{"id":129,"date":"2026-04-27T17:43:19","date_gmt":"2026-04-27T09:43:19","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=129"},"modified":"2026-04-23T17:44:38","modified_gmt":"2026-04-23T09:44:38","slug":"ffn-%e4%b8%8e-swiglu-%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/27\/ffn-%e4%b8%8e-swiglu-%e8%a7%a3%e6%9e%90\/","title":{"rendered":"FFN \u4e0e SwiGLU \u89e3\u6790"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u4e4b\u524d\u7684\u7cfb\u5217\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u4ece Tokenizer \u8bb2\u5230 Embedding \u518d\u5230 RoPE\uff0c\u5df2\u7ecf\u8d70\u5b8c\u4e86\u6570\u636e\u4ece\u201c\u4eba\u7c7b\u8bed\u8a00\u201d\u5230\u201c\u6a21\u578b\u53ef\u8ba1\u7b97\u7684\u5411\u91cf\u201d\u7684\u5b8c\u6574\u8def\u7a0b\u3002\u73b0\u5728\uff0c\u6211\u4eec\u6765\u5230\u4e86 LLM \u6700\u201c\u539a\u91cd\u201d\u7684\u90e8\u5206\u2014\u2014<strong>FFN\uff08Feed-Forward Network\uff0c\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff09<\/strong> \u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\u8bf4\u5b83\u201c\u539a\u91cd\u201d\uff1f\u56e0\u4e3a\u5728\u4e00\u5c42 Transformer Block \u4e2d\uff0cAttention \u8d1f\u8d23\u8ba9 Token \u4e4b\u95f4\u76f8\u4e92\u201c\u4ea4\u6d41\u201d\uff0c\u800c FFN \u624d\u662f\u771f\u6b63<strong>\u5b58\u50a8\u548c\u5904\u7406\u77e5\u8bc6\u7684\u573a\u6240<\/strong>\u3002\u5982\u679c\u628a Transformer \u6bd4\u4f5c\u4e00\u4e2a\u5de5\u5382\uff0cAttention \u662f\u6d41\u6c34\u7ebf\u4e0a\u7684\u201c\u8c03\u914d\u5de5\u4eba\u201d\uff0cFFN \u5219\u662f\u4e00\u53f0\u53f0\u201c\u52a0\u5de5\u673a\u5668\u201d\u2014\u2014\u7edd\u5927\u90e8\u5206\u7684\u201c\u751f\u4ea7\u8bbe\u5907\u201d\u90fd\u5728\u8fd9\u91cc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">FFN \u7684\u53c2\u6570\u91cf\u901a\u5e38\u5360\u636e\u6574\u4e2a\u6a21\u578b\u53c2\u6570\u7684 <strong>2\/3<\/strong> \u5de6\u53f3\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5f53\u4e0b\u8f7d\u4e00\u4e2a 7B \u7684\u6a21\u578b\uff0c\u5176\u4e2d\u5927\u7ea6\u6709 4.5B \u7684\u53c2\u6570\u90fd\u8eba\u5728 FFN \u91cc\u3002\u7406\u89e3 FFN\uff0c\u5c31\u662f\u7406\u89e3 LLM \u7684\u201c\u8bb0\u5fc6\u201d\u548c\u201c\u601d\u8003\u201d\u662f\u600e\u4e48\u53d1\u751f\u7684\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001FFN \u5728 Transformer \u4e2d\u7684\u5b9a\u4f4d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 Transformer Block \u7684\u201c\u53cc\u5f15\u64ce\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e00\u4e2a\u6807\u51c6\u7684 Transformer Block \u90fd\u7531\u4e24\u4e2a\u6838\u5fc3\u6a21\u5757\u7ec4\u6210\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u8f93\u5165 x\n   \u2193\nPre-Norm \u2192 Attention \u2192 \u6b8b\u5dee\u8fde\u63a5\uff08x + Attention(x)\uff09\n   \u2193\nPre-Norm \u2192 FFN \u2192 \u6b8b\u5dee\u8fde\u63a5\uff08x + FFN(x)\uff09\n   \u2193\n\u8f93\u51fa<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u8ba9\u5e8f\u5217\u4e2d\u7684 Token \u4e92\u76f8\u201c\u770b\u5230\u201d\u5bf9\u65b9\uff0c\u5b8c\u6210\u4fe1\u606f\u7684\u878d\u5408\u548c\u4ea4\u4e92\u3002\u4f46 Attention \u672c\u8eab\u662f<strong>\u7ebf\u6027\u53d8\u6362 + \u52a0\u6743\u6c42\u548c<\/strong>\u2014\u2014\u672c\u8d28\u4e0a\u8fd8\u662f\u7ebf\u6027\u7684\u3002\u5982\u679c\u6ca1\u6709 FFN\uff0c\u518d\u591a\u7684 Attention \u5c42\u53e0\u52a0\u8d77\u6765\u4e5f\u53ea\u662f\u4e00\u4e2a\u590d\u6742\u7684\u7ebf\u6027\u6a21\u578b\uff0c\u8868\u8fbe\u80fd\u529b\u4e25\u91cd\u53d7\u9650\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">FFN \u7684\u4f5c\u7528\u5c31\u662f<strong>\u5f15\u5165\u975e\u7ebf\u6027<\/strong>\uff0c\u8ba9\u6a21\u578b\u80fd\u591f\u62df\u5408\u66f4\u590d\u6742\u7684\u51fd\u6570\uff0c\u4ece\u800c\u5b58\u50a8\u548c\u5904\u7406\u66f4\u4e30\u5bcc\u7684\u77e5\u8bc6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 FFN \u4e3a\u4ec0\u4e48\u80fd\u201c\u5b58\u50a8\u77e5\u8bc6\u201d\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u4e2a\u6709\u8da3\u7684\u89c2\u5bdf\u662f\uff1a\u5728 LLM \u4e2d\uff0cAttention \u5c42\u66f4\u591a\u8d1f\u8d23\u201c\u63d0\u53d6\u201d\u548c\u201c\u6574\u5408\u201d\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u800c<strong>\u4e8b\u5b9e\u6027\u77e5\u8bc6\u548c\u63a8\u7406\u89c4\u5219\u4e3b\u8981\u5b58\u50a8\u5728 FFN \u7684\u53c2\u6570\u4e2d<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u80cc\u540e\u7684\u76f4\u89c2\u7406\u89e3\u662f\uff1aAttention \u7684\u6743\u91cd\u77e9\u9635\uff08Q\u3001K\u3001V\uff09\u8d1f\u8d23\u201c\u5bf9\u8c01\u611f\u5174\u8da3\u201d\uff0c\u5b83\u4eec\u51b3\u5b9a\u4e86\u4fe1\u606f\u5728 Token \u4e4b\u95f4\u7684\u6d41\u52a8\u8def\u5f84\uff1b\u800c FFN \u7684\u6743\u91cd\u77e9\u9635\uff08W1\u3001W2\uff09\u5219\u50cf\u4e00\u4e2a\u5de8\u5927\u7684\u201c\u952e\u503c\u5bf9\u5b58\u50a8\u5668\u201d\uff0c\u5c06\u8f93\u5165\u6a21\u5f0f\u6620\u5c04\u5230\u8f93\u51fa\u6a21\u5f0f\u3002\u7814\u7a76\u8868\u660e\uff0c<strong>Transformer \u4e2d\u7684 FFN \u5c42\u53ef\u4ee5\u88ab\u89c6\u4e3a\u67d0\u79cd\u5f62\u5f0f\u7684\u952e\u503c\u8bb0\u5fc6\u7f51\u7edc<\/strong>\u2014\u2014\u7b2c\u4e00\u5c42\u6743\u91cd\u77e9\u9635\u5145\u5f53\u201c\u952e\u201d\uff0c\u7b2c\u4e8c\u5c42\u5145\u5f53\u201c\u503c\u201d\uff0c\u5c06\u8f93\u5165\u7684\u8bed\u4e49\u6a21\u5f0f\u8f6c\u6362\u4e3a\u7279\u5b9a\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e5f\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48 FFN \u7684\u53c2\u6570\u91cf\u5982\u6b64\u5de8\u5927\uff1a\u8981\u5b58\u50a8\u6d77\u91cf\u7684\u4e16\u754c\u77e5\u8bc6\u548c\u8bed\u8a00\u89c4\u5219\uff0c\u9700\u8981\u8db3\u591f\u5927\u7684\u201c\u8bb0\u5fc6\u7a7a\u95f4\u201d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u4ece ReLU \u5230 SwiGLU\uff1aFFN \u6fc0\u6d3b\u51fd\u6570\u7684\u6f14\u8fdb<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4 FFN \u7684\u77e9\u9635\u4e58\u6cd5\u662f\u6a21\u578b\u7684\u201c\u9aa8\u67b6\u201d\uff0c\u90a3\u4e48\u6fc0\u6d3b\u51fd\u6570\u5c31\u662f\u8d4b\u4e88\u5176\u201c\u751f\u547d\u529b\u201d\u7684\u5173\u952e\u3002\u5b83\u7684\u9009\u62e9\u76f4\u63a5\u5f71\u54cd\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u3001\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6536\u655b\u901f\u5ea6\u3002\u8fc7\u53bb\u51e0\u5e74\uff0cFFN \u6fc0\u6d3b\u51fd\u6570\u7684\u6f14\u8fdb\u7ecf\u5386\u4e86\u4e09\u4e2a\u91cd\u8981\u9636\u6bb5\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 \u7b2c\u4e00\u9636\u6bb5\uff1aReLU \u7684\u7b80\u5355\u65f6\u4ee3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u65e9\u671f Transformer\uff08\u5982 2017 \u5e74\u539f\u7248\u8bba\u6587\uff09\u4f7f\u7528\u7684\u662f\u6700\u7b80\u5355\u7684 <strong>ReLU<\/strong> \u6fc0\u6d3b\u51fd\u6570\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>ReLU(x) = max(0, x)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6807\u51c6 FFN \u7684\u7ed3\u6784\u4e3a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>FFN(x) = W2 \u00b7 ReLU(W1 \u00b7 x)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d\uff0cW1 \u5c06\u8f93\u5165\u4ece d_model \u7ef4\u5ea6\u6620\u5c04\u5230 d_ffn\uff08\u901a\u5e38\u4e3a d_model \u7684 4 \u500d\uff09\uff0cW2 \u518d\u6620\u5c04\u56de d_model\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">ReLU \u7684\u4f18\u52bf\u975e\u5e38\u660e\u663e\uff1a<strong>\u8ba1\u7b97\u7b80\u5355\uff0c\u68af\u5ea6\u4e0d\u9971\u548c<\/strong>\u3002\u4f46\u5b83\u6709\u4e00\u4e2a\u8457\u540d\u7684\u7f3a\u9677\u2014\u2014<strong>\u201cDead ReLU\u201d\u95ee\u9898<\/strong>\uff1a\u5f53\u8f93\u5165\u5c0f\u4e8e 0 \u65f6\uff0cReLU \u7684\u8f93\u51fa\u4e3a 0\uff0c\u68af\u5ea6\u4e5f\u4e3a 0\u3002\u4e00\u65e6\u67d0\u4e2a\u795e\u7ecf\u5143\u201c\u6b7b\u6389\u201d\uff0c\u5b83\u53ef\u80fd\u6c38\u8fdc\u65e0\u6cd5\u6062\u590d\uff0c\u9020\u6210\u6a21\u578b\u5bb9\u91cf\u7684\u6c38\u4e45\u635f\u5931\u3002\u8fd9\u79cd\u786c\u96f6\u8fb9\u754c\u5728\u6df1\u5c42\u7f51\u7edc\u4e2d\u5c24\u4e3a\u81f4\u547d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u7b2c\u4e8c\u9636\u6bb5\uff1aGELU \u7684\u5e73\u6ed1\u66ff\u4ee3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u89e3\u51b3 ReLU \u7684\u201c\u6b7b\u4ea1\u201d\u95ee\u9898\uff0c\u7814\u7a76\u8005\u4eec\u5f15\u5165\u4e86 <strong>GELU\uff08Gaussian Error Linear Unit\uff09<\/strong> \uff0cBERT \u7cfb\u5217\u6a21\u578b\u5c31\u662f\u5b83\u7684\u8457\u540d\u4f7f\u7528\u8005\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>GELU(x) = x \u00b7 \u03a6(x)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d \u03a6(x) \u662f\u6807\u51c6\u6b63\u6001\u5206\u5e03\u7684\u7d2f\u79ef\u5206\u5e03\u51fd\u6570\u3002\u76f4\u89c2\u7406\u89e3\uff1aGELU \u4e0d\u662f\u786c\u6027\u5730\u628a\u8d1f\u503c\u622a\u65ad\u4e3a 0\uff0c\u800c\u662f\u6839\u636e x \u843d\u5728\u6b63\u6001\u5206\u5e03\u4e2d\u7684\u6982\u7387\u6765\u51b3\u5b9a\u201c\u653e\u591a\u5c11\u4fe1\u606f\u8fc7\u53bb\u201d\u2014\u2014\u6709\u70b9\u50cf\u4e00\u4e2a\u201c\u6982\u7387\u95e8\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GELU \u5728\u5b9e\u8df5\u4e2d\u901a\u5e38\u4f7f\u7528\u4e00\u4e2a\u8fd1\u4f3c\u516c\u5f0f\u4ee5\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>GELU(x) \u2248 0.5x \u00b7 (1 + tanh(\u221a(2\/\u03c0) \u00b7 (x + 0.044715\u00b7x\u00b3)))<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">GELU \u5728\u8d1f\u503c\u533a\u57df\u4fdd\u7559\u4e86\u5e73\u6ed1\u7684\u975e\u96f6\u8f93\u51fa\uff0c\u663e\u8457\u7f13\u89e3\u4e86 Dead ReLU \u95ee\u9898\u3002\u4f46\u5b83\u4ecd\u7136\u53ea\u662f\u4e00\u4e2a<strong>\u9010\u5143\u7d20\u7684\u975e\u7ebf\u6027\u53d8\u6362<\/strong>\u2014\u2014\u5bf9\u6bcf\u4e2a\u795e\u7ecf\u5143\u201c\u4e00\u89c6\u540c\u4ec1\u201d\u5730\u5e94\u7528\u540c\u4e00\u4e2a\u51fd\u6570\u66f2\u7ebf\uff0c\u6ca1\u6709\u8003\u8651\u5230\u4e0d\u540c\u7279\u5f81\u4e4b\u95f4\u7684\u4ea4\u4e92\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 \u7b2c\u4e09\u9636\u6bb5\uff1aGLU \u5bb6\u65cf\u4e0e\u95e8\u63a7\u9769\u547d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u771f\u6b63\u6539\u53d8\u6e38\u620f\u89c4\u5219\u7684\uff0c\u662f<strong>\u95e8\u63a7\u7ebf\u6027\u5355\u5143\uff08Gated Linear Unit, GLU\uff09<\/strong> \u7684\u5f15\u5165\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GLU \u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a<strong>\u8ba9\u6a21\u578b\u81ea\u5df1\u5b66\u4e60\u201c\u54ea\u4e9b\u4fe1\u606f\u53ef\u4ee5\u901a\u8fc7\u201d<\/strong> \uff0c\u800c\u4e0d\u662f\u7528\u4e00\u4e2a\u56fa\u5b9a\u7684\u51fd\u6570\u4e00\u5200\u5207\u3002GLU \u5c06\u8f93\u5165\u5206\u6210\u4e24\u8def\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>GLU(x) = (x \u00b7 W1 + b1) \u2299 \u03c3(x \u00b7 W2 + b2)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u503c\u8def\uff08Value Path\uff09<\/strong>\uff1a<code>x \u00b7 W1 + b1<\/code> \u627f\u8f7d\u5b9e\u9645\u7684\u4fe1\u606f\u5185\u5bb9<\/li>\n\n\n\n<li><strong>\u95e8\u8def\uff08Gate Path\uff09<\/strong>\uff1a<code>x \u00b7 W2 + b2<\/code> \u7ecf\u8fc7 Sigmoid \u6fc0\u6d3b\uff0c\u8f93\u51fa 0 \u5230 1 \u4e4b\u95f4\u7684\u201c\u95e8\u63a7\u4fe1\u53f7\u201d<\/li>\n\n\n\n<li><code>\u2299<\/code> \u8868\u793a\u9010\u5143\u7d20\u4e58\u6cd5\u2014\u2014\u95e8\u63a7\u4fe1\u53f7\u51b3\u5b9a\u4e86\u503c\u8def\u4e2d\u7684\u6bcf\u4e2a\u7ef4\u5ea6\u6709\u591a\u5c11\u80fd\u901a\u8fc7<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u8bbe\u8ba1\u8ba9 FFN \u62e5\u6709\u4e86<strong>\u52a8\u6001\u7684\u3001\u6570\u636e\u4f9d\u8d56\u7684\u4fe1\u606f\u8def\u7531\u80fd\u529b<\/strong>\u3002\u4e0d\u662f\u6240\u6709\u8f93\u5165\u90fd\u8d70\u540c\u4e00\u6761\u8def\u5f84\uff0c\u800c\u662f\u6839\u636e\u8f93\u5165\u672c\u8eab\u6765\u51b3\u5b9a\u4fe1\u606f\u6d41\u7684\u201c\u5f00\u5173\u201d\u548c\u201c\u5f3a\u5ea6\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GLU \u7684\u6210\u529f\u50ac\u751f\u4e86\u4e00\u7cfb\u5217\u53d8\u4f53\u3002Noam Shazeer \u5728\u8bba\u6587\u300aGLU Variants Improve Transformer\u300b\u4e2d\u7cfb\u7edf\u5bf9\u6bd4\u4e86\u591a\u79cd\u95e8\u63a7\u6fc0\u6d3b\u51fd\u6570\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u53d8\u4f53<\/th><th>\u95e8\u8def\u6fc0\u6d3b\u51fd\u6570<\/th><th>\u7279\u70b9<\/th><\/tr><\/thead><tbody><tr><td><strong>GLU<\/strong><\/td><td>Sigmoid<\/td><td>\u539f\u59cb\u7248\u672c\uff0c\u95e8\u63a7\u503c\u5728 0~1 \u4e4b\u95f4<\/td><\/tr><tr><td><strong>ReGLU<\/strong><\/td><td>ReLU<\/td><td>\u8ba1\u7b97\u7b80\u5355\uff0c\u4f46\u4ecd\u6709\u786c\u96f6\u8fb9\u754c<\/td><\/tr><tr><td><strong>GEGLU<\/strong><\/td><td>GELU<\/td><td>\u5e73\u6ed1\u95e8\u63a7\uff0c\u6548\u679c\u4f18\u4e8e ReGLU<\/td><\/tr><tr><td><strong>SwiGLU<\/strong><\/td><td>Swish\/SiLU<\/td><td><strong>\u6700\u7ec8\u80dc\u51fa\u7684\u65b9\u6848<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u8bc1\u660e\uff0c<strong>SwiGLU \u5728\u5927\u591a\u6570\u4efb\u52a1\u4e0a\u8868\u73b0\u6700\u4f18<\/strong>\uff0c\u6210\u4e3a\u73b0\u4ee3 LLM \u7684\u4e8b\u5b9e\u6807\u51c6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.4 SwiGLU\uff1a\u80dc\u51fa\u7684\u201c\u51a0\u519b\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SwiGLU\uff08Swish Gated Linear Unit\uff09<\/strong> \u662f GLU \u5bb6\u65cf\u4e2d\u6700\u7ec8\u8131\u9896\u800c\u51fa\u7684\u65b9\u6848\u3002\u5b83\u7531 Meta \u7684 LLaMA \u56e2\u961f\u7387\u5148\u5927\u89c4\u6a21\u91c7\u7528\uff0c\u968f\u540e\u88ab Qwen\u3001Mistral\u3001PaLM \u7b49\u51e0\u4e4e\u6240\u6709\u4e3b\u6d41\u6a21\u578b\u91c7\u7eb3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SwiGLU \u7684\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>SwiGLU(x) = (x \u00b7 W_gate) \u2299 Swish(x \u00b7 W_up)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d <code>Swish(x) = x \u00b7 \u03c3(x)<\/code>\uff08\u03c3 \u4e3a Sigmoid \u51fd\u6570\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba9\u6211\u4eec\u62c6\u89e3\u4e00\u4e0b\u8fd9\u4e2a\u516c\u5f0f\u7684\u542b\u4e49\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u8f93\u5165\u6295\u5f71<\/strong>\uff1a\u8f93\u5165 x \u88ab\u5206\u522b\u6295\u5f71\u5230\u4e24\u6761\u8def\u5f84\u2014\u2014\u95e8\u8def\uff08W_gate\uff09\u548c\u4e0a\u8def\uff08W_up\uff09\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u4e3a\u4e86\u53c2\u6570\u6548\u7387\uff0c\u8fd9\u4e24\u4e2a\u6295\u5f71\u901a\u5e38\u88ab\u5408\u5e76\u4e3a\u4e00\u4e2a\u77e9\u9635\uff0c\u518d\u62c6\u5206\u7ed3\u679c\u3002<\/li>\n\n\n\n<li><strong>\u95e8\u63a7\u751f\u6210<\/strong>\uff1a\u95e8\u8def\u7ecf\u8fc7 Swish \u6fc0\u6d3b\u3002Swish \u7684\u7279\u70b9\u662f\uff1a\u5728\u6b63\u533a\u57df\u8fd1\u4f3c\u7ebf\u6027\uff0c\u5728\u8d1f\u533a\u57df\u5e73\u6ed1\u8870\u51cf\u5230\u96f6\u9644\u8fd1\u4f46\u4e0d\u5b8c\u5168\u4e3a\u96f6\uff0c\u4e14\u662f\u975e\u5355\u8c03\u7684\uff08\u6709\u4e00\u4e2a\u8f7b\u5fae\u7684\u201c\u51f9\u9677\u201d\uff09\u3002\u8fd9\u79cd\u5e73\u6ed1\u3001\u975e\u5355\u8c03\u7684\u7279\u6027\u8ba9\u5b83\u80fd\u591f\u6355\u6349\u66f4\u590d\u6742\u7684\u975e\u7ebf\u6027\u5173\u7cfb\u3002<\/li>\n\n\n\n<li><strong>\u9010\u5143\u7d20\u4e58\u6cd5<\/strong>\uff1a\u6fc0\u6d3b\u540e\u7684\u95e8\u63a7\u4fe1\u53f7\u4e0e\u4e0a\u8def\u7684\u503c\u4fe1\u53f7\u9010\u5143\u7d20\u76f8\u4e58\uff0c\u5b9e\u73b0\u201c\u9009\u62e9\u6027\u901a\u8fc7\u201d\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48 SwiGLU \u6bd4 GELU \u548c ReLU \u66f4\u597d\uff1f<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u95e8\u63a7\u673a\u5236<\/strong>\uff1a\u8ba9\u6a21\u578b\u80fd\u591f\u52a8\u6001\u51b3\u5b9a\u4fe1\u606f\u6d41\uff0c\u800c\u4e0d\u662f\u88ab\u56fa\u5b9a\u7684\u51fd\u6570\u675f\u7f1a<\/li>\n\n\n\n<li><strong>\u5e73\u6ed1\u6027<\/strong>\uff1aSwish \u5728\u8d1f\u503c\u533a\u57df\u4e0d\u662f\u76f4\u63a5\u7f6e\u96f6\uff0c\u4fdd\u7559\u4e86\u68af\u5ea6\uff0c\u907f\u514d\u4e86\u201c\u6b7b\u795e\u7ecf\u5143\u201d<\/li>\n\n\n\n<li><strong>\u975e\u5355\u8c03\u6027<\/strong>\uff1aSwish \u7684\u8f7b\u5fae\u201c\u51f9\u9677\u201d\u8ba9\u5b83\u5728\u67d0\u4e9b\u8f93\u5165\u8303\u56f4\u5185\u53ef\u4ee5\u6709\u66f4\u5f3a\u7684\u975e\u7ebf\u6027\u8868\u8fbe<\/li>\n\n\n\n<li><strong>\u8ba1\u7b97\u6548\u7387<\/strong>\uff1aSwish \u6bd4 GELU \u8ba1\u7b97\u66f4\u7b80\u5355\uff0c\u5728\u5b9e\u8df5\u4e2d\u6548\u679c\u5dee\u5f02\u4e0d\u5927\u4f46\u5de5\u7a0b\u4e0a\u66f4\u53cb\u597d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u76ee\u524d\uff0cSwiGLU \u5df2\u7ecf\u6210\u4e3a LLM FFN \u7684\u201c\u6807\u51c6\u914d\u7f6e\u201d\uff0c\u51e0\u4e4e\u6240\u6709\u65b0\u53d1\u5e03\u7684\u5f00\u6e90\u5927\u6a21\u578b\u90fd\u5728\u4f7f\u7528\u5b83\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001FFN \u4e2d\u7684\u7ef4\u5ea6\u6269\u5f20\uff1a\u4e3a\u4ec0\u4e48\u662f 4 \u500d\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u770b\u8fc7\u6a21\u578b\u914d\u7f6e\u6587\u4ef6\u7684\u670b\u53cb\u53ef\u80fd\u6ce8\u610f\u8fc7\u4e00\u4e2a\u53c2\u6570\uff1a<code>intermediate_size<\/code>\uff08\u6216 <code>ffn_dim<\/code>\uff09\uff0c\u901a\u5e38\u662f <code>hidden_size<\/code> \u7684 <strong>4 \u500d<\/strong>\u3002\u6bd4\u5982 LLaMA-7B \u7684 hidden_size \u662f 4096\uff0cintermediate_size \u5219\u662f 11008\uff08\u7ea6\u4e3a 4096 \u00d7 2.69\uff0cSwiGLU \u4e0b\u7684\u8c03\u6574\uff09\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u6807\u51c6 FFN \u7684 4 \u500d\u89c4\u5219<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u7ecf\u5178 Transformer \u4e2d\uff0cFFN \u7684\u7ed3\u6784\u662f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u8f93\u5165\u7ef4\u5ea6 d \u2192 \u6295\u5f71\u5230 4d \u2192 \u6fc0\u6d3b\u51fd\u6570 \u2192 \u6295\u5f71\u56de d<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u201c4 \u500d\u201d\u4e0d\u662f\u4e00\u4e2a\u968f\u610f\u9009\u7684\u6570\u5b57\uff0c\u800c\u662f\u7ecf\u8fc7\u5927\u91cf\u5b9e\u9a8c\u9a8c\u8bc1\u7684\u7ecf\u9a8c\u6700\u4f18\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\u8981\u6269\u5927\u4e2d\u95f4\u7ef4\u5ea6\uff1f\u56e0\u4e3a FFN \u9700\u8981\u8db3\u591f\u7684\u201c\u5bb9\u91cf\u201d\u6765\u5b58\u50a8\u77e5\u8bc6\u548c\u5b8c\u6210\u975e\u7ebf\u6027\u53d8\u6362\u3002\u5982\u679c\u4e2d\u95f4\u7ef4\u5ea6\u8fc7\u5c0f\uff0cFFN \u4f1a\u6210\u4e3a\u4e00\u4e2a\u4fe1\u606f\u74f6\u9888\uff0c\u9650\u5236\u6a21\u578b\u7684\u5b66\u4e60\u80fd\u529b\u3002\u5982\u679c\u8fc7\u5927\uff0c\u53c2\u6570\u91cf\u548c\u8ba1\u7b97\u91cf\u4f1a\u6025\u5267\u81a8\u80c0\uff0c\u8fb9\u9645\u6536\u76ca\u9012\u51cf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 SwiGLU \u4e0b\u7684\u7ef4\u5ea6\u8c03\u6574<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53 FFN \u4ece GELU \u5347\u7ea7\u5230 SwiGLU \u65f6\uff0c\u51fa\u73b0\u4e86\u4e00\u4e2a\u7ec6\u8282\u95ee\u9898\uff1a<strong>SwiGLU \u9700\u8981\u4e09\u4e2a\u6743\u91cd\u77e9\u9635<\/strong>\uff08\u8f93\u5165\u5230\u95e8\u8def\u3001\u8f93\u5165\u5230\u503c\u8def\u3001\u5408\u5e76\u540e\u5230\u8f93\u51fa\uff09\uff0c\u6bd4\u6807\u51c6 FFN \u591a\u4e86\u4e00\u4e2a\u3002\u5982\u679c\u4fdd\u6301\u540c\u6837\u7684 4d \u4e2d\u95f4\u7ef4\u5ea6\uff0c\u53c2\u6570\u91cf\u4f1a\u589e\u52a0 50%\uff0cFLOPs \u4e5f\u4f1a\u76f8\u5e94\u589e\u957f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u516c\u5e73\u5bf9\u6bd4\uff0cSwiGLU \u8bba\u6587\u5efa\u8bae\u5c06\u4e2d\u95f4\u7ef4\u5ea6\u8c03\u6574\u4e3a\u539f\u6765\u7684 <strong>2\/3 \u00d7 4d \u2248 2.67d<\/strong>\uff0c\u4ee5\u4fdd\u8bc1\u53c2\u6570\u91cf\u548c\u8ba1\u7b97\u91cf\u4e0e\u6807\u51c6 FFN \u5927\u81f4\u76f8\u5f53\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48 LLaMA \u7684 intermediate_size \u662f 11008\uff08\u7ea6\u4e3a 4096 \u00d7 2.69\uff09\u800c\u4e0d\u662f 16384 \u7684\u539f\u56e0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u63d0\u9192\u6211\u4eec\uff1a\u770b\u6a21\u578b\u53c2\u6570\u91cf\u7684\u65f6\u5019\uff0c\u4e0d\u80fd\u53ea\u770b\u5c42\u6570\u548c hidden_size\uff0cFFN \u7684\u7ed3\u6784\u548c\u7ef4\u5ea6\u8bbe\u7f6e\u540c\u6837\u91cd\u8981\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001FFN \u4e0e\u53c2\u6570\u91cf\uff1a\u6a21\u578b\u201c\u4f53\u91cd\u201d\u7684\u4e3b\u8981\u6765\u6e90<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 FFN \u7684\u53c2\u6570\u5360\u6bd4<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u4e00\u4e2a\u6807\u51c6\u7684 Transformer \u5c42\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Attention \u90e8\u5206\uff1a4 \u4e2a\u6743\u91cd\u77e9\u9635\uff08Q\u3001K\u3001V\u3001O\uff09\uff0c\u6bcf\u4e2a\u5c3a\u5bf8\u4e3a d \u00d7 d\uff0c\u53c2\u6570\u91cf\u7ea6\u4e3a 4d\u00b2\uff08\u4e0d\u8ba1\u504f\u7f6e\uff09<\/li>\n\n\n\n<li>FFN \u90e8\u5206\uff1a2 \u4e2a\u6743\u91cd\u77e9\u9635\uff08\u6807\u51c6 FFN\uff09\u6216 3 \u4e2a\uff08SwiGLU\uff09\uff0c\u5c3a\u5bf8\u5206\u522b\u4e3a d \u00d7 4d \u548c 4d \u00d7 d\uff0c\u53c2\u6570\u91cf\u7ea6\u4e3a 8d\u00b2\uff08\u6807\u51c6 FFN\uff09\u6216\u7ea6 8d\u00b2\uff08SwiGLU \u8c03\u6574\u540e\uff09<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c<strong>\u5355\u5c42 FFN \u7684\u53c2\u6570\u91cf\u5927\u7ea6\u662f Attention \u7684 2 \u500d<\/strong>\u3002\u53e0\u52a0\u591a\u5c42\u7684 Embedding \u548c\u6700\u7ec8\u8f93\u51fa\u5c42\u540e\uff0cFFN \u7684\u603b\u4f53\u5360\u6bd4\u7ea6\u4e3a <strong>2\/3<\/strong> \u5de6\u53f3\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u4e3a\u4ec0\u4e48 FFN \u9700\u8981\u8fd9\u4e48\u591a\u53c2\u6570\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u95ee\u9898\u89e6\u53ca\u4e86 LLM \u8bbe\u8ba1\u7684\u6838\u5fc3\u6743\u8861\u3002Attention \u8d1f\u8d23\u201c\u4fe1\u606f\u8def\u7531\u201d\u2014\u2014\u51b3\u5b9a\u54ea\u4e9b Token \u4e4b\u95f4\u8981\u4ea4\u6362\u4fe1\u606f\u3002\u5b83\u7684\u590d\u6742\u5ea6\u662f O(n\u00b2)\uff08n \u4e3a\u5e8f\u5217\u957f\u5ea6\uff09\uff0c\u6240\u4ee5 Attention \u7684\u53c2\u6570\u4e0d\u5b9c\u8fc7\u591a\uff0c\u5426\u5219\u8ba1\u7b97\u91cf\u4f1a\u7206\u70b8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">FFN \u5219\u4e0d\u540c\u3002\u5b83\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4e0e\u5e8f\u5217\u957f\u5ea6\u662f<strong>\u7ebf\u6027\u7684<\/strong>\uff08\u6bcf\u4e2a Token \u72ec\u7acb\u5904\u7406\uff09\uff0c\u56e0\u6b64\u53ef\u4ee5\u5728\u4e0d\u663e\u8457\u589e\u52a0\u5e8f\u5217\u7ea7\u8ba1\u7b97\u8d1f\u62c5\u7684\u524d\u63d0\u4e0b\uff0c\u5927\u91cf\u5806\u53e0\u53c2\u6570\u6765\u63d0\u5347\u6a21\u578b\u7684\u201c\u77e5\u8bc6\u5bb9\u91cf\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f Transformer \u67b6\u6784\u8bbe\u8ba1\u7684\u7cbe\u5999\u4e4b\u5904\uff1a<strong>\u7528\u53c2\u6570\u5c11\u7684 Attention \u505a\u201c\u7a00\u758f\u8def\u7531\u201d\uff0c\u7528\u53c2\u6570\u591a\u7684 FFN \u505a\u201c\u5bc6\u96c6\u5b58\u50a8\u201d<\/strong>\u3002\u4e24\u8005\u914d\u5408\uff0c\u65e2\u4fdd\u8bc1\u4e86\u957f\u5e8f\u5217\u7684\u5904\u7406\u6548\u7387\uff0c\u53c8\u4fdd\u8bc1\u4e86\u8db3\u591f\u7684\u6a21\u578b\u5bb9\u91cf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u4ece FFN \u5230 MoE\uff1a\u5f53\u201c\u4e00\u4e2a\u4e13\u5bb6\u201d\u4e0d\u591f\u7528\u65f6<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u7a20\u5bc6 FFN \u7684\u56f0\u5883<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6807\u51c6 FFN \u867d\u7136\u5f3a\u5927\uff0c\u4f46\u6709\u4e00\u4e2a\u6839\u672c\u5c40\u9650\uff1a<strong>\u65e0\u8bba\u8f93\u5165\u662f\u4ec0\u4e48\uff0c\u6240\u6709\u53c2\u6570\u90fd\u4f1a\u88ab\u6fc0\u6d3b<\/strong>\u3002\u8fd9\u610f\u5473\u7740\uff0c\u6a21\u578b\u89c4\u6a21\uff08\u603b\u53c2\u6570\u91cf\uff09\u548c\u6bcf\u6b21\u63a8\u7406\u7684\u8ba1\u7b97\u91cf\uff08\u6fc0\u6d3b\u53c2\u6570\u91cf\uff09\u662f<strong>\u4e25\u683c\u7ed1\u5b9a\u7684<\/strong>\u2014\u2014\u66f4\u5927\u7684\u6a21\u578b\u5fc5\u7136\u5e26\u6765\u66f4\u6162\u7684\u63a8\u7406\u548c\u66f4\u9ad8\u7684\u663e\u5b58\u6d88\u8017\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740\u6a21\u578b\u89c4\u6a21\u4ece 7B \u589e\u957f\u5230 70B \u751a\u81f3 400B\uff0c\u8fd9\u79cd\u201c\u7b49\u6bd4\u4f8b\u201d\u7684\u589e\u957f\u53d8\u5f97\u8d8a\u6765\u8d8a\u96be\u4ee5\u627f\u53d7\u3002\u7814\u7a76\u8005\u4eec\u5f00\u59cb\u601d\u8003\uff1a\u80fd\u4e0d\u80fd\u8ba9\u6a21\u578b\u62e5\u6709\u5de8\u5927\u7684\u603b\u53c2\u6570\u91cf\uff0c\u4f46\u6bcf\u6b21\u63a8\u7406\u65f6\u53ea\u6fc0\u6d3b\u5176\u4e2d\u4e00\u5c0f\u90e8\u5206\uff1f<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 MoE \u7684\u6838\u5fc3\u601d\u60f3\uff1a\u4e13\u5bb6\u5206\u5de5\uff0c\u6309\u9700\u6fc0\u6d3b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>MoE\uff08Mixture of Experts\uff0c\u6df7\u5408\u4e13\u5bb6\uff09<\/strong> \u6b63\u662f\u5bf9\u8fd9\u4e00\u95ee\u9898\u7684\u56de\u5e94\u3002MoE \u5c42\u7684\u7ed3\u6784\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u591a\u4e2a\u4e13\u5bb6 FFN<\/strong>\uff1aMoE \u5c42\u4e0d\u518d\u53ea\u6709\u4e00\u4e2a FFN\uff0c\u800c\u662f\u5305\u542b N \u4e2a\u72ec\u7acb\u7684 FFN\uff08\u79f0\u4e3a\u201c\u4e13\u5bb6\u201d\uff09\uff0c\u6bcf\u4e2a\u4e13\u5bb6\u90fd\u6709\u81ea\u5df1\u7684\u5b8c\u6574\u53c2\u6570\u3002<\/li>\n\n\n\n<li><strong>\u8def\u7531\u5668\uff08Router\/Gating Network\uff09<\/strong> \uff1a\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u7f51\u7edc\uff0c\u8d1f\u8d23\u4e3a\u6bcf\u4e2a\u8f93\u5165 Token \u9009\u62e9\u6700\u5408\u9002\u7684\u4e13\u5bb6\u3002<\/li>\n\n\n\n<li><strong>\u7a00\u758f\u6fc0\u6d3b<\/strong>\uff1a\u5bf9\u4e8e\u6bcf\u4e2a Token\uff0c\u8def\u7531\u5668\u53ea\u9009\u62e9 Top-K \u4e2a\u4e13\u5bb6\uff08\u901a\u5e38 K=1 \u6216 2\uff09\uff0c\u53ea\u6709\u88ab\u9009\u4e2d\u7684\u4e13\u5bb6\u88ab\u6fc0\u6d3b\u53c2\u4e0e\u8ba1\u7b97\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u4e00\u6765\uff0c<strong>\u6a21\u578b\u7684\u603b\u53c2\u6570\u91cf = \u6240\u6709\u4e13\u5bb6\u7684\u53c2\u6570\u4e4b\u548c<\/strong>\uff0c\u800c<strong>\u6bcf\u6b21\u63a8\u7406\u7684\u6fc0\u6d3b\u53c2\u6570\u91cf \u2248 \u88ab\u6fc0\u6d3b\u4e13\u5bb6\u6570 \/ \u603b\u4e13\u5bb6\u6570 \u00d7 \u603b\u53c2\u6570<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4e2a\u4f8b\u5b50\uff1a\u4e00\u4e2a MoE \u6a21\u578b\u6709 64 \u4e2a\u4e13\u5bb6\uff0c\u6bcf\u6b21\u6fc0\u6d3b Top-2\uff0c\u90a3\u4e48\u5b83\u7684\u603b\u53c2\u6570\u91cf\u662f\u7a20\u5bc6\u6a21\u578b\u7684 64 \u500d\uff0c\u4f46\u6bcf\u6b21\u63a8\u7406\u7684\u8ba1\u7b97\u91cf\u53ea\u662f\u7a20\u5bc6\u6a21\u578b\u7684 2 \u500d\u3002Voyage AI \u7684\u5b9e\u8df5\u8868\u660e\uff0c\u4f7f\u7528 MoE \u67b6\u6784\u53ef\u4ee5\u5728\u51e0\u4e4e\u76f8\u540c\u7684\u68c0\u7d22\u7cbe\u5ea6\u4e0b\uff0c\u5c06\u6fc0\u6d3b\u53c2\u6570\u91cf\u51cf\u5c11 75%\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">MoE \u7684\u76f4\u89c9\u662f\uff1a<strong>\u5c06\u201c\u77e5\u8bc6\u5bb9\u91cf\u201d\u4e0e\u201c\u8ba1\u7b97\u6210\u672c\u201d\u89e3\u8026<\/strong>\u3002\u6a21\u578b\u53ef\u4ee5\u50cf\u4e00\u5ea7\u5de8\u578b\u56fe\u4e66\u9986\u4e00\u6837\u5b58\u50a8\u6d77\u91cf\u77e5\u8bc6\uff0c\u4f46\u6bcf\u6b21\u67e5\u8d44\u6599\u65f6\u53ea\u9700\u8981\u8d70\u8fdb\u5176\u4e2d\u4e00\u4e2a\u4e66\u67b6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 \u7ecf\u5178 MoE \u7684\u4e24\u5927\u95ee\u9898<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">MoE \u770b\u4f3c\u5b8c\u7f8e\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d\u66b4\u9732\u4e86\u4e24\u4e2a\u6839\u672c\u95ee\u9898\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u95ee\u9898\u4e00\uff1a\u77e5\u8bc6\u6df7\u6742\uff08Knowledge Hybridity\uff09<\/strong> \u3002\u7ecf\u5178 MoE \u901a\u5e38\u53ea\u6709 8 \u5230 16 \u4e2a\u4e13\u5bb6\u3002\u4e13\u5bb6\u6570\u91cf\u5c11\u610f\u5473\u7740\u6bcf\u4e2a\u4e13\u5bb6\u5fc5\u987b\u5904\u7406\u6781\u5176\u591a\u6837\u5316\u7684\u8f93\u5165\u2014\u2014\u4e00\u4e2a\u4e13\u5bb6\u53ef\u80fd\u540c\u65f6\u8981\u5904\u7406\u201c\u6570\u5b66\u63a8\u7406\u201d\u548c\u201c\u60c5\u611f\u5206\u6790\u201d\u7684 Token\u3002\u7ed3\u679c\u5c31\u662f\uff1a\u4e13\u5bb6\u88ab\u8feb\u5b66\u4e60\u6df7\u5408\u5728\u4e00\u8d77\u7684\u4e0d\u540c\u9886\u57df\u77e5\u8bc6\uff0c\u96be\u4ee5\u771f\u6b63\u201c\u4e13\u7cbe\u201d\u67d0\u4e00\u65b9\u5411\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u95ee\u9898\u4e8c\uff1a\u77e5\u8bc6\u5197\u4f59\uff08Knowledge Redundancy\uff09<\/strong> \u3002\u4e0d\u540c\u4e13\u5bb6\u90fd\u9700\u8981\u638c\u63e1\u4e00\u4e9b\u5171\u6709\u7684\u57fa\u7840\u80fd\u529b\uff0c\u6bd4\u5982\u201c\u57fa\u672c\u8bed\u6cd5\u89c4\u5219\u201d\u201c\u5e38\u89c1\u8bcd\u6c47\u542b\u4e49\u201d\u7b49\u3002\u7ed3\u679c\u662f\uff0c\u8fd9\u4e9b\u57fa\u7840\u80fd\u529b\u88ab\u591a\u4e2a\u4e13\u5bb6\u91cd\u590d\u5b58\u50a8\uff0c\u9020\u6210\u4e86\u53c2\u6570\u6d6a\u8d39\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.4 DeepSeekMoE\uff1a\u7ec6\u7c92\u5ea6\u4e13\u5bb6\u4e0e\u5171\u4eab\u4e13\u5bb6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek \u56e2\u961f\u9488\u5bf9\u4e0a\u8ff0\u95ee\u9898\u63d0\u51fa\u4e86\u4e24\u5927\u521b\u65b0\u7b56\u7565\uff0c\u6210\u4e3a MoE \u67b6\u6784\u6f14\u8fdb\u4e2d\u7684\u91cc\u7a0b\u7891\u5f0f\u5de5\u4f5c\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b56\u7565\u4e00\uff1a\u7ec6\u7c92\u5ea6\u4e13\u5bb6\u5206\u5272\uff08Fine-Grained Expert Segmentation\uff09<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeekMoE \u5c06\u6bcf\u4e2a\u4e13\u5bb6\u6309\u7167 FFN \u7684\u4e2d\u95f4\u9690\u85cf\u7ef4\u5ea6\u8fdb\u4e00\u6b65\u62c6\u5206\u4e3a\u591a\u4e2a\u66f4\u5c0f\u7684\u4e13\u5bb6\u3002\u4f8b\u5982\uff0c\u539f\u6765\u7684 16 \u4e2a\u4e13\u5bb6\uff0c\u6bcf\u4e2a\u62c6\u5206\u4e3a 4 \u4e2a\uff0c\u5f97\u5230 64 \u4e2a\u7ec6\u7c92\u5ea6\u4e13\u5bb6\uff0c\u540c\u65f6\u6fc0\u6d3b\u6570\u91cf\u4e5f\u76f8\u5e94\u589e\u52a0\uff08\u5982\u4ece 2 \u4e2a\u589e\u81f3 8 \u4e2a\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u597d\u5904\u662f\u4ec0\u4e48\uff1f\u4e13\u5bb6\u6570\u91cf\u5927\u5e45\u589e\u52a0\uff0c\u6bcf\u4e2a\u4e13\u5bb6\u53ef\u4ee5\u66f4\u52a0\u201c\u4e13\u7cbe\u201d\uff0c\u77e5\u8bc6\u6df7\u6742\u95ee\u9898\u5f97\u5230\u663e\u8457\u7f13\u89e3\u3002\u6548\u679c\u975e\u5e38\u60ca\u8273\uff1aDeepSeekMoE \u7684 16B \u6a21\u578b\u4ec5\u7528 40% \u7684\u8ba1\u7b97\u91cf\u5c31\u8fbe\u5230\u4e86 DeepSeek 7B \u548c LLaMA2 7B \u7684\u6027\u80fd\u6c34\u5e73\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b56\u7565\u4e8c\uff1a\u5171\u4eab\u4e13\u5bb6\u9694\u79bb\uff08Shared Expert Isolation\uff09<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeekMoE \u5728\u8def\u7531\u4e13\u5bb6\u4e4b\u5916\uff0c\u989d\u5916\u8bbe\u7f6e\u4e86<strong>\u5171\u4eab\u4e13\u5bb6<\/strong>\u2014\u2014\u8fd9\u4e9b\u4e13\u5bb6\u7684\u8f93\u51fa<strong>\u59cb\u7ec8\u88ab\u4fdd\u7559<\/strong>\uff0c\u4e0d\u7ecf\u8fc7\u8def\u7531\u9009\u62e9\u3002\u5171\u4eab\u4e13\u5bb6\u8d1f\u8d23\u5b58\u50a8\u6240\u6709 Token \u90fd\u9700\u8981\u7684\u201c\u901a\u7528\u77e5\u8bc6\u201d\uff0c\u800c\u8def\u7531\u4e13\u5bb6\u5219\u53ef\u4ee5\u4e13\u6ce8\u4e8e\u5404\u81ea\u7684\u201c\u4e13\u4e1a\u9886\u57df\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u65b9\u6cd5\u5b8c\u7f8e\u89e3\u51b3\u4e86\u77e5\u8bc6\u5197\u4f59\u95ee\u9898\uff1a\u901a\u7528\u80fd\u529b\u7531\u5171\u4eab\u4e13\u5bb6\u7edf\u4e00\u63d0\u4f9b\uff0c\u5404\u8def\u7531\u4e13\u5bb6\u4e0d\u518d\u9700\u8981\u201c\u91cd\u590d\u9020\u8f6e\u5b50\u201d\uff0c\u53c2\u6570\u6548\u7387\u5927\u5e45\u63d0\u5347\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u5916\uff0cDeepSeekMoE \u8fd8\u91c7\u7528\u4e86<strong>\u5f52\u4e00\u5316 Sigmoid \u95e8\u63a7\uff08Normalized Sigmoid Gating\uff09<\/strong> \u6765\u66ff\u4ee3\u4f20\u7edf\u7684 Softmax \u95e8\u63a7\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u4e13\u5bb6\u5229\u7528\u7387\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u5de5\u7a0b\u6548\u679c\u6765\u770b\uff0cDeepSeek-V3 \u5c55\u793a\u4e86 MoE \u67b6\u6784\u7684\u6781\u81f4\u6f5c\u529b\uff1a671B \u603b\u53c2\u6570\uff0c\u63a8\u7406\u65f6\u4ec5\u6fc0\u6d3b 37B\uff0c\u5374\u80fd\u5b9e\u73b0 GPT-4 \u7ea7\u522b\u7684\u6027\u80fd\u3002\u5230 2025 \u5e74\u5e95\uff0cMoE \u67b6\u6784\u5df2\u7ecf\u5360\u636e\u4e86\u8d85\u8fc7 60% \u7684\u5f00\u6e90 AI \u6a21\u578b\u53d1\u5e03\u91cf\u3002\u76ee\u524d\uff0c\u6240\u6709\u62ab\u9732\u67b6\u6784\u7684 Top-10 \u524d\u6cbf\u6a21\u578b\u5168\u90e8\u90fd\u662f MoE\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001FFN \u4e0e\u8bad\u7ec3\u7a33\u5b9a\u6027\uff1aPre-Norm \u7684\u5fc5\u8981\u6027<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u4e0a\u4e00\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u63d0\u5230\u4e86 Pre-Norm \u662f\u73b0\u4ee3 LLM \u7684\u6807\u914d\u2014\u2014\u5f52\u4e00\u5316\u653e\u5728 Attention \u548c FFN <strong>\u4e4b\u524d<\/strong>\uff0c\u800c\u975e\u539f\u59cb Transformer \u7684 Post-Norm\uff08\u653e\u5728\u4e4b\u540e\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Pre-Norm \u5bf9 FFN \u5c42\u540c\u6837\u91cd\u8981\u3002\u968f\u7740\u5c42\u6570\u589e\u52a0\u5230\u51e0\u5341\u751a\u81f3\u4e0a\u767e\u5c42\uff0cPost-Norm \u5bb9\u6613\u5bfc\u81f4\u68af\u5ea6\u5728\u53cd\u5411\u4f20\u64ad\u4e2d\u9010\u6e10\u6d88\u5931\uff0c\u4f7f\u5f97\u6df1\u5c42\u7f51\u7edc\u7684\u8bad\u7ec3\u6781\u4e0d\u7a33\u5b9a\u3002\u800c Pre-Norm \u901a\u8fc7\u5728\u6bcf\u4e2a\u5b50\u5c42\uff08Attention \u548c FFN\uff09\u4e4b\u524d\u505a\u5f52\u4e00\u5316\uff0c\u8ba9\u68af\u5ea6\u4f20\u64ad\u66f4\u52a0\u5e73\u6ed1\uff0c\u4ece\u800c\u652f\u6301\u66f4\u6df1\u7684\u7f51\u7edc\u7ed3\u6784\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u70b9\u5728 FFN \u5c42\u5c24\u4e3a\u5173\u952e\uff0c\u56e0\u4e3a FFN \u7684\u53c2\u6570\u91cf\u6700\u5927\uff0c\u5bf9\u68af\u5ea6\u4e0d\u7a33\u5b9a\u4e5f\u66f4\u52a0\u654f\u611f\u3002\u91c7\u7528 Pre-Norm \u540e\uff0c\u5373\u4f7f\u662f 70B\u300180 \u5c42\u7684\u6a21\u578b\u4e5f\u80fd\u7a33\u5b9a\u8bad\u7ec3\u6536\u655b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u6700\u65b0\u7814\u7a76\u524d\u6cbf\uff1aFFN \u67b6\u6784\u7684\u201c\u8fdb\u5316\u6811\u201d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u9664\u4e86 MoE \u4e4b\u5916\uff0cFFN \u7684\u7814\u7a76\u5728\u8fc7\u53bb\u4e00\u5e74\u4e2d\u6d8c\u73b0\u51fa\u4e86\u591a\u4e2a\u4ee4\u4eba\u5174\u594b\u7684\u65b0\u65b9\u5411\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.1 MoC\uff1a\u5229\u7528 SwiGLU \u539f\u751f\u7a00\u758f\u6027<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Mixture-of-Channels\uff08MoC\uff09\u662f\u4e00\u7bc7 2025 \u5e74\u7684\u5de5\u4f5c\uff0c\u5b83\u63d0\u51fa\u4e86\u4e00\u4e2a\u975e\u5e38\u806a\u660e\u7684\u60f3\u6cd5\uff1a<strong>\u5229\u7528 SwiGLU \u81ea\u5e26\u7684\u95e8\u63a7\u4fe1\u53f7\u6765\u76f4\u63a5\u5b9e\u73b0 FFN \u7684\u7a00\u758f\u5316<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u8005\u9996\u5148\u5bf9 LLM \u8fdb\u884c\u4e86\u8be6\u7ec6\u7684\u5185\u5b58\u5206\u6790\uff0c\u53d1\u73b0\u5728\u5f15\u5165 FlashAttention \u540e\uff0c<strong>FFN \u7684\u6fc0\u6d3b\u503c\u5df2\u6210\u4e3a\u663e\u5b58\u7684\u4e3b\u8981\u74f6\u9888<\/strong>\u3002MoC \u7684\u505a\u6cd5\u662f\uff1a\u5728\u6bcf\u4e2a Token \u7684 FFN \u8ba1\u7b97\u4e2d\uff0c\u53ea\u6fc0\u6d3b\u7531 SwiGLU \u95e8\u63a7\u4fe1\u53f7\u51b3\u5b9a\u7684 Top-K \u6700\u76f8\u5173\u901a\u9053\uff0c\u4ece\u800c\u5927\u5e45\u51cf\u5c11\u6fc0\u6d3b\u5185\u5b58\u7684\u5360\u7528\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u9700\u8981\u989d\u5916\u7684\u8def\u7531\u7f51\u7edc\uff0c\u76f4\u63a5\u201c\u5c31\u5730\u53d6\u6750\u201d\uff0c\u5728\u9884\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u90fd\u80fd\u663e\u8457\u63d0\u5347\u6548\u7387\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.2 \u4ece\u201c\u4e13\u5bb6\u201d\u5230\u201c\u8f68\u8ff9\u201d\uff1aMoE \u53ef\u89e3\u91ca\u6027\u7684\u65b0\u53d1\u73b0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u7bc7 2026 \u5e74\u7684\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u98a0\u8986\u6027\u7684\u89c2\u70b9\uff1a<strong>MoE \u4e2d\u771f\u6b63\u6709\u610f\u4e49\u7684\u6700\u5c0f\u5355\u5143\u4e0d\u662f\u5355\u4e2a\u4e13\u5bb6\uff0c\u800c\u662f\u4e13\u5bb6\u5728\u4e0d\u540c\u5c42\u4e4b\u95f4\u5f62\u6210\u7684\u201c\u8f68\u8ff9\u201d<\/strong> \u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u53d1\u73b0\uff0c\u867d\u7136\u5355\u4e2a\u4e13\u5bb6\u662f\u201c\u591a\u4e49\u7684\u201d\u2014\u2014\u540c\u4e00\u4e2a\u4e13\u5bb6\u53ef\u80fd\u5904\u7406\u591a\u4e2a\u5b8c\u5168\u4e0d\u76f8\u5173\u7684\u6982\u5ff5\u2014\u2014\u4f46\u4e00\u4e2a Token \u5728\u591a\u4e2a\u5c42\u4e4b\u95f4\u4f9d\u6b21\u8bbf\u95ee\u7684\u4e13\u5bb6\u5e8f\u5217\uff0c\u5374\u8868\u73b0\u51fa\u4e86\u9ad8\u5ea6\u4e00\u81f4\u7684\u8bed\u4e49\u6a21\u5f0f\u3002\u540c\u4e00 Token \u5728\u4e0d\u540c\u4e0a\u4e0b\u6587\u4e2d\uff08\u4f8b\u5982\u5192\u53f7\u201c:\u201d\u5728\u4e0d\u540c\u7528\u6cd5\u4e0b\uff09\u4f1a\u6cbf\u7740\u622a\u7136\u4e0d\u540c\u7684\u201c\u4e13\u5bb6\u8f68\u8ff9\u201d\u524d\u8fdb\uff0c\u800c\u76f8\u540c\u8bed\u4e49\u529f\u80fd\u7684 Token \u4f1a\u805a\u5408\u6210\u6e05\u6670\u7684\u8f68\u8ff9\u7c07\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u9879\u7814\u7a76\u4e3a MoE \u7684\u53ef\u89e3\u91ca\u6027\u63d0\u4f9b\u4e86\u5168\u65b0\u7684\u89c6\u89d2\uff1a\u4e0d\u8981\u95ee\u201c\u8fd9\u4e2a\u4e13\u5bb6\u662f\u505a\u4ec0\u4e48\u7684\u201d\uff0c\u800c\u8981\u95ee\u201c\u8fd9\u6761\u4e13\u5bb6\u8def\u5f84\u901a\u5411\u54ea\u91cc\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.3 FIRM-MoE\uff1a\u4e13\u5bb6\u7ec6\u7c92\u5ea6\u5378\u8f7d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740 MoE \u6a21\u578b\u89c4\u6a21\u7ee7\u7eed\u81a8\u80c0\uff08\u6570\u767e\u4ebf\u5230\u6570\u5343\u4ebf\u53c2\u6570\uff09\uff0c\u5355\u4e2a GPU \u751a\u81f3\u5355\u4e2a\u8282\u70b9\u7684\u663e\u5b58\u90fd\u4e0d\u591f\u7528\u3002FIRM-MoE\uff08Fine-Grained Expert Decomposition for Resource-Adaptive MoE Inference\uff09\u63d0\u51fa\u4e86\u4e00\u5957\u7ec6\u7c92\u5ea6\u7684\u4e13\u5bb6\u5378\u8f7d\u6846\u67b6\uff0c\u5141\u8bb8\u6a21\u578b\u6839\u636e\u786c\u4ef6\u8d44\u6e90\u52a8\u6001\u51b3\u5b9a\u54ea\u4e9b\u4e13\u5bb6\u653e\u5728 GPU \u663e\u5b58\u4e2d\u3001\u54ea\u4e9b\u653e\u5728 CPU \u5185\u5b58\u4e2d\uff0c\u5728\u4fdd\u6301\u7cbe\u5ea6\u7684\u540c\u65f6\u7075\u6d3b\u9002\u5e94\u4e0d\u540c\u7684\u90e8\u7f72\u73af\u5883\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u5de5\u7a0b\u5b9e\u8df5\uff1a\u5982\u4f55\u914d\u7f6e FFN\uff1f<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">8.1 \u573a\u666f\u9a71\u52a8\u7684\u9009\u62e9<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u573a\u666f<\/th><th>\u63a8\u8350\u65b9\u6848<\/th><th>\u539f\u56e0<\/th><\/tr><\/thead><tbody><tr><td><strong>7B \u4ee5\u4e0b\u5c0f\u6a21\u578b<\/strong><\/td><td>\u6807\u51c6 FFN + SwiGLU<\/td><td>MoE \u7684\u8def\u7531\u5f00\u9500\u5728\u5c0f\u6a21\u578b\u4e0a\u4e0d\u5212\u7b97<\/td><\/tr><tr><td><strong>\u4e2d\u5927\u89c4\u6a21\uff0813B-70B\uff09<\/strong><\/td><td>MoE + SwiGLU\uff08\u6fc0\u6d3b\u6bd4 1\/10\uff09<\/td><td>\u7528\u7a00\u758f\u6fc0\u6d3b\u6362\u53d6\u66f4\u5927\u603b\u53c2\u6570\u91cf<\/td><\/tr><tr><td><strong>\u591a\u8bed\u8a00\/\u591a\u9886\u57df<\/strong><\/td><td>DeepSeekMoE \u98ce\u683c<\/td><td>\u5171\u4eab\u4e13\u5bb6 + \u7ec6\u7c92\u5ea6\u4e13\u5bb6\uff0c\u9002\u914d\u591a\u6837\u77e5\u8bc6<\/td><\/tr><tr><td><strong>\u6781\u81f4\u63a8\u7406\u6548\u7387<\/strong><\/td><td>MoC\uff08Mixture-of-Channels\uff09<\/td><td>\u5229\u7528 SwiGLU \u539f\u751f\u7a00\u758f\u6027\uff0c\u65e0\u9700\u989d\u5916\u8def\u7531<\/td><\/tr><tr><td><strong>\u8d44\u6e90\u53d7\u9650\u90e8\u7f72<\/strong><\/td><td>FIRM-MoE \u5378\u8f7d\u7b56\u7565<\/td><td>\u7075\u6d3b\u8c03\u914d GPU\/CPU \u5185\u5b58<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">8.2 SwiGLU \u7684\u5b9e\u73b0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 PyTorch \u4e2d\u5b9e\u73b0 SwiGLU FFN \u975e\u5e38\u7b80\u5355\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch.nn as nn\nimport torch.nn.functional as F\n\nclass SwiGLU_FFN(nn.Module):\n    def __init__(self, d_model, d_ffn):\n        super().__init__()\n        # \u6807\u51c6 SwiGLU\uff1a\u4e24\u4e2a\u6295\u5f71\u77e9\u9635\n        self.gate_proj = nn.Linear(d_model, d_ffn, bias=False)\n        self.up_proj = nn.Linear(d_model, d_ffn, bias=False)\n        self.down_proj = nn.Linear(d_ffn, d_model, bias=False)\n\n    def forward(self, x):\n        # SiLU = Swish\n        gate = F.silu(self.gate_proj(x))\n        up = self.up_proj(x)\n        return self.down_proj(gate * up)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8ffd\u6c42\u6781\u81f4\u6548\u7387\uff0c\u8fd8\u53ef\u4ee5\u5c06 <code>gate_proj<\/code> \u548c <code>up_proj<\/code> \u5408\u5e76\u4e3a\u4e00\u4e2a\u77e9\u9635\uff0c\u51cf\u5c11\u4e00\u6b21 Kernel \u542f\u52a8\u5f00\u9500\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001\u603b\u7ed3\u4e0e\u5c55\u671b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">9.1 FFN \u7684\u201c\u8fdb\u5316\u8def\u7ebf\u56fe\u201d<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u9636\u6bb5<\/th><th>\u4ee3\u8868\u6280\u672f<\/th><th>\u6838\u5fc3\u8d21\u732e<\/th><th>\u4ee3\u8868\u6a21\u578b<\/th><\/tr><\/thead><tbody><tr><td><strong>1.0 \u57fa\u7840\u65f6\u4ee3<\/strong><\/td><td>ReLU FFN<\/td><td>\u5f15\u5165\u975e\u7ebf\u6027<\/td><td>\u539f\u59cb Transformer<\/td><\/tr><tr><td><strong>2.0 \u5e73\u6ed1\u65f6\u4ee3<\/strong><\/td><td>GELU<\/td><td>\u7f13\u89e3 Dead ReLU<\/td><td>BERT, GPT-2<\/td><\/tr><tr><td><strong>3.0 \u95e8\u63a7\u65f6\u4ee3<\/strong><\/td><td>SwiGLU<\/td><td>\u52a8\u6001\u4fe1\u606f\u8def\u7531<\/td><td>LLaMA, Qwen, Mistral<\/td><\/tr><tr><td><strong>4.0 \u7a00\u758f\u65f6\u4ee3<\/strong><\/td><td>MoE\uff08Top-K \u8def\u7531\uff09<\/td><td>\u5bb9\u91cf\u4e0e\u8ba1\u7b97\u89e3\u8026<\/td><td>Mixtral, GPT-4<\/td><\/tr><tr><td><strong>5.0 \u7cbe\u7ec6\u65f6\u4ee3<\/strong><\/td><td>DeepSeekMoE\uff08\u5171\u4eab+\u7ec6\u7c92\u5ea6\uff09<\/td><td>\u89e3\u51b3\u77e5\u8bc6\u6df7\u6742\u4e0e\u5197\u4f59<\/td><td>DeepSeek-V2\/V3<\/td><\/tr><tr><td><strong>6.0 \u539f\u751f\u7a00\u758f\u65f6\u4ee3<\/strong><\/td><td>MoC\uff08Mixture-of-Channels\uff09<\/td><td>\u5229\u7528 SwiGLU \u81ea\u5e26\u7a00\u758f\u6027<\/td><td>\u524d\u6cbf\u7814\u7a76<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">9.2 FFN \u5728\u6574\u4e2a LLM \u67b6\u6784\u4e2d\u7684\u5b9a\u4f4d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u6211\u4eec\u6574\u4e2a\u7cfb\u5217\u7684\u6570\u636e\u6d41\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Tokenizer<\/strong> \u2192 Token ID \u5e8f\u5217<\/li>\n\n\n\n<li><strong>Embedding<\/strong> \u2192 \u9ad8\u7ef4\u8bed\u4e49\u5411\u91cf<\/li>\n\n\n\n<li><strong>RoPE<\/strong> \u2192 \u6ce8\u5165\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f<\/li>\n\n\n\n<li><strong>Attention\uff08GQA\uff09<\/strong> \u2192 Token \u95f4\u4fe1\u606f\u4ea4\u4e92<\/li>\n\n\n\n<li><strong>FFN\uff08SwiGLU \/ MoE\uff09<\/strong> \u2192 <strong>\u77e5\u8bc6\u5b58\u50a8\u4e0e\u975e\u7ebf\u6027\u53d8\u6362<\/strong><\/li>\n\n\n\n<li>\u56de\u5230\u6b65\u9aa4 4\uff0c\u5faa\u73af N \u5c42<\/li>\n\n\n\n<li>\u8f93\u51fa\u5c42 \u2192 \u9884\u6d4b\u4e0b\u4e00\u4e2a Token<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">FFN \u5728\u5176\u4e2d\u7684\u89d2\u8272\u662f <strong>\u201c\u77e5\u8bc6\u7684\u52a0\u5de5\u5382\u548c\u5b58\u50a8\u5668\u201d<\/strong> \u3002Attention \u8d1f\u8d23\u201c\u627e\u8c01\u4ea4\u6d41\u201d\uff0cFFN \u8d1f\u8d23\u201c\u4ea4\u6d41\u540e\u600e\u4e48\u6d88\u5316\u548c\u8bb0\u5fc6\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">9.3 \u672a\u6765\u8d8b\u52bf<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u66f4\u7ec6\u7c92\u5ea6\u7684 MoE<\/strong>\uff1a\u4e13\u5bb6\u6570\u91cf\u6301\u7eed\u589e\u52a0\uff0c\u6fc0\u6d3b\u673a\u5236\u66f4\u52a0\u667a\u80fd\uff0cDeepSeekMoE \u7684\u6210\u529f\u8bc1\u660e\u4e86\u8fd9\u4e00\u65b9\u5411\u7684\u5de8\u5927\u6f5c\u529b<\/li>\n\n\n\n<li><strong>\u539f\u751f\u7a00\u758f FFN<\/strong>\uff1a\u4e0d\u4f9d\u8d56\u989d\u5916\u8def\u7531\uff0c\u76f4\u63a5\u5229\u7528\u6fc0\u6d3b\u51fd\u6570\u81ea\u8eab\u7684\u7a00\u758f\u7279\u6027\uff0c\u5982 MoC \u6240\u793a<\/li>\n\n\n\n<li><strong>FFN \u4e0e Attention \u7684\u6df1\u5ea6\u878d\u5408<\/strong>\uff1a\u6253\u7834\u4f20\u7edf\u7684\u201cAttention \u2192 FFN\u201d\u56fa\u5b9a\u987a\u5e8f\uff0c\u63a2\u7d22\u66f4\u7075\u6d3b\u7684\u6a21\u5757\u7f16\u6392\u65b9\u5f0f<\/li>\n\n\n\n<li><strong>FFN \u7684\u53ef\u89e3\u91ca\u6027<\/strong>\uff1a\u4ece\u201c\u4e13\u5bb6\u201d\u5230\u201c\u8f68\u8ff9\u201d\u7684\u89c6\u89d2\u8f6c\u53d8\uff0c\u6b63\u5728\u6253\u5f00 MoE \u6a21\u578b\u5185\u90e8\u8fd0\u4f5c\u7684\u201c\u9ed1\u7bb1\u201d<\/li>\n\n\n\n<li><strong>\u7eff\u8272 AI<\/strong>\uff1aFFN \u5360\u636e 2\/3 \u7684\u53c2\u6570\u548c\u5927\u91cf\u8ba1\u7b97\uff0c\u4f18\u5316 FFN \u6548\u7387\u76f4\u63a5\u5173\u7cfb\u5230\u964d\u4f4e\u5927\u6a21\u578b\u7684\u78b3\u6392\u653e<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 \u5728\u4e4b\u524d\u7684\u7cfb\u5217\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u4ece Tokenizer \u8bb2\u5230 Embedding \u518d\u5230 RoPE\uff0c\u5df2\u7ecf\u8d70\u5b8c\u4e86 [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":131,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[41,37],"class_list":["post-129","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-ffn","tag-llm"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/129","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=129"}],"version-history":[{"count":1,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/129\/revisions"}],"predecessor-version":[{"id":130,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/129\/revisions\/130"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/131"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=129"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=129"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=129"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}