{"id":119,"date":"2026-04-24T17:32:57","date_gmt":"2026-04-24T09:32:57","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=119"},"modified":"2026-04-23T17:35:20","modified_gmt":"2026-04-23T09:35:20","slug":"%e4%bb%8e%e9%9b%b6%e5%bc%80%e5%a7%8b%ef%bc%8c%e6%90%9e%e6%87%82%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e4%ba%94%e8%84%8f%e5%85%ad%e8%85%91-llm","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/24\/%e4%bb%8e%e9%9b%b6%e5%bc%80%e5%a7%8b%ef%bc%8c%e6%90%9e%e6%87%82%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e4%ba%94%e8%84%8f%e5%85%ad%e8%85%91-llm\/","title":{"rendered":"\u4ece\u96f6\u5f00\u59cb\uff0c\u641e\u61c2\u5927\u8bed\u8a00\u6a21\u578b\u7684\u201c\u4e94\u810f\u516d\u8151\u201d\u2014\u2014LLM\u67b6\u6784\u5168\u666f\u89e3\u6790"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">AI\u65e0\u7591\u662f\u6700\u8fd1\u51e0\u5e74\u6700\u706b\u7684\u4e1c\u897f\u4e86\uff0c\u5176\u80fd\u529b\u4e0e\u53d1\u5c55\u901f\u5ea6\u90fd\u8ba9\u4eba\u60ca\u53f9\uff0c\u4ece17\u5e74\u7684Transformer\u5230\u5982\u4eca\u7684\u5404\u79cd\u5927\u6a21\u578b\uff0c\u6211\u4eec\u7684\u4e16\u754c\u5728\u77ed\u65f6\u95f4\u5185\u53d1\u751f\u4e86\u7ffb\u5929\u8986\u5730\u7684\u53d8\u5316\u3002\u6a21\u578b\u7684\u76f8\u5173\u6280\u672f\u548c\u77e5\u8bc6\u4e5f\u5728\u9ad8\u901f\u66f4\u65b0\u548c\u53d1\u5c55\uff0c\u5176\u67b6\u6784\u4ece\u539f\u59cbtransformer\uff0c\u5230\u5982\u4eca\u7684\u73b0\u4ee3\u5316\u53d8\u4f53\uff0c\u5176\u5e94\u7528\u4eceprompt\u5de5\u7a0b\u5230RAG\u518d\u5230Agent\u4e43\u81f3\u5982\u4eca\u7684Harness\u5de5\u7a0b\uff0c\u5373\u4f7f\u4e13\u95e8\u8bb0\u5f55\u4e5f\u8981\u82b1\u5f88\u591a\u7bc7\u5e45\u548c\u65f6\u95f4\uff0c\u6240\u4ee5\u6211\u6253\u7b97\u7528\u5341\u51e0\u7bc7\u535a\u5ba2\u5feb\u901f\u6982\u8ff0\u4e00\u4e9b\u8fc7\u53bb\u7684\u6280\u672f\u4ee5\u53ca\u4e00\u4e9b\u5e95\u5c42\u6280\u672f\uff0c\u7136\u540e\u4f1a\u5f00\u59cb\u4ecb\u7ecd\u81ea\u5df1\u7684\u8fd1\u671f\u5b66\u4e60\u60c5\u51b5\uff08\u4e5f\u5c31\u662f\u5f53\u524d\u7684\u6a21\u578b\u6280\u672f\u5185\u5bb9\uff09\u3002\u8fd9\u7bc7\u535a\u5ba2\u5c31\u662f\u8fd9\u4e2a\u7cfb\u5217\u7684\u9996\u7bc7\u535a\u5ba2\uff0c\u6982\u8ff0\u539f\u59cbtransformer\u7684\u4e00\u4e9b\u60c5\u51b5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fc7\u53bb\u4e24\u5e74\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Model, LLM\uff09\u7684\u8fed\u4ee3\u901f\u5ea6\u4ee4\u4eba\u77a0\u76ee\u3002\u4ece\u6700\u521d\u7684 GPT \u5230\u5982\u4eca\u7684 LLaMA\u3001Qwen\u3001DeepSeek \u7b49\u660e\u661f\u6a21\u578b\uff0c\u5b83\u4eec\u770b\u4f3c\u5f62\u6001\u5404\u5f02\uff0c\u4f46\u6380\u5f00\u5916\u58f3\uff0c\u5185\u90e8\u7684\u9aa8\u67b6\u5176\u5b9e\u662f\u540c\u4e00\u5957\u903b\u8f91\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u800c\uff0c\u771f\u6b63\u963b\u788d\u521d\u5b66\u8005\u5165\u95e8\u7684\uff0c\u5f80\u5f80\u4e0d\u662f\u6280\u672f\u672c\u8eab\u6709\u591a\u96be\u61c2\uff0c\u800c\u662f\u90a3\u5835\u7531\u751f\u50fb\u8bcd\u548c\u82f1\u6587\u7f29\u5199\u780c\u6210\u7684\u201c\u540d\u8bcd\u9ad8\u5899\u201d\u2014\u2014<strong>Token\u3001Embedding\u3001RoPE\u3001MHA\u3001GQA\u3001FFN\u3001Pre-Norm\u2026\u2026<\/strong> \u6bcf\u4e2a\u8bcd\u5355\u72ec\u62ce\u51fa\u6765\u90fd\u8ba9\u4eba\u4e00\u5934\u96fe\u6c34\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u7684\u76ee\u6807\u5f88\u7b80\u5355\uff1a<strong>\u628a\u5927\u6a21\u578b\u7684\u201c\u4e94\u810f\u516d\u8151\u201d\u62c6\u5f00\u6765\u770b\u6e05\u695a<\/strong>\u3002\u4ece\u6570\u636e\u5982\u4f55\u8f93\u5165\uff0c\u5230\u6a21\u578b\u5185\u90e8\u5982\u4f55\u201c\u601d\u8003\u201d\uff0c\u518d\u5230\u6700\u7ec8\u5982\u4f55\u8f93\u51fa\uff0c\u6211\u4f1a\u6cbf\u7740\u4e00\u6761\u6e05\u6670\u7684\u6570\u636e\u6d41\uff0c\u628a\u6bcf\u4e00\u4e2a\u6838\u5fc3\u7ec4\u4ef6\u8bb2\u660e\u767d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u51c6\u5907\u597d\u4e86\u5417\uff1f\u6211\u4eec\u4ece\u6700\u57fa\u7840\u7684\u201c\u6570\u636e\u600e\u4e48\u8fdb\u5165\u6a21\u578b\u201d\u5f00\u59cb\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6574\u4f53\u6982\u89c8\uff1a\u4e00\u5f20\u56fe\u770b\u61c2 LLM \u7684\u201c\u9aa8\u67b6\u201d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6df1\u5165\u7ec6\u8282\u4e4b\u524d\uff0c\u5148\u7ed9\u5927\u5bb6\u4e00\u4e2a\u5168\u5c40\u89c6\u89d2\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u51e0\u4e4e\u6240\u6709\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u5e95\u5c42\u90fd\u57fa\u4e8e\u540c\u4e00\u4e2a\u67b6\u6784\u2014\u2014<strong>Transformer<\/strong>\u3002\u4f46\u8bf7\u6ce8\u610f\uff0c\u4eca\u5929\u7684 LLM \u5df2\u7ecf\u4e0d\u518d\u662f 2017 \u5e74\u8bba\u6587\u300aAttention is All You Need\u300b\u91cc\u90a3\u4e2a\u539f\u6c41\u539f\u5473\u7684 Transformer \u4e86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u8005\u4eec\u5728\u5b9e\u8df5\u4e2d\u53d1\u73b0\uff0c\u5bf9\u5176\u4e2d\u67d0\u4e9b\u6a21\u5757\u8fdb\u884c\u66ff\u6362\u548c\u4f18\u5316\uff0c\u53ef\u4ee5\u5e26\u6765\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002\u4e8e\u662f\uff0c\u4e00\u5957\u201c\u73b0\u4ee3 LLM \u9ed8\u8ba4\u914d\u7f6e\u201d\u9010\u6e10\u6210\u5f62\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7ec4\u4ef6<\/th><th>\u7ecf\u5178 Transformer<\/th><th>\u73b0\u4ee3 LLM\uff08\u4ee5 LLaMA \u4e3a\u4ee3\u8868\uff09<\/th><\/tr><\/thead><tbody><tr><td>\u6ce8\u610f\u529b\u673a\u5236<\/td><td>MHA\uff08Multi-Head Attention\uff09<\/td><td><strong>GQA<\/strong>\uff08Grouped Query Attention\uff09<\/td><\/tr><tr><td>\u4f4d\u7f6e\u7f16\u7801<\/td><td>\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801 \/ Learned PE<\/td><td><strong>RoPE<\/strong>\uff08\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff09<\/td><\/tr><tr><td>\u524d\u9988\u7f51\u7edc\u6fc0\u6d3b<\/td><td>ReLU \/ GELU<\/td><td><strong>SwiGLU<\/strong><\/td><\/tr><tr><td>\u5f52\u4e00\u5316<\/td><td>LayerNorm + Post-Norm<\/td><td><strong>RMSNorm + Pre-Norm<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f\u76ee\u524d LLaMA\u3001Qwen\u3001Mistral\u3001DeepSeek \u7b49\u4e3b\u6d41\u6a21\u578b\u5171\u540c\u9075\u5faa\u7684\u201c\u6807\u914d\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a\u8fd9\u4e9b\u7ec4\u4ef6\u4e32\u8d77\u6765\uff0c\u4e00\u6761\u5b8c\u6574\u7684\u6570\u636e\u5904\u7406\u94fe\u8def\u662f\u8fd9\u6837\u7684\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u539f\u59cb\u6587\u672c \u2192 Tokenizer\uff08\u5206\u8bcd\uff09 \u2192 Embedding\uff08\u5411\u91cf\u5316\uff09 \u2192 + \u4f4d\u7f6e\u7f16\u7801\uff08RoPE\uff09\n       \u2193\n   &#91;Transformer Block] \u00d7 N \u5c42\n       \u2193\n   Pre-Norm \u2192 Attention\uff08GQA\uff09 \u2192 \u6b8b\u5dee\u8fde\u63a5\n       \u2193\n   Pre-Norm \u2192 FFN\uff08SwiGLU\uff09 \u2192 \u6b8b\u5dee\u8fde\u63a5\n       \u2193\n   \u6700\u7ec8\u8f93\u51fa\u5c42 \u2192 \u9884\u6d4b\u4e0b\u4e00\u4e2a Token<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c31\u6cbf\u7740\u8fd9\u6761\u94fe\u8def\uff0c\u4e00\u4e2a\u6a21\u5757\u4e00\u4e2a\u6a21\u5757\u5730\u62c6\u89e3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Tokenizer\uff1a\u628a\u4eba\u7c7b\u8bed\u8a00\u7ffb\u8bd1\u6210\u6a21\u578b\u80fd\u61c2\u7684\u201c\u6570\u5b57\u5bc6\u7801\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 Tokenizer \u662f\u505a\u4ec0\u4e48\u7684\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5927\u6a21\u578b\u4e0d\u8ba4\u8bc6\u6c49\u5b57\u3001\u5b57\u6bcd\u6216\u6807\u70b9\u7b26\u53f7\uff0c\u5b83\u53ea\u201c\u5403\u201d\u6570\u5b57\u3002Tokenizer \u7684\u5de5\u4f5c\uff0c\u5c31\u662f\u628a\u4e00\u6bb5\u4eba\u7c7b\u5199\u7684\u6587\u672c\uff0c\u5207\u5206\u6210\u4e00\u4e2a\u4e2a\u5c0f\u788e\u7247\uff0c\u7136\u540e\u6620\u5c04\u6210\u6a21\u578b\u8bcd\u8868\u4e2d\u7684\u6570\u5b57 ID\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\uff0c\u8f93\u5165\u4e00\u53e5\u8bdd\uff1a\u201c\u6211\u7231\u5927\u8bed\u8a00\u6a21\u578b\u201d\uff0cTokenizer \u53ef\u80fd\u4f1a\u628a\u5b83\u5207\u6210\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;\"\u6211\", \"\u7231\", \"\u5927\", \"\u8bed\u8a00\", \"\u6a21\u578b\"] \u2192 &#91;231, 452, 789, 1023, 88]<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u6570\u5b57\u5e8f\u5217\uff0c\u5c31\u662f\u6a21\u578b\u771f\u6b63\u63a5\u6536\u7684\u201c\u539f\u6599\u201d\u3002\u800c\u6bcf\u4e00\u4e2a\u88ab\u5207\u51fa\u6765\u7684\u788e\u7247\uff0c\u5c31\u53eb\u505a\u4e00\u4e2a <strong>Token\uff08\u8bcd\u5143\uff09<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u4e3b\u6d41\u5206\u8bcd\u7b97\u6cd5\u6709\u54ea\u4e9b\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u76ee\u524d\u5927\u6a21\u578b\u4e2d\u5e38\u7528\u7684\u5206\u8bcd\u7b97\u6cd5\u4e3b\u8981\u6709\u4e09\u79cd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>BPE\uff08Byte Pair Encoding\uff09<\/strong> \uff1a\u901a\u8fc7\u7edf\u8ba1\u8bed\u6599\u5e93\u4e2d\u9ad8\u9891\u51fa\u73b0\u7684\u5b57\u7b26\u5bf9\uff0c\u4e0d\u65ad\u5408\u5e76\uff0c\u9010\u6b65\u6784\u5efa\u8bcd\u6c47\u8868\u3002\u4f18\u70b9\u662f\u80fd\u6709\u6548\u5904\u7406\u672a\u767b\u5f55\u8bcd\uff08OOV\uff09\uff0c\u8bcd\u6c47\u91cf\u53ef\u63a7\u3002\u7f3a\u70b9\u662f\u5bf9\u4e2d\u6587\u8fd9\u7c7b\u65e0\u7a7a\u683c\u8bed\u8a00\u7684\u652f\u6301\u8f83\u5f31\u3002<\/li>\n\n\n\n<li><strong>WordPiece<\/strong>\uff1aGoogle \u4e3a BERT \u5f00\u53d1\u7684\u7b97\u6cd5\uff0c\u539f\u7406\u4e0e BPE \u7c7b\u4f3c\uff0c\u4f46\u5408\u5e76\u7b56\u7565\u57fa\u4e8e\u6700\u5927\u4f3c\u7136\u4f30\u8ba1\u3002\u5b83\u5728\u5b50\u8bcd\u524d\u6dfb\u52a0 <code>##<\/code> \u6807\u8bb0\uff0c\u66f4\u9002\u5408\u82f1\u6587\u5904\u7406\u3002<\/li>\n\n\n\n<li><strong>SentencePiece<\/strong>\uff1a\u4e00\u79cd\u8bed\u8a00\u65e0\u5173\u7684\u5206\u8bcd\u5668\uff0c\u5c06\u6587\u672c\u89c6\u4e3a Unicode \u5b57\u7b26\u5e8f\u5217\uff0c\u4e0d\u4f9d\u8d56\u7a7a\u683c\u5206\u8bcd\u3002\u5bf9\u4e2d\u6587\u3001\u65e5\u6587\u7b49\u65e0\u7a7a\u683c\u8bed\u8a00\u7279\u522b\u53cb\u597d\uff0c\u800c\u4e14\u652f\u6301\u53ef\u9006\u8f6c\u6362\u3002LLaMA\u3001Mistral\u3001Gemma \u7b49\u4e3b\u6d41\u6a21\u578b\u90fd\u91c7\u7528\u4e86 SentencePiece\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 \u4e3a\u4ec0\u4e48\u5206\u8bcd\u8fd9\u4e48\u91cd\u8981\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5206\u8bcd\u5668\u7684\u8bbe\u8ba1\u76f4\u63a5\u5f71\u54cd\u6a21\u578b\u7684\u6027\u80fd\u548c\u6548\u7387\u3002\u4e00\u4e2a\u4e0d\u597d\u7684\u5206\u8bcd\u5668\u53ef\u80fd\u5bfc\u81f4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>OOV\uff08\u672a\u767b\u5f55\u8bcd\uff09\u95ee\u9898<\/strong>\uff1a\u9047\u5230\u8bcd\u8868\u4e2d\u6ca1\u6709\u7684\u8bcd\uff0c\u53ea\u80fd\u6620\u5c04\u5230 <code>[UNK]<\/code>\uff0c\u4fe1\u606f\u76f4\u63a5\u4e22\u5931\u3002<\/li>\n\n\n\n<li><strong>\u4e2d\u6587\u5206\u8bcd\u707e\u96be<\/strong>\uff1a\u628a\u201c\u5927\u8bed\u8a00\u6a21\u578b\u201d\u5207\u6210\u201c\u5927\u201d\u3001\u201c\u8bed\u201d\u3001\u201c\u8a00\u201d\u3001\u201c\u6a21\u201d\u3001\u201c\u578b\u201d\uff0c\u8bed\u4e49\u5b8c\u5168\u7834\u788e\u3002<\/li>\n\n\n\n<li><strong>\u8bcd\u6c47\u8868\u7206\u70b8<\/strong>\uff1a\u8bcd\u6c47\u8868\u592a\u5927\uff0c\u6a21\u578b\u53c2\u6570\u91cf\u6025\u5267\u81a8\u80c0\uff0c\u8bad\u7ec3\u548c\u63a8\u7406\u6210\u672c\u6fc0\u589e\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c\u9009\u62e9\u4e00\u4e2a\u5408\u9002\u7684\u5206\u8bcd\u5668\uff0c\u662f\u6784\u5efa\u9ad8\u8d28\u91cf LLM \u7684\u7b2c\u4e00\u6b65\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001Embedding\uff1a\u8ba9\u8bcd\u8bed\u5728\u6570\u5b66\u7a7a\u95f4\u91cc\u201c\u5b89\u5bb6\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u4ece Token ID \u5230\u5411\u91cf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ecf\u8fc7 Tokenizer\uff0c\u6211\u4eec\u5f97\u5230\u4e86\u4e00\u4e32\u6570\u5b57 ID\uff0c\u4f46\u8fd9\u4e9b ID \u672c\u8eab\u6ca1\u6709\u8bed\u4e49\u4fe1\u606f\u3002ID 231 \u548c ID 452 \u4e4b\u95f4\uff0c\u6709\u4ec0\u4e48\u5185\u5728\u8054\u7cfb\u5417\uff1f\u6a21\u578b\u65e0\u4ece\u77e5\u6653\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Embedding\uff08\u8bcd\u5d4c\u5165\uff09<\/strong> \u7684\u4f5c\u7528\uff0c\u5c31\u662f\u628a\u6bcf\u4e00\u4e2a Token ID\uff0c\u6620\u5c04\u5230\u4e00\u4e2a\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u5411\u91cf\uff08\u901a\u5e38\u662f\u51e0\u767e\u5230\u51e0\u5343\u7ef4\u7684\u6d6e\u70b9\u6570\u6570\u7ec4\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5047\u8bbe Embedding \u7ef4\u5ea6\u662f 4096\uff08LLaMA-7B \u7684\u914d\u7f6e\uff09\uff0c\u90a3\u4e48 Token ID 231 \u4f1a\u88ab\u8f6c\u6362\u6210\u4e00\u4e2a\u5305\u542b 4096 \u4e2a\u6570\u5b57\u7684\u5411\u91cf\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Token ID: 231 \u2192 Embedding: &#91;0.023, -0.451, 0.789, ..., 0.112]\uff084096 \u7ef4\uff09<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 \u4e3a\u4ec0\u4e48 Embedding \u8fd9\u4e48\u795e\u5947\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Embedding \u7684\u7f8e\u5999\u4e4b\u5904\u5728\u4e8e\uff1a\u5728\u8fd9\u4e2a\u9ad8\u7ef4\u6570\u5b66\u7a7a\u95f4\u91cc\uff0c<strong>\u8bed\u4e49\u76f8\u8fd1\u7684\u8bcd\uff0c\u5b83\u4eec\u7684\u5411\u91cf\u5728\u51e0\u4f55\u7a7a\u95f4\u4e2d\u4e5f\u9760\u5f97\u5f88\u8fd1<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\uff0c\u201c\u56fd\u738b\u201d\u548c\u201c\u738b\u540e\u201d\u7684\u5411\u91cf\u8ddd\u79bb\u4f1a\u5f88\u8fd1\uff0c\u201c\u82f9\u679c\u201d\u548c\u201c\u6c34\u679c\u201d\u7684\u5411\u91cf\u8ddd\u79bb\u4e5f\u5f88\u8fd1\u3002\u800c\u201c\u82f9\u679c\u201d\u548c\u201c\u6c7d\u8f66\u201d\u7684\u5411\u91cf\u5219\u76f8\u8ddd\u751a\u8fdc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u6027\u8d28\u4e0d\u662f\u4eba\u5de5\u5b9a\u4e49\u7684\uff0c\u800c\u662f\u901a\u8fc7\u5728\u6d77\u91cf\u8bed\u6599\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u8ba9\u6a21\u578b\u81ea\u5df1\u201c\u5b66\u4f1a\u201d\u7684\u3002Embedding \u672c\u8d28\u4e0a\u662f\u4e00\u79cd<strong>\u5206\u5e03\u5f0f\u7684\u8bed\u4e49\u8868\u793a<\/strong>\uff0c\u5b83\u628a\u79bb\u6563\u7684\u7b26\u53f7\uff08\u8bcd\u8bed\uff09\u6620\u5c04\u5230\u4e86\u8fde\u7eed\u7684\u5411\u91cf\u7a7a\u95f4\uff0c\u4f7f\u5f97\u540e\u7eed\u7684\u795e\u7ecf\u7f51\u7edc\u80fd\u591f\u8fdb\u884c\u9ad8\u6548\u7684\u6570\u5b66\u8fd0\u7b97\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u4e00\u4e2a\u5c0f\u95ee\u9898\uff1a\u4f4d\u7f6e\u4fe1\u606f\u53bb\u54ea\u4e86\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ed4\u7ec6\u60f3\u60f3\uff0cEmbedding \u53ea\u5305\u542b\u4e86\u8bcd\u8bed\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u4f46\u5b8c\u5168\u6ca1\u6709\u544a\u8bc9\u6a21\u578b\u8fd9\u4e9b\u8bcd\u8bed\u5728\u53e5\u5b50\u4e2d\u7684\u4f4d\u7f6e\u5173\u7cfb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u201c\u6211\u7231\u4f60\u201d\u548c\u201c\u4f60\u7231\u6211\u201d\uff0c\u7531\u5b8c\u5168\u76f8\u540c\u7684\u4e09\u4e2a\u8bcd\u7ec4\u6210\uff0c\u4f46\u610f\u601d\u622a\u7136\u4e0d\u540c\u3002\u5982\u679c\u6a21\u578b\u4e0d\u77e5\u9053\u8bcd\u8bed\u7684\u987a\u5e8f\uff0c\u5b83\u5c31\u65e0\u6cd5\u7406\u89e3\u8fd9\u79cd\u5dee\u5f02\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f<strong>\u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09<\/strong> \u8981\u89e3\u51b3\u7684\u95ee\u9898\u2014\u2014\u6211\u4eec\u63a5\u4e0b\u6765\u91cd\u70b9\u8bb2\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001RoPE\uff1a\u8ba9\u6a21\u578b\u77e5\u9053\u201c\u8c01\u5728\u524d\uff0c\u8c01\u5728\u540e\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u4e3a\u4ec0\u4e48\u9700\u8981\u4f4d\u7f6e\u7f16\u7801\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Transformer \u7684\u6838\u5fc3\u673a\u5236\u662f Attention\uff08\u540e\u9762\u4f1a\u8be6\u7ec6\u8bb2\uff09\uff0c\u800c Attention \u672c\u8eab\u5bf9\u8f93\u5165\u5e8f\u5217\u7684\u987a\u5e8f\u662f<strong>\u4e0d\u654f\u611f<\/strong>\u7684\u3002\u5982\u679c\u4f60\u628a\u4e00\u53e5\u8bdd\u7684\u8bcd\u5e8f\u6253\u4e71\uff0c\u7136\u540e\u4e22\u7ed9\u4e00\u4e2a\u6ca1\u6709\u4f4d\u7f6e\u4fe1\u606f\u7684 Transformer\uff0c\u5b83\u770b\u5230\u7684 Embedding \u96c6\u5408\u662f\u4e00\u6837\u7684\uff0c\u81ea\u7136\u65e0\u6cd5\u533a\u5206\u987a\u5e8f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c\u5fc5\u987b\u5728 Embedding \u7684\u57fa\u7840\u4e0a\uff0c<strong>\u6ce8\u5165\u4f4d\u7f6e\u4fe1\u606f<\/strong>\uff0c\u544a\u8bc9\u6a21\u578b\u6bcf\u4e2a Token \u5728\u53e5\u5b50\u4e2d\u7684\u4f4d\u7f6e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u4ece\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801\u5230 RoPE<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u65e9\u671f\u7684 Transformer \u4f7f\u7528\u7684\u662f<strong>\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801<\/strong>\u2014\u2014\u7ed9\u6bcf\u4e2a\u4f4d\u7f6e\u5206\u914d\u4e00\u4e2a\u56fa\u5b9a\u7684\u5411\u91cf\uff0c\u76f4\u63a5\u52a0\u5230 Embedding \u4e0a\u3002\u6bd4\u5982\u4f4d\u7f6e 0 \u52a0\u5411\u91cf A\uff0c\u4f4d\u7f6e 1 \u52a0\u5411\u91cf B\u2026\u2026\u8fd9\u79cd\u65b9\u6cd5\u7684\u7f3a\u9677\u662f\uff1a\u6a21\u578b\u5f88\u96be\u6cdb\u5316\u5230\u8bad\u7ec3\u65f6\u6ca1\u89c1\u8fc7\u7684\u957f\u5e8f\u5217\u957f\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u540e\u6765\uff0c\u7814\u7a76\u8005\u53d1\u73b0\uff0c<strong>\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f<\/strong> \u66f4\u91cd\u8981\u2014\u2014\u6a21\u578b\u5e94\u8be5\u5173\u5fc3\u201c\u8bcd A \u548c\u8bcd B \u4e4b\u95f4\u76f8\u9694\u591a\u5c11\u8ddd\u79bb\u201d\uff0c\u800c\u4e0d\u662f\u201c\u8bcd A \u5728\u7edd\u5bf9\u7684\u7b2c\u51e0\u4e2a\u4f4d\u7f6e\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\uff0c<strong>RoPE\uff08Rotary Position Embedding\uff0c\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff09<\/strong> \u5e94\u8fd0\u800c\u751f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 RoPE \u7684\u6838\u5fc3\u601d\u60f3\uff1a\u7528\u65cb\u8f6c\u6765\u7f16\u7801\u76f8\u5bf9\u4f4d\u7f6e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">RoPE \u7684\u6838\u5fc3\u60f3\u6cd5\u975e\u5e38\u4f18\u96c5\uff1a<strong>\u901a\u8fc7\u5bf9 Query \u548c Key \u5411\u91cf\u65bd\u52a0\u4e00\u4e2a\u4e0e\u4f4d\u7f6e\u76f8\u5173\u7684\u65cb\u8f6c\u64cd\u4f5c\uff0c\u4f7f\u5f97\u4e24\u4e2a Token \u7684\u70b9\u79ef\u7ed3\u679c\u5929\u7136\u5305\u542b\u5b83\u4eec\u7684\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6765\u8bf4\uff0cRoPE \u5c06\u6bcf\u4e2a Token \u7684\u5411\u91cf\u770b\u4f5c\u590d\u6570\uff0c\u7136\u540e\u6839\u636e\u5b83\u5728\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\uff0c\u5c06\u5176\u65cb\u8f6c\u4e00\u5b9a\u7684\u89d2\u5ea6\u3002\u4f4d\u7f6e\u8d8a\u9760\u540e\uff0c\u65cb\u8f6c\u7684\u89d2\u5ea6\u8d8a\u5927\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u505a\u7684\u795e\u5947\u4e4b\u5904\u5728\u4e8e\uff1a\u4e24\u4e2a Token \u7684 Query \u548c Key \u505a\u5185\u79ef\u65f6\uff0c\u7ed3\u679c\u53ea\u4f9d\u8d56\u4e8e\u5b83\u4eec\u4e4b\u95f4\u7684<strong>\u76f8\u5bf9\u4f4d\u7f6e\u5dee<\/strong>\uff08m &#8211; n\uff09\uff0c\u800c\u4e0d\u662f\u5404\u81ea\u7684\u7edd\u5bf9\u4f4d\u7f6e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u516c\u5f0f\u6765\u8868\u8fbe\uff0c\u5047\u8bbe Token m \u7684 Query \u5411\u91cf\u4e3a (q_m)\uff0cToken n \u7684 Key \u5411\u91cf\u4e3a (k_n)\uff0c\u7ecf\u8fc7 RoPE \u65cb\u8f6c\u540e\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>q_m \\cdot k_n = f(x_m, x_n, m &#8211; n)<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d (x_m)\u3001(x_n) \u662f\u5404\u81ea\u7684\u8bcd\u5d4c\u5165\u5411\u91cf\uff0c(m &#8211; n) \u5c31\u662f\u76f8\u5bf9\u4f4d\u7f6e\u5dee\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 \u4e3a\u4ec0\u4e48 RoPE \u6210\u4e3a\u201c\u6807\u914d\u201d\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">RoPE \u6709\u51e0\u4e2a\u663e\u8457\u4f18\u52bf\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5929\u7136\u7f16\u7801\u76f8\u5bf9\u4f4d\u7f6e<\/strong>\uff1a\u4e0d\u9700\u8981\u989d\u5916\u5b66\u4e60\u4f4d\u7f6e\u53c2\u6570\uff0c\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f\u76f4\u63a5\u5185\u5d4c\u5728\u5411\u91cf\u8fd0\u7b97\u4e2d\u3002<\/li>\n\n\n\n<li><strong>\u957f\u5ea6\u5916\u63a8\u80fd\u529b\u597d<\/strong>\uff1a\u7406\u8bba\u4e0a\u53ef\u4ee5\u5904\u7406\u6bd4\u8bad\u7ec3\u65f6\u66f4\u957f\u7684\u5e8f\u5217\u3002<\/li>\n\n\n\n<li><strong>\u8ba1\u7b97\u9ad8\u6548<\/strong>\uff1a\u5b9e\u73b0\u7b80\u5355\uff0c\u5bf9 Attention \u8ba1\u7b97\u7684\u6539\u52a8\u5f88\u5c0f\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b63\u56e0\u5982\u6b64\uff0cLLaMA\u3001Gemini\u3001DeepSeek\u3001Qwen \u7b49\u51e0\u4e4e\u6240\u6709\u4e3b\u6d41 LLM \u90fd\u91c7\u7528\u4e86 RoPE \u4f5c\u4e3a\u4f4d\u7f6e\u7f16\u7801\u65b9\u6848\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001Attention \u5168\u5bb6\u6876\uff1aMHA \u2192 MQA \u2192 GQA<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ec8\u4e8e\u6765\u5230\u4e86 LLM \u6700\u6838\u5fc3\u7684\u6a21\u5757\u2014\u2014<strong>Attention\uff08\u6ce8\u610f\u529b\u673a\u5236\uff09<\/strong> \u3002\u8fd9\u662f Transformer \u7684\u7075\u9b42\u6240\u5728\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 Attention \u5728\u505a\u4ec0\u4e48\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u7684\u6838\u5fc3\u601d\u60f3\u53ef\u4ee5\u7528\u4e00\u53e5\u8bdd\u6982\u62ec\uff1a<strong>\u5728\u5904\u7406\u5f53\u524d\u8bcd\u65f6\uff0c\u8ba9\u6a21\u578b\u80fd\u591f\u201c\u6ce8\u610f\u201d\u5230\u4e0a\u4e0b\u6587\u4e2d\u6240\u6709\u76f8\u5173\u7684\u8bcd\uff0c\u5e76\u7ed9\u5b83\u4eec\u5206\u914d\u4e0d\u540c\u7684\u91cd\u8981\u6027\u6743\u91cd<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u53e5\u5b50\u201c\u82f9\u679c\u5f88\u597d\u5403\u201d\u4e2d\uff0c\u5904\u7406\u201c\u5f88\u201d\u8fd9\u4e2a\u5b57\u65f6\uff0c\u6a21\u578b\u5e94\u8be5\u66f4\u591a\u5730\u5173\u6ce8\u201c\u82f9\u679c\u201d\u548c\u201c\u597d\u5403\u201d\uff0c\u800c\u8f83\u5c11\u5173\u6ce8\u53e5\u53f7\u3002\u8fd9\u79cd\u201c\u9009\u62e9\u6027\u5173\u6ce8\u201d\u5c31\u662f Attention \u7684\u9b54\u529b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 Self-Attention \u7684\u8ba1\u7b97\u6d41\u7a0b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6807\u51c6\u7684 Self-Attention \u8ba1\u7b97\u5206\u4e3a\u4ee5\u4e0b\u51e0\u6b65\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\u6b65\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u8f93\u5165 Token\uff0c\u901a\u8fc7\u4e09\u4e2a\u4e0d\u540c\u7684\u6743\u91cd\u77e9\u9635 (W_Q)\u3001(W_K)\u3001(W_V) \u6295\u5f71\uff0c\u5f97\u5230\u4e09\u4e2a\u5411\u91cf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Query\uff08\u67e5\u8be2\u5411\u91cf\uff09<\/strong> \uff1a\u4ee3\u8868\u201c\u6211\u5728\u627e\u4ec0\u4e48\u201d\u3002<\/li>\n\n\n\n<li><strong>Key\uff08\u952e\u5411\u91cf\uff09<\/strong> \uff1a\u4ee3\u8868\u201c\u6211\u662f\u4ec0\u4e48\u201d\u3002<\/li>\n\n\n\n<li><strong>Value\uff08\u503c\u5411\u91cf\uff09<\/strong> \uff1a\u4ee3\u8868\u201c\u6211\u7684\u5185\u5bb9\u662f\u4ec0\u4e48\u201d\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\u6b65\uff0c\u8ba1\u7b97 Query \u548c\u6240\u6709 Key \u7684\u76f8\u4f3c\u5ea6\uff08\u70b9\u79ef\uff09\uff0c\u5f97\u5230\u6ce8\u610f\u529b\u5206\u6570\u77e9\u9635\uff0c\u7136\u540e\u901a\u8fc7 Softmax \u5f52\u4e00\u5316\u4e3a\u6743\u91cd\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^\\top}{\\sqrt{d_k}}\\right)V<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d (d_k) \u662f\u5411\u91cf\u7684\u7ef4\u5ea6\uff0c\u9664\u4ee5 (\\sqrt{d_k}) \u662f\u4e3a\u4e86\u9632\u6b62\u70b9\u79ef\u503c\u8fc7\u5927\u5bfc\u81f4 Softmax \u68af\u5ea6\u6d88\u5931\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e09\u6b65\uff0c\u7528\u8fd9\u4e9b\u6743\u91cd\u5bf9 Value \u8fdb\u884c\u52a0\u6743\u6c42\u548c\uff0c\u5f97\u5230\u6700\u7ec8\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 MHA\uff1a\u591a\u5934\u6ce8\u610f\u529b\u2014\u2014\u591a\u89d2\u5ea6\u201c\u89c2\u5bdf\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u53ea\u6709\u4e00\u4e2a\u6ce8\u610f\u529b\u5934\uff0c\u6a21\u578b\u53ef\u80fd\u53ea\u80fd\u6355\u6349\u5230\u4e00\u79cd\u8bed\u4e49\u5173\u7cfb\uff08\u6bd4\u5982\u53ea\u5173\u6ce8\u8bed\u6cd5\uff0c\u5ffd\u7565\u4e86\u8bed\u4e49\uff09\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cTransformer \u63d0\u51fa\u4e86<strong>MHA\uff08Multi-Head Attention\uff0c\u591a\u5934\u6ce8\u610f\u529b\uff09<\/strong> \u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">MHA \u7684\u505a\u6cd5\u662f\uff1a\u628a Query\u3001Key\u3001Value \u62c6\u5206\u6210\u591a\u4e2a\u201c\u5934\u201d\uff0c\u6bcf\u4e2a\u5934\u72ec\u7acb\u8ba1\u7b97 Attention\uff0c\u6700\u540e\u628a\u6240\u6709\u5934\u7684\u8f93\u51fa\u62fc\u63a5\u8d77\u6765\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>\\text{MultiHead}(Q, K, V) = \\text{Concat}(\\text{head}_1, \\text{head}_2, \u2026, \\text{head}_h)W^O<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u5934\u53ef\u4ee5\u5173\u6ce8\u4e0d\u540c\u7684\u8bed\u8a00\u7279\u5f81\u2014\u2014\u6709\u7684\u5934\u5173\u6ce8\u53e5\u6cd5\uff0c\u6709\u7684\u5934\u5173\u6ce8\u6307\u4ee3\uff0c\u6709\u7684\u5934\u5173\u6ce8\u60c5\u611f\u3002\u591a\u4e2a\u5934\u5408\u5728\u4e00\u8d77\uff0c\u6a21\u578b\u5c31\u6709\u4e86\u66f4\u4e30\u5bcc\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.4 \u63a8\u7406\u74f6\u9888\uff1aKV Cache \u5e26\u6765\u7684\u663e\u5b58\u538b\u529b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u4e00\u4e2a\u5de5\u7a0b\u4e0a\u7684\u5173\u952e\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u81ea\u56de\u5f52\u751f\u6210\uff08\u9010\u8bcd\u751f\u6210\uff09\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u6bcf\u6b21\u751f\u6210\u4e00\u4e2a\u65b0\u8bcd\uff0c\u90fd\u9700\u8981\u8ba1\u7b97\u5b83\u4e0e\u4e4b\u524d\u6240\u6709\u8bcd\u7684 Attention\u3002\u5982\u679c\u6bcf\u6b21\u4ece\u5934\u7b97\u4e00\u904d\uff0c\u8ba1\u7b97\u91cf\u4f1a\u4ee5\u5e73\u65b9\u7ea7\u589e\u957f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\uff0c\u5de5\u7a0b\u5e08\u4eec\u53d1\u660e\u4e86 <strong>KV Cache<\/strong> \u6280\u672f\uff1a\u628a\u5df2\u7ecf\u7b97\u8fc7\u7684 Key \u548c Value \u5411\u91cf\u7f13\u5b58\u8d77\u6765\uff0c\u751f\u6210\u4e0b\u4e00\u4e2a\u8bcd\u65f6\u76f4\u63a5\u590d\u7528\uff0c\u907f\u514d\u91cd\u590d\u8ba1\u7b97\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u95ee\u9898\u6765\u4e86\uff1aMHA \u4e2d\u6bcf\u4e2a\u5934\u90fd\u6709\u81ea\u5df1\u72ec\u7acb\u7684 K \u548c V\u3002\u5e8f\u5217\u8d8a\u957f\u3001\u5934\u6570\u8d8a\u591a\uff0cKV Cache \u5360\u7528\u7684\u663e\u5b58\u5c31\u8d8a\u5927\u3002\u8fd9\u6210\u4e86\u5927\u6a21\u578b\u63a8\u7406\u6548\u7387\u7684\u74f6\u9888\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.5 MQA\uff1a\u8ba9\u6240\u6709\u5934\u5171\u4eab\u540c\u4e00\u4efd K \u548c V<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u538b\u7f29 KV Cache\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86 <strong>MQA\uff08Multi-Query Attention\uff0c\u591a\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">MQA \u7684\u6838\u5fc3\u601d\u60f3\u5f88\u7b80\u5355\uff1a<strong>\u8ba9\u6240\u6709\u6ce8\u610f\u529b\u5934\u5171\u4eab\u540c\u4e00\u4efd Key \u548c Value\uff0c\u53ea\u6709 Query \u4fdd\u6301\u72ec\u7acb<\/strong>\u3002\u8fd9\u6837 KV Cache \u7684\u5927\u5c0f\u76f4\u63a5\u7f29\u5c0f\u5230\u539f\u6765\u7684 1\/h\uff08h \u662f\u5934\u6570\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee3\u4ef7\u662f\uff1a\u8868\u8fbe\u80fd\u529b\u6709\u6240\u4e0b\u964d\u3002\u6bd5\u7adf\u6240\u6709\u5934\u90fd\u5728\u7528\u76f8\u540c\u7684 K \u548c V \u6765\u8ba1\u7b97 Attention\uff0c\u6a21\u578b\u6355\u6349\u591a\u6837\u5316\u8bed\u4e49\u7684\u80fd\u529b\u53d7\u5230\u4e86\u9650\u5236\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.6 GQA\uff1a\u5728\u6548\u7387\u548c\u6548\u679c\u4e4b\u95f4\u627e\u5230\u5e73\u8861<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GQA\uff08Grouped Query Attention\uff0c\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \u662f MHA \u548c MQA \u4e4b\u95f4\u7684\u201c\u6298\u4e2d\u65b9\u6848\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GQA \u7684\u505a\u6cd5\u662f\uff1a\u5c06 Query \u5934\u5206\u6210\u82e5\u5e72\u7ec4\uff08\u6bd4\u5982 8 \u4e2a\u5934\u5206\u6210 2 \u7ec4\uff0c\u6bcf\u7ec4 4 \u4e2a\u5934\uff09\uff0c<strong>\u6bcf\u7ec4\u5185\u7684 Query \u5934\u5171\u4eab\u540c\u4e00\u4efd Key \u548c Value<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u65e2\u4fdd\u6301\u4e86\u6bd4 MQA \u66f4\u5f3a\u7684\u8868\u8fbe\u80fd\u529b\uff08\u591a\u4e2a\u7ec4\u53ef\u4ee5\u5b66\u4e60\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\uff09\uff0c\u53c8\u663e\u8457\u538b\u7f29\u4e86 KV Cache\uff08\u76f8\u6bd4 MHA \u7f29\u5c0f\u4e86\u6570\u500d\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9645\u6548\u679c\u975e\u5e38\u51fa\u8272\uff1a<strong>GQA \u5728\u4fdd\u6301\u63a5\u8fd1 MHA \u8d28\u91cf\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u4e86\u63a5\u8fd1 MQA \u7684\u63a8\u7406\u901f\u5ea6<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LLaMA-2\/3 \u7cfb\u5217\u5728 70B \u5927\u6a21\u578b\u4e0a\u91c7\u7528\u4e86 GQA\uff0cQwen\u3001Mistral \u7b49\u6a21\u578b\u4e5f\u7eb7\u7eb7\u8ddf\u8fdb\uff0c\u4f7f\u5176\u6210\u4e3a\u73b0\u4ee3 LLM \u7684\u6807\u914d\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\ud83d\udca1 <strong>\u8fdb\u9636\u5ef6\u4f38<\/strong>\uff1aDeepSeek \u5728 GQA \u7684\u57fa\u7840\u4e0a\u66f4\u8fdb\u4e00\u6b65\uff0c\u63d0\u51fa\u4e86 <strong>MLA\uff08Multi-head Latent Attention\uff0c\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff09<\/strong> \uff0c\u901a\u8fc7\u5f15\u5165\u6f5c\u5728\u7a7a\u95f4\u6295\u5f71\u548c\u52a8\u6001\u8def\u7531\uff0c\u8fdb\u4e00\u6b65\u538b\u7f29 KV Cache \u5e76\u63d0\u5347\u63a8\u7406\u6548\u7387\u3002\u8fd9\u4e2a\u8bdd\u9898\u6211\u4eec\u4f1a\u5728\u540e\u7eed\u7cfb\u5217\u6587\u7ae0\u4e2d\u4e13\u95e8\u5c55\u5f00\u3002<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001FFN \u4e0e SwiGLU\uff1a\u6a21\u578b\u7684\u201c\u77e5\u8bc6\u5b58\u50a8\u5668\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">6.1 FFN \u7684\u4f5c\u7528<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u8d1f\u8d23\u8ba9\u6a21\u578b\u77e5\u9053\u201c\u54ea\u4e9b\u8bcd\u4e4b\u95f4\u6709\u5173\u7cfb\u201d\uff0c\u4f46\u5149\u9760 Attention \u662f\u4e0d\u591f\u7684\u2014\u2014\u6a21\u578b\u7684\u201c\u77e5\u8bc6\u201d\u548c\u201c\u63a8\u7406\u80fd\u529b\u201d\u4e3b\u8981\u5b58\u50a8\u5728 <strong>FFN\uff08Feed-Forward Network\uff0c\u524d\u9988\u7f51\u7edc\uff09<\/strong> \u4e2d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">FFN \u7684\u7ed3\u6784\u975e\u5e38\u7b80\u5355\uff0c\u5c31\u662f\u4e00\u4e2a\u4e24\u5c42\u7684\u5168\u8fde\u63a5\u7f51\u7edc\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>\\text{FFN}(x) = W_2 \\cdot \\sigma(W_1 \\cdot x + b_1) + b_2<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d (\\sigma) \u662f\u6fc0\u6d3b\u51fd\u6570\uff0c(W_1) \u5c06\u8f93\u5165\u7ef4\u5ea6\u653e\u5927\uff08\u901a\u5e38\u653e\u5927 4 \u500d\uff09\uff0c(W_2) \u518d\u7f29\u56de\u539f\u6765\u7684\u7ef4\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u770b\u4f3c\u7b80\u5355\u7684\u7ed3\u6784\uff0c\u5b9e\u9645\u4e0a\u627f\u62c5\u4e86\u6a21\u578b\u7684\u5927\u90e8\u5206\u53c2\u6570\uff08\u7ea6 2\/3\uff09\uff0c\u53ef\u4ee5\u8bf4\u662f LLM \u7684\u201c\u77e5\u8bc6\u5b58\u50a8\u5668\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6.2 \u4ece ReLU\/GELU \u5230 SwiGLU<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u65e9\u671f Transformer \u7684 FFN \u4f7f\u7528\u7684\u662f ReLU \u6216 GELU \u4f5c\u4e3a\u6fc0\u6d3b\u51fd\u6570\u3002\u4f46\u8fd1\u5e74\u6765\uff0c\u4e00\u79cd\u53eb\u505a <strong>SwiGLU<\/strong> \u7684\u6fc0\u6d3b\u51fd\u6570\u9010\u6e10\u6210\u4e3a\u4e3b\u6d41\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SwiGLU \u7684\u5168\u79f0\u662f <strong>Swish Gated Linear Unit<\/strong>\uff0c\u5b83\u7684\u6838\u5fc3\u521b\u65b0\u662f\u5f15\u5165\u4e86\u4e00\u4e2a<strong>\u95e8\u63a7\u673a\u5236<\/strong>\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>\\text{SwiGLU}(x) = (\\text{Swish}(x W_1) \\odot x V) W_2<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u53d1\u751f\u4e86\u4ec0\u4e48\uff1f\u8f93\u5165 x \u88ab\u6295\u5f71\u6210\u4e24\u8def\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e00\u8def\u7ecf\u8fc7 <strong>Swish\uff08\u4e5f\u53eb SiLU\uff09<\/strong> \u6fc0\u6d3b\u51fd\u6570\uff0c\u4f5c\u4e3a\u201c\u95e8\u63a7\u4fe1\u53f7\u201d\uff0c\u51b3\u5b9a\u54ea\u4e9b\u4fe1\u606f\u53ef\u4ee5\u901a\u8fc7\u3002<\/li>\n\n\n\n<li>\u53e6\u4e00\u8def\u76f4\u63a5\u4f5c\u4e3a\u201c\u503c\u4fe1\u53f7\u201d\uff0c\u627f\u8f7d\u5b9e\u9645\u7684\u5185\u5bb9\u4fe1\u606f\u3002<\/li>\n\n\n\n<li>\u4e24\u8def\u505a\u9010\u5143\u7d20\u4e58\u6cd5\uff0c\u518d\u6295\u5f71\u56de\u539f\u7ef4\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u201c\u95e8\u63a7 + \u503c\u201d\u7684\u53cc\u8def\u8bbe\u8ba1\uff0c\u8ba9\u7f51\u7edc\u80fd\u591f<strong>\u52a8\u6001\u5730\u9009\u62e9\u6027\u5730\u4f20\u9012\u4fe1\u606f<\/strong>\uff0c\u800c\u4e0d\u662f\u50cf\u4f20\u7edf FFN \u90a3\u6837\u628a\u8f93\u5165\u4e00\u80a1\u8111\u5730\u4e22\u8fdb\u6fc0\u6d3b\u51fd\u6570\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u6709\u4e86\u66f4\u5f3a\u7684\u975e\u7ebf\u6027\u8868\u8fbe\u80fd\u529b\u548c\u66f4\u597d\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6.3 \u4e3a\u4ec0\u4e48 SwiGLU \u66f4\u597d\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f8\u6bd4 ReLU\uff0cSwish \u51fd\u6570\u5728\u8d1f\u503c\u533a\u57df\u4e0d\u662f\u76f4\u63a5\u7f6e\u96f6\uff0c\u800c\u662f\u4fdd\u7559\u4e86\u5e73\u6ed1\u7684\u975e\u96f6\u503c\u3002\u8fd9\u5e26\u6765\u4e86\u51e0\u4e2a\u597d\u5904\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u68af\u5ea6\u6d41\u52a8\u66f4\u987a\u7545<\/strong>\uff1a\u6ca1\u6709\u201c\u6b7b\u795e\u7ecf\u5143\u201d\u95ee\u9898\uff0c\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3002<\/li>\n\n\n\n<li><strong>\u8868\u8fbe\u80fd\u529b\u66f4\u5f3a<\/strong>\uff1a\u95e8\u63a7\u673a\u5236\u8ba9\u6a21\u578b\u80fd\u6839\u636e\u8f93\u5165\u52a8\u6001\u8c03\u6574\u4fe1\u606f\u6d41\u3002<\/li>\n\n\n\n<li><strong>\u6536\u655b\u66f4\u5feb<\/strong>\uff1a\u5728\u5b9e\u9645\u8bad\u7ec3\u4e2d\uff0cSwiGLU \u901a\u5e38\u80fd\u66f4\u5feb\u8fbe\u5230\u66f4\u4f4e\u7684 loss\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">Meta \u7684 LLaMA \u7cfb\u5217\u3001\u963f\u91cc\u7684 Qwen \u7cfb\u5217\u3001Mistral \u7b49\u4e3b\u6d41\u6a21\u578b\uff0c\u90fd\u5df2\u7ecf\u5c06 FFN \u5347\u7ea7\u4e3a SwiGLU\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u5f52\u4e00\u5316\u4e0e\u6b8b\u5dee\uff1a\u8ba9\u6df1\u5c42\u7f51\u7edc\u201c\u7a33\u5f97\u4f4f\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">7.1 \u4e3a\u4ec0\u4e48\u9700\u8981\u5f52\u4e00\u5316\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">LLM \u7684\u5c42\u6570\u52a8\u8f84\u51e0\u5341\u4e0a\u767e\u5c42\uff08LLaMA-70B \u6709 80 \u5c42\uff09\u3002\u968f\u7740\u7f51\u7edc\u53d8\u6df1\uff0c\u4e00\u4e2a\u7ecf\u5178\u95ee\u9898\u51fa\u73b0\u4e86\uff1a<strong>\u68af\u5ea6\u7206\u70b8\u548c\u68af\u5ea6\u6d88\u5931<\/strong>\u2014\u2014\u524d\u9762\u7684\u5c42\u68af\u5ea6\u592a\u5927\u5bfc\u81f4\u53c2\u6570\u9707\u8361\uff0c\u540e\u9762\u7684\u5c42\u68af\u5ea6\u592a\u5c0f\u5bfc\u81f4\u5b66\u4e0d\u52a8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5f52\u4e00\u5316\uff08Normalization\uff09<\/strong> \u5c31\u662f\u7528\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u7684\u3002\u5b83\u901a\u8fc7\u5bf9\u6bcf\u4e00\u5c42\u7684\u8f93\u51fa\u8fdb\u884c\u6807\u51c6\u5316\uff0c\u8ba9\u6570\u503c\u5206\u5e03\u4fdd\u6301\u7a33\u5b9a\uff0c\u4ece\u800c\u8ba9\u8bad\u7ec3\u80fd\u591f\u987a\u5229\u8fdb\u884c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.2 LayerNorm vs RMSNorm<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ecf\u5178\u7684 Transformer \u4f7f\u7528\u7684\u662f <strong>LayerNorm\uff08\u5c42\u5f52\u4e00\u5316\uff09<\/strong> \uff0c\u5176\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>\\text{LayerNorm}(x) = \\frac{x &#8211; \\mu}{\\sqrt{\\sigma^2 + \\epsilon}} \\cdot \\gamma + \\beta<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d (\\mu) \u662f\u5747\u503c\uff0c(\\sigma^2) \u662f\u65b9\u5dee\uff0c(\\gamma) \u548c (\\beta) \u662f\u53ef\u5b66\u4e60\u7684\u7f29\u653e\u548c\u504f\u79fb\u53c2\u6570\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LayerNorm \u6548\u679c\u4e0d\u9519\uff0c\u4f46\u8ba1\u7b97\u5747\u503c\u548c\u65b9\u5dee\u6709\u4e00\u5b9a\u5f00\u9500\u3002\u4e8e\u662f\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u4e2a\u7b80\u5316\u7248\u2014\u2014<strong>RMSNorm\uff08Root Mean Square Normalization\uff09<\/strong> \uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>\\text{RMSNorm}(x) = \\frac{x}{\\text{RMS}(x)} \\cdot \\gamma<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">RMSNorm \u53ea\u8ba1\u7b97<strong>\u5747\u65b9\u6839<\/strong>\uff08RMS\uff09\uff0c\u53bb\u6389\u4e86\u5747\u503c\u7684\u8ba1\u7b97\u3002\u867d\u7136\u727a\u7272\u4e86\u201c\u4e2d\u5fc3\u5316\u201d\u7684\u80fd\u529b\uff0c\u4f46\u5b9e\u8df5\u8868\u660e\u6700\u7ec8\u6548\u679c\u548c LayerNorm \u5dee\u4e0d\u591a\uff0c\u800c\u8ba1\u7b97\u6548\u7387\u66f4\u9ad8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0cLLaMA\u3001Qwen\u3001DeepSeek \u7b49\u6a21\u578b\u90fd\u91c7\u7528\u4e86 RMSNorm \u4f5c\u4e3a\u5f52\u4e00\u5316\u65b9\u5f0f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.3 Pre-Norm vs Post-Norm\uff1a\u8c01\u5728\u524d\uff0c\u8c01\u5728\u540e\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f52\u4e00\u5316\u653e\u5728\u54ea\u91cc\uff0c\u540c\u6837\u81f3\u5173\u91cd\u8981\u3002\u6709\u4e24\u79cd\u9009\u62e9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Post-Norm<\/strong>\uff1a\u5f52\u4e00\u5316\u653e\u5728\u5b50\u5c42\uff08Attention \u6216 FFN\uff09<strong>\u4e4b\u540e<\/strong>\u3002\u8fd9\u662f 2017 \u5e74\u539f\u7248 Transformer \u7684\u8bbe\u8ba1\u3002<\/li>\n\n\n\n<li><strong>Pre-Norm<\/strong>\uff1a\u5f52\u4e00\u5316\u653e\u5728\u5b50\u5c42<strong>\u4e4b\u524d<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5b9e\u8df5\u4e2d\uff0c\u7814\u7a76\u8005\u53d1\u73b0 Post-Norm \u5728\u6df1\u5c42\u7f51\u7edc\uff08\u51e0\u5341\u5c42\u4ee5\u4e0a\uff09\u4e2d\u5bb9\u6613\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff0c\u68af\u5ea6\u4f20\u64ad\u4e0d\u591f\u5e73\u6ed1\u3002\u800c Pre-Norm \u5728\u6b8b\u5dee\u8fde\u63a5\u4e4b\u524d\u505a\u5f52\u4e00\u5316\uff0c\u80fd\u591f\u663e\u8457\u6539\u5584\u68af\u5ea6\u7684\u4f20\u64ad\uff0c\u8ba9\u6df1\u5c42\u7f51\u7edc\u7684\u8bad\u7ec3\u66f4\u52a0\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c<strong>\u73b0\u4ee3\u5927\u6a21\u578b\u51e0\u4e4e\u6e05\u4e00\u8272\u5730\u91c7\u7528 Pre-Norm + RMSNorm \u7684\u7ec4\u5408<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u603b\u7ed3\u4e0e\u5c55\u671b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">8.1 \u4e00\u5f20\u5168\u666f\u201c\u6807\u914d\u201d\u6e05\u5355<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u5168\u6587\uff0c\u73b0\u4ee3 LLM \u7684\u67b6\u6784\u201c\u6807\u914d\u201d\u53ef\u4ee5\u603b\u7ed3\u4e3a\u4e0b\u8868\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7ec4\u4ef6<\/th><th>\u7ecf\u5178\u65b9\u6848<\/th><th>\u73b0\u4ee3 LLM \u6807\u914d<\/th><th>\u6838\u5fc3\u4f18\u52bf<\/th><\/tr><\/thead><tbody><tr><td>Tokenizer<\/td><td>BPE\/WordPiece<\/td><td>SentencePiece\uff08BPE\uff09<\/td><td>\u8bed\u8a00\u65e0\u5173\uff0c\u4e2d\u6587\u53cb\u597d<\/td><\/tr><tr><td>\u4f4d\u7f6e\u7f16\u7801<\/td><td>\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801<\/td><td><strong>RoPE<\/strong><\/td><td>\u76f8\u5bf9\u4f4d\u7f6e\uff0c\u957f\u5e8f\u5217\u6cdb\u5316<\/td><\/tr><tr><td>\u6ce8\u610f\u529b<\/td><td>MHA<\/td><td><strong>GQA<\/strong><\/td><td>\u5e73\u8861\u6548\u679c\u4e0e\u6548\u7387<\/td><\/tr><tr><td>FFN \u6fc0\u6d3b<\/td><td>ReLU\/GELU<\/td><td><strong>SwiGLU<\/strong><\/td><td>\u95e8\u63a7\u673a\u5236\uff0c\u66f4\u5f3a\u8868\u8fbe<\/td><\/tr><tr><td>\u5f52\u4e00\u5316<\/td><td>LayerNorm + Post-Norm<\/td><td><strong>RMSNorm + Pre-Norm<\/strong><\/td><td>\u8bad\u7ec3\u7a33\u5b9a\uff0c\u8ba1\u7b97\u9ad8\u6548<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">LLaMA\u3001Qwen\u3001Mistral\u3001DeepSeek \u7b49\u4e3b\u6d41\u6a21\u578b\uff0c\u5728\u67b6\u6784\u4e0a\u90fd\u9075\u5faa\u4e86\u8fd9\u5957\u201c\u9ec4\u91d1\u914d\u65b9\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8.2 \u6570\u636e\u6d41\u5168\u666f\u56de\u987e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba9\u6211\u4eec\u518d\u628a\u6574\u6761\u6570\u636e\u6d41\u4e32\u8d77\u6765\u590d\u4e60\u4e00\u904d\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Tokenizer<\/strong> \u5c06\u539f\u59cb\u6587\u672c\u5207\u6210 Token\uff0c\u6620\u5c04\u4e3a\u6570\u5b57 ID\u3002<\/li>\n\n\n\n<li><strong>Embedding<\/strong> \u5c06\u6bcf\u4e2a Token ID \u8f6c\u6362\u4e3a\u9ad8\u7ef4\u5411\u91cf\u3002<\/li>\n\n\n\n<li><strong>RoPE<\/strong> \u5bf9\u5411\u91cf\u65bd\u52a0\u65cb\u8f6c\u64cd\u4f5c\uff0c\u6ce8\u5165\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f\u3002<\/li>\n\n\n\n<li><strong>Transformer Block<\/strong>\uff08\u91cd\u590d N \u5c42\uff09\uff1a<\/li>\n<\/ol>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Pre-Norm\uff08RMSNorm\uff09<\/strong> \u7a33\u5b9a\u6570\u503c\u5206\u5e03\u3002<\/li>\n\n\n\n<li><strong>GQA\uff08\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \u8ba9\u6a21\u578b\u201c\u6ce8\u610f\u201d\u5230\u4e0a\u4e0b\u6587\u4e2d\u76f8\u5173\u7684 Token\u3002<\/li>\n\n\n\n<li><strong>\u6b8b\u5dee\u8fde\u63a5<\/strong> \u4fdd\u7559\u539f\u59cb\u4fe1\u606f\uff0c\u9632\u6b62\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n\n\n\n<li><strong>Pre-Norm\uff08RMSNorm\uff09<\/strong> \u518d\u6b21\u5f52\u4e00\u5316\u3002<\/li>\n\n\n\n<li><strong>FFN\uff08SwiGLU\uff09<\/strong> \u5bf9\u4fe1\u606f\u8fdb\u884c\u975e\u7ebf\u6027\u53d8\u6362\uff0c\u5b58\u50a8\u77e5\u8bc6\u3002<\/li>\n\n\n\n<li><strong>\u6b8b\u5dee\u8fde\u63a5<\/strong> \u518d\u6b21\u4fdd\u7559\u4fe1\u606f\u3002<\/li>\n<\/ul>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u6700\u540e\u901a\u8fc7\u8f93\u51fa\u5c42\u9884\u6d4b\u4e0b\u4e00\u4e2a Token\uff0c\u5982\u6b64\u5faa\u73af\u76f4\u5230\u751f\u6210\u5b8c\u6574\u56de\u7b54\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 AI\u65e0\u7591\u662f\u6700\u8fd1\u51e0\u5e74\u6700\u706b\u7684\u4e1c\u897f\u4e86\uff0c\u5176\u80fd\u529b\u4e0e\u53d1\u5c55\u901f\u5ea6\u90fd\u8ba9\u4eba\u60ca\u53f9\uff0c\u4ece17\u5e74\u7684Transformer\u5230\u5982\u4eca\u7684 [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":121,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[37],"class_list":["post-119","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-llm"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/119","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=119"}],"version-history":[{"count":1,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/119\/revisions"}],"predecessor-version":[{"id":120,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/119\/revisions\/120"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/121"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=119"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=119"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=119"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}