{"id":137,"date":"2026-04-29T17:55:44","date_gmt":"2026-04-29T09:55:44","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=137"},"modified":"2026-04-23T17:56:01","modified_gmt":"2026-04-23T09:56:01","slug":"%e6%ae%8b%e5%b7%ae%e8%bf%9e%e6%8e%a5%e3%80%81pre-norm%e4%b8%8epost-norm","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/29\/%e6%ae%8b%e5%b7%ae%e8%bf%9e%e6%8e%a5%e3%80%81pre-norm%e4%b8%8epost-norm\/","title":{"rendered":"\u6b8b\u5dee\u8fde\u63a5\u3001Pre-Norm\u4e0ePost-Norm"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u524d\u51e0\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u50cf\u62c6\u89e3\u4e00\u4e2a\u7cbe\u5bc6\u7684\u673a\u68b0\u8868\u4e00\u6837\uff0c\u4e00\u6b65\u6b65\u89e3\u5256\u4e86 LLM \u7684\u5404\u4e2a\u6838\u5fc3\u7ec4\u4ef6\uff1a\u4ece Tokenizer \u5230 Embedding\uff0c\u4ece Attention \u5230 FFN\u3002\u6211\u4eec\u770b\u7740\u6570\u636e\u5728\u6a21\u578b\u4e2d\u6e38\u8d70\uff0cToken \u4e4b\u95f4\u4e0d\u65ad\u4ea4\u6d41\uff0c\u77e5\u8bc6\u88ab\u4e00\u5c42\u5c42\u52a0\u5de5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u8fd9\u4e2a\u6d41\u7a0b\u4e2d\u6709\u4e00\u4e2a\u5171\u540c\u70b9\uff1a<strong>\u6bcf\u4e2a\u6838\u5fc3\u6a21\u5757\uff08Attention \u548c FFN\uff09\u90fd\u88ab\u4e00\u5c42\u201c\u795e\u5947\u7684\u5149\u73af\u201d\u6240\u5305\u88f9\uff0c\u8fd9\u5149\u73af\u5c31\u662f\u6b8b\u5dee\u8fde\u63a5\u4e0e\u5c42\u5f52\u4e00\u5316<\/strong>\u3002\u5b83\u4eec\u4e0d\u521b\u9020\u65b0\u7684\u5185\u5bb9\uff0c\u4e5f\u4e0d\u6539\u53d8\u4fe1\u606f\u7684\u65b9\u5411\uff0c\u53ea\u662f\u9ed8\u9ed8\u5730\u5728\u5e55\u540e\u5de5\u4f5c\uff0c\u7ef4\u6301\u7740\u6a21\u578b\u8fd9\u4e2a\u5e9e\u7136\u5927\u7269\u7684\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u800c\uff0c\u5c31\u662f\u8fd9\u6837\u4e00\u4e2a\u770b\u4f3c\u5fae\u5c0f\u3001\u901a\u5e38\u53ea\u9700\u51e0\u884c\u4ee3\u7801\u5b9a\u4e49\u7684\u7ec4\u4ef6\uff0c\u5176<strong>\u4f4d\u7f6e\u9009\u62e9<\/strong>\u2014\u2014\u5e94\u8be5\u653e\u5728\u54ea\u91cc\uff1f\u2014\u2014\u5374\u5728\u8fc7\u53bb\u51e0\u5e74\u91cc\u5f15\u53d1\u4e86\u5927\u89c4\u6a21\u7684\u201c\u67b6\u6784\u4e4b\u4e89\u201d\u3002\u8fd9\u76f4\u63a5\u51b3\u5b9a\u4e86\u6211\u4eec\u80fd\u5426\u8ba9 Transformer \u4ece\u51e0\u5341\u5c42\uff0c\u7a33\u5b9a\u5730\u6269\u5c55\u5230\u6570\u767e\u751a\u81f3\u4e0a\u5343\u5c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4eca\u5929\uff0c\u6211\u4eec\u5c06\u63ed\u5f00\u8fd9\u5c42\u9762\u7eb1\uff0c\u6df1\u5165\u63a2\u8ba8\u6b8b\u5dee\u8fde\u63a5\u7684\u7cbe\u5999\u4e4b\u5904\uff0c\u91cd\u8d70 Pre-Norm \u4e0e Post-Norm \u7684\u201c\u6069\u6028\u60c5\u4ec7\u201d\uff0c\u5e76\u4e00\u540c\u5c55\u671b\u6700\u65b0\u7684\u5f52\u4e00\u5316\u7b56\u7565\u5982\u4f55\u7a81\u7834\u6781\u9650\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6b8b\u5dee\u8fde\u63a5\uff1a\u4e3a\u6df1\u5c42\u7f51\u7edc\u5288\u5f00\u7684\u201c\u7eff\u8272\u901a\u9053\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u4e3a\u4f55\u201c\u6df1\u201d\u5374\u6210\u4e86\u95ee\u9898\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6df1\u5ea6\u5b66\u4e60\u7684\u65e9\u671f\uff0c\u4e00\u4e2a\u76f4\u89c2\u7684\u60f3\u6cd5\u662f\uff1a\u7f51\u7edc\u8d8a\u6df1\uff0c\u80fd\u529b\u8d8a\u5f3a\u3002\u4f46\u5b9e\u8df5\u5374\u7ed9\u4e86\u7814\u7a76\u8005\u4eec\u4e00\u8bb0\u91cd\u51fb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6ca1\u6709\u6b8b\u5dee\u8fde\u63a5\u7684\u4f20\u7edf\u7f51\u7edc\uff08\u5982 VGG\uff09\u4e2d\uff0c\u5f53\u7f51\u7edc\u5c42\u6570\u8d85\u8fc7 20 \u5c42\u65f6\uff0c\u7814\u7a76\u8005\u4eec\u89c2\u5bdf\u5230\u4e86\u4e00\u4e2a\u5947\u7279\u7684\u73b0\u8c61\uff1a\u4e0d\u4ec5\u6d4b\u8bd5\u96c6\u7684\u9519\u8bef\u7387\u4e0a\u5347\uff0c\u8fde\u8bad\u7ec3\u96c6\u7684\u9519\u8bef\u7387\u4e5f\u540c\u6b65\u5347\u9ad8\u3002\u8fd9\u4e0d\u662f\u8fc7\u62df\u5408\uff0c\u800c\u662f <strong>\u201c\u7f51\u7edc\u9000\u5316\u201d<\/strong> \u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b80\u5355\u6765\u8bf4\uff0c\u6df1\u5c42\u7f51\u7edc\u7684\u4f18\u5316\u96be\u5ea6\u968f\u5c42\u6570\u589e\u52a0\u5448\u6307\u6570\u7ea7\u589e\u957f\u3002\u68af\u5ea6\u5728\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u7ecf\u8fc7\u5c42\u5c42\u4f20\u9012\uff0c\u5c31\u50cf\u4f20\u8bdd\u6e38\u620f\u91cc\u7684\u4fe1\u606f\uff0c\u8981\u4e48\u9010\u6e10\u5fae\u5f31\u5230\u6d88\u5931\uff08<strong>\u68af\u5ea6\u6d88\u5931<\/strong>\uff09\uff0c\u8981\u4e48\u88ab\u4e0d\u5408\u7406\u5730\u653e\u5927\u5230\u5931\u771f\uff08<strong>\u68af\u5ea6\u7206\u70b8<\/strong>\uff09\u3002\u6700\u7ec8\uff0c\u6d45\u5c42\u7684\u6743\u91cd\u51e0\u4e4e\u65e0\u6cd5\u66f4\u65b0\uff0c\u7f51\u7edc\u7684\u6f5c\u529b\u88ab\u6df1\u5ea6\u672c\u8eab\u627c\u6740\u4e86\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 \u201c\u9ad8\u901f\u516c\u8def\u201d\u7684\u8bde\u751f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b8b\u5dee\u8fde\u63a5\uff08Residual Connection\uff09\u7684\u63d0\u51fa\uff0c\u6b63\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u96be\u9898\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u5f02\u5e38\u7b80\u6d01\uff1a<strong>\u5728\u7f51\u7edc\u7684\u8f93\u5165\u548c\u8f93\u51fa\u4e4b\u95f4\uff0c\u5efa\u7acb\u4e00\u6761\u201c\u8df3\u8dc3\u8fde\u63a5\u201d\uff08Shortcut Connection\uff09<\/strong> \uff0c\u8ba9\u4fe1\u606f\u53ef\u4ee5\u7ed5\u8fc7\u4e2d\u95f4\u7684\u53d8\u6362\u5c42\uff0c\u76f4\u63a5\u4f20\u9012\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u516c\u5f0f\u8868\u793a\uff0c\u4e00\u4e2a\u6807\u51c6\u7684\u6b8b\u5dee\u5757\u53ef\u4ee5\u5199\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">[<br>y = x + F(x)<br>]<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\uff0c(x) \u662f\u539f\u59cb\u7684\u8f93\u5165\uff0c(F(x)) \u662f\u5b50\u5c42\u8981\u5b66\u4e60\u7684\u201c\u6b8b\u5dee\u51fd\u6570\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8bbe\u8ba1\u753b\u9f99\u70b9\u775b\u4e4b\u7b14\u5728\u4e8e\uff1a<strong>\u5b83\u91cd\u6784\u4e86\u7f51\u7edc\u7684\u5b66\u4e60\u76ee\u6807\u3002<\/strong> \u5728\u8fc7\u53bb\uff0c\u7f51\u7edc\u9700\u8981\u8d39\u529b\u5730\u5b66\u4e60\u4e00\u4e2a\u5b8c\u6574\u7684\u6620\u5c04 (H(x))\u3002\u73b0\u5728\uff0c\u5b83\u53ea\u9700\u8981\u5b66\u4e60\u8f93\u5165\u4e0e\u8f93\u51fa\u4e4b\u95f4\u7684\u201c\u53d8\u5316\u91cf\u201d (F(x) = H(x) &#8211; x)\u3002\u5b66\u4e60\u6052\u7b49\u6620\u5c04\uff08\u5373\u8f93\u51fa\u7b49\u4e8e\u8f93\u5165\uff09\u53d8\u5f97\u5f02\u5e38\u7b80\u5355\uff1a\u53ea\u9700\u5c06 (F(x)) \u903c\u5411\u96f6\u5373\u53ef\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\uff0c\u68af\u5ea6\u591a\u4e86\u4e00\u6761\u4ece\u6df1\u5c42\u6d41\u5411\u6d45\u5c42\u7684\u201c\u9ad8\u901f\u516c\u8def\u201d\u3002\u5373\u4f7f\u4e3b\u8def\u5f84\u4e0a\u7684\u68af\u5ea6\u8870\u51cf\u4e86\uff0c\u6b8b\u5dee\u8fde\u63a5\u4e5f\u80fd\u4fdd\u8bc1\u6d45\u5c42\u81f3\u5c11\u5f97\u5230\u4e00\u4e2a\u4fdd\u5e95\u7684\u68af\u5ea6\u4fe1\u53f7\uff0c\u4ece\u800c\u907f\u514d\u4e86\u68af\u5ea6\u6d88\u5931\uff0c\u8ba9\u8bad\u7ec3\u4e0a\u767e\u5c42\u3001\u4e0a\u5343\u5c42\u7684\u7f51\u7edc\u6210\u4e3a\u53ef\u80fd\u3002\u6ca1\u6709\u5b83\uff0c\u6211\u4eec\u6240\u8ba8\u8bba\u7684\u201c\u5927\u6a21\u578b\u201d\u5c06\u4e0d\u590d\u5b58\u5728\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Normalization\uff1a\u7ed9\u68af\u5ea6\u6234\u4e0a\u201c\u7a33\u5b9a\u5668\u201d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4\u6b8b\u5dee\u8fde\u63a5\u662f\u4fe1\u606f\u9ad8\u901f\u516c\u8def\uff0c\u90a3\u5f52\u4e00\u5316\uff08Normalization\uff09\u5c31\u662f\u8def\u4e0a\u7684\u201c\u4ea4\u901a\u89c4\u5219\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u4fe1\u606f\u5728\u9ad8\u901f\u516c\u8def\u4e0a\u5954\u9a70\u65f6\uff0c\u6bcf\u4e00\u5c42\u7f51\u7edc\u7684\u6743\u91cd\u8ba1\u7b97\uff0c\u90fd\u53ef\u80fd\u5bfc\u81f4\u6570\u636e\u7684\u5206\u5e03\u53d1\u751f\u5267\u70c8\u53d8\u5316\u3002\u4e0a\u4e00\u5c42\u7684\u8f93\u51fa\uff0c\u7ecf\u8fc7\u77e9\u9635\u4e58\u6cd5\u540e\uff0c\u53ef\u80fd\u4f1a\u53d8\u5f97\u201c\u9762\u76ee\u5168\u975e\u201d\uff0c\u5c3a\u5ea6\u98d8\u5ffd\u4e0d\u5b9a\u3002\u8fd9\u79cd\u73b0\u8c61\u5728\u51e0\u767e\u4e0a\u5343\u5c42\u7684\u5de8\u578b\u7f51\u7edc\u91cc\u4f1a\u8fc5\u901f\u6076\u5316\uff0c\u5bfc\u81f4\u6574\u4e2a\u6a21\u578b\u6570\u503c\u6ea2\u51fa\uff0c\u8bad\u7ec3\u5d29\u6e83\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c42\u5f52\u4e00\u5316\uff08Layer Normalization, LN\uff09\u7684\u4f5c\u7528\uff0c\u4fbf\u662f\u5728\u6bcf\u4e00\u4e2a\u6b8b\u5dee\u5757\u7684\u5185\u90e8\uff0c\u5bf9\u8f93\u51fa\u8fdb\u884c\u4e00\u6b21\u201c\u6807\u51c6\u5316\u201d\u3002\u5b83\u4f1a\u5c06\u6570\u636e\u62c9\u56de\u5230\u4e00\u4e2a\u5747\u503c\u4e3a 0\u3001\u65b9\u5dee\u4e3a 1 \u7684\u7a33\u5b9a\u5206\u5e03\uff0c\u4ece\u800c\u786e\u4fdd\u68af\u5ea6\u6d41\u5728\u4efb\u4f55\u65f6\u5019\u90fd\u5904\u4e8e\u4e00\u4e2a\u5065\u5eb7\u5408\u7406\u7684\u533a\u95f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u4ee3 LLM \u5219\u66f4\u504f\u7231\u4f7f\u7528 <strong>RMSNorm\uff08Root Mean Square Normalization\uff09<\/strong> \u3002\u5b83\u662f LayerNorm \u7684\u7cbe\u7b80\u7248\uff0c\u53bb\u6389\u4e86\u8ba1\u7b97\u5747\u503c\u7684\u6b65\u9aa4\uff0c\u53ea\u4fdd\u7559\u5bf9\u6570\u636e\u7684\u201c\u5c3a\u5ea6\u201d\u8fdb\u884c\u7ea6\u675f\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\">\u522b\u5c0f\u770b\u8fd9\u4e2a\u201c\u5077\u61d2\u201d\u7684\u64cd\u4f5c\u3002\u5b83\u6362\u6765\u7684\uff0c\u662f\u66f4\u5c11\u7684\u8ba1\u7b97\u91cf\u548c\u66f4\u5feb\u7684\u63a8\u7406\u901f\u5ea6\uff0c\u800c\u5728\u6a21\u578b\u6700\u7ec8\u6548\u679c\u4e0a\uff0c\u548c LayerNorm \u51e0\u4e4e\u4e0d\u76f8\u4e0a\u4e0b\u3002\u8fd9\u6b63\u662f LLaMA\u3001Qwen\u3001DeepSeek \u7b49\u4e00\u7ebf\u6a21\u578b\u90fd\u91c7\u7528\u5b83\u7684\u539f\u56e0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001Post-Norm vs Pre-Norm\uff1a\u767e\u5e74\u201c\u7ad9\u4f4d\u201d\u4e4b\u4e89<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f52\u4e00\u5316\u548c\u6b8b\u5dee\u8fde\u63a5\u8fd9\u5bf9\u9ec4\u91d1\u642d\u6863\uff0c\u5b83\u4eec\u4e4b\u95f4\u5fae\u5999\u7684\u7a7a\u95f4\u5173\u7cfb\uff0c\u5374\u5f15\u53d1\u4e86 Transformer \u67b6\u6784\u53f2\u4e0a\u6700\u8457\u540d\u7684\u4e00\u573a\u201c\u7ad9\u4f4d\u201d\u4e4b\u4e89\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u539f\u6559\u65e8\u4e3b\u4e49\uff1aPost-Norm \u7684\u7ed3\u6784<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u539f\u59cb\u7684 Transformer \u8bba\u6587\u300aAttention Is All You Need\u300b\u91cc\uff0c\u91c7\u7528\u7684\u662f <strong>Post-Norm\uff08\u540e\u5f52\u4e00\u5316\uff09<\/strong> \u7ed3\u6784\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u8ba1\u7b97\u987a\u5e8f\u662f\uff1a<strong>\u5b50\u5c42\uff08Attention\/FFN\uff09\u2192 \u6b8b\u5dee\u8fde\u63a5 \u2192 \u5c42\u5f52\u4e00\u5316<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4ee3\u7801\u4f2a\u4ee3\u7801\u6765\u4f53\u73b0\uff0c\u8fd9\u5c31\u662f\u7ecf\u5178\u7684 Post-Norm Block\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># Post-Norm\ndef post_norm_block(x):\n    # 1. \u5148\u8ba1\u7b97\u5b50\u5c42 (Attention \u6216 FFN)\n    sub_output = sub_layer(x)  \n    # 2. \u518d\u4e0e\u8f93\u5165\u6b8b\u5dee\u76f8\u52a0\n    x = x + sub_output\n    # 3. \u6700\u540e\u8fdb\u884c\u5f52\u4e00\u5316\n    x = layer_norm(x)          \n    return x<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u201c\u8bed\u4e49\u201d\u4e0a\u8bb2\uff0c\u8fd9\u79cd\u8bbe\u8ba1\u975e\u5e38\u81ea\u7136\u3002\u7f51\u7edc\u5757\u5904\u7406\u4fe1\u606f\uff0c\u7136\u540e\u5c06\u65b0\u65e7\u4fe1\u606f\u878d\u5408\uff0c\u6700\u540e\u628a\u878d\u5408\u540e\u7684\u7ed3\u679c\u6807\u51c6\u5316\uff0c\u518d\u4ea4\u7ed9\u4e0b\u4e00\u5757\u5904\u7406\u3002\u4e07\u7269\u6709\u5e8f\uff0c\u7b26\u5408\u76f4\u89c9\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 \u201c\u6b7b\u4ea1\u201d\u771f\u76f8\uff1aPost-Norm \u4e3a\u4f55\u4e0d\u53ef\u6df1\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u800c\uff0c\u4e00\u65e6\u5c06 Post-Norm \u7684\u7f51\u7edc\u5806\u53e0\u5230\u51e0\u5341\u5c42\u4ee5\u4e0a\uff0c\u9ebb\u70e6\u5c31\u6765\u4e86\u3002\u8bad\u7ec3\u53d8\u5f97\u6781\u4e0d\u7a33\u5b9a\uff0c\u6a21\u578b\u5728\u542f\u52a8\u9636\u6bb5\u6781\u6613\u201c\u7206\u70b8\u201d\u6216\u8005\u201c\u8dd1\u98de\u201d\uff0c\u8fd9\u79cd\u73b0\u8c61\u5728\u5f53\u65f6\u51e0\u4e4e\u662f\u65e0\u89e3\u7684\u3002\u4e8e\u662f\uff0c\u7814\u7a76\u8005\u4e0d\u5f97\u4e0d\u5f15\u5165\u5404\u79cd\u590d\u6742\u7684\u201c\u62d0\u6756\u201d\uff1a\u6bd4\u5982<strong>\u590d\u6742\u7684\u5b66\u4e60\u7387\u9884\u70ed\uff08warm-up\uff09\u7b56\u7565<\/strong>\uff0c\u8ba9\u5b66\u4e60\u7387\u4ece\u6781\u5c0f\u7684\u503c\u5f00\u59cb\u7f13\u6162\u589e\u52a0\uff0c\u6765\u201c\u54c4\u7740\u201d\u7f51\u7edc\u5ea6\u8fc7\u6700\u521d\u7684\u5371\u9669\u671f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ecf\u8fc7\u7406\u8bba\u5206\u6790\uff0c\u80cc\u540e\u7684\u539f\u56e0\u6d6e\u51fa\u6c34\u9762\uff1aPost-Norm \u7ed3\u6784\u4e0b\uff0c\u9760\u8fd1\u8f93\u51fa\u7aef\u7684\u5c42\u7684\u53c2\u6570\uff0c\u5176\u68af\u5ea6\u5728\u8bad\u7ec3\u521d\u671f\u60ca\u4eba\u5730\u5927\u3002\u4e00\u65e6\u5b66\u4e60\u7387\u7a0d\u6709\u4e0d\u614e\uff0c\u68af\u5ea6\u5c31\u4f1a\u50cf\u6d2a\u6c34\u4e00\u6837\u51b2\u57ae\u6574\u4e2a\u6a21\u578b\u7684\u4f18\u5316\u3002\u5c42\u6570\u8d8a\u6df1\uff0c\u8fd9\u79cd\u73b0\u8c61\u8d8a\u4e25\u91cd\uff0c\u76f4\u63a5\u628a\u5927\u89c4\u6a21\u6df1\u5ea6\u8bad\u7ec3\u7684\u8def\u5835\u6b7b\u4e86\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u5b9e\u7528\u4e3b\u4e49\u80dc\u51fa\uff1aPre-Norm \u7684\u5d1b\u8d77<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u89e3\u51b3 Post-Norm \u7684\u4e0d\u7a33\u5b9a\u95ee\u9898\uff0c\u7814\u7a76\u8005\u4eec\u5fae\u8c03\u4e86 LN \u7684\u4f4d\u7f6e\uff0c\u63d0\u51fa <strong>Pre-Norm\uff08\u524d\u5f52\u4e00\u5316\uff09<\/strong> \u7ed3\u6784\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba1\u7b97\u987a\u5e8f\u53d8\u4e3a\uff1a<strong>\u5c42\u5f52\u4e00\u5316 \u2192 \u5b50\u5c42\uff08Attention\/FFN\uff09\u2192 \u6b8b\u5dee\u8fde\u63a5<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4ee3\u7801\u6765\u8868\u793a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># Pre-Norm\ndef pre_norm_block(x):\n    # 1. \u5148\u5bf9\u8f93\u5165\u8fdb\u884c\u5f52\u4e00\u5316\n    x_norm = layer_norm(x)\n    # 2. \u518d\u8ba1\u7b97\u5b50\u5c42\n    sub_output = sub_layer(x_norm)\n    # 3. \u6700\u540e\u8fdb\u884c\u6b8b\u5dee\u8fde\u63a5\n    x = x + sub_output\n    return x<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u5c0f\u5c0f\u7684\u987a\u5e8f\u8c03\u6574\uff0c\u5e26\u6765\u7684\u662f\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u8d28\u7684\u98de\u8dc3\u3002\u7814\u7a76\u8bc1\u660e\uff0cPre-Norm \u80fd\u5929\u7136\u5730\u8ba9\u6574\u4e2a\u7f51\u7edc\u7684\u68af\u5ea6\u5206\u5e03\u66f4\u52a0\u5e73\u6ed1\uff0c\u5373\u4f7f\u5728\u8bad\u7ec3\u521d\u671f\u4e5f\u4e0d\u9700\u8981\u5c0f\u5fc3\u7ffc\u7ffc\u7684\u5b66\u4e60\u7387\u9884\u70ed\u3002\u4e00\u53e5\u8bdd\u6982\u62ec\uff0c\u5b83\u8ba9\u8bad\u7ec3\u6df1\u5c42\u7f51\u7edc\u4ece\u201c\u8d70\u94a2\u4e1d\u201d\u53d8\u6210\u4e86\u201c\u5f00\u9ad8\u901f\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8be5\u8bbe\u8ba1\u5f88\u5feb\u53d7\u5230\u5de5\u4e1a\u754c\u548c\u5b66\u672f\u754c\u7684\u72c2\u70ed\u91c7\u7eb3\uff0c\u8fc5\u901f\u6210\u4e3a\u4e86 GPT \u7cfb\u5217\u3001LLaMA \u7cfb\u5217\u7b49\u51e0\u4e4e\u6240\u6709\u4e3b\u6d41\u5927\u6a21\u578b\u7684\u9ed8\u8ba4\u67b6\u6784\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.4 \u65b0\u65e7\u9635\u8425\u4e00\u89c8<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">\u5f52\u4e00\u5316\u4f4d\u7f6e<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u4ee3\u8868\u6a21\u578b<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u4f18\u70b9<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u7f3a\u70b9<\/th><\/tr><\/thead><tbody><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>Post-Norm<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">BERT, RoBERTa, ALBERT<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u540c\u7b49\u89c4\u6a21\u4e0b\uff0c\u6a21\u578b\u6700\u7ec8\u4e0a\u9650\u53ef\u80fd\u66f4\u9ad8<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8bad\u7ec3\u6781\u4e0d\u7a33\u5b9a\uff0c\u65e0\u6cd5\u6df1\u5ea6\u6269\u5c55<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>Pre-Norm<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">GPT, LLaMA, PaLM, Qwen<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8bad\u7ec3\u6781\u5176\u7a33\u5b9a\uff0c\u53ef\u8f7b\u677e\u6269\u5c55\u81f3\u5343\u4ebf\u53c2\u6570<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6df1\u5c42\u7f51\u7edc\u53ef\u80fd\u5b58\u5728\u8868\u5f81\u574d\u7f29\u98ce\u9669<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">3.5 Pre-Norm \u7684\u201c\u963f\u5580\u7409\u65af\u4e4b\u8e35\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Pre-Norm \u867d\u7136\u8d62\u4e86\uff0c\u4f46\u5e76\u975e\u5b8c\u7f8e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u65b0\u7814\u7a76\u53d1\u73b0\uff0cPre-Norm \u6709\u4e00\u4e2a\u6f5c\u5728\u7684\u4ee3\u4ef7\uff1a\u5b83\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6df1\u5c42 Transformer \u51fa\u73b0 <strong>\u201c\u8868\u5f81\u574d\u7f29\u201d\uff08Representation Collapse\uff09<\/strong> \u3002\u901a\u4fd7\u8bb2\uff0c\u7f51\u7edc\u5c42\u6570\u592a\u591a\u65f6\uff0c\u4e0d\u540c Token \u4e4b\u95f4\u7684\u5dee\u5f02\u6027\u7ecf\u8fc7\u51e0\u5341\u4e0a\u767e\u6b21 Pre-Norm \u6807\u51c6\u5316\u540e\uff0c\u53ef\u80fd\u88ab\u201c\u78e8\u5e73\u201d\u4e86\u3002\u6df1\u5c42\u4f3c\u4e4e\u5931\u53bb\u4e86\u8db3\u591f\u7684\u533a\u5206\u5ea6\uff0c\u6240\u6709\u4e1c\u897f\u90fd\u8d8b\u540c\u4e86\uff0c\u963b\u788d\u4e86\u6a21\u578b\u901a\u8fc7\u589e\u52a0\u6df1\u5ea6\u6765\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e5f\u5c31\u662f\u4e3a\u4ec0\u4e48\u6700\u7ec8\u4eba\u4eec\u53d1\u73b0\uff0cLaaS \u7684\u4e00\u4e9b\u6781\u5176\u6df1\u7684 Pre-Norm \u6a21\u578b\uff0c\u5176\u6027\u80fd\u65e0\u6cd5\u6301\u7eed\u7ebf\u6027\u589e\u5f3a\uff0c\u751a\u81f3\u4f1a\u51fa\u73b0\u8fb9\u9645\u6548\u76ca\u9012\u51cf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u7a81\u7834\u8fb9\u754c\uff1a\u4e3a\u4e86\u8ffd\u6c42\u6781\u81f4\u800c\u767b\u573a\u7684\u201c\u65b0\u661f\u201d\u4eec<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Pre-Norm \u548c Post-Norm \u5404\u6709\u4f18\u52a3\uff0c\u79d1\u7814\u4eba\u5458\u81ea\u7136\u4e0d\u4f1a\u6b62\u6b65\u4e8e\u201c\u4e8c\u9009\u4e00\u201d\u3002\u4ed6\u4eec\u5f00\u59cb\u901a\u8fc7\u5404\u79cd\u7cbe\u5999\u7684\u8bbe\u8ba1\uff0c\u5c1d\u8bd5\u201c\u65e2\u8981\u53c8\u8981\u201d\uff1a\u65e2\u8981 Pre-Norm \u7684\u7a33\u5b9a\uff0c\u53c8\u8981 Post-Norm \u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 DeepNorm\uff1a\u5fae\u8f6f\u5c06Transformer\u63a8\u52301000\u5c42\u7684\u79d8\u5bc6\u6b66\u5668<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">2022 \u5e74\uff0c\u5fae\u8f6f\u7814\u7a76\u9662\u56e2\u961f\u5e26\u7740\u4ed6\u4eec\u7684 <strong>DeepNorm<\/strong> \u9707\u64bc\u4e86\u5b66\u672f\u754c\u3002\u4ed6\u4eec\u63d0\u51fa\u7684\u53e3\u53f7\u7b80\u5355\u76f4\u63a5\uff1a\u8ba9 Post-Norm \u4e5f\u53ef\u4ee5\u7a33\u5b9a\u8bad\u7ec3\u5230\u4e0a\u5343\u5c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DeepNorm \u7684\u601d\u8def\u5f88\u5de7\u5999\uff0c\u5b83\u8ba4\u4e3a Post-Norm \u7684\u4e0d\u7a33\u5b9a\u6e90\u4e8e\u6a21\u578b\u66f4\u65b0\u6df1\u5ea6\uff08model update\uff09\u968f\u7740\u7f51\u7edc\u52a0\u6df1\u800c\u5931\u63a7\u3002\u56e0\u6b64\uff0c\u5b83\u5728<strong>\u4e0d\u6539\u53d8\u5f52\u4e00\u5316\u4f4d\u7f6e<\/strong>\u7684\u524d\u63d0\u4e0b\uff0c\u5728\u4e24\u4e2a\u5730\u65b9\u52a8\u4e86\u624b\u811a\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5728\u6b8b\u5dee\u8fde\u63a5\u4e0a\u52a0\u4e86\u4e00\u4e2a\u7f29\u653e\u56e0\u5b50 (\\alpha > 1)<\/strong>\uff1a<code>x = LayerNorm(\u03b1 * x + sub_layer(x))<\/code>\uff0c\u5bf9\u539f\u59cb\u8f93\u5165\u7684\u4fe1\u53f7\u8fdb\u884c\u4e86\u201c\u63d0\u6743\u201d\u5904\u7406\u3002<\/li>\n\n\n\n<li><strong>\u5728\u521d\u59cb\u5316\u9636\u6bb5\u7f29\u5c0f\u7279\u5b9a\u53c2\u6570\u7684\u521d\u59cb\u503c<\/strong>\uff1a\u6bd4\u5982\u5728 Xavier \u521d\u59cb\u5316\u540e\uff0c\u989d\u5916\u4e58\u4e0a\u4e00\u4e2a\u7f29\u653e\u56e0\u5b50 (\\beta &lt; 1)\uff0c\u8ba9\u6a21\u578b\u4e00\u5f00\u59cb\u7684\u884c\u4e3a\u66f4\u201c\u6e29\u987a\u201d\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u5927\u4e00\u5c0f\u7684\u56e0\u5b50\uff0c\u8054\u624b\u5c06\u6a21\u578b\u66f4\u65b0\u7262\u7262\u675f\u7f1a\u5728\u4e86\u4e00\u4e2a\u53ef\u63a7\u8303\u56f4\u5185\uff0c\u65e0\u8bba\u7f51\u7edc\u6df1\u6d45\u3002\u4ece\u6027\u80fd\u4e0a\u770b\uff0c\u5fae\u8f6f\u7528 32 \u4ebf\u53c2\u6570\u7684 200 \u5c42 DeepNet\uff0c\u5728\u7ffb\u8bd1\u4efb\u52a1\u4e0a\u51fb\u8d25\u4e86 120 \u4ebf\u53c2\u6570\u7684 48 \u5c42\u6700\u4f18\u6a21\u578b\uff0c\u5f3a\u529b\u8bc1\u660e\u4e86\u901a\u8fc7\u63d0\u5347\u6df1\u5ea6\u80fd\u591f\u6316\u6398\u51fa\u6781\u5f3a\u7684\u6f5c\u529b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u5404\u8def\u82f1\u8c6a\uff1a\u53d6\u957f\u8865\u77ed\u7684\u7b56\u7565<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">DeepNorm \u8bc1\u660e\u4e86\u201c\u7528 Post-Norm \u4e4b\u9b42 + \u989d\u5916\u7a33\u5b9a\u672f\u201d\u662f\u53ef\u4ee5\u9a6f\u670d\u6781\u6df1\u7f51\u7edc\u7684\uff0c\u4f46\u5b83\u7684\u6839\u57fa\u4ecd\u7136\u662f Post-Norm\uff0c\u7ed3\u6784\u672c\u8d28\u4e0a\u8fd8\u662f\u5355\u4e00\u8def\u5f84\u3002\u4e8e\u662f\uff0c\u66f4\u5177\u98a0\u8986\u6027\u7684\u7ed3\u6784\u6539\u8fdb\u6d8c\u73b0\u51fa\u6765\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Mix-LN\uff1a\u5206\u5c42\u6cbb\u7406<\/strong>\uff1a2024 \u5e74\u5e95\u63d0\u51fa\u7684 Mix-LN \u8ba4\u4e3a\uff0c\u65e2\u7136 Pre-LN \u5728\u6d45\u5c42\u597d\u7528\u3001Post-LN \u5728\u6df1\u5c42\u6f5c\u529b\u5927\uff0c\u90a3\u4f55\u4e0d\u6309\u7167\u5c42\u6570\u52a8\u6001\u5206\u914d\uff1f\u4e8e\u662f\uff0c\u5b83\u5c06\u524d\u51e0\u5c42\u8bbe\u4e3a Post-LN \u4ee5\u4fdd\u6301\u6df1\u5c42\u68af\u5ea6\uff0c\u540e\u51e0\u5c42\u8bbe\u4e3a Pre-LN \u4ee5\u4fdd\u8bc1\u6574\u4f53\u7a33\u5b9a\u3002\u8fd9\u79cd\u65b9\u6cd5\u51e0\u4e4e\u96f6\u6210\u672c\u63d0\u5347\u4e86\u6a21\u578b\u6f5c\u80fd\uff0c\u8ba9\u65b0\u65e7\u7ed3\u5408\u7684\u4f18\u52bf\u5f97\u4ee5\u91ca\u653e\u3002<\/li>\n\n\n\n<li><strong>HybridNorm\uff1a\u533a\u5757\u81ea\u6cbb<\/strong>\uff1a2025 \u5e74\u7684 HybridNorm \u505a\u5f97\u66f4\u201c\u7edd\u201d\uff0c\u5b83\u5c06\u7b56\u7565\u7cbe\u7ec6\u5230\u4e86\u5757\u7ea7\u522b\u3002\u5b83\u91c7\u7528 <strong>QKV \u5f52\u4e00\u5316 + FFN \u7684 Post-Norm<\/strong> \u6df7\u5408\u67b6\u6784\u3002\u6b64\u4e3e\u65e2\u7a33\u4f4f\u4e86\u81ea\u6ce8\u610f\u529b\u7684\u5b66\u4e60\u52a8\u6001\uff0c\u53c8\u5145\u5206\u53d1\u6325\u4e86 FFN \u5728\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u6f5c\u80fd\uff0c\u663e\u793a\u4e86\u5728\u5927\u578b\u7a00\u758f\u548c\u7a20\u5bc6\u6a21\u578b\u4e2d\u5747\u80fd\u5168\u9762\u8d85\u8d8a Pre\/Post-LN \u7684\u5b9e\u529b\u3002<\/li>\n\n\n\n<li><strong>Keel\uff1aPost-LN \u7684\u201c\u91cd\u751f\u201d<\/strong>\uff1a\u5982\u679c DeepNorm \u53ea\u662f\u201c\u6280\u672f\u4fee\u8f66\u201d\uff0c\u90a3\u5b57\u8282\u8df3\u52a8\u56e2\u961f 2026 \u5e74\u7684 <strong>Keel<\/strong> \u4fbf\u662f\u76f4\u63a5\u7ed9\u5b83\u201c\u6362\u4e86\u53d1\u52a8\u673a\u201d\u3002Keel \u6838\u5fc3\u8bca\u65ad\u51fa\u201c\u539f\u7f6a\u201d\u5728\u4e8e <strong>ResNet \u5f0f\u7684\u6b8b\u5dee\u8def\u5f84<\/strong>\u4f1a\u5bfc\u81f4\u6df1\u5c42\u68af\u5ea6\u6d88\u5931\u3002\u56e0\u6b64\uff0c\u5b83\u76f4\u63a5\u66ff\u6362\u4e3a <strong>Highway \u98ce\u683c\u8fde\u63a5<\/strong>\uff0c\u8ba9\u68af\u5ea6\u4e0d\u53d7\u963b\u788d\u5730\u8d2f\u7a7f\u7f51\u7edc\u3002\u6548\u679c\u51fa\u5947\u60ca\u4eba\uff1a<strong>\u8d85\u8fc7 1000 \u5c42\u7684\u7a33\u5b9a\u8bad\u7ec3<\/strong>\u6210\u4e3a\u53ef\u80fd\uff0c\u4e14\u6027\u80fd\u968f\u65f6\u5ef6\u5c55\u800c\u6301\u7eed\u653e\u5927\u3002\u8fd9\u7ed9\u4e86\u4e1a\u754c\u4e00\u4e2a\u5f3a\u70c8\u7684\u4fe1\u53f7\uff1a<strong>\u6df1\u5ea6 Scaling \u7684\u65f6\u4ee3\uff0c\u4e5f\u8bb8\u624d\u521a\u62c9\u5f00\u5e37\u5e55<\/strong>\u3002<\/li>\n\n\n\n<li><strong>ResiDual\uff1a\u53cc\u8def\u5f84\u4fdd\u9669<\/strong>\uff1a\u65e2\u7136 Pre \u548c Post \u5404\u6709\u597d\u574f\uff0c\u4e3a\u4ec0\u4e48\u53ea\u80fd\u505a\u5355\u9009\u9898\uff1f\u5fae\u8f6f 2023 \u5e74\u7684 <strong>ResiDual<\/strong> \u52c7\u6562\u6253\u7834\u5e38\u89c4\uff0c\u901a\u8fc7 <strong>Pre-Post-LN (PPLN) \u53cc\u91cd\u6b8b\u5dee\u8fde\u63a5<\/strong> \u5c06\u4e24\u8005\u7f1d\u5408\u5728\u4e00\u8d77\u3002\u4ece\u6570\u636e\u8def\u5f84\u4e0a\uff0c\u8fd9\u5c31\u50cf\u540c\u65f6\u4e70\u4e86\u4e24\u79cd\u4fdd\u9669\uff0c\u786e\u4fdd\u65e2\u4e0d\u4f1a\u51fa\u73b0\u68af\u5ea6\u6d88\u5931\uff0c\u4e5f\u4e0d\u4f1a\u5411\u8868\u5f81\u574d\u7f29\u9000\u5316\uff0c\u5728\u5b9e\u9645\u7ffb\u8bd1\u4efb\u52a1\u4e0a\u4e5f\u663e\u8457\u66f4\u4f18\u3002<\/li>\n\n\n\n<li><strong>SiameseNorm\uff1a\u53cc\u80de\u80ce\u7f51\u7edc<\/strong>\uff1a2026 \u5e74 2 \u6708 ArXiv \u53d1\u8868\u7684 <strong>SiameseNorm\uff08\u5f52\u4e00\u5316\u8fde\u4f53\u5a74\uff09<\/strong> \uff0c\u582a\u79f0\u89e3\u51b3\u8fd9\u4e00\u96be\u9898\u7684\u6700\u5927\u80c6\u8bbe\u8ba1\u3002\u5b83\u4e0d\u5355\u662f\u201c\u53cc\u8def\u5f84\u201d\uff0c\u800c\u662f\u76f4\u63a5\u9020\u4e86\u4e00\u5bf9 <strong>Pre-Norm \u6d41\u548c Post-Norm \u6d41\u7684\u53cc\u80de\u80ce\u7f51\u7edc\uff08Siamese\uff09<\/strong> \uff0c\u5e76\u8ba9\u4e24\u8005<strong>\u5171\u4eab\u53c2\u6570<\/strong>\uff0c\u540c\u6b65\u6d41\u52a8\u540e\u5408\u5e76\u68af\u5ea6\u3002\u8fd9\u4e00\u8bbe\u8ba1\u6253\u7834\u4e86\u5fc5\u987b\u53d6\u820d\u5355\u4e00\u8def\u5f84\u7684\u9b54\u5492\uff0c\u5c06\u201c\u65e2\u8981\u53c8\u8981\u201d\u53d8\u6210\u4e86\u5929\u7136\u7684\u5e76\u96c6\u3002\u5728 13 \u4ebf\u53c2\u6570\u89c4\u6a21\u5b9e\u9a8c\u4e2d\uff0cSiameseNorm \u5c55\u73b0\u51fa\u4e86\u6781\u5f3a\u7684\u9c81\u68d2\u6027\u3002<\/li>\n\n\n\n<li><strong>SpanNorm\uff1a\u8de8\u5757\u4e4b\u6865<\/strong>\uff1a\u7d27\u63a5\u7740\uff0c2026 \u5e74\u521d ArXiv \u4e0a\u7684 <strong>SpanNorm<\/strong> \u53c8\u63d0\u4f9b\u4e86\u4e00\u79cd\u65b0\u601d\u8def\u3002\u5b83\u653e\u5f03\u4e86\u4f20\u7edf\u6b8b\u5dee\u53ea\u8de8\u5b50\u5c42\u7684\u6a21\u5f0f\uff0c\u5f7b\u5e95\u5efa\u7acb\u4e86\u4e00\u6761<strong>\u8de8\u8d8a\u6574\u4e2a Transformer \u5757\u7684\u5e72\u51c0\u6b8b\u5dee\u8fde\u63a5<\/strong>\uff0c\u5e76\u5728\u5927\u7ec8\u70b9\u4f7f\u7528\u4e86 PostNorm \u5f0f\u805a\u5408\u5f52\u4e00\u5316\u3002\u4e00\u79cd\u5927\u4e00\u7edf\u7684\u6574\u5408\u4f18\u52bf\u53e0\u52a0\uff0c\u5728\u7406\u8bba\u4e0a\u4fdd\u8bc1\u4e86\u8de8\u6df1\u5ea6\u4fe1\u53f7\u7684\u65e0\u635f\u4f20\u9012\uff0c\u5e76\u5bf9\u8868\u5f81\u574d\u7f29\u901a\u75c5\u63d0\u51fa\u4e86\u7f13\u89e3\u65b9\u6848\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u603b\u7ed3\u4e0e\u5c55\u671b\uff1a\u4ece\u201c\u7ad9\u4f4d\u201d\u5230\u201c\u878d\u5408\u201d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u81f3\u6b64\uff0c\u6211\u4eec\u53ef\u4ee5\u6e05\u6670\u5730\u770b\u5230\u4e00\u6761\u6280\u672f\u6f14\u8fdb\u7684\u4e3b\u7ebf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">\u9636\u6bb5<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u4ee3\u8868\u65b9\u6848<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u6838\u5fc3\u601d\u60f3<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u89e3\u51b3\u7684\u95ee\u9898<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u5f15\u5165\u7684\u65b0\u6311\u6218<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u53d1\u5e03\u65f6\u95f4\u7ea6<\/th><\/tr><\/thead><tbody><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>1. \u539f\u59cb\u65f6\u4ee3<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">Post-Norm<\/td><td class=\"has-text-align-left\" data-align=\"left\">LN \u653e\u5728\u6b8b\u5dee\u6c42\u548c\u540e<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u5b9a\u4e49\u4e86 Transformer \u7ecf\u5178\u7ed3\u6784<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6df1\u5c42\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff0c\u6781\u5ea6\u4f9d\u8d56\u5b66\u4e60\u7387\u9884\u70ed<\/td><td class=\"has-text-align-left\" data-align=\"left\">2017<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>2. \u7a33\u5b9a\u65f6\u4ee3<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">Pre-Norm<\/td><td class=\"has-text-align-left\" data-align=\"left\">LN \u653e\u5728\u5b50\u5c42\u524d\u3001\u6b8b\u5dee\u6c42\u548c\u524d<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6781\u81f4\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u5b9e\u73b0\u8f7b\u677e\u6df1\u5ea6\u6269\u5c55<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6df1\u5c42\u51fa\u73b0\u8868\u5f81\u574d\u7f29\uff0c\u6df1\u5ea6\u6536\u76ca\u53d7\u9650<\/td><td class=\"has-text-align-left\" data-align=\"left\">2019-2020<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>3. \u6df1\u5ea6\u65f6\u4ee3<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\"><strong>DeepNorm<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6b8b\u5dee\u4fa7\u7f29\u653e + \u521d\u59cb\u5316\u7a00\u758f\u5316<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u9996\u6b21\u8ba9\u6781\u6df1\uff08\u5343\u5c42\uff09\u6a21\u578b\u80fd\u7a33\u5b9a\u8bad\u7ec3<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8c03\u53c2\u590d\u6742\uff0c\u5de5\u7a0b\u5b9e\u73b0\u7a0d\u663e\u7e41\u7410<\/td><td class=\"has-text-align-left\" data-align=\"left\">2022<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>4. \u878d\u5408\u65f6\u4ee3<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\"><strong>Mix-LN<\/strong>, <strong>HybridNorm<\/strong>, <strong>Keel<\/strong>, <strong>ResiDual<\/strong>, <strong>SiameseNorm<\/strong>, <strong>SpanNorm<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">Pre\/Post \u4f53\u7cfb\u6df7\u5408\uff0c\u6216\u91cd\u6784\u6b8b\u5dee\u8fde\u63a5<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u7edf\u4e00\u6781\u6df1\u7f51\u7edc\u7684\u8bad\u7ec3\u6548\u7387\u4e0e\u6027\u80fd\u8868\u73b0<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u65b0\u8303\u5f0f\u4ecd\u5728\u63a2\u7d22\uff0c\u5de5\u7a0b\u6846\u67b6\u5c1a\u5f85\u6210\u719f\u9002\u914d<\/td><td class=\"has-text-align-left\" data-align=\"left\">2023-2026<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u671b\u8fc7\u53bb\uff0c\u6b8b\u5dee\u8fde\u63a5\u4e0e\u5f52\u4e00\u5316\u8fd9\u5bf9\u8001\u642d\u6863\uff0c\u6700\u5f00\u59cb\u53ea\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u6570\u5b66\u516c\u5f0f\u3002\u4e3a\u4e86\u8ffd\u6c42\u66f4\u6df1\u3001\u66f4\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u65e0\u6570\u7814\u7a76\u8005\u4e0d\u5f97\u4e0d\u56de\u8fc7\u5934\u6765\uff0c\u5bf9\u8fd9\u4e00\u516c\u5f0f\u8fdb\u884c\u7ec6\u81f4\u7684\u8c03\u6574\u548c\u9769\u65b0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u7ecf\u5386\u4e86\u4e00\u5f00\u59cb\u7684\u201c\u7ecf\u9a8c\u76f4\u89c9\u201d\uff08Post-Norm\uff09\uff0c\u8f6c\u5411\u4e86\u201c\u7a33\u624e\u7a33\u6253\u201d\uff08Pre-Norm\uff09\uff0c\u770b\u8fc7\u4e86\u901a\u8fc7\u53c2\u6570\u8c03\u8282\u5b9e\u73b0\u201c\u65e0\u4e0a\u7a81\u7834\u201d\uff08DeepNorm\uff09\uff0c\u5982\u4eca\u6b63\u5728\u52a0\u901f\u8fdb\u51fb\u201c\u5168\u80fd\u6574\u4f53\u201d\uff08Hybrid\/ResiDual\/SiameseNorm\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e0d\u4ec5\u662f\u4e00\u6b21\u6280\u672f\u8fed\u4ee3\uff0c\u66f4\u662f\u4e00\u6b21\u8bbe\u8ba1\u7406\u5ff5\u7684\u8f6c\u53d8\u3002\u6211\u4eec\u4e0d\u518d\u628a\u6b8b\u5dee\u8fde\u63a5\u548c\u5f52\u4e00\u5316\u770b\u4f5c\u9759\u6001\u3001\u5b64\u7acb\u7684\u90e8\u4ef6\uff0c\u800c\u5f00\u59cb\u628a\u5b83\u4eec\u5f53\u4f5c\u4e00\u4e2a<strong>\u52a8\u6001\u7684\u3001\u9700\u8981\u4e0e\u5176\u5b83\u7ec4\u4ef6\u534f\u4f5c\u7684\u7cfb\u7edf<\/strong>\u3002\u4e5f\u8bb8\uff0c\u5728\u672a\u6765\uff0c\u8fd9\u79cd\u6446\u653e\u4f4d\u7f6e\u5c06\u4e0d\u518d\u662f\u9700\u8981\u7ea0\u7ed3\u7684\u201c\u56fa\u5b9a\u8bbe\u8ba1\u201d\uff0c\u800c\u4f1a\u6210\u4e3a\u795e\u7ecf\u7f51\u7edc <strong>\u201c\u81ea\u9002\u5e94\u5b66\u4e60\u201d<\/strong> \u7684\u4e00\u90e8\u5206\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 \u5728\u524d\u51e0\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u50cf\u62c6\u89e3\u4e00\u4e2a\u7cbe\u5bc6\u7684\u673a\u68b0\u8868\u4e00\u6837\uff0c\u4e00\u6b65\u6b65\u89e3\u5256\u4e86 LLM \u7684\u5404\u4e2a\u6838\u5fc3\u7ec4\u4ef6\uff1a\u4ece Token [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":138,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[37,43],"class_list":["post-137","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-llm","tag-pre-norm"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/137","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=137"}],"version-history":[{"count":2,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/137\/revisions"}],"predecessor-version":[{"id":140,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/137\/revisions\/140"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/138"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=137"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=137"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=137"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}