresources.com\/?p=32399","title":{"rendered":"\u8fc4\u4eca\u4e3a\u6b62\u6700\u5f3a\u5927\u7684\u5f00\u6e90 LLM\uff1a META LLAMA 3.1-405B"},"content":{"rendered":"

Llama 3.1-405B<\/u><\/a>\u7531 Meta AI \u5f00\u53d1\u7684 Llama 3.1-405B \u4ee3\u8868\u4e86\u5f00\u6e90\u8bed\u8a00\u6a21\u578b\u7684\u91cd\u5927\u98de\u8dc3\u3002 \u5b83\u62e5\u6709 4050 \u4ebf\u4e2a\u53c2\u6570\uff0c\u662f\u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u7684\u516c\u5f00\u53ef\u7528\u8bed\u8a00\u6a21\u578b\uff0c\u5728\u5404\u79cd\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u53ef\u4e0e\u4e00\u4e9b\u6700\u5148\u8fdb\u7684\u4e13\u6709\u6a21\u578b\u76f8\u5ab2\u7f8e\uff0c\u751a\u81f3\u8d85\u8d8a\u5b83\u4eec\u3002<\/p>\n

\u4e3b\u8981\u529f\u80fd<\/b><\/strong><\/h3>\n
\n
<\/b>4050 \u4ebf\u4e2a\u53c2\u6570<\/b><\/strong><\/li>\n
<\/b>128K \u6807\u8bb0\u4e0a\u4e0b\u6587\u957f\u5ea6<\/b><\/strong><\/li>\n
\u652f\u6301\u591a\u79cd\u8bed\u8a00\uff088 \u79cd\u8bed\u8a00\uff09<\/li>\n
<\/b>\u7ecf\u8fc7\u8bf4\u660e\u8c03\u6574\u7684<\/b><\/strong>\u7248\u672c\u53ef\u7528<\/li>\n
<\/b>\u5f00\u653e\u6e90\u4ee3\u7801<\/b><\/strong>\u5177\u6709\u8bb8\u53ef\u534f\u8bae<\/li>\n<\/ol>\n
\u5728\u5f00\u6e90\u9886\u57df\u53d1\u5e03\u8fd9\u6837\u4e00\u4e2a\u529f\u80fd\u5f3a\u5927\u7684\u6a21\u578b\u6539\u53d8\u4e86\u6e38\u620f\u89c4\u5219\uff0c\u4f7f\u6700\u5148\u8fdb\u7684\u4eba\u5de5\u667a\u80fd\u529f\u80fd\u7684\u83b7\u53d6\u53d8\u5f97\u66f4\u52a0\u6c11\u4e3b\uff0c\u5e76\u4fc3\u8fdb\u4e86\u6574\u4e2a\u884c\u4e1a\u7684\u521b\u65b0\u3002<\/p>\n
\u6a21\u578b\u67b6\u6784\u548c\u8bad\u7ec3<\/b><\/strong><\/h2>\n
\u8fd9\u4e00\u8fc7\u7a0b\u4ece\u5c06\u8f93\u5165\u6587\u672c\u6807\u8bb0\u8f6c\u6362\u4e3a\u6807\u8bb0\u5d4c\u5165\u5f00\u59cb\u3002 \u8fd9\u4e9b\u5d4c\u5165\u7b26\u901a\u8fc7\u591a\u5c42\u81ea\u6ce8\u610f\u548c\u524d\u9988\u7f51\u7edc\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u6355\u6349\u6587\u672c\u4e2d\u7684\u590d\u6742\u5173\u7cfb\u548c\u4f9d\u8d56\u6027\u3002 \u7136\u540e\uff0c\u81ea\u56de\u5f52\u89e3\u7801\u673a\u5236\u751f\u6210\u8f93\u51fa\u6587\u672c\u6807\u8bb0\uff0c\u5b8c\u6210\u6574\u4e2a\u8fc7\u7a0b\u3002<\/p>\n
1.\u00a0<\/b>\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b (GQA)<\/b><\/strong><\/h3>\n
\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f<\/p>\n
Llama 3.1 \u91c7\u7528\u4e86\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u6280\u672f\uff0c\u8fd9\u662f\u4e00\u9879\u91cd\u8981\u7684\u4f18\u5316\u6280\u672f\uff0c\u4f46\u5728\u524d\u9762\u7684\u56de\u7b54\u4e2d\u5e76\u672a\u5b8c\u5168\u6d89\u53ca\u3002 \u8ba9\u6211\u4eec\u6765\u8be6\u7ec6\u63a2\u8ba8\u4e00\u4e0b\uff1a<\/p>\n
\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff08GQA\uff09\u662f\u591a\u5934\u6ce8\u610f\u529b\u7684\u4e00\u79cd\u53d8\u4f53\uff0c\u65e8\u5728\u51cf\u5c11\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u6210\u672c\u548c\u5185\u5b58\u4f7f\u7528\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u957f\u5e8f\u5217\u3002 \u5728 Llama 3.1 405B \u6a21\u578b\u4e2d\uff0cGQA \u662f\u901a\u8fc7 8 \u4e2a\u952e\u503c\u5934\u5b9e\u73b0\u7684\u3002<\/p>\n
\u4ee5\u4e0b\u662f GQA \u7684\u5de5\u4f5c\u539f\u7406\uff1a<\/p>\n
\n
GQA \u5c06\u591a\u4e2a\u67e5\u8be2\u5934\u5206\u7ec4\uff0c\u5171\u4eab\u76f8\u540c\u7684\u952e\u548c\u503c\u5934\uff0c\u800c\u4e0d\u662f\u4e3a\u6bcf\u4e2a\u6ce8\u610f\u5934\u5206\u522b\u5efa\u7acb\u952e\u548c\u503c\u9884\u6d4b\u3002<\/li>\n
\u8fd9\u79cd\u5206\u7ec4\u65b9\u5f0f\u5927\u5927\u51cf\u5c11\u4e86\u952e\u548c\u503c\u6295\u5f71\u4e2d\u7684\u53c2\u6570\u6570\u91cf\uff0c\u4ece\u800c\u7f29\u5c0f\u4e86\u6a21\u578b\u89c4\u6a21\uff0c\u52a0\u5feb\u4e86\u63a8\u7406\u901f\u5ea6\u3002<\/li>\n
\u6ce8\u610f\u529b\u8ba1\u7b97\u53ef\u8868\u793a\u4e3a<\/li>\n<\/ol>\n
\u6ce8\u610f\u529b\uff08Q\u3001K\u3001V\uff09= softmax(QK^T \/ sqrt(d_k))V<\/p>\n
\u5176\u4e2d Q \u88ab\u5206\u4e3a g \u7ec4\uff0cK \u548c V \u7684\u5934\u6570\u5c11\u4e8e Q\u3002<\/p>\n
Llama 3.1 405B \u4e2d GQA \u7684\u4f18\u70b9\u5305\u62ec<\/p>\n
\n
<\/b>\u51cf\u5c11\u5185\u5b58\u5360\u7528<\/b><\/strong>\uff1a \u66f4\u5c11\u7684\u952e\u548c\u503c\u6295\u5f71\u610f\u5473\u7740\u5b58\u50a8\u6a21\u578b\u53c2\u6570\u6240\u9700\u7684\u5185\u5b58\u66f4\u5c11\u3002<\/li>\n
<\/b>\u63a8\u7406\u901f\u5ea6\u66f4\u5feb<\/b><\/strong>\uff1a \u7531\u4e8e\u952e\u548c\u503c\u9884\u6d4b\u6240\u9700\u7684\u8ba1\u7b97\u91cf\u51cf\u5c11\uff0c\u63a8\u7406\u901f\u5ea6\u5f97\u5230\u63d0\u9ad8\u3002<\/li>\n
<\/b>\u4fdd\u6301\u6027\u80fd<\/b><\/strong>\uff1a \u5c3d\u7ba1\u53c2\u6570\u6709\u6240\u51cf\u5c11\uff0c\u4f46\u5728\u8bb8\u591a\u4efb\u52a1\u4e2d\uff0cGQA \u4ecd\u80fd\u4fdd\u6301\u4e0e\u6807\u51c6\u591a\u5934\u6ce8\u610f\u529b\u76f8\u5f53\u7684\u6027\u80fd\u3002<\/li>\n<\/ol>\n
1.\u00a0<\/b>\u6269\u5c55\u8bed\u5883\u7684\u4e24\u9636\u6bb5\u9884\u8bad\u7ec3<\/b><\/strong><\/h3>\n
\u6587\u7ae0\u63d0\u5230\u4e86\u5b9e\u73b0 128K \u6807\u8bb0\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u4e24\u9636\u6bb5\u9884\u8bad\u7ec3\u8fc7\u7a0b\u3002 \u8fd9\u662f Llama 3.1 405B \u80fd\u529b\u7684\u4e00\u4e2a\u91cd\u8981\u65b9\u9762\uff1a<\/p>\n
\u7b2c 1 \u9636\u6bb5\uff1a\u5bf9 8K \u4e2a\u8bcd\u7ec4\u8fdb\u884c\u521d\u59cb\u9884\u8bad\u7ec3<\/b><\/strong><\/p>\n
\n
\u9996\u5148\u5728\u591a\u8fbe 8K \u4e2a\u8bcd\u7ec4\u7684\u5e8f\u5217\u4e0a\u5bf9\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\u3002<\/li>\n
\u8fd9\u4e00\u9636\u6bb5\u53ef\u8ba9\u6a21\u578b\u5b66\u4e60\u4e00\u822c\u8bed\u8a00\u7406\u89e3\u548c\u751f\u6210\u80fd\u529b\u3002<\/li>\n<\/ol>\n
\u7b2c\u4e8c\u9636\u6bb5\uff1a\u4e3a\u6269\u5c55\u8bed\u5883\u7ee7\u7eed\u8fdb\u884c\u9884\u8bad\u7ec3<\/b><\/strong><\/p>\n
\n
\u521d\u59cb\u8bad\u7ec3\u7ed3\u675f\u540e\uff0c\u6a21\u578b\u5c06\u7ee7\u7eed\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u4ee5\u5c06\u4e0a\u4e0b\u6587\u957f\u5ea6\u589e\u52a0\u5230 128K \u5b57\u8282\u3002<\/li>\n
\u8fd9\u4e00\u9636\u6bb5\u9700\u8981\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u8bad\u7ec3\u65b9\u6cd5\uff0c\u4ee5\u5e2e\u52a9\u6a21\u578b\u6cdb\u5316\u5230\u66f4\u957f\u7684\u5e8f\u5217\uff0c\u540c\u65f6\u53c8\u4e0d\u4e27\u5931\u5904\u7406\u8f83\u77ed\u4e0a\u4e0b\u6587\u7684\u80fd\u529b\u3002<\/li>\n<\/ol>\n
1.\u00a0<\/b>\u591a\u6a21\u6001\u80fd\u529b<\/b><\/strong><\/h3>\n
\u867d\u7136\u524d\u9762\u7684\u56de\u7b54\u6d89\u53ca\u4e86\u591a\u6a21\u6001\u529f\u80fd\uff0c\u4f46\u6211\u4eec\u53ef\u4ee5\u8fdb\u4e00\u6b65\u8bf4\u660e Llama 3.1 405B \u662f\u5982\u4f55\u5b9e\u73b0\u591a\u6a21\u6001\u529f\u80fd\u7684\uff1a<\/p>\n
\u5408\u6210\u65b9\u6cd5\uff1a<\/b><\/strong><\/p>\n
\n
Llama 3.1 405B \u5bf9\u4e0d\u540c\u7684\u6a21\u5f0f\uff08\u5982\u56fe\u50cf\u3001\u8bed\u97f3\uff09\u4f7f\u7528\u4e0d\u540c\u7684\u7f16\u7801\u5668\u3002<\/li>\n
\u8fd9\u4e9b\u7f16\u7801\u5668\u5c06\u6765\u81ea\u4e0d\u540c\u6a21\u6001\u7684\u8f93\u5165\u8f6c\u5316\u4e3a\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u7406\u89e3\u7684\u5171\u4eab\u5d4c\u5165\u7a7a\u95f4\u3002<\/li>\n<\/ol>\n
\u4e0e\u8bed\u8a00\u6a21\u578b\u6574\u5408\uff1a<\/b><\/strong><\/p>\n
\n
\u8fd9\u4e9b\u4e13\u7528\u7f16\u7801\u5668\u7684\u8f93\u51fa\u7ed3\u679c\u4f1a\u88ab\u8f93\u5165\u4e3b\u8bed\u8a00\u6a21\u578b\u3002<\/li>\n
\u8fd9\u6837\uff0cLlama 3.1 405B \u5c31\u80fd\u540c\u65f6\u5904\u7406\u548c\u7406\u89e3\u4e0d\u540c\u7c7b\u578b\u7684\u6570\u636e\uff0c\u6267\u884c\u6d89\u53ca\u591a\u79cd\u6a21\u5f0f\u7684\u4efb\u52a1\u3002<\/li>\n<\/ol>\n
\u4ea4\u53c9\u6ce8\u610f\u673a\u5236\uff1a<\/b><\/strong><\/p>\n
\n
\u4e3a\u4e86\u5904\u7406\u4e0d\u540c\u6a21\u5f0f\u7684\u6574\u5408\u95ee\u9898\uff0cLlama 3.1 405B \u5f88\u53ef\u80fd\u91c7\u7528\u4e86\u4ea4\u53c9\u6ce8\u610f\u673a\u5236\u3002<\/li>\n
\u8fd9\u4e9b\u673a\u5236\u5141\u8bb8\u6a21\u578b\u5728\u751f\u6210\u6587\u672c\u6216\u6267\u884c\u5176\u4ed6\u4efb\u52a1\u65f6\u5173\u6ce8\u6765\u81ea\u4e0d\u540c\u6a21\u6001\u7684\u76f8\u5173\u4fe1\u606f\u3002<\/li>\n<\/ol>\n
Llama 3.1 405B \u7684\u591a\u6a21\u6001\u529f\u80fd\u5f00\u8f9f\u4e86\u5e7f\u6cdb\u7684\u5e94\u7528\u9886\u57df\uff0c\u4f8b\u5982<\/p>\n
\n
\u56fe\u50cf\u5b57\u5e55\u548c\u89c6\u89c9\u95ee\u9898\u89e3\u7b54<\/li>\n
\u8bed\u97f3\u5230\u6587\u672c\u8f6c\u5f55\u4e0e\u4e0a\u4e0b\u6587\u7406\u89e3<\/li>\n
\u7ed3\u5408\u6587\u672c\u3001\u56fe\u50cf\u548c\u53ef\u80fd\u7684\u5176\u4ed6\u6570\u636e\u7c7b\u578b\u7684\u591a\u6a21\u5f0f\u63a8\u7406\u4efb\u52a1<\/li>\n<\/ol>\n
\u8bad\u7ec3\u8be6\u60c5<\/b><\/strong><\/h3>\n
\n
\u5728\u8d85\u8fc715 \u4e07\u4ebf<\/b><\/strong>\u6807\u8bb0\u4e0a\u8fdb\u884c\u8bad\u7ec3<\/li>\n
\u4e3a 405B \u6a21\u578b\u5b9a\u5236\u7684 GPU \u96c6\u7fa4\u5177\u67093930 \u4e07 GPU \u5c0f\u65f6\u6570<\/b><\/strong>\u3002<\/li>\n
\u9488\u5bf9\u591a\u8bed\u8a00\u80fd\u529b\u7684\u591a\u6837\u5316\u6570\u636e\u96c6\u7b56\u5212<\/li>\n<\/ol>\n
\u7ecf\u8fc7\u6559\u5b66\u8c03\u6574\u7684\u7248\u672c\u7ecf\u8fc7\u4e86\u989d\u5916\u7684\u57f9\u8bad\uff1a<\/p>\n
\n
\u5728\u516c\u5f00\u7684\u6307\u4ee4\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u5fae\u8c03<\/li>\n
\u8d85\u8fc72500 \u4e07<\/b><\/strong>\u5408\u6210\u793a\u4f8b<\/li>\n
\u76d1\u7763\u5fae\u8c03<\/u><\/a>\uff08SFT\uff09\u548c\u00a0\u6709\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60<\/u><\/a>\u00a0\uff08RLHF\uff09<\/li>\n<\/ol>\n
\u6027\u80fd\u57fa\u51c6<\/b><\/strong><\/h2>\n
\u4e0b\u8868\u6bd4\u8f83\u4e86 Llama 3.1 405B\u3001Nemotron 4 340B Instruct\u3001GPT-4 (0125)\u3001GPT-4 Omni \u548c Claude 3.5 Sonnet\u3002 \u4e3b\u8981\u57fa\u51c6\u5305\u62ec MMLU \u548c IFEval \u7b49\u4e00\u822c\u4efb\u52a1\u3001HumanEval \u548c GSM8K \u7b49\u4ee3\u7801\u4efb\u52a1\u4ee5\u53ca ARC Challenge \u7b49\u63a8\u7406\u4efb\u52a1\u3002 \u6bcf\u4e2a\u57fa\u51c6\u5f97\u5206\u90fd\u53cd\u6620\u4e86\u6a21\u578b\u5728\u7406\u89e3\u548c\u751f\u6210\u7c7b\u4eba\u6587\u672c\u3001\u89e3\u51b3\u590d\u6742\u95ee\u9898\u548c\u6267\u884c\u4ee3\u7801\u65b9\u9762\u7684\u80fd\u529b\u3002 \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cLlama 3.1 405B \u548c Claude 3.5 Sonnet \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u5c55\u793a\u4e86\u5b83\u4eec\u5728\u4e00\u822c\u4efb\u52a1\u548c\u7279\u5b9a\u9886\u57df\u4efb\u52a1\u4e2d\u7684\u5148\u8fdb\u80fd\u529b\u3002<\/p>\n
Llama 3.1-405B \u7684\u5185\u5b58\u8981\u6c42<\/b><\/strong><\/h3>\n
\u8fd0\u884c Llama 3.1-405B \u9700\u8981\u5927\u91cf\u5185\u5b58\u548c\u8ba1\u7b97\u8d44\u6e90\uff1a<\/p>\n
\n
<\/b>GPU \u5185\u5b58<\/b><\/strong>\uff1a405B \u6a21\u578b\u53ef\u5229\u7528\u6bcf\u4e2a A100 GPU \u9ad8\u8fbe 80GB \u7684 GPU \u5185\u5b58\u8fdb\u884c\u9ad8\u6548\u63a8\u7406\u3002 \u4f7f\u7528\u5f20\u91cf\u5e76\u884c\u6280\u672f\u53ef\u4ee5\u5728\u591a\u4e2a GPU \u4e4b\u95f4\u5206\u914d\u8d1f\u8f7d\u3002<\/b><\/strong><\/li>\n
<\/b>\u5185\u5b58<\/b><\/strong>\uff1a \u5efa\u8bae\u81f3\u5c11\u4f7f\u7528 512GB \u7684\u7cfb\u7edf\u5185\u5b58\u6765\u5904\u7406\u6a21\u578b\u7684\u5185\u5b58\u5360\u7528\uff0c\u5e76\u786e\u4fdd\u6570\u636e\u5904\u7406\u7684\u6d41\u7545\u6027\u3002<\/b><\/strong><\/li>\n
<\/b>\u5b58\u50a8<\/b><\/strong>\uff1a\u786e\u4fdd\u4e3a\u6a21\u578b\u6743\u91cd\u548c\u76f8\u5173\u6570\u636e\u96c6\u63d0\u4f9b\u6570 TB \u7684 SSD \u5b58\u50a8\u7a7a\u95f4\u3002 \u9ad8\u901f\u56fa\u6001\u786c\u76d8\u5bf9\u4e8e\u7f29\u77ed\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6570\u636e\u8bbf\u95ee\u65f6\u95f4\u81f3\u5173\u91cd\u8981<\/b><\/strong><\/b>(<\/b><\/strong>Llama Ai Model<\/b><\/strong><\/a>)<\/b><\/strong>\u200b\u200b<\/b><\/strong>\u00a0<\/b><\/strong>(<\/b><\/strong>Groq<\/b><\/strong><\/a>)<\/b><\/strong>\u3002<\/b><\/strong><\/li>\n<\/ol>\n
Llama 3.1-405B \u7684\u63a8\u7406\u4f18\u5316\u6280\u672f<\/b><\/strong><\/h3>\n
\u6709\u6548\u8fd0\u884c Llama 3.1 \u8fd9\u6837\u7684 405B \u53c2\u6570\u6a21\u578b\u9700\u8981\u591a\u79cd\u4f18\u5316\u6280\u672f\u3002 \u4ee5\u4e0b\u662f\u786e\u4fdd\u6709\u6548\u63a8\u65ad\u7684\u5173\u952e\u65b9\u6cd5\uff1a<\/p>\n
\n
a) \u91cf\u5316\uff1a<\/b><\/strong>\u91cf\u5316\u5305\u62ec\u964d\u4f4e\u6a21\u578b\u6743\u91cd\u7684\u7cbe\u5ea6\uff0c\u4ece\u800c\u51cf\u5c11\u5185\u5b58\u4f7f\u7528\u91cf\u5e76\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\uff0c\u800c\u4e0d\u4f1a\u660e\u663e\u727a\u7272\u7cbe\u5ea6\u3002 Llama 3.1 \u652f\u6301\u4f7f\u7528 QLoRA\uff08Quantized Low-Rank Adaptation\uff09\u7b49\u6280\u672f\u5c06\u91cf\u5316\u7cbe\u5ea6\u63d0\u9ad8\u5230 FP8 \u751a\u81f3\u66f4\u4f4e\uff0c\u4ee5\u4f18\u5316 GPU \u4e0a\u7684\u6027\u80fd\u3002<\/li>\n<\/ol>\n
b\uff09\u5f20\u91cf\u5e76\u884c\uff1a<\/b><\/strong>\u5f20\u91cf\u5e76\u884c\u6d89\u53ca\u5728\u591a\u4e2a GPU \u4e0a\u5206\u5272\u6a21\u578b\u5c42\u4ee5\u5e76\u884c\u8ba1\u7b97\u3002 \u8fd9\u5bf9\u4e8e\u50cf Llama 3.1 \u8fd9\u6837\u7684\u5927\u578b\u6a21\u578b\u5c24\u5176\u6709\u7528\uff0c\u53ef\u4ee5\u6709\u6548\u5229\u7528\u8d44\u6e90\u3002<\/p>\n
\n
c) KV \u7f13\u5b58\u4f18\u5316\uff1a\u952e\u503c\uff08KV\uff09\u7f13\u5b58\u7684\u9ad8\u6548\u7ba1\u7406\u5bf9\u4e8e\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u81f3\u5173\u91cd\u8981\u3002 Llama 3.1 \u652f\u6301\u6269\u5c55\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u53ef\u4f7f\u7528\u4f18\u5316\u7684 KV \u7f13\u5b58\u6280\u672f\u5bf9\u5176\u8fdb\u884c\u6709\u6548\u7ba1\u7406\uff1a<\/b><\/strong><\/li>\n<\/ol>\n
\u90e8\u7f72\u7b56\u7565<\/b><\/strong><\/h3>\n
\u90e8\u7f72 Llama 3.1-405B \u9700\u8981\u4ed4\u7ec6\u8003\u8651\u786c\u4ef6\u8d44\u6e90\u3002 \u4ee5\u4e0b\u662f\u4e00\u4e9b\u9009\u9879\uff1a<\/p>\n
\n
a) \u57fa\u4e8e\u4e91\u7684\u90e8\u7f72\uff1a<\/b><\/strong>\u5229\u7528 AWS\uff08P4d \u5b9e\u4f8b\uff09\u6216 Google Cloud\uff08TPU v4\uff09\u7b49\u4e91\u63d0\u4f9b\u5546\u63d0\u4f9b\u7684\u9ad8\u5185\u5b58 GPU \u5b9e\u4f8b\u3002<\/li>\n<\/ol>\n
b\uff09\u5185\u90e8\u90e8\u7f72\uff1a<\/b><\/strong>\u5bf9\u4e8e\u5177\u6709\u9ad8\u6027\u80fd\u8ba1\u7b97\u80fd\u529b\u7684\u7ec4\u7ec7\uff0c\u5728\u5185\u90e8\u90e8\u7f72 Llama 3.1 \u53ef\u63d0\u4f9b\u66f4\u591a\u63a7\u5236\uff0c\u5e76\u53ef\u80fd\u964d\u4f4e\u957f\u671f\u6210\u672c\u3002<\/p>\n
\u793a\u4f8b\u8bbe\u7f6e\uff1a<\/b><\/strong><\/p>\n
\n
c) \u5206\u5e03\u5f0f\u63a8\u7406\uff1a<\/b><\/strong>\u5bf9\u4e8e\u5927\u578b\u90e8\u7f72\uff0c\u53ef\u8003\u8651\u5728\u591a\u4e2a\u8282\u70b9\u4e0a\u5206\u5e03\u6a21\u578b\u3002<\/li>\n<\/ol>\n
\u4f7f\u7528\u6848\u4f8b\u548c\u5e94\u7528<\/b><\/strong><\/h3>\n
Llama 3.1-405B \u7684\u5f3a\u5927\u529f\u80fd\u548c\u7075\u6d3b\u6027\u5e26\u6765\u4e86\u65e0\u6570\u53ef\u80fd\u6027\uff1a<\/p>\n
\n
a) \u751f\u6210\u5408\u6210\u6570\u636e\uff1a<\/b><\/strong>\u751f\u6210\u9ad8\u8d28\u91cf\u3001\u7279\u5b9a\u9886\u57df\u7684\u6570\u636e\uff0c\u7528\u4e8e\u8bad\u7ec3\u66f4\u5c0f\u7684\u6a21\u578b\u3002<\/li>\n<\/ol>\n
b\uff09\u77e5\u8bc6\u63d0\u70bc\uff1a<\/b><\/strong>\u5c06 405B \u6a21\u578b\u7684\u77e5\u8bc6\u8f6c\u79fb\u5230\u66f4\u5c0f\u3001\u66f4\u6613\u4e8e\u90e8\u7f72\u7684\u6a21\u578b\u4e2d\u3002<\/p>\n
\n
c) \u9488\u5bf9\u7279\u5b9a\u9886\u57df\u7684\u5fae\u8c03\uff1a<\/b><\/strong>\u9488\u5bf9\u4e13\u4e1a\u4efb\u52a1\u6216\u884c\u4e1a\u8c03\u6574\u6a21\u578b\u3002<\/li>\n<\/ol>\n
\u8fd9\u4e9b\u6280\u672f\u548c\u7b56\u7565\u5c06\u5e2e\u52a9\u60a8\u5145\u5206\u53d1\u6325 Llama 3.1-405B \u7684\u6f5c\u529b\uff0c\u786e\u4fdd\u9ad8\u6548\u3001\u53ef\u6269\u5c55\u548c\u4e13\u4e1a\u5316\u7684\u4eba\u5de5\u667a\u80fd\u5e94\u7528\u3002<\/p>\n
\u672a\u6765\u65b9\u5411<\/b><\/strong><\/h3>\n
Llama 3.1-405B \u7684\u53d1\u5e03\u53ef\u80fd\u4f1a\u52a0\u901f\u591a\u4e2a\u9886\u57df\u7684\u521b\u65b0\uff1a<\/p>\n
\n
\u6539\u8fdb\u4e13\u95e8\u9886\u57df\u7684\u5fae\u8c03\u6280\u672f<\/li>\n
\u5f00\u53d1\u66f4\u9ad8\u6548\u7684\u63a8\u7406\u65b9\u6cd5<\/li>\n
\u6a21\u578b\u538b\u7f29\u548c\u63d0\u70bc\u65b9\u9762\u7684\u8fdb\u6b65<\/li>\n<\/ol>\n
\u539f\u6587\u94fe\u63a5<\/u><\/a><\/p>\n
\uff08\u673a\u5668\u7ffb\u8bd1\uff0c\u8f7b\u5ea6\u8bd1\u540e\u7f16\u8f91\uff0c\u4ec5\u4f9b\u53c2\u8003\uff09<\/p>\n
\u7f16\u8f91\uff1a\u80e1\u8dc3<\/p>\n","protected":false},"excerpt":{"rendered":"
Llama 3.1-405B\u7531 Meta AI \u5f00\u53d1\u7684 Llama 3.1-405B \u4ee3\u8868\u4e86\u5f00\u6e90\u8bed\u8a00\u6a21\u578b\u7684\u91cd\u5927 […]<\/p>\n","protected":false},"author":1,"featured_media":32400,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[391],"tags":[],"class_list":["post-32399","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\/32399","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=32399"}],"version-history":[{"count":1,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/32399\/revisions"}],"predecessor-version":[{"id":32401,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/32399\/revisions\/32401"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/media\/32400"}],"wp:attachment":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=32399"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=32399"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=32399"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}