{"id":19956,"date":"2026-01-08T14:20:34","date_gmt":"2026-01-08T06:20:34","guid":{"rendered":"https:\/\/92it.top\/?p=19956"},"modified":"2026-01-08T14:22:50","modified_gmt":"2026-01-08T06:22:50","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%86%85%e9%83%a8%e7%bb%93%e6%9e%84%e6%8f%ad%e7%a7%98-%e4%bb%8etoken%e5%88%b0transformer%e7%9a%84%e5%ae%8c%e6%95%b4%e6%8c%87%e5%8d%97%ef%bc%81","status":"publish","type":"post","link":"https:\/\/92it.top\/?p=19956","title":{"rendered":"\u5927\u6a21\u578b\u5185\u90e8\u7ed3\u6784\u63ed\u79d8\u2014\u2014\u4eceToken\u5230Transformer\u7684\u5b8c\u6574\u6307\u5357\uff01"},"content":{"rendered":"\n

\u8f6c\u8f7d\uff1a\u5c0f\u767d\u5fc5\u770b\uff1a\u5927\u6a21\u578b\u5185\u90e8\u7ed3\u6784\u63ed\u79d8\u2014\u2014\u4eceToken\u5230Transformer\u7684\u5b8c\u6574\u6307\u5357\uff01<\/a><\/p>\n\n\n\n

\u524d\u8a00\ud83d\udd16<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u5728 ChatGPT \u5e2d\u5377\u4e16\u754c\u4e4b\u540e\uff0c\u5f88\u591a\u4eba\u90fd\u542c\u8fc7\u300c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u300d\u8fd9\u4e2a\u8bcd\u3002
\u4f46\u5f53\u6211\u4eec\u95ee\u81ea\u5df1\uff1a\u201cLLM \u7684\u5185\u90e8\u7ed3\u6784\u7a76\u7adf\u662f\u4ec0\u4e48\uff1f\u201d \u65f6\uff0c\u5f80\u5f80\u53ea\u6709\u6a21\u7cca\u7684\u5370\u8c61\u3002<\/p>\n\n\n\n

\u5b83\u4f3c\u4e4e\u662f\u67d0\u79cd\u201c\u5de8\u5927\u7684\u9ed1\u7bb1\u201d\uff0c\u80fd\u7406\u89e3\u4eba\u7c7b\u8bed\u8a00\u3001\u751f\u6210\u56de\u7b54\u3001\u751a\u81f3\u63a8\u7406\u521b\u4f5c\u3002
\u7136\u800c\uff0c\u6253\u5f00\u90a3\u5c42\u9ed1\u7bb1\uff0c\u4f60\u4f1a\u53d1\u73b0\u2014\u2014\u5b83\u5176\u5b9e\u662f\u4e00\u4e2a\u5c42\u5c42\u6784\u5efa\u7684\u8bed\u8a00\u201c\u601d\u8003\u673a\u5668\u201d\u3002<\/p>\n\n\n\n

\u4eca\u5929\uff0c\u6211\u4eec\u4ece\u7ed3\u6784\u89d2\u5ea6\uff0c\u5e26\u4f60\u770b\u6e05 LLM \u7684\u6838\u5fc3\u7ec4\u6210\uff1a
\u5206\u8bcd\uff08Tokenization\uff09 \u2192 \u8bcd\u5d4c\u5165\uff08Embedding\uff09 \u2192 \u4f4d\u7f6e\u7f16\u7801 \u2192 N \u5c42 Transformer\u67b6\u6784 \u2192 \u8f93\u51fa\u6982\u7387\u3002<\/p>\n\n\n

\n
\"\"<\/figure><\/div>\n\n\n

\u4e0a\u56fe\u662f\u7ecf\u5178\u7684 Transformer \u6574\u4f53\u6846\u67b6\u7ed3\u6784<\/strong>\uff0c\u5b83\u7684\u539f\u59cb\u51fa\u5904<\/strong>\u662f\u53d1\u8868\u4e8e 2017 \u5e74<\/strong>\u7684\u5f00\u521b\u6027\u8bba\u6587”Attention Is All You Need”\uff0c\u8bfb\u5b8c\u6574\u7bc7\u6587\u7ae0\u60a8\u5c31\u53ef\u4ee5\u7406\u89e3<\/p>\n\n\n\n

\u3000\u3000<\/p>\n\n\n\n

\u7c97\u6d45\u7406\u89e3 \u5927\u6a21\u578b\u5c31\u662f\u4e00\u4e2a\u51fd\u6570\u662f\u5417\uff1f \u63d0\u793a\u8bcd\u662f\u8c03\u7528\u7684\u53c2\u6570\uff1f\ud83d\udd16<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u5927\u6a21\u578b\u672c\u8d28\u5c31\u662f\u4e00\u4e2a \u201c\u8f93\u5165\u2192\u8ba1\u7b97\u2192\u8f93\u51fa\u201d \u7684\u6570\u5b66\u51fd\u6570\uff0c\u63d0\u793a\u8bcd\u5c31\u662f\u8c03\u7528\u8fd9\u4e2a\u51fd\u6570\u7684\u6838\u5fc3\u8f93\u5165\u53c2\u6570\u3002<\/p>\n\n\n\n

\u6211\u4eec\u53ef\u4ee5\u7528\u66f4\u901a\u4fd7\u7684\u65b9\u5f0f\u62c6\u89e3\u8fd9\u4e2a\u7c7b\u6bd4\uff1a<\/p>\n\n\n\n

\u51fd\u6570\u7684\u57fa\u672c\u5f62\u5f0f\u662f\uff1a<\/p>\n\n\n\n

\u8f93\u51fa = \u6a21\u578b\u51fd\u6570(\u8f93\u5165, \u63a8\u7406\u53c2\u6570)\n       \u2191\n       \u7531\u4e24\u90e8\u5206\u5171\u540c\u6784\u6210\uff1a\n       1. \u67b6\u6784\u4ee3\u7801 \u2192 \u51fd\u6570\u7684\u201c\u8ba1\u7b97\u89c4\u5219\u201d\uff08\u6bd4\u5982\u6ce8\u610f\u529b\u600e\u4e48\u7b97\u3001\u5411\u91cf\u600e\u4e48\u6620\u5c04\uff09\n       2. \u6743\u91cd\u6587\u4ef6 \u2192 \u51fd\u6570\u7684\u201c\u7cfb\u6570\u53c2\u6570\u201d\uff08\u6bd4\u5982Q\u77e9\u9635\u76840.0123\u3001\u8bcd\u5d4c\u5165\u77e9\u9635\u7684\u6570\u503c\uff09<\/pre>\n\n\n\n
    \n
  • \u51fd\u6570\u7684 \u201c\u8ba1\u7b97\u89c4\u5219\u201d\uff1a\u7531\u8bad\u7ec3\u597d\u7684\u6743\u91cd\u53c2\u6570\u51b3\u5b9a\uff08\u5c31\u662f\u90a3\u4e9b\u4f60\u4e0b\u8f7d\u7684.bin\/.safetensors\u6587\u4ef6\u91cc\u7684\u6d6e\u70b9\u6570\uff09\u3002\u8fd9\u4e9b\u6743\u91cd\u662f\u51fd\u6570\u7684 \u201c\u56fa\u5b9a\u903b\u8f91\u201d\uff0c\u8bad\u7ec3\u5b8c\u5c31\u4e0d\u4f1a\u53d8\uff0c\u76f8\u5f53\u4e8e\u51fd\u6570\u91cc\u7684\u7cfb\u6570\u3002<\/li>\n\n\n\n
  • \u51fd\u6570\u7684 \u201c\u8f93\u5165\u201d\uff1a\u4e3b\u8981\u5c31\u662f\u63d0\u793a\u8bcd\uff08Prompt\uff09\uff0c\u6bd4\u5982 \u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u8bd7\u201d\u201c\u628a\u8fd9\u5f20\u56fe\u53d8\u6210\u52a8\u6001\u89c6\u9891\u201d\u3002<\/li>\n\n\n\n
  • \u51fd\u6570\u7684 \u201c\u63a7\u5236\u53c2\u6570\u201d\uff1a\u5c31\u662f\u4f60\u4e4b\u524d\u4e86\u89e3\u7684temperature\u3001top_k\u3001max_new_tokens\u7b49\u63a8\u7406\u53c2\u6570\uff0c\u7528\u6765\u8c03\u8282\u51fd\u6570\u8f93\u51fa\u7684\u98ce\u683c\uff08\u6bd4\u5982\u4e25\u8c28 \/ \u521b\u610f\uff09\u548c\u957f\u5ea6\u3002<\/li>\n\n\n\n
  • \u51fd\u6570\u7684 \u201c\u8f93\u51fa\u201d\uff1a\u5c31\u662f\u6a21\u578b\u751f\u6210\u7684\u6587\u672c\u3001\u89c6\u9891\u3001\u4ee3\u7801\u7b49\u7ed3\u679c\u3002<\/li>\n<\/ul>\n\n\n\n
    \u7528\u66f4\u7cbe\u51c6\u7684\u65b9\u5f0f\u62c6\u89e3\u8fd9\u4e2a\u51fd\u6570\u7684\u6784\u6210\uff1a<\/p>\n\n\n\n
      \n
    • \u67b6\u6784\u4ee3\u7801 = \u51fd\u6570\u7684 \u201c\u8ba1\u7b97\u89c4\u5219\u201d<\/li>\n<\/ul>\n\n\n\n
      \u76f8\u5f53\u4e8e\u6570\u5b66\u91cc\u7684f(x) = ax\u00b2 + bx + c\u8fd9\u4e2a\u516c\u5f0f\u672c\u8eab\uff1a
      \u5b83\u5b9a\u4e49\u4e86 \u201c\u8f93\u5165 x \u8981\u5148\u5e73\u65b9\u3001\u518d\u4e58 a\u3001\u518d\u52a0 bx\u3001\u6700\u540e\u52a0 c\u201d \u7684\u6b65\u9aa4\uff1b
      \u5bf9\u5e94 LLM \u7684\u67b6\u6784\u4ee3\u7801\uff0c\u5c31\u662f\u5b9a\u4e49\u4e86 \u201c\u8f93\u5165\u63d0\u793a\u8bcd\u8981\u5148\u5206\u8bcd\u3001\u518d\u505a\u8bcd\u5d4c\u5165\u3001\u518d\u8dd1 Transformer\u3001\u6700\u540e\u7b97\u6982\u7387\u201d \u7684\u6b65\u9aa4\u3002
      <\/pre>\n\n\n\n
        \n
      • \u6743\u91cd\u6587\u4ef6 = \u51fd\u6570\u7684 \u201c\u7cfb\u6570\uff08a\/b\/c\uff09\u201d<\/li>\n<\/ul>\n\n\n\n
        \u76f8\u5f53\u4e8ef(x) = 2x\u00b2 + 3x + 5\u91cc\u76842\/3\/5\u8fd9\u4e9b\u5177\u4f53\u6570\u503c\uff1a
        \u6ca1\u6709\u8fd9\u4e9b\u6570\u503c\uff0c\u516c\u5f0f\u53ea\u662f\u7a7a\u67b6\u5b50\uff08\u6bd4\u5982f(x)=ax\u00b2+bx+c\u7b97\u4e0d\u51fa\u4efb\u4f55\u7ed3\u679c\uff09\uff1b
        \u5bf9\u5e94 LLM \u7684\u6743\u91cd\u6587\u4ef6\uff0c\u5c31\u662f.safetensors\u91cc\u7684\u6d6e\u70b9\u6570 \u2014\u2014 \u6ca1\u6709\u8fd9\u4e9b\u6570\u503c\uff0c\u67b6\u6784\u4ee3\u7801\u7684 \u201c\u6ce8\u610f\u529b\u8ba1\u7b97\u201d\u201c\u8bcd\u5d4c\u5165\u201d \u90fd\u53ea\u662f\u7a7a\u7684\u6570\u5b66\u516c\u5f0f\uff0c\u65e0\u6cd5\u751f\u6210\u4efb\u4f55\u6709\u610f\u4e49\u7684\u8f93\u51fa\u3002
        <\/pre>\n\n\n\n
          \n
        • \u5173\u952e\u8865\u5145\uff1a<\/li>\n<\/ul>\n\n\n\n
          \u5f53\u4f60\u52a0\u8f7d\u4e00\u4e2a\u5927\u6a21\u578b\u65f6\uff0c\u672c\u8d28\u662f\u628a\u6743\u91cd\u6587\u4ef6\u7684\u6570\u503c \u201c\u586b\u5145\u201d \u5230\u67b6\u6784\u4ee3\u7801\u7684\u516c\u5f0f\u91cc\uff0c\u8ba9\u7a7a\u7684 \u201c\u6a21\u578b\u51fd\u6570\u201d \u53d8\u6210\u6709\u5177\u4f53\u8ba1\u7b97\u80fd\u529b\u7684 \u201c\u53ef\u7528\u51fd\u6570\u201d\uff1b
          \u63a8\u7406\u53c2\u6570\uff08temperature\/top_k \u7b49\uff09\u662f\u8fd9\u4e2a\u51fd\u6570\u7684\u53ef\u9009\u53c2\u6570\u2014\u2014 \u4e0d\u6539\u53d8\u51fd\u6570\u7684\u6838\u5fc3\u8ba1\u7b97\u903b\u8f91\uff0c\u53ea\u8c03\u6574\u8f93\u51fa\u7684\u98ce\u683c\uff08\u6bd4\u5982\u66f4\u968f\u673a \/ \u66f4\u786e\u5b9a\uff09\u3002<\/pre>\n\n\n\n
            \n
          • \u603b\u7ed3<\/li>\n<\/ul>\n\n\n\n
            \u201c\u6a21\u578b\u51fd\u6570\u201d \u786e\u5b9e\u540c\u65f6\u5305\u542b\u4e86\uff1a
            \u67b6\u6784\u4ee3\u7801\uff08\u51fd\u6570\u7684\u8ba1\u7b97\u903b\u8f91 \/ \u516c\u5f0f\uff09\uff1b
            \u6743\u91cd\u6587\u4ef6\uff08\u51fd\u6570\u7684\u7cfb\u6570 \/ \u5177\u4f53\u6570\u503c\uff09\u3002
            \u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 \u201c\u53ea\u7ed9\u6743\u91cd\u6587\u4ef6\u7528\u4e0d\u4e86\u201d\u2014\u2014 \u5c31\u50cf\u53ea\u7ed9\u4f602\/3\/5\uff0c\u5374\u4e0d\u7ed9\u4f60f(x)=ax\u00b2+bx+c\u8fd9\u4e2a\u516c\u5f0f\uff0c\u4f60\u6c38\u8fdc\u4e0d\u77e5\u9053\u8be5\u600e\u4e48\u7528\u8fd9\u4e9b\u6570\u5b57\u8ba1\u7b97\u3002
            \u3000\u3000<\/pre>\n\n\n\n
            \u3000\u3000<\/p>\n\n\n\n
            \u5206\u8bcd\uff08Tokenization\uff09\uff1a\u628a\u6587\u5b57\u62c6\u6210\u6a21\u578b\u80fd\u61c2\u7684\u300c\u8bed\u8a00\u9897\u7c92\u300d\ud83d\udd16<\/strong><\/p>\n\n\n\n
            \n\n\n\n
            \u4eba\u7c7b\u7528\u201c\u5b57\u201d\u6216\u201c\u8bcd\u201d\u4ea4\u6d41\uff0c\u4f46\u8ba1\u7b97\u673a\u53ea\u61c2\u6570\u5b57\u3002
            Tokenization\uff08\u5206\u8bcd\uff09<\/strong> \u7684\u4efb\u52a1\uff0c\u5c31\u662f\u628a\u81ea\u7136\u8bed\u8a00\u5207\u6210\u4e00\u4e2a\u4e2a\u6a21\u578b\u80fd\u7406\u89e3\u7684\u6700\u5c0f\u5355\u4f4d\u2014\u2014Token<\/strong>\u3002<\/p>\n\n\n\n
            \u4e3e\u4e2a\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
            \u53e5\u5b50\uff1a\u6211\u559c\u6b22\u5b66\u4e60AI<\/pre>\n\n\n\n
            \u53ef\u80fd\u4f1a\u88ab\u5207\u5206\u6210\uff1a<\/p>\n\n\n\n
            [\"\u6211\", \"\u559c\u6b22\", \"\u5b66\u4e60\", \"A\", \"I\"]<\/pre>\n\n\n\n
            \u6bcf\u4e2a token \u90fd\u4f1a\u88ab\u6620\u5c04\u6210\u4e00\u4e2a\u552f\u4e00\u7f16\u53f7\uff08\u6bd4\u5982 104\u30012075\u3001\u2026\uff09\u3002\u6a21\u578b\u4e4b\u540e\u5904\u7406\u7684\uff0c\u5176\u5b9e\u5c31\u662f\u8fd9\u4e9b\u7f16\u53f7\uff0c\u800c\u975e\u6c49\u5b57\u672c\u8eab\u3002<\/p>\n\n\n\n
            \u3000\u3000<\/p>\n\n\n\n
            \u4e3a\u4ec0\u4e48\u8981\u5206\u8bcd\uff1f\ud83d\udd16<\/strong><\/strong><\/p>\n\n\n\n
            \n\n\n\n
            \u56e0\u4e3a\u8fd9\u6837\u80fd\u628a\u8bed\u8a00\u201c\u79bb\u6563\u5316\u201d\u4e3a\u6570\u5b57\u5e8f\u5217\uff0c\u65b9\u4fbf\u8f93\u5165\u795e\u7ecf\u7f51\u7edc\uff0c\u540c\u65f6\u63a7\u5236\u8ba1\u7b97\u91cf\u3002
            LLM \u4e0d\u80fd\u628a\u6574\u53e5\u8bdd\u76f4\u63a5\u5f53\u53c2\u6570\uff0c\u5fc5\u987b\u5206\u8bcd\uff0c\u672c\u8d28\u662f\u4e3a\u4e86 \u201c\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6 + \u5b9e\u73b0\u8bed\u8a00\u7684\u590d\u7528\u4e0e\u6cdb\u5316\u201d<\/strong>\uff0c\u6211\u4eec\u7528\u901a\u4fd7\u7684\u903b\u8f91\u62c6\u89e3\u6e05\u695a\u3002<\/p>\n\n\n\n
            \ud83d\udd39\u5148\u60f3\u6e05\u695a\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u628a \u201c\u6574\u53e5\u8bdd\u201d \u5f53\u53c2\u6570\uff1f<\/strong><\/p>\n\n\n\n
            \u5047\u8bbe\u4f60\u8f93\u5165 \u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u8bd7\u201d \u8fd9\u53e5\u8bdd\uff0c\u5982\u679c\u76f4\u63a5\u628a\u5b83\u5f53\u4e00\u4e2a \u201c\u6574\u4f53\u53c2\u6570\u201d \u5582\u7ed9\u6a21\u578b\uff1a<\/p>\n\n\n\n
              \n
            1. \u53c2\u6570\u6570\u91cf\u4f1a\u7206\u70b8<\/strong>\uff1a\u4eba\u7c7b\u7684\u8bed\u8a00\u7ec4\u5408\u662f\u65e0\u9650\u7684 \u2014\u2014\u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u8bd7\u201d\u201c\u5199\u4e24\u9996\u6625\u5929\u7684\u8bd7\u201d\u201c\u5199\u4e00\u9996\u590f\u5929\u7684\u8bd7\u201d\u2026\u2026 \u5149\u662f\u548c \u201c\u5199\u8bd7\u201d \u76f8\u5173\u7684\u53e5\u5b50\u5c31\u6709\u4e0a\u4ebf\u79cd\uff0c\u6a21\u578b\u4e0d\u53ef\u80fd\u4e3a\u6bcf\u4e00\u53e5\u8bdd\u90fd\u5355\u72ec\u5206\u914d\u4e00\u4e2a \u201c\u53c2\u6570\u5411\u91cf\u201d\uff08\u76f8\u5f53\u4e8e\u8981\u8bb0\u4f4f\u4e0a\u4ebf\u4e2a\u72ec\u7acb\u7684\u6570\u5b57\u7ec4\u5408\uff09\uff0c7B \u6a21\u578b\u7684\u53c2\u6570\u90fd\u4e0d\u591f\u7528\u3002<\/li>\n\n\n\n
            2. \u65e0\u6cd5\u6cdb\u5316<\/strong>\uff1a\u5982\u679c\u6a21\u578b\u53ea\u8ba4\u8bc6 \u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u8bd7\u201d \u8fd9\u4e2a\u6574\u53e5\uff0c\u5f53\u4f60\u8f93\u5165 \u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u6b4c\u201d\uff0c\u6a21\u578b\u4f1a\u5b8c\u5168\u4e0d\u8ba4\u8bc6 \u2014\u2014 \u56e0\u4e3a\u8fd9\u662f\u4e00\u4e2a \u201c\u65b0\u6574\u53e5\u201d\uff0c\u6ca1\u6709\u5bf9\u5e94\u7684\u53c2\u6570\uff0c\u5c31\u50cf\u5b57\u5178\u91cc\u6ca1\u6709\u8fd9\u4e2a\u8bcd\uff0c\u6839\u672c\u67e5\u4e0d\u5230\u3002<\/li>\n\n\n\n
            3. \u8ba1\u7b97\u6548\u7387\u6781\u4f4e<\/strong>\uff1a\u6574\u53e5\u8bdd\u6ca1\u6709 \u201c\u7ed3\u6784\u201d\uff0c\u6a21\u578b\u65e0\u6cd5\u62c6\u89e3\u8bed\u4e49\uff08\u6bd4\u5982\u5206\u4e0d\u6e05 \u201c\u6625\u5929\u201d \u662f\u6838\u5fc3\u3001\u201c\u8bd7\u201d \u662f\u76ee\u6807\uff09\uff0c\u53ea\u80fd\u50cf\u80cc\u7b54\u6848\u4e00\u6837\u6b7b\u8bb0\u786c\u80cc\uff0c\u65e2\u8017\u663e\u5b58\u53c8\u6162\uff0c\u5b8c\u5168\u4e0d\u5177\u5907\u5b9e\u7528\u4ef7\u503c\u3002<\/li>\n<\/ol>\n\n\n\n
              \u3000\u3000<\/p>\n\n\n\n
              \ud83d\udd39\u5206\u8bcd\u7684\u6838\u5fc3\u4ef7\u503c\uff1a\u628a \u201c\u65e0\u9650\u53e5\u5b50\u201d \u53d8\u6210 \u201c\u6709\u9650\u7ec4\u4ef6\u201d<\/strong><\/p>\n\n\n\n
              \u5206\u8bcd\uff08Tokenization\uff09\u672c\u8d28\u662f\u628a\u65e0\u9650\u7684\u81ea\u7136\u8bed\u8a00\u53e5\u5b50<\/strong>\uff0c\u62c6\u89e3\u6210\u6709\u9650\u7684\u57fa\u7840\u7ec4\u4ef6\uff08Token\uff09<\/strong>\uff0c\u5c31\u50cf\u642d\u4e50\u9ad8\uff1a<\/p>\n\n\n\n
                \n
              • \u4e50\u9ad8\u7684\u6838\u5fc3\uff1a\u7528 \u201c\u65b9\u5757\u3001\u957f\u677f\u3001\u8f6e\u5b50\u201d \u8fd9\u51e0\u767e\u79cd\u57fa\u7840\u96f6\u4ef6\uff0c\u80fd\u62fc\u51fa\u65e0\u9650\u79cd\u9020\u578b\uff08\u6c7d\u8f66\u3001\u623f\u5b50\u3001\u673a\u5668\u4eba\uff09\uff1b<\/li>\n\n\n\n
              • LLM \u7684\u6838\u5fc3\uff1a\u7528 \u201c\u4f60\u3001\u597d\u3001\u6625\u3001\u5929\u3001\u7684\u3001\u8bd7\u201d \u8fd9\u51e0\u4e07\u79cd\u57fa\u7840 Token\uff08\u6bd4\u5982 Llama 2 \u7684\u8bcd\u8868\u662f 32000 \u4e2a\uff09\uff0c\u80fd\u7ec4\u5408\u51fa\u65e0\u9650\u79cd\u53e5\u5b50\u3002<\/li>\n<\/ul>\n\n\n\n
                \u3000\u3000<\/p>\n\n\n\n
                \ud83d\udd39\u5177\u4f53\u6765\u8bf4\uff0c\u5206\u8bcd\u5e26\u6765 3 \u4e2a\u5173\u952e\u597d\u5904\uff1a<\/strong><\/p>\n\n\n\n
                1. \u964d\u4f4e\u8ba1\u7b97\u6210\u672c\uff08\u6838\u5fc3\u539f\u56e0\uff09<\/p>\n\n\n\n
                \u6a21\u578b\u53ea\u9700\u8981\u4e3a3-5 \u4e07\u4e2a Token<\/strong> \u5206\u522b\u8bad\u7ec3\u8bcd\u5d4c\u5165\u5411\u91cf\uff08\u6bd4\u5982 \u201c\u6625\u201d \u5bf9\u5e94\u4e00\u4e2a 4096 \u7ef4\u7684\u5411\u91cf\uff0c\u201c\u5929\u201d \u5bf9\u5e94\u53e6\u4e00\u4e2a\uff09\uff0c\u800c\u4e0d\u662f\u4e3a\u65e0\u9650\u53e5\u5b50\u8bad\u7ec3\u5411\u91cf\u3002<\/p>\n\n\n\n
                  \n
                • \u4e3e\u4f8b\uff1aLlama 2-7B \u7684\u8bcd\u5d4c\u5165\u77e9\u9635\u662f32000\u00d74096<\/code>\uff08\u7ea6 1.3 \u4ebf\u53c2\u6570\uff09\uff0c\u5982\u679c\u4e3a\u6574\u53e5\u8bdd\u8bad\u7ec3\uff0c\u8fd9\u4e2a\u77e9\u9635\u4f1a\u53d8\u6210 \u201c\u65e0\u9650 \u00d74096\u201d\uff0c\u5b8c\u5168\u4e0d\u53ef\u80fd\u5b9e\u73b0\u3002<\/li>\n<\/ul>\n\n\n\n
                  2. \u5b9e\u73b0\u8bed\u4e49\u6cdb\u5316<\/p>\n\n\n\n
                  Token \u662f \u201c\u8bed\u4e49\u6700\u5c0f\u5355\u5143\u201d\uff0c\u6a21\u578b\u80fd\u901a\u8fc7 Token \u7684\u7ec4\u5408\u7406\u89e3\u65b0\u53e5\u5b50\uff1a<\/p>\n\n\n\n
                    \n
                  • \u6a21\u578b\u5b66\u8fc7 \u201c\u6625\u201d\u201c\u5929\u201d\u201c\u7684\u201d\u201c\u8bd7\u201d \u8fd9\u4e9b Token \u7684\u8bed\u4e49\uff0c\u5f53\u5b83\u7b2c\u4e00\u6b21\u770b\u5230 \u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u8bd7\u201d\uff0c\u80fd\u901a\u8fc7 Token \u7684\u7ec4\u5408\uff08\u201c\u6625 + \u5929 = \u6625\u5929\u201d\u201c\u6625\u5929 + \u7684 + \u8bd7 = \u6625\u5929\u7684\u8bd7\u201d\uff09\u7406\u89e3\u6574\u53e5\u8bdd\u7684\u610f\u601d\uff1b<\/li>\n\n\n\n
                  • \u54ea\u6015\u4f60\u8f93\u5165 \u201c\u6625\u5929\u7684\u8bd7\u5f88\u7f8e\u201d \u8fd9\u4e2a\u65b0\u53e5\u5b50\uff0c\u6a21\u578b\u4e5f\u80fd\u9760\u5df2\u5b66\u7684 Token \u7ec4\u5408\u51fa\u8bed\u4e49\uff0c\u4e0d\u7528\u91cd\u65b0\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n\n\n\n
                    3. \u9002\u914d\u6a21\u578b\u7684\u8ba1\u7b97\u903b\u8f91<\/p>\n\n\n\n
                    Transformer \u7684\u6ce8\u610f\u529b\u673a\u5236\u662f\u6309 Token \u7ef4\u5ea6\u8ba1\u7b97<\/strong>\u7684 \u2014\u2014 \u6a21\u578b\u9700\u8981\u4e3a\u6bcf\u4e2a Token \u8ba1\u7b97\u548c\u5176\u4ed6 Token \u7684\u5173\u8054\uff08\u6bd4\u5982 \u201c\u5b83\u201d \u5173\u6ce8 \u201c\u732b\u201d \u8fd8\u662f \u201c\u72d7\u201d\uff09\uff0c\u5982\u679c\u6ca1\u6709\u5206\u8bcd\uff0c\u6ce8\u610f\u529b\u673a\u5236\u5c31\u6ca1\u6709 \u201c\u8ba1\u7b97\u5355\u5143\u201d\uff0c\u65e0\u6cd5\u5b8c\u6210\u8bed\u4e49\u5173\u8054\u5206\u6790\u3002<\/p>\n\n\n\n
                    \u3000\u3000<\/p>\n\n\n\n
                    \ud83d\udd39\u901a\u4fd7\u7c7b\u6bd4\uff1a\u5206\u8bcd = \u628a\u6587\u7ae0\u62c6\u6210 \u201c\u5b57 \/ \u8bcd\u201d<\/strong><\/p>\n\n\n\n
                    \u4f60\u53ef\u4ee5\u628a LLM \u7684\u5206\u8bcd\u7c7b\u6bd4\u6210\u6211\u4eec\u8bfb\u4e66\u65f6\u7684 \u201c\u8bc6\u5b57\u8fc7\u7a0b\u201d\uff1a<\/p>\n\n\n\n
                      \n
                    • \u5a74\u513f\u5b66\u8bf4\u8bdd\uff1a\u5148\u8ba4 \u201c\u7238\u3001\u5988\u3001\u6c34\u3001\u996d\u201d \u8fd9\u4e9b\u5355\u5b57\uff08Token\uff09\uff0c\u518d\u5b66 \u201c\u7238\u7238\u3001\u5988\u5988\u3001\u559d\u6c34\u3001\u5403\u996d\u201d \u8fd9\u4e9b\u8bcd\uff0c\u6700\u540e\u624d\u80fd\u7ec4\u5408\u51fa \u201c\u7238\u7238\u559d\u6c34\u201d\u201c\u5988\u5988\u5403\u996d\u201d \u8fd9\u4e9b\u53e5\u5b50\uff1b<\/li>\n\n\n\n
                    • LLM \u5b66\u8bed\u8a00\uff1a\u5148\u5b66\u6bcf\u4e2a Token \u7684\u8bed\u4e49\uff08\u201c\u6625\u201d= \u6e29\u6696\u3001\u82b1\u5f00\uff0c\u201c\u8bd7\u201d= \u97f5\u5f8b\u3001\u6587\u5b57\uff09\uff0c\u518d\u5b66 Token \u7684\u7ec4\u5408\u89c4\u5219\uff0c\u6700\u540e\u624d\u80fd\u751f\u6210 \u201c\u5199\u4e00\u9996\u6625\u5929\u7684\u8bd7\u201d\u3002<\/li>\n<\/ul>\n\n\n\n
                      \u5982\u679c\u8df3\u8fc7\u5206\u8bcd\u76f4\u63a5\u5b66\u6574\u53e5\uff0c\u5c31\u50cf\u5a74\u513f\u76f4\u63a5\u5b66 \u201c\u7238\u7238\u559d\u6c34\u201d \u8fd9\u4e2a\u6574\u53e5\uff0c\u5374\u4e0d\u8ba4\u8bc6 \u201c\u7238\u3001\u7238\u3001\u559d\u3001\u6c34\u201d\u2014\u2014 \u6362\u4e00\u53e5 \u201c\u5988\u5988\u559d\u6c34\u201d \u5c31\u5b8c\u5168\u4e0d\u61c2\u4e86\u3002<\/p>\n\n\n\n
                      \u3000\u3000<\/p>\n\n\n\n
                      \ud83d\udd39\u603b\u7ed3<\/strong><\/p>\n\n\n\n
                        \n
                      • \u274c \u6574\u53e5\u8bdd\u5f53\u53c2\u6570\uff1a\u65e0\u9650\u7ec4\u5408\u3001\u65e0\u6cd5\u6cdb\u5316\u3001\u8ba1\u7b97\u7206\u70b8\uff0c\u5b8c\u5168\u4e0d\u53ef\u884c\uff1b<\/li>\n\n\n\n
                      • \u2705 \u5206\u8bcd\u5f53\u53c2\u6570\uff1a\u628a\u65e0\u9650\u53e5\u5b50\u62c6\u89e3\u6210\u6709\u9650 Token\uff0c\u964d\u4f4e\u8ba1\u7b97\u6210\u672c\uff0c\u8ba9\u6a21\u578b\u80fd\u7406\u89e3\u3001\u6cdb\u5316\u3001\u751f\u6210\u65e0\u9650\u7684\u81ea\u7136\u8bed\u8a00\u3002<\/li>\n<\/ul>\n\n\n\n
                        \u7b80\u5355\u8bf4\uff1a\u5206\u8bcd\u662f LLM \u80fd \u201c\u7528\u6709\u9650\u7684\u53c2\u6570\uff0c\u5904\u7406\u65e0\u9650\u7684\u8bed\u8a00\u201d \u7684\u6838\u5fc3\u524d\u63d0\u3002<\/p>\n\n\n\n
                        \u3000\u3000<\/p>\n\n\n\n
                        \u8bcd\u5d4c\u5165\uff08Embedding\uff09\uff1a\u8ba9\u7b26\u53f7\u62e5\u6709\u300c\u8bed\u4e49\u6e29\u5ea6\u300d\ud83d\udd16<\/strong><\/p>\n\n\n\n
                        \n\n\n\n
                        \u5728\u8bed\u8a00\u6a21\u578b\u4e2d\uff0c\u8ba1\u7b97\u673a\u6700\u521d\u770b\u5230\u7684\u53ea\u662f\u4e00\u4e2a\u4e2a\u79bb\u6563\u7684\u7b26\u53f7\uff08token\uff09\uff0c\u4f8b\u5982\u201c\u732b\u201d\u201c\u94f6\u884c\u201d\u201crun\u201d\u3002
                        \u8fd9\u4e9b\u7b26\u53f7\u5bf9\u8ba1\u7b97\u673a\u6765\u8bf4\u662f\u6ca1\u6709\u610f\u4e49\u7684\uff0c\u5b83\u4eec\u53ea\u662f\u6570\u5b57 ID\u3002\u800c Embedding \u5c42 \u7684\u4f5c\u7528\uff0c\u5c31\u662f\u8ba9\u8fd9\u4e9b\u51b0\u51b7\u7684\u7b26\u53f7\u62e5\u6709\u201c\u8bed\u4e49\u6e29\u5ea6\u201d\u2014\u2014\u4e5f\u5c31\u662f\uff0c\u628a\u5b83\u4eec\u6620\u5c04\u5230\u4e00\u4e2a\u53ef\u4ee5\u8ba1\u7b97\u8bed\u4e49\u8ddd\u79bb\u7684\u5411\u91cf\u7a7a\u95f4\u4e2d\u3002<\/p>\n\n\n\n
                        \ud83d\udd39\u4ece\u7b26\u53f7\u5230\u5411\u91cf<\/strong><\/p>\n\n\n\n
                        Embedding \u53ef\u4ee5\u7406\u89e3\u4e3a\u4e00\u4e2a\u201c\u8bcd\u5178\u77e9\u9635\u201d\uff1a
                        \u6bcf\u4e2a\u8bcd\uff08token\uff09\u90fd\u88ab\u8868\u793a\u4e3a\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff08\u5982 1024 \u7ef4\uff09\u3002
                        \u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u4f1a\u4e0d\u65ad\u8c03\u6574\u8fd9\u4e9b\u5411\u91cf\uff0c\u4f7f\u5f97\u8bed\u4e49\u76f8\u8fd1\u7684\u8bcd\u5728\u7a7a\u95f4\u4e2d\u9760\u5f97\u66f4\u8fd1\u3002<\/p>\n\n\n\n
                        \u4f8b\u5982\uff1a<\/p>\n\n\n\n
                        \u2022 \u201c\u732b\uff08cat\uff09\u201d\u548c\u201c\u72d7\uff08dog\uff09\u201d\u7684\u5411\u91cf\u8ddd\u79bb\u4f1a\u975e\u5e38\u63a5\u8fd1\uff1b
                        \u2022 \u800c\u201c\u732b\uff08cat\uff09\u201d\u548c\u201c\u91d1\u878d\uff08finance\uff09\u201d\u7684\u8ddd\u79bb\u5219\u4f1a\u5f88\u8fdc\u3002<\/pre>\n\n\n\n
                        \u4e8e\u662f\uff0c\u6a21\u578b\u4e0d\u4ec5\u201c\u770b\u5230\u201d\u4e86\u8bcd\u672c\u8eab\uff0c\u8fd8\u80fd\u201c\u611f\u77e5\u201d\u5b83\u4e0e\u5176\u4ed6\u8bcd\u7684\u8bed\u4e49\u5173\u7cfb\u3002
                        \u8fd9\u4e3a\u540e\u7eed\u7684\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09\u63d0\u4f9b\u4e86\u57fa\u7840\u7684\u201c\u8bed\u4e49\u51e0\u4f55\u7a7a\u95f4\u201d\u3002<\/p>\n\n\n\n
                        \u3000\u3000<\/p>\n\n\n\n
                        \ud83d\udd39\u5411\u91cf\u7a7a\u95f4\u7684\u8bed\u4e49\u89c4\u5f8b<\/strong><\/p>\n\n\n\n
                        \u5728\u8fd9\u4e2a\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\uff0c\u8bed\u4e49\u5173\u7cfb\u4f1a\u4ee5\u51e0\u4f55\u7ed3\u6784\u7684\u65b9\u5f0f\u81ea\u7136\u663e\u73b0\u3002
                        \u6bd4\u5982\u7ecf\u5178\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
                        king - man + woman \u2248 queen<\/pre>\n\n\n\n
                        \u8fd9\u8bf4\u660e\u6a21\u578b\u201c\u7406\u89e3\u201d\u4e86\u6027\u522b\u3001\u89d2\u8272\u7b49\u62bd\u8c61\u8bed\u4e49\u5dee\u5f02\u3002<\/p>\n\n\n
                        \n
                        \"\"<\/figure><\/div>\n\n\n
                        \u3000\u3000<\/p>\n\n\n\n
                        \ud83d\udd39\u8c01\u6765\u8bad\u7ec3\u8fd9\u4e9b\u5411\u91cf\uff1f<\/strong><\/p>\n\n\n\n
                        \u8bcd\u5d4c\u5165\u7684\u751f\u6210\u5e76\u4e0d\u662f\u624b\u5de5\u8bbe\u5b9a\u7684\uff0c\u800c\u662f\u901a\u8fc7\u5927\u89c4\u6a21\u65e0\u76d1\u7763\u5b66\u4e60\u81ea\u52a8\u83b7\u5f97\u7684\u3002
                        \u6700\u65e9\u7684\u6a21\u578b\u5982 word2vec\u3001GloVe \u4e13\u95e8\u7528\u6765\u8bad\u7ec3\u8bcd\u5411\u91cf\uff1b
                        \u800c\u5728\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\uff08\u5982 GPT\u3001Qwen\u3001DeepSeek\uff09\u4e2d\uff0cEmbedding \u5c42\u7684\u53c2\u6570\u4e0e\u6a21\u578b\u4e00\u540c\u8bad\u7ec3\uff0c\u5e76\u5728\u66f4\u5e9e\u5927\u7684\u8bed\u6599\u4e2d\u6301\u7eed\u66f4\u65b0\u3002<\/p>\n\n\n\n
                        \u4f8b\u5982\uff0cQwen-1.5 \u7cfb\u5217\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u5c31\u5305\u542b\u4e86\u4e0a\u767e\u4ebf\u6761\u8bed\u6599\uff0cEmbedding \u5c42\u4f1a\u968f\u7740\u6a21\u578b\u5b66\u4e60\u8bed\u4e49\u5173\u7cfb\u800c\u4e0d\u65ad\u6f14\u5316\uff0c\u4f7f\u5176\u80fd\u7cbe\u7ec6\u5730\u6355\u6349\u4e0d\u540c\u8bed\u8a00\u3001\u9886\u57df\u3001\u98ce\u683c\u7684\u8bed\u4e49\u7ec6\u8282\u3002<\/p>\n\n\n\n
                        \u6362\u53e5\u8bdd\u8bf4\uff1a<\/p>\n\n\n\n
                        \u8bcd\u5d4c\u5165\u662f\u8bed\u8a00\u6a21\u578b\u7684\u300c\u8bed\u4e49\u5730\u57fa\u300d\uff0c
                        \u5b83\u5c06\u79bb\u6563\u7684\u8bed\u8a00\u7b26\u53f7\u8f6c\u5316\u4e3a\u8fde\u7eed\u7684\u8bed\u4e49\u5750\u6807\uff0c
                        \u4e3a Transformer \u7406\u89e3\u3001\u63a8\u7406\u548c\u751f\u6210\u8bed\u8a00\u6253\u4e0b\u57fa\u7840\u3002<\/pre>\n\n\n\n
                        \u3000\u3000<\/p>\n\n\n\n
                        \u4f4d\u7f6e\u7f16\u7801\uff1a\u8d4b\u4e88\u8bed\u8a00\u300c\u987a\u5e8f\u611f\u300d\ud83d\udd16<\/strong><\/p>\n\n\n\n
                        \n\n\n\n
                        \u5230\u8fd9\u91cc\uff0c\u6a21\u578b\u77e5\u9053\u6bcf\u4e2a\u8bcd\u201c\u662f\u4ec0\u4e48\u201d\uff0c\u4f46\u8fd8\u4e0d\u77e5\u9053\u5b83\u201c\u5728\u53e5\u5b50\u91cc\u7684\u4f4d\u7f6e\u201d\u3002
                        \u4f8b\u5982\uff1a<\/p>\n\n\n\n
                        \u201c\u732b\u8ffd\u72d7\u201d vs \u201c\u72d7\u8ffd\u732b\u201d
                        \u8bcd\u76f8\u540c\uff0c\u4f46\u987a\u5e8f\u98a0\u5012\uff0c\u610f\u4e49\u5b8c\u5168\u4e0d\u540c\u3002<\/pre>\n\n\n\n
                        Transformer \u7ed3\u6784\u6ca1\u6709\u50cf RNN \u90a3\u6837\u7684\u201c\u65f6\u95f4\u8bb0\u5fc6\u201d\uff0c\u6240\u4ee5\u5fc5\u987b\u663e\u5f0f\u544a\u8bc9\u5b83\u8bcd\u8bed\u7684\u987a\u5e8f\u3002\u8fd9\u5c31\u662f Position Encoding\uff08\u4f4d\u7f6e\u7f16\u7801\uff09 \u7684\u4f5c\u7528\u3002<\/p>\n\n\n\n
                        \u5b83\u4f1a\u4e3a\u6bcf\u4e2a token \u589e\u52a0\u4e00\u4e2a\u4f4d\u7f6e\u5411\u91cf\uff0c\u6bd4\u5982\u7b2c\u4e00\u4e2a\u8bcd\u52a0 [sin(1), cos(1)]\uff0c\u7b2c\u4e8c\u4e2a\u52a0 [sin(2), cos(2)]\u2026\u2026
                        \u8fd9\u6837\u6a21\u578b\u5728\u5b66\u4e60\u65f6\u80fd\u533a\u5206\u51fa\uff1a<\/p>\n\n\n\n
                        \u201c\u524d\u540e\u987a\u5e8f\u4e0d\u540c \u2192 \u542b\u4e49\u4e5f\u4e0d\u540c\u201d\u3002<\/pre>\n\n\n\n
                        \u73b0\u5728\uff0c\u6a21\u578b\u62e5\u6709\u4e86\u8bed\u4e49\uff08Embedding\uff09+ \u987a\u5e8f\uff08Position\uff09\u4e24\u79cd\u57fa\u7840\u4fe1\u606f\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u8fdb\u5165\u771f\u6b63\u7684\u6838\u5fc3\u2014\u2014Transformer Block\u3002<\/p>\n\n\n\n
                        \u3000\u3000<\/p>\n\n\n\n
                        Transformer Block\uff1a\u8bed\u8a00\u7406\u89e3\u7684\u300c\u5fc3\u810f\u300d\ud83d\udd16<\/strong><\/p>\n\n\n\n
                        \n\n\n\n
                        \u6211\u4eec\u5e0c\u671b LLM \u4e0d\u4ec5\u662f\u7b80\u5355\u7684\u8bcd\u8bed\u62fc\u63a5\u673a\u5668\uff0c\u800c\u662f\u5177\u5907\u5f3a\u5927\u7684\u63a8\u7406\u548c\u8bed\u4e49\u7406\u89e3\u80fd\u529b\u7684\u201c\u8bed\u8a00\u5927\u8111\u201d\u3002 \u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u4ec5\u4ec5\u5806\u53e0\u795e\u7ecf\u5143\u662f\u4e0d\u591f\u7684\u3002Transformer Block \u4f5c\u4e3a LLM \u7684\u57fa\u672c\u8ba1\u7b97\u5355\u5143\uff0c\u5fc5\u987b\u5305\u542b\u4e00\u5957\u7cbe\u5de7\u7684\u534f\u4f5c\u673a\u5236\uff0c\u624d\u80fd\u6709\u6548\u5730\u5b8c\u6210\u4ee5\u4e0b\u5173\u952e\u4efb\u52a1\uff1a<\/p>\n\n\n\n
                          \n
                        • \u6355\u6349\u4e0a\u4e0b\u6587\u7684\u52a8\u6001\u5173\u7cfb\uff08\u89e3\u51b3\u591a\u4e49\u8bcd\u6b67\u4e49\uff09\u3002<\/li>\n\n\n\n
                        • \u540c\u65f6\u4ece\u591a\u4e2a\u7ef4\u5ea6\u89e3\u6790\u8bed\u8a00\uff08\u89e3\u51b3\u590d\u6742\u6027\u548c\u591a\u91cd\u542b\u4e49\uff09\u3002<\/li>\n\n\n\n
                        • \u786e\u4fdd\u751f\u6210\u8fc7\u7a0b\u7684\u903b\u8f91\u548c\u65f6\u95f4\u6b63\u786e\u6027\uff08\u89e3\u51b3\u81ea\u56de\u5f52\u751f\u6210\u95ee\u9898\uff09\u3002<\/li>\n\n\n\n
                        • \u5bf9\u62bd\u8c61\u7279\u5f81\u8fdb\u884c\u6df1\u5ea6\u52a0\u5de5\u548c\u63d0\u70bc\uff08\u89e3\u51b3\u9ad8\u5c42\u8bed\u4e49\u7406\u89e3\uff09\u3002<\/li>\n<\/ul>\n\n\n\n
                          \u63a5\u4e0b\u6765\u7684\u5404\u4e2a\u7ec4\u4ef6\uff08\u81ea\u6ce8\u610f\u529b\u3001\u591a\u5934\u3001\u63a9\u7801\u3001\u524d\u9988\u7f51\u7edc\u7b49\uff09\uff0c\u6b63\u662f\u4e3a\u4e86\u534f\u540c\u5b8c\u6210\u4e0a\u8ff0\u590d\u6742\u4efb\u52a1\u800c\u8bbe\u8ba1\uff0c\u5b83\u4eec\u5171\u540c\u6784\u6210\u4e86\u771f\u6b63\u52a0\u5de5\u8bed\u8a00\u7684\u201c\u8f66\u95f4\u201d\u7684\u5185\u90e8\u6d41\u7a0b\u3002<\/p>\n\n\n\n
                          \u4e00\u4e2a LLM \u901a\u5e38\u7531\u51e0\u5341\u4e0a\u767e\u5c42\uff08\u6bd4\u5982 GPT-4 \u7ea6 120 \u5c42<\/strong>\uff09\u8fd9\u6837\u7684 Block \u5806\u53e0\u800c\u6210\u3002\u6bcf\u4e00\u5c42\u90fd\u8d1f\u8d23\u5bf9\u8bed\u8a00\u7684\u7406\u89e3\u66f4\u8fdb\u4e00\u6b65\u3002<\/p>\n\n\n\n
                          \u3000\u3000<\/p>\n\n\n\n
                          \ud83d\udd39\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09\uff1a\u7406\u89e3\u4e0a\u4e0b\u6587\u7684\u7075\u9b42\u673a\u5236<\/strong><\/p>\n\n\n\n
                          \u6838\u5fc3\u601d\u60f3\u662f\uff1a<\/p>\n\n\n\n
                          \u201c\u4e00\u4e2a\u8bcd\u7684\u610f\u4e49\uff0c\u53d6\u51b3\u4e8e\u5b83\u4e0e\u4e0a\u4e0b\u6587\u7684\u5173\u7cfb\u3002\u201d<\/pre>\n\n\n\n
                          \u4f8b\u5982\u8fd9\u53e5\u8bdd\uff1a<\/p>\n\n\n\n
                          \u201c\u5979\u628a\u5361\u91cc\u7684\u94b1\u5b58\u5230\u4e86A\u884c\u3002\u201d<\/pre>\n\n\n\n
                          \u8fd9\u91cc\u7684\u201cA\u884c\u201d\uff0c\u5230\u5e95\u6307\u7684\u662f\u94f6\u884c<\/strong>\u8fd8\u662f\u516c\u53f8\u5206\u884c<\/strong>\uff1f\u6a21\u578b\u4e0d\u80fd\u51ed\u201c\u884c\u201d\u8fd9\u4e2a\u8bcd\u672c\u8eab\u53bb\u5224\u65ad\uff0c\u5b83\u8981\u6839\u636e\u4e0a\u4e0b\u6587\u7684\u7ebf\u7d22\u2014\u2014\u201c\u5361\u91cc\u7684\u94b1\u201d\u201c\u5b58\u5230\u201d\u2014\u2014\u53bb\u7406\u89e3\u8bed\u4e49\u3002<\/p>\n\n\n\n
                          Self-Attention \u5c31\u662f\u8ba9\u6a21\u578b\u5b66\u4f1a\u8fd9\u6837\u201c\u73af\u987e\u56db\u5468\u201d\u3002
                          \u6bcf\u4e2a\u8bcd\u5728\u8fd9\u4e00\u5c42\u90fd\u4f1a\u751f\u6210\u4e09\u4e2a\u5411\u91cf\uff1a<\/p>\n\n\n\n
                            \n
                          • Query\uff08\u67e5\u8be2\uff09\uff1a\u8868\u793a\u201c\u6211\u73b0\u5728\u60f3\u627e\u548c\u6211\u6709\u5173\u7684\u4fe1\u606f\u201d\uff1b<\/li>\n\n\n\n
                          • Key\uff08\u952e\uff09\uff1a\u8868\u793a\u201c\u6211\u80fd\u88ab\u522b\u4eba\u5173\u8054\u5230\u54ea\u4e9b\u4fe1\u606f\u201d\uff1b<\/li>\n\n\n\n
                          • Value\uff08\u503c\uff09\uff1a\u8868\u793a\u201c\u5982\u679c\u522b\u4eba\u627e\u5230\u6211\uff0c\u6211\u80fd\u63d0\u4f9b\u4ec0\u4e48\u4fe1\u606f\u201d\u3002<\/li>\n<\/ul>\n\n\n\n
                            \u4f60\u53ef\u4ee5\u628a\u5b83\u60f3\u8c61\u6210\u4e00\u4e2a\u4f1a\u8bae\u573a\u666f\uff1a<\/p>\n\n\n\n
                            \u6bcf\u4e2a\u8bcd\uff08\u4e0e\u4f1a\u8005\uff09\u90fd\u6709\u4e00\u4e2a\u63d0\u95ee\u5361\uff08Query\uff09\uff0c\u4e00\u5f20\u540d\u7247\uff08Key\uff09\uff0c\u548c\u4e00\u4efd\u8d44\u6599\uff08Value\uff09\u3002
                            \u5f53\u201c\u884c\u201d\u8fd9\u4e2a\u8bcd\u60f3\u5f04\u6e05\u81ea\u5df1\u662f\u4ec0\u4e48\u610f\u601d\u65f6\uff0c\u5b83\u4f1a\u628a\u81ea\u5df1\u7684 Query \u62ff\u53bb\u548c\u6240\u6709\u8bcd\u7684 Key \u5bf9\u6bd4\u3002
                            \u76f8\u4f3c\u5ea6\u8d8a\u9ad8\uff0c\u8bf4\u660e\u5bf9\u65b9\u63d0\u4f9b\u7684\u4fe1\u606f\u8d8a\u76f8\u5173\uff0c\u5c31\u4f1a\u7ed9\u4e88\u66f4\u9ad8\u7684\u201c\u6ce8\u610f\u529b\u6743\u91cd\u201d\u3002
                            \u6700\u540e\uff0c\u201c\u884c\u201d\u4f1a\u6839\u636e\u8fd9\u4e9b\u6743\u91cd\uff0c\u628a\u522b\u4eba\u7684 Value \u6309\u6bd4\u4f8b\u6574\u5408\u8d77\u6765\uff0c
                            \u4e8e\u662f\u5b83\u5bf9\u81ea\u5df1\u7684\u7406\u89e3\u53d8\u5f97\u66f4\u6e05\u6670\u2014\u2014\u5728\u8fd9\u4e2a\u8bed\u5883\u4e0b\uff0c\u5b83\u5c31\u662f\u201c\u94f6\u884c\u201d\u7684\u201c\u884c\u201d\u3002<\/pre>\n\n\n
                            \n
                            \"\"<\/figure><\/div>\n\n\n
                            \u3000\u3000<\/p>\n\n\n\n
                            \ud83d\udd39\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff08Multi-Head Attention\uff09\uff1a\u4ece\u591a\u4e2a\u89c6\u89d2\u7406\u89e3\u8bed\u8a00<\/strong><\/p>\n\n\n\n
                            \u4ec5\u9760\u4e00\u4e2a\u6ce8\u610f\u529b\u901a\u9053\u662f\u4e0d\u591f\u7684\u3002<\/span>\u8bed\u8a00\u4e2d\u7684\u8054\u7cfb\u662f\u591a\u7ef4\u7684\u2014\u2014\u6709\u7684\u8bcd\u5728\u8bed\u6cd5\u4e0a\u76f8\u5173\uff0c\u6709\u7684\u5728\u8bed\u4e49\u4e0a\u547c\u5e94\u3002<\/span>\u4e3a\u4e86\u540c\u65f6\u6355\u6349\u8fd9\u4e9b\u4e0d\u540c\u5c42\u9762\u7684\u4fe1\u606f\uff0c<\/span>Transformer \u5f15\u5165\u4e86 <\/span>\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff08Multi-Head Attention\uff09<\/span>\u3002<\/span><\/p>\n\n\n\n
                            \u6a21\u578b\u4f1a\u5e76\u884c\u5730\u5efa\u7acb\u591a\u4e2a\u72ec\u7acb\u7684\u6ce8\u610f\u529b\u901a\u9053\uff0c\u6bcf\u4e2a\u201c\u5934\u201d\u90fd\u4e13\u6ce8\u4e8e\u4e0d\u540c\u7684\u8bed\u8a00\u5173\u7cfb\u3002<\/pre>\n\n\n\n
                            \u53ef\u4ee5\u628a\u5b83\u60f3\u8c61\u6210\u4e00\u4e2a\u8ba8\u8bba\u7ec4\uff1a\u6bcf\u4e2a\u6210\u5458\u4ece\u81ea\u5df1\u7684\u89d2\u5ea6\u9605\u8bfb\u53e5\u5b50\uff0c\u6700\u540e\u518d\u7efc\u5408\u610f\u89c1\u3002<\/p>\n\n\n\n
                            \u4f8b\u5982\uff1a<\/p>\n\n\n\n
                              \n
                            • \u7b2c\u4e00\u4e2a\u5934\u5173\u6ce8\u4e3b\u8bed\u548c\u52a8\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb<\/li>\n\n\n\n
                            • \u7b2c\u4e8c\u4e2a\u5934\u5173\u6ce8\u65f6\u95f4\u526f\u8bcd\u548c\u52a8\u4f5c\u4e4b\u95f4\u7684\u8054\u7cfb<\/li>\n\n\n\n
                            • \u7b2c\u4e09\u4e2a\u5934\u53ef\u80fd\u4e13\u95e8\u6355\u6349\u60c5\u611f\u8272\u5f69\u6216\u8bed\u6c14\u53d8\u5316<\/li>\n<\/ul>\n\n\n\n
                              \u6240\u6709\u8fd9\u4e9b\u201c\u89c6\u89d2\u201d\u6700\u540e\u88ab\u878d\u5408\u5728\u4e00\u8d77\uff0c\u5f62\u6210\u4e86\u5bf9\u6574\u53e5\u8bdd\u66f4\u7acb\u4f53\u3001\u66f4\u5168\u9762\u7684\u7406\u89e3\u3002\u8fd9\u5c31\u662f\u591a\u5934\u6ce8\u610f\u529b\u7684\u610f\u4e49\u6240\u5728\u3002<\/p>\n\n\n\n
                              \u3000\u3000<\/p>\n\n\n\n
                              \ud83d\udd39\u5e26\u63a9\u7801\u7684\u591a\u5934\u6ce8\u610f\u529b\uff08Masked Multi-Head Attention\uff09\uff1a\u9632\u6b62\u201c\u5077\u770b\u672a\u6765\u201d<\/strong><\/p>\n\n\n\n
                              \u5728\u8bed\u8a00\u751f\u6210\uff08Decoder\uff09\u9636\u6bb5\uff0c\u6a21\u578b\u5fc5\u987b\u4e00\u8fb9\u751f\u6210\u3001\u4e00\u8fb9\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u3002\u5b83\u4e0d\u80fd\u201c\u5077\u770b\u201d\u672a\u6765\u8fd8\u6ca1\u751f\u6210\u7684\u90e8\u5206\u3002<\/p>\n\n\n\n
                              \u4e3a\u6b64\uff0c\u5f15\u5165\u4e86 Mask\uff08\u63a9\u7801\uff09\u673a\u5236\uff1a<\/p>\n\n\n\n
                              \u5bf9\u5f53\u524d\u4f4d\u7f6e\u4e4b\u540e\u7684\u8bcd\uff0c\u6a21\u578b\u4f1a\u81ea\u52a8\u5c4f\u853d\uff0c\u4e0d\u53bb\u8ba1\u7b97\u5b83\u4eec\u7684\u6ce8\u610f\u529b\u3002<\/pre>\n\n\n\n
                              \u8fd9\u6837\uff0c\u6a21\u578b\u5728\u751f\u6210\u53e5\u5b50\u65f6\uff0c\u53ea\u80fd\u53c2\u8003\u5df2\u7ecf\u751f\u6210\u7684\u5185\u5bb9\uff0c\u5c31\u50cf\u4eba\u7c7b\u5199\u4f5c\u65f6\u4e00\u8fb9\u60f3\u4e00\u8fb9\u5199\uff0c\u4fdd\u8bc1\u8bed\u8a00\u7684\u65f6\u95f4\u987a\u5e8f\u662f\u6b63\u786e\u7684\u3002<\/p>\n\n\n\n
                              \u5728Transformer \u6a21\u578b\u7684Decoder \u4e2d\u51fa\u73b0\u5728Masked Multi-Head Attention\u3002\u4e3a\u4ec0\u4e48\u9700\u8981masked,\u201c\u63a9\u7801\u201d\u7684\u76ee\u7684\u662f\u5728\u8ba1\u7b97\u6ce8\u610f\u529b\u65f6\uff0c\u963b\u6b62\u6a21\u578b\u201c\u770b\u5230\u201d\u67d0\u4e9b\u4e0d\u5e94\u8be5\u770b\u5230\u7684\u4fe1\u606f\u3002\u5728\u7ed9\u51fa\u7b2c\u4e00\u4e2a\u8bcd\u7684\u57fa\u7840\u4e0a\uff0c\u53bb\u731c\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u662f\u4ec0\u4e48\u3002\u4e0d\u6dfb\u52a0\u63a9\u7801\uff0c\u53ef\u4ee5\u8bf4\u662f\u770b\u7740\u7b54\u6848\u7ed3\u8bba\u7684\u4e00\u4e2a\u64cd\u4f5c\u3002<\/p>\n\n\n\n
                              \u8fd9\u662f\u751f\u6210\u5f0f\u6a21\u578b\uff08\u5982GPT\uff09\u7684\u5173\u952e<\/strong>\u3002\u5728\u8bad\u7ec3\u6216\u751f\u6210\u65f6\uff0c\u89e3\u7801\u5668\u5e94\u8be5\u53ea\u80fd\u201c\u770b\u5230\u201d\u5f53\u524d\u8bcd\u53ca\u4e4b\u524d<\/strong>\u7684\u8bcd\uff0c\u800c\u4e0d\u80fd\u201c\u5077\u770b\u201d\u672a\u6765\u7684\u8bcd\u3002\u8fd9\u786e\u4fdd\u4e86\u81ea\u56de\u5f52\u751f\u6210\u7684\u7279\u6027\u3002<\/p>\n\n\n\n
                              \u4ee5\u7ffb\u8bd1\u4e3a\u4f8b\uff1a<\/p>\n\n\n
                              \n
                              \"\"<\/figure><\/div>\n\n\n
                              Decoder \u7684\u8f93\u5165\u77e9\u9635\u548c Mask \u77e9\u9635, \u8f93\u5165\u77e9\u9635\u5305\u542b \u201c I have a cat\u201d (0, 1, 2, 3, 4) \u4e94\u4e2a\u5355\u8bcd\u7684\u8868\u793a\u5411\u91cf, Mask \u662f\u4e00\u4e2a 5×5 \u7684\u77e9\u9635. \u5728 Mask \u53ef\u4ee5\u53d1\u73b0\u5355\u8bcd 0 \u53ea\u80fd\u4f7f\u7528\u5355\u8bcd 0 \u7684\u4fe1\u606f, \u800c\u5355\u8bcd 1 \u53ef\u4ee5\u4f7f\u7528\u5355\u8bcd 0, 1 \u7684\u4fe1\u606f, \u5373\u53ea\u80fd\u4f7f\u7528\u4e4b\u524d\u7684\u4fe1\u606f\u3002<\/p>\n\n\n
                              \n
                              \"\"<\/figure><\/div>\n\n\n
                              \u3000\u3000<\/p>\n\n\n\n
                              \ud83d\udd39\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08Feed Forward\uff09\uff1a\u63d0\u53d6\u66f4\u9ad8\u5c42\u8bed\u4e49<\/strong><\/p>\n\n\n\n
                              Self-Attention \u5e2e\u6a21\u578b\u7406\u89e3\u4e86\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u4f46\u6bcf\u4e2a\u8bcd\u81ea\u8eab\u7684\u7279\u5f81\u4e5f\u9700\u8981\u8fdb\u4e00\u6b65\u52a0\u5de5\u3002\u8fd9\u65f6\u5c31\u8f6e\u5230 \u524d\u9988\u7f51\u7edc\uff08Feed Forward Network\uff09 \u51fa\u573a\u4e86\u3002<\/p>\n\n\n\n
                              \u5b83\u76f8\u5f53\u4e8e\u7ed9\u6bcf\u4e2a\u8bcd\u5355\u72ec\u914d\u5907\u4e00\u6bb5\u5c0f\u578b\u7684\u201c\u601d\u8003\u56de\u8def\u201d\uff0c\u5e2e\u52a9\u6a21\u578b\u8fdb\u4e00\u6b65\u63d0\u70bc\u548c\u91cd\u7ec4\u8bed\u4e49\u3002
                              \u7ecf\u8fc7\u8fd9\u4e00\u6b65\uff0c\u6a21\u578b\u80fd\u6355\u6349\u5230\u66f4\u6df1\u5c42\u6b21\u3001\u66f4\u62bd\u8c61\u7684\u7279\u5f81\uff0c\u8ba9\u8bed\u8a00\u7406\u89e3\u66f4\u52a0\u7ec6\u817b\u3002<\/p>\n\n\n\n
                              \u3000\u3000<\/p>\n\n\n\n
                              \ud83d\udd39\u6b8b\u5dee\u8fde\u63a5\u4e0e\u5c42\u5f52\u4e00\u5316\uff1a\u8ba9\u5b66\u4e60\u66f4\u7a33\u5b9a<\/strong><\/p>\n\n\n\n
                              Transformer \u7684\u6bcf\u4e00\u5c42\u8fd8\u5305\u542b\u4e24\u9879\u5173\u952e\u8bbe\u8ba1\uff1a
                              \u6b8b\u5dee\u8fde\u63a5\uff08Residual Connection\uff09 \u548c \u5c42\u5f52\u4e00\u5316\uff08Layer Normalization\uff09\u3002<\/p>\n\n\n\n
                                \n
                              • \u6b8b\u5dee\u8fde\u63a5\u8ba9\u8f93\u5165\u4fe1\u606f\u80fd\u591f\u7ed5\u8fc7\u5f53\u524d\u5c42\u76f4\u63a5\u4f20\u9012\u4e0b\u53bb\uff0c\u9632\u6b62\u7f51\u7edc\u592a\u6df1\u5bfc\u81f4\u201c\u8bb0\u4e0d\u4f4f\u201d\u539f\u59cb\u4fe1\u606f\u3002<\/li>\n\n\n\n
                              • \u5c42\u5f52\u4e00\u5316\u5219\u8ba9\u6bcf\u5c42\u7684\u6570\u503c\u4fdd\u6301\u7a33\u5b9a\uff0c\u907f\u514d\u8bad\u7ec3\u8fc7\u7a0b\u53d1\u6563\u6216\u9707\u8361\u3002\u901a\u8fc7\u8fd9\u4e24\u4e2a\u673a\u5236\uff0cTransformer \u53ef\u4ee5\u7a33\u5b9a\u5730\u5806\u53e0\u51e0\u5341\u5c42\u751a\u81f3\u4e0a\u767e\u5c42\uff0c\u4ecd\u7136\u4fdd\u6301\u4fe1\u606f\u6d41\u52a8\u987a\u7545\u3002<\/li>\n<\/ul>\n\n\n\n
                                \u3000\u3000<\/p>\n\n\n\n
                                 \ud83d\udd39\u5c0f\u7ed3<\/strong><\/p>\n\n\n\n
                                Transformer Block \u7684\u5185\u90e8\u7ed3\u6784\uff0c\u5c31\u50cf\u4e00\u53f0\u201c\u8bed\u8a00\u7406\u89e3\u5f15\u64ce\u201d\uff1a<\/p>\n\n\n\n
                                  \n
                                • Self-Attention\uff1a\u7406\u89e3\u4e0a\u4e0b\u6587\uff0c\u5efa\u7acb\u8bcd\u4e0e\u8bcd\u7684\u5173\u7cfb\uff1b<\/li>\n\n\n\n
                                • Multi-Head\uff1a\u591a\u89d2\u5ea6\u5e76\u884c\u601d\u8003\uff1b<\/li>\n\n\n\n
                                • Mask\uff1a\u4fdd\u8bc1\u751f\u6210\u65f6\u4e0d\u201c\u5077\u770b\u672a\u6765\u201d\uff1b<\/li>\n\n\n\n
                                • Feed Forward\uff1a\u63d0\u70bc\u66f4\u9ad8\u5c42\u8bed\u4e49\uff1b<\/li>\n\n\n\n
                                • \u6b8b\u5dee\u4e0e\u5f52\u4e00\u5316\uff1a\u4fdd\u8bc1\u6df1\u5c42\u7f51\u7edc\u7a33\u5b9a\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n\n\n\n
                                  \u5f53\u8fd9\u6837\u7684 Block \u5c42\u5c42\u5806\u53e0\u65f6\uff0c\u6a21\u578b\u5c31\u80fd\u4ece\u6d45\u5c42\u8bed\u6cd5\u3001\u5230\u6df1\u5c42\u8bed\u4e49\u3001\u518d\u5230\u903b\u8f91\u63a8\u7406\uff0c\u4e00\u6b65\u6b65\u6784\u5efa\u51fa\u5bf9\u8bed\u8a00\u7684\u6df1\u5ea6\u7406\u89e3\u3002<\/p>\n\n\n\n
                                  \u3000\u3000<\/p>\n\n\n\n
                                  \u8f93\u51fa\u5c42\uff1a\u9884\u6d4b\u4e0b\u4e00\u4e2a Token \u7684\u6982\u7387\ud83d\udd16<\/strong><\/p>\n\n\n\n
                                  \n\n\n\n
                                  \u6240\u6709 Transformer \u5c42\u7684\u8f93\u51fa\u6700\u7ec8\u6c47\u603b\u6210\u4e00\u4e2a\u9690\u85cf\u5411\u91cf\uff0c\u8fd9\u4e2a\u5411\u91cf\u4f1a\u88ab\u6295\u5f71\uff08Linear Projection\uff09\u56de\u5230\u8bcd\u8868\u7ef4\u5ea6\uff0c\u5f97\u5230\u6bcf\u4e2a\u53ef\u80fd\u8bcd\u7684\u6982\u7387\u5206\u5e03\uff1a<\/p>\n\n\n\n
                                  \u201c\u6211\u4eca\u5929\u5f88 \u2014\u2014\u201d
                                  \u2192 \u9ad8\u6982\u7387\u8f93\u51fa\uff1a\u201c\u5f00\u5fc3\u201d\u3001\u201c\u7d2f\u201d\u3001\u201c\u5fd9\u201d\u2026\u2026
                                  \u2192 \u4f4e\u6982\u7387\u8f93\u51fa\uff1a\u201c\u9999\u8549\u201d\u3001\u201c\u6c7d\u8f66\u201d\u3002<\/pre>\n\n\n\n
                                  \u6a21\u578b\u901a\u8fc7 Softmax \u5c06\u8fd9\u4e9b\u5206\u6570\u8f6c\u6210\u6982\u7387\uff0c\u4ece\u4e2d\u91c7\u6837\u6216\u9009\u53d6\u6700\u9ad8\u503c\uff0c\u5c31\u751f\u6210\u4e0b\u4e00\u4e2a\u8bcd\u3002
                                  \u8fd9\u5c31\u662f\u8bed\u8a00\u751f\u6210\u7684\u57fa\u672c\u8fc7\u7a0b\u3002<\/p>\n\n\n\n
                                  \u3000\u3000<\/p>\n\n\n\n
                                  \u6a21\u578b\u7684\u53c2\u6570\uff1a\u54ea\u4e9b\u4f1a\u88ab\u8fed\u4ee3\uff1f\ud83d\udd16<\/strong><\/p>\n\n\n\n
                                  \n\n\n\n
                                  \u5728\u6574\u4e2a\u6a21\u578b\u7ed3\u6784\u4e2d\uff0c\u4ee5\u4e0b\u90e8\u5206\u7684\u53c2\u6570\u4f1a\u5728\u8bad\u7ec3\u4e2d\u4e0d\u65ad\u66f4\u65b0\uff1a<\/p>\n\n\n\n
                                  \u6a21\u5757<\/th>\u53c2\u6570\u7c7b\u578b<\/th><\/tr><\/thead>
                                  Embedding \u5c42<\/td>\u6bcf\u4e2a token \u7684\u8bcd\u5411\u91cf\u77e9\u9635<\/td><\/tr>
                                  Attention \u5c42<\/td>Q, K, V, O \u56db\u4e2a\u6295\u5f71\u77e9\u9635<\/td><\/tr>
                                  Feed Forward \u5c42<\/td>\u4e24\u5c42\u7ebf\u6027\u6743\u91cd\u77e9\u9635<\/td><\/tr>
                                  \u5c42\u5f52\u4e00\u5316<\/td>\u7f29\u653e\u548c\u504f\u79fb\u53c2\u6570<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                                  \u8fd9\u4e9b\u53c2\u6570\u7684\u6570\u91cf\u5f80\u5f80\u4ee5 \u767e\u4ebf\u751a\u81f3\u4e07\u4ebf<\/strong> \u8ba1\uff0c
                                  \u8bad\u7ec3\u7684\u8fc7\u7a0b\uff0c\u5c31\u662f\u901a\u8fc7\u68af\u5ea6\u4e0b\u964d\u4e0d\u65ad\u5fae\u8c03\u8fd9\u4e9b\u53c2\u6570\uff0c\u8ba9\u6a21\u578b\u201c\u66f4\u61c2\u8bed\u8a00\u201d\u3002<\/p>\n\n\n\n
                                  \u3000\u3000<\/p>\n\n\n\n
                                  \u603b\u7ed3\uff1a\u8bed\u8a00\u7684\u673a\u5668\u5fc3\u810f\ud83d\udd16<\/strong><\/p>\n\n\n\n
                                  \n\n\n\n
                                  \u6211\u4eec\u53ef\u4ee5\u7528\u4e00\u53e5\u8bdd\u6982\u62ec\u6574\u4e2a\u903b\u8f91\uff1a<\/p>\n\n\n\n
                                    \n
                                  • Tokenization \u8ba9\u6587\u5b57\u88ab\u770b\u89c1<\/li>\n\n\n\n
                                  • Embedding \u8ba9\u6587\u5b57\u6709\u610f\u4e49<\/li>\n\n\n\n
                                  • Position \u8ba9\u8bed\u8a00\u6709\u987a\u5e8f<\/li>\n\n\n\n
                                  • Transformer \u8ba9\u8bed\u4e49\u88ab\u7406\u89e3<\/li>\n\n\n\n
                                  • Output \u8ba9\u601d\u60f3\u5f97\u4ee5\u8868\u8fbe<\/li>\n<\/ul>\n\n\n\n
                                    \u8fd9\uff0c\u5c31\u662f\u4e00\u4e2a LLM \u7684\u5b8c\u6574\u751f\u547d\u7ebf\u3002\u5b83\u5e76\u975e\u9b54\u6cd5\uff0c\u800c\u662f\u6570\u5b66\u4e0e\u7ed3\u6784\u7684\u5947\u8ff9\u3002\u5728\u65e0\u6570\u5c42\u77e9\u9635\u8ba1\u7b97\u7684\u80cc\u540e\uff0c\u662f\u4eba\u7c7b\u8ba9\u8bed\u8a00\u81ea\u6211\u7406\u89e3\u7684\u58ee\u4e3e\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
                                    \u8f6c\u8f7d\uff1a\u5c0f\u767d\u5fc5\u770b\uff1a\u5927\u6a21\u578b\u5185\u90e8\u7ed3\u6784\u63ed\u79d8\u2014\u2014\u4eceToken\u5230Transformer\u7684\u5b8c\u6574\u6307\u5357\uff01 \u524d\u8a00\ud83d\udd16 \u5728 ChatG […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[48],"tags":[],"_links":{"self":[{"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts\/19956"}],"collection":[{"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=19956"}],"version-history":[{"count":14,"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts\/19956\/revisions"}],"predecessor-version":[{"id":19979,"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts\/19956\/revisions\/19979"}],"wp:attachment":[{"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=19956"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=19956"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=19956"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}