{"id":132,"date":"2026-04-28T17:50:41","date_gmt":"2026-04-28T09:50:41","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=132"},"modified":"2026-04-23T17:52:44","modified_gmt":"2026-04-23T09:52:44","slug":"132","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/28\/132\/","title":{"rendered":"\u6ce8\u610f\u529b\u5373\u4e00\u5207\uff1a\u4ece MHA \u5230 GQA"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ec8\u4e8e\uff0c\u6211\u4eec\u6765\u5230\u4e86 Transformer \u6700\u6838\u5fc3\u3001\u6700\u201c\u7075\u9b42\u201d\u7684\u90e8\u5206\u2014\u2014<strong>Attention\uff08\u6ce8\u610f\u529b\u673a\u5236\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u524d\u56db\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u4e00\u8def\u8d70\u8fc7 Tokenizer\u3001Embedding\u3001RoPE \u548c FFN\uff0c\u5df2\u7ecf\u5927\u81f4\u6478\u6e05\u4e86\u6570\u636e\u5728 LLM \u4e2d\u7684\u6d41\u52a8\u8def\u5f84\u3002\u4f46\u6709\u4e00\u4e2a\u95ee\u9898\u59cb\u7ec8\u60ac\u800c\u672a\u51b3\uff1a<strong>\u8fd9\u4e9b Token \u4e4b\u95f4\uff0c\u5230\u5e95\u662f\u600e\u4e48\u201c\u4ea4\u6d41\u201d\u7684\uff1f<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b54\u6848\u5c31\u662f Attention\u3002\u5982\u679c\u8bf4 FFN \u662f\u6a21\u578b\u7684\u201c\u77e5\u8bc6\u5e93\u201d\uff0c\u90a3\u4e48 Attention \u5c31\u662f\u6a21\u578b\u7684\u201c\u4ea4\u6d41\u7f51\u201d\u2014\u2014\u5b83\u51b3\u5b9a\u4e86\u5728\u5904\u7406\u5f53\u524d\u8bcd\u65f6\uff0c\u6a21\u578b\u5e94\u8be5\u201c\u6ce8\u610f\u201d\u5230\u4e0a\u4e0b\u6587\u4e2d\u7684\u54ea\u4e9b\u8bcd\uff0c\u4ee5\u53ca\u7ed9\u6bcf\u4e2a\u8bcd\u5206\u914d\u591a\u5927\u7684\u201c\u4ea4\u6d41\u6743\u91cd\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b63\u662f\u56e0\u4e3a\u6709\u4e86 Attention\uff0cLLM \u624d\u80fd\u7406\u89e3\u201c\u82f9\u679c\u5f88\u597d\u5403\u201d\u4e2d\u7684\u201c\u5f88\u201d\u662f\u5728\u4fee\u9970\u201c\u597d\u5403\u201d\uff0c\u800c\u4e0d\u662f\u5728\u4fee\u9970\u201c\u82f9\u679c\u201d\uff1b\u624d\u80fd\u5728\u505a\u7ffb\u8bd1\u65f6\uff0c\u628a\u82f1\u6587\u7684\u201cbank\u201d\u6839\u636e\u4e0a\u4e0b\u6587\u6b63\u786e\u5730\u8bd1\u4e3a\u201c\u94f6\u884c\u201d\u6216\u201c\u6cb3\u5cb8\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4eca\u5929\u8fd9\u7bc7\u6587\u7ae0\uff0c\u6211\u4eec\u5c06\u4e00\u9053\u4e00\u9053\u5730\u5256\u5f00 Attention \u7684\u201c\u9ed1\u7bb1\u201d\uff0c\u4ece\u6700\u57fa\u7840\u7684 <strong>MHA\uff08\u591a\u5934\u6ce8\u610f\u529b\uff09<\/strong> \u5f00\u59cb\uff0c\u8bb2\u6e05\u695a\u5b83\u5185\u90e8\u7684 Q\u3001K\u3001V \u5230\u5e95\u5728\u505a\u4ec0\u4e48\uff1b\u7136\u540e\u6df1\u5165\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u6700\u91cd\u8981\u7684\u4e24\u4e2a\u53d8\u4f53\u2014\u2014<strong>MQA\uff08\u591a\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \u548c <strong>GQA\uff08\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \uff1b\u6700\u540e\uff0c\u6211\u4eec\u4f1a\u89e6\u53ca\u8fd9\u4e2a\u9886\u57df\u7684\u6700\u524d\u6cbf\u2014\u2014<strong>MLA\uff08\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff09<\/strong> \u548c\u6700\u65b0\u7684\u6ce8\u610f\u529b\u4f18\u5316\u6280\u672f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7bc7\u6587\u7ae0\u53ef\u80fd\u4f1a\u662f\u6574\u4e2a\u7cfb\u5217\u4e2d\u4fe1\u606f\u5bc6\u5ea6\u6700\u9ad8\u7684\u4e00\u7bc7\uff0c\u56e0\u4e3a\u5b83\u6d89\u53ca LLM \u63a8\u7406\u4f18\u5316\u4e2d\u6700\u5173\u952e\u7684\u4e24\u4e2a\u6982\u5ff5\uff1a<strong>KV Cache<\/strong> \u548c<strong>\u663e\u5b58\u74f6\u9888<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u4e3a\u4ec0\u4e48 Attention \u662f Transformer \u7684\u7075\u9b42\uff1f<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u5982\u679c\u6ca1\u6709 Attention\uff1a\u4e00\u4e2a\u201c\u5065\u5fd8\u201d\u7684\u6a21\u578b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u8c61\u4e00\u4e0b\uff0c\u4f60\u8981\u5199\u4e00\u4e2a\u7a0b\u5e8f\uff0c\u8ba9\u5b83\u8bfb\u5b8c\u4e00\u53e5\u8bdd\u540e\uff0c\u56de\u7b54\u201c\u8fd9\u53e5\u8bdd\u7684\u4e3b\u8bed\u662f\u8c01\uff1f\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u6ca1\u6709 Attention\uff0c\u4f60\u80fd\u600e\u4e48\u505a\uff1f\u4f60\u53ef\u80fd\u4f1a\u628a\u6574\u4e2a\u53e5\u5b50\u538b\u7f29\u6210\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\uff08\u6bd4\u5982 LSTM \u6216 GRU \u7684\u505a\u6cd5\uff09\uff0c\u7136\u540e\u7528\u8fd9\u4e2a\u5411\u91cf\u6765\u56de\u7b54\u95ee\u9898\u3002\u4f46\u95ee\u9898\u5728\u4e8e\uff1a<strong>\u628a\u4e00\u53e5\u8bdd\u538b\u7f29\u6210\u4e00\u4e2a\u5411\u91cf\uff0c\u5fc5\u7136\u4f1a\u4e22\u5931\u5927\u91cf\u7ec6\u8282<\/strong>\u3002\u5f53\u53e5\u5b50\u53d8\u957f\u65f6\uff0c\u6a21\u578b\u5c31\u50cf\u4e00\u4e2a\u4eba\u62fc\u547d\u60f3\u8bb0\u4f4f\u4e00\u4e2a\u957f\u6545\u4e8b\uff0c\u5374\u53ea\u6709\u4e00\u5f20\u4fbf\u7b7e\u7eb8\u2014\u2014\u5199\u7740\u5199\u7740\uff0c\u524d\u9762\u7684\u5185\u5bb9\u5c31\u5168\u5fd8\u4e86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Attention \u7684\u9769\u547d\u6027\u5728\u4e8e\uff1a\u5b83\u653e\u5f03\u4e86\u201c\u538b\u7f29\u201d\u7684\u6267\u5ff5\u3002<\/strong> Attention \u5141\u8bb8\u6a21\u578b\u5728\u5904\u7406\u6bcf\u4e00\u4e2a\u8bcd\u65f6\uff0c\u76f4\u63a5\u201c\u56de\u770b\u201d\u6574\u4e2a\u8f93\u5165\u5e8f\u5217\uff0c\u5e76\u6839\u636e\u9700\u8981\u4ece\u5e8f\u5217\u4e2d\u7684\u4efb\u610f\u4f4d\u7f6e\u63d0\u53d6\u4fe1\u606f\u3002\u5b83\u4e0d\u662f\u628a\u6545\u4e8b\u8bb0\u5728\u4fbf\u7b7e\u7eb8\u4e0a\uff0c\u800c\u662f\u8ba9\u6a21\u578b\u80fd\u591f\u968f\u65f6\u628a\u76ee\u5149\u6295\u5411\u6545\u4e8b\u7684\u4efb\u610f\u4e00\u9875\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 Attention \u7684\u76f4\u89c9\uff1a\u76f8\u4f3c\u5ea6\u9a71\u52a8\u7684\u201c\u56de\u6eaf\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u8ba1\u7b97\u7684\u6838\u5fc3\u53ef\u4ee5\u7528\u4e00\u53e5\u8bdd\u6982\u62ec\uff1a<strong>\u5bf9\u4e8e\u5f53\u524d\u7684 Token\uff0c\u627e\u51fa\u4e0a\u4e0b\u6587\u4e2d\u6240\u6709\u4e0e\u5b83\u201c\u76f8\u4f3c\u201d\u7684 Token\uff0c\u7136\u540e\u628a\u8fd9\u4e9b Token \u7684\u4fe1\u606f\u52a0\u6743\u878d\u5408\u8fc7\u6765\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684\u201c\u76f8\u4f3c\u201d\uff0c\u4e0d\u662f\u8bed\u4e49\u4e0a\u7684\u201c\u82f9\u679c\u548c\u6c34\u679c\u76f8\u4f3c\u201d\uff0c\u800c\u662f\u7ecf\u8fc7\u6295\u5f71\u540e\uff0c\u4e24\u4e2a\u5411\u91cf\u5728\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684<strong>\u70b9\u79ef\u76f8\u4f3c\u5ea6<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5f53\u6a21\u578b\u5904\u7406\u53e5\u5b50\u201c\u5c0f\u732b\u5728\u8ffd\u4e00\u53ea\u8001\u9f20\uff0c\u5b83\u8dd1\u5f97\u5f88\u5feb\u201d\u4e2d\u7684\u201c\u5b83\u201d\u5b57\u65f6\uff0cAttention \u4f1a\u81ea\u52a8\u53d1\u73b0\uff1a\u201c\u5b83\u201d\u7684 Query \u5411\u91cf\uff0c\u4e0e\u524d\u9762\u201c\u5c0f\u732b\u201d\u7684 Key \u5411\u91cf\u4e4b\u95f4\u7684\u70b9\u79ef\u975e\u5e38\u5927\u3002\u4e8e\u662f\uff0c\u201c\u5c0f\u732b\u201d\u7684\u4fe1\u606f\u4f1a\u88ab\u5206\u914d\u5f88\u9ad8\u7684\u6743\u91cd\uff0c\u878d\u5165\u5230\u201c\u5b83\u201d\u7684\u8868\u793a\u4e2d\uff0c\u8fd9\u6837\u6a21\u578b\u5c31\u77e5\u9053\u201c\u5b83\u201d\u6307\u7684\u662f\u201c\u5c0f\u732b\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a Attention \u673a\u5236\u5c31\u50cf\u662f\u7ed9\u6bcf\u4e2a Token \u914d\u4e86\u4e00\u90e8\u201c\u65f6\u7a7a\u7a7f\u68ad\u673a\u201d\uff1a\u6bcf\u5904\u7406\u4e00\u4e2a\u65b0\u8bcd\uff0c\u90fd\u53ef\u4ee5\u81ea\u7531\u5730\u56de\u5230\u8fc7\u53bb\u7684\u4efb\u610f\u65f6\u523b\uff0c\u628a\u90a3\u91cc\u7684\u4fe1\u606f\u201c\u63d0\u53d6\u201d\u8fc7\u6765\u7528\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Self-Attention \u7684\u6570\u5b66\u62c6\u89e3<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 Q\u3001K\u3001V\uff1a\u4e09\u79cd\u8eab\u4efd\uff0c\u5404\u53f8\u5176\u804c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6807\u51c6 Self-Attention \u7684\u8ba1\u7b97\uff0c\u4f9d\u8d56\u4e8e\u4e09\u4e2a\u6838\u5fc3\u77e9\u9635\uff1a<strong>Query\uff08Q\uff09\u3001Key\uff08K\uff09\u3001Value\uff08V\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e00\u4e2a Token\uff0c\u5b83\u7684 Embedding \u5411\u91cf x \u4f1a\u901a\u8fc7\u4e09\u4e2a\u4e0d\u540c\u7684\u6743\u91cd\u77e9\u9635\uff08W_Q\u3001W_K\u3001W_V\uff09\u6295\u5f71\uff0c\u751f\u6210\u4e09\u4e2a\u5168\u65b0\u7684\u5411\u91cf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Q\uff08Query\uff0c\u67e5\u8be2\u5411\u91cf\uff09<\/strong> \uff1a\u4ee3\u8868\u201c\u6211\u6b63\u5728\u5bfb\u627e\u4ec0\u4e48\u201d\u3002\u8fd9\u662f\u5f53\u524d Token \u53d1\u51fa\u7684\u201c\u9700\u6c42\u4fe1\u53f7\u201d\u3002<\/li>\n\n\n\n<li><strong>K\uff08Key\uff0c\u952e\u5411\u91cf\uff09<\/strong> \uff1a\u4ee3\u8868\u201c\u6211\u662f\u4ec0\u4e48\u201d\u3002\u8fd9\u662f\u6bcf\u4e2a Token \u53d1\u51fa\u7684\u201c\u8eab\u4efd\u6807\u7b7e\u201d\u3002<\/li>\n\n\n\n<li><strong>V\uff08Value\uff0c\u503c\u5411\u91cf\uff09<\/strong> \uff1a\u4ee3\u8868\u201c\u6211\u7684\u5185\u5bb9\u662f\u4ec0\u4e48\u201d\u3002\u8fd9\u662f\u6bcf\u4e2a Token \u5b9e\u9645\u8981\u4f20\u9012\u7684\u4fe1\u606f\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4e00\u4e2a\u7b80\u5355\u7684\u6bd4\u55bb\u6765\u7406\u89e3\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u4f60\u8d70\u8fdb\u4e00\u5bb6\u56fe\u4e66\u9986\uff08\u8f93\u5165\u5e8f\u5217\uff09\uff0c\u60f3\u627e\u5173\u4e8e\u201c\u7f57\u9a6c\u5e1d\u56fd\u201d\u7684\u4e66\uff08Query\uff09\u3002\u6bcf\u672c\u4e66\u7684\u5c01\u9762\u4e0a\u90fd\u8d34\u7740\u5173\u952e\u8bcd\u6807\u7b7e\uff08Key\uff09\u3002\u4f60\u626b\u4e00\u773c\u6240\u6709\u6807\u7b7e\uff0c\u53d1\u73b0\u67d0\u672c\u4e66\u7684\u6807\u7b7e\u662f\u201c\u7f57\u9a6c\u5e1d\u56fd\u201d\uff0c\u4e0e\u4f60\u7684 Query \u9ad8\u5ea6\u5339\u914d\uff0c\u4e8e\u662f\u4f60\u628a\u5b83\u4ece\u4e66\u67b6\u4e0a\u53d6\u4e0b\u6765\uff0c\u9605\u8bfb\u5b83\u7684\u5185\u5bb9\uff08Value\uff09\u3002\u4f60\u6700\u7ec8\u83b7\u5f97\u7684\u77e5\u8bc6\uff0c\u5c31\u662f\u6240\u6709\u5339\u914d\u7a0b\u5ea6\u52a0\u6743\u540e\u7684 Value \u4e4b\u548c\u3002<\/p>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 Attention \u8ba1\u7b97\u56db\u6b65\u8d70<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\u6b65\uff1a<strong>\u751f\u6210 Q\u3001K\u3001V<\/strong>\u3002\u5bf9\u4e8e\u6bcf\u4e2a Token\uff0c\u7528\u4e09\u4e2a\u77e9\u9635\u6295\u5f71\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Q = x \u00b7 W_Q\nK = x \u00b7 W_K\nV = x \u00b7 W_V<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\u6b65\uff1a<strong>\u8ba1\u7b97\u6ce8\u610f\u529b\u5206\u6570<\/strong>\u3002\u7528\u5f53\u524d Token \u7684 Q \u53bb\u548c\u6240\u6709 Token \u7684 K \u505a\u70b9\u79ef\uff0c\u5f97\u5230\u4e00\u4e2a\u201c\u5339\u914d\u7a0b\u5ea6\u201d\u77e9\u9635\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Score = Q \u00b7 K^T<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e09\u6b65\uff1a<strong>\u5f52\u4e00\u5316<\/strong>\u3002\u7528 Softmax \u5c06\u5206\u6570\u8f6c\u6362\u4e3a 0 \u5230 1 \u4e4b\u95f4\u7684\u6982\u7387\u5206\u5e03\uff08\u5168\u90e8\u52a0\u8d77\u6765\u7b49\u4e8e 1\uff09\u3002\u4e3a\u4e86\u9632\u6b62\u70b9\u79ef\u503c\u8fc7\u5927\u5bfc\u81f4 Softmax \u68af\u5ea6\u6d88\u5931\uff0c\u901a\u5e38\u8fd8\u4f1a\u9664\u4ee5\u7ef4\u5ea6\u7684\u5e73\u65b9\u6839\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Attention_Weights = softmax(Score \/ \u221ad_k)<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u56db\u6b65\uff1a<strong>\u52a0\u6743\u6c42\u548c<\/strong>\u3002\u7528\u8fd9\u4e9b\u6743\u91cd\uff0c\u5bf9\u6240\u6709 Token \u7684 V \u8fdb\u884c\u52a0\u6743\u6c42\u548c\uff0c\u5f97\u5230\u5f53\u524d Token \u7684\u6700\u7ec8\u8f93\u51fa\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Output = Attention_Weights \u00b7 V<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u56db\u6b65\uff0c\u5c31\u662f Attention \u7684\u5168\u90e8\u6570\u5b66\u3002\u867d\u7136\u770b\u8d77\u6765\u7b80\u5355\uff0c\u4f46\u6b63\u5982\u90a3\u53e5\u540d\u8a00\u6240\u8bf4\uff1a\u201c\u7b80\u5355\uff0c\u4f46\u4e0d\u7b80\u964b\u3002\u201d<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 Masked Self-Attention\uff1a\u8ba9\u6a21\u578b\u5b66\u4f1a\u201c\u9884\u77e5\u672a\u6765\u662f\u4f5c\u5f0a\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8bad\u7ec3 LLM \u505a\u6587\u672c\u751f\u6210\u65f6\uff0c\u6211\u4eec\u4f1a\u523b\u610f\u906e\u4f4f\u672a\u6765\u7684\u4fe1\u606f\u3002\u56e0\u4e3a\u5982\u679c\u6a21\u578b\u80fd\u770b\u5230\u540e\u9762\u7684\u8bcd\uff0c\u90a3\u201c\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u201d\u5c31\u53d8\u6210\u5f00\u5377\u8003\u8bd5\u4e86\u2014\u2014\u6beb\u65e0\u610f\u4e49\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Masked Self-Attention<\/strong> \u7684\u505a\u6cd5\u662f\uff1a\u5728\u8ba1\u7b97 Softmax \u4e4b\u524d\uff0c\u5c06 Q \u4e0e\u672a\u6765\u4f4d\u7f6e\u7684 K \u7684\u70b9\u79ef\u5206\u6570\u8bbe\u4e3a\u8d1f\u65e0\u7a77\uff08-\u221e\uff09\u3002\u8fd9\u6837\u7ecf\u8fc7 Softmax \u540e\uff0c\u672a\u6765\u4f4d\u7f6e\u7684\u6ce8\u610f\u529b\u6743\u91cd\u5c31\u53d8\u6210 0\uff0c\u6a21\u578b\u53ea\u80fd\u770b\u5230\u5f53\u524d\u53ca\u4e4b\u524d\u7684\u5185\u5bb9\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u53eb\u505a<strong>\u56e0\u679c\u6ce8\u610f\u529b\uff08Causal Attention\uff09<\/strong>\uff0c\u4e5f\u662f\u201c\u81ea\u56de\u5f52\u751f\u6210\u201d\u7684\u6838\u5fc3\u7ea6\u675f\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001MHA\uff1a\u8ba9\u6a21\u578b\u6234\u4e0a\u201c\u591a\u526f\u773c\u955c\u201d\u770b\u4e16\u754c<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u4e00\u4e2a\u5934\u7684\u5c40\u9650<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5355\u72ec\u7684 Self-Attention \u6709\u4e00\u4e2a\u5929\u7136\u7684\u95ee\u9898\uff1a<strong>\u5b83\u53ea\u80fd\u6355\u6349\u4e00\u79cd\u201c\u76f8\u5173\u6027\u6a21\u5f0f\u201d<\/strong>\u3002\u6bd4\u5982\u5728\u5904\u7406\u201c\u94f6\u884c\u201d\u8fd9\u4e2a\u8bcd\u65f6\uff0c\u4e00\u4e2a\u6ce8\u610f\u529b\u5934\u53ef\u80fd\u9700\u8981\u540c\u65f6\u5224\u65ad\u5b83\u662f\u6307\u201c\u91d1\u878d\u673a\u6784\u201d\u8fd8\u662f\u201c\u6cb3\u5cb8\u201d\uff0c\u8fd8\u8981\u540c\u65f6\u5904\u7406\u8bed\u6cd5\u5173\u7cfb\u548c\u6307\u4ee3\u6d88\u89e3\u2014\u2014\u628a\u6240\u6709\u4efb\u52a1\u538b\u5728\u4e00\u4e2a\u5934\u4e0a\uff0c\u6a21\u578b\u4f1a\u975e\u5e38\u5403\u529b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 MHA \u7684\u65b9\u6848\uff1a\u591a\u5934\uff0c\u5404\u770b\u5404\u7684<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>MHA\uff08Multi-Head Attention\uff0c\u591a\u5934\u6ce8\u610f\u529b\uff09<\/strong> \u7684\u89e3\u51b3\u65b9\u6848\u975e\u5e38\u76f4\u89c9\uff1a<strong>\u5e76\u884c\u5730\u8ba1\u7b97\u591a\u4e2a\u72ec\u7acb\u7684 Attention\uff0c\u6bcf\u4e2a\u5934\u4f7f\u7528\u81ea\u5df1\u7684\u4e00\u5957 W_Q\u3001W_K\u3001W_V \u6743\u91cd\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u505a\u6cd5\u662f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5c06 Q\u3001K\u3001V \u62c6\u5206\u6210 h \u4e2a\u5934\uff08\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\u4e3a d_head = d_model \/ h\uff09<\/li>\n\n\n\n<li>\u6bcf\u4e2a\u5934\u72ec\u7acb\u6267\u884c\u4e00\u6b21\u5b8c\u6574\u7684 Self-Attention \u8ba1\u7b97<\/li>\n\n\n\n<li>\u5c06\u6240\u6709\u5934\u7684\u8f93\u51fa\u62fc\u63a5\u8d77\u6765\uff0c\u518d\u7ecf\u8fc7\u4e00\u4e2a\u7ebf\u6027\u53d8\u6362\uff08W_O\uff09\u878d\u5408<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u516c\u5f0f\u8868\u793a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>MultiHead(Q, K, V) = Concat(head_1, head_2, ..., head_h) \u00b7 W_O<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 MHA \u7684\u4f18\u52bf\uff1a\u591a\u89d2\u5ea6\u89c2\u5bdf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u5c31\u50cf\u6234\u4e86\u4e00\u526f\u4e0d\u540c\u5ea6\u6570\u6216\u4e0d\u540c\u989c\u8272\u7684\u773c\u955c\u6765\u89c2\u5bdf\u540c\u4e00\u4e2a\u53e5\u5b50\u3002\u6709\u7684\u5934\u5173\u6ce8\u53e5\u6cd5\uff08\u6bd4\u5982\u540d\u8bcd\u548c\u5f62\u5bb9\u8bcd\u7684\u5173\u8054\uff09\uff0c\u6709\u7684\u5934\u5173\u6ce8\u6307\u4ee3\uff08\u6bd4\u5982\u4ee3\u8bcd\u548c\u5b83\u6307\u4ee3\u7684\u540d\u8bcd\uff09\uff0c\u6709\u7684\u5934\u5173\u6ce8\u957f\u8ddd\u79bb\u4f9d\u8d56\uff08\u6bd4\u5982\u6bb5\u843d\u9996\u5c3e\u7684\u547c\u5e94\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u201c\u591a\u89d2\u5ea6\u89c2\u5bdf\u201d\u8ba9 MHA \u62e5\u6709\u4e86\u8fdc\u8d85\u5355\u5934\u6ce8\u610f\u529b\u7684\u8868\u8fbe\u80fd\u529b\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u539f\u7248 Transformer \u548c\u65e9\u671f GPT \u90fd\u91c7\u7528\u4e86 MHA\u2014\u2014\u5b83\u786e\u5b9e\u597d\u7528\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.4 MHA \u7684\u4ee3\u4ef7\uff1aKV Cache \u7684\u9690\u5f62\u6210\u672c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u800c\uff0cMHA \u6709\u4e00\u4e2a\u5728\u63a8\u7406\u65f6\u4f1a\u53d8\u6210\u201c\u663e\u5b58\u602a\u517d\u201d\u7684\u5de5\u7a0b\u95ee\u9898\u2014\u2014<strong>KV Cache<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u81ea\u56de\u5f52\u751f\u6210\u4e2d\uff0c\u6bcf\u751f\u6210\u4e00\u4e2a\u65b0 Token\uff0c\u6211\u4eec\u90fd\u9700\u8981\u8ba1\u7b97\u5b83\u548c\u4e4b\u524d\u6240\u6709 Token \u7684 Attention\u3002\u5982\u679c\u6bcf\u6b21\u4ece\u5934\u7b97\u4e00\u904d\uff0c\u8ba1\u7b97\u91cf\u4f1a\u4ee5 O(n\u00b2) \u7684\u901f\u5ea6\u7206\u70b8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>KV Cache<\/strong> \u5c31\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u800c\u751f\u7684\uff1a<strong>\u628a\u5df2\u7ecf\u7b97\u8fc7\u7684\u6240\u6709 Token \u7684 Key \u548c Value \u5411\u91cf\u4fdd\u5b58\u4e0b\u6765\uff0c\u751f\u6210\u4e0b\u4e00\u4e2a Token \u65f6\u76f4\u63a5\u590d\u7528\uff0c\u53ea\u8ba1\u7b97\u5f53\u524d Token \u7684 Q\u3001K\u3001V\uff0c\u7136\u540e\u7528\u5f53\u524d\u7684 Q \u53bb\u548c\u7f13\u5b58\u7684 K \u8ba1\u7b97 Attention\uff0c\u6700\u540e\u66f4\u65b0 KV Cache\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u95ee\u9898\u5728\u4e8e\uff1aMHA \u6709\u591a\u5c11\u4e2a\u5934\uff0c\u6bcf\u4e2a Token \u5c31\u9700\u8981\u7f13\u5b58\u591a\u5c11\u7ec4 K \u548c V\u3002\u5bf9\u4e8e LLaMA-7B \u6765\u8bf4\uff0c\u5b83\u6709 32 \u4e2a\u5934\uff0c\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\u662f 128\uff0c\u6240\u4ee5\u6bcf\u4e2a Token \u7684 KV Cache \u5927\u5c0f\u4e3a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>32 \u5934 \u00d7 128 \u7ef4 \u00d7 2\uff08K \u548c V\uff09\u00d7 2 \u5b57\u8282\uff08FP16\uff09= 16 KB \/ Token<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u770b\u8d77\u6765\u4e0d\u591a\u3002\u4f46\u60f3\u8c61\u4e00\u4e0b\uff0c\u5f53\u6279\u5904\u7406\u5927\u5c0f\uff08batch size\uff09\u53d8\u5927\u3001\u5e8f\u5217\u957f\u5ea6\u53d8\u6210 128K \u751a\u81f3 1M\u3001\u5c42\u6570\u53e0\u52a0\u5230\u51e0\u5341\u5c42\u65f6\uff0cKV Cache \u4f1a\u8fc5\u901f\u81a8\u80c0\u6210\u63a8\u7406\u663e\u5b58\u7684\u6700\u5927\u5934\u2014\u2014<strong>\u5728\u67d0\u4e9b\u573a\u666f\u4e0b\uff0cKV Cache \u751a\u81f3\u5360\u4e86\u603b\u663e\u5b58\u7684 90% \u4ee5\u4e0a<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\uff0c\u538b\u7f29 KV Cache \u5c31\u6210\u4e86 LLM \u63a8\u7406\u4f18\u5316\u7684\u6838\u5fc3\u6218\u573a\u3002MQA \u548c GQA \u6b63\u662f\u5728\u8fd9\u79cd\u9700\u6c42\u4e0b\u8bde\u751f\u7684\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001MQA\uff1a\u6781\u81f4\u7684 KV Cache \u538b\u7f29<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 MQA \u7684\u6838\u5fc3\u601d\u60f3\uff1a\u6240\u6709\u5934\u5171\u4eab KV<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>MQA\uff08Multi-Query Attention\uff0c\u591a\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \u7684\u60f3\u6cd5\u975e\u5e38\u7b80\u5355\u76f4\u63a5\uff1a<strong>\u8ba9\u6240\u6709\u6ce8\u610f\u529b\u5934\u5171\u4eab\u540c\u4e00\u4efd Key \u548c Value\uff0c\u53ea\u6709 Query \u4fdd\u6301\u72ec\u7acb\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 MQA \u4e2d\uff0c\u6bcf\u4e2a Token \u4ecd\u7136\u6709\u591a\u4e2a Q \u5934\uff08\u6bcf\u4e2a\u5934\u770b\u5230\u7684\u4e16\u754c\u4e0d\u540c\uff09\uff0c\u4f46\u53ea\u6709\u4e00\u4efd K \u548c V\u3002\u65e2\u7136 K \u548c V \u662f\u591a\u5934\u5171\u4eab\u7684\uff0c\u90a3 KV Cache \u7684\u5927\u5c0f\u5c31\u53ea\u5269\u4e0b MHA \u7684 <strong>1\/h<\/strong>\uff08h \u4e3a\u5934\u6570\uff09\u3002\u5982\u679c\u5934\u6570\u662f 32\uff0cKV Cache \u7684\u5360\u7528\u76f4\u63a5\u7f29\u6c34 32 \u500d\uff01<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5728\u5b9e\u9645\u5de5\u7a0b\u4e2d\u5e26\u6765\u7684\u5f71\u54cd\u662f\u5de8\u5927\u7684\uff1a\u63a8\u7406\u901f\u5ea6\u63d0\u5347\uff0c\u663e\u5b58\u5360\u7528\u5927\u5e45\u4e0b\u964d\uff0c\u80fd\u591f\u652f\u6301\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u548c\u66f4\u5927\u7684\u6279\u5904\u7406\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 MQA \u7684\u4ee3\u4ef7\uff1a\u8868\u8fbe\u80fd\u529b\u7684\u9000\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5929\u4e0b\u6ca1\u6709\u514d\u8d39\u7684\u5348\u9910\u3002MQA \u867d\u7136\u538b\u7f29\u4e86\u663e\u5b58\uff0c\u4f46\u4e5f\u4ed8\u51fa\u4e86\u4ee3\u4ef7\uff1a<strong>\u6240\u6709\u5934\u5728\u770b\u540c\u4e00\u4e2a K \u548c V \u65f6\uff0c\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u4f1a\u4e0b\u964d\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u6bcf\u4e2a\u5934\u4e0d\u518d\u6709\u81ea\u5df1\u7684\u4e13\u5c5e\u201c\u89c6\u89d2\u201d\u6765\u5ba1\u89c6 K \u548c V\uff0c\u5b83\u4eec\u88ab\u8feb\u4ece\u540c\u4e00\u4efd\u4fe1\u606f\u4e2d\u5bfb\u627e\u5404\u81ea\u9700\u8981\u7684\u90e8\u5206\u3002\u5b9e\u9a8c\u8868\u660e\uff0cMQA \u5728\u540c\u7b49\u53c2\u6570\u91cf\u4e0b\uff0c\u6548\u679c\u666e\u904d\u7565\u900a\u4e8e MHA\u3002\u867d\u7136\u5dee\u8ddd\u4e0d\u5927\uff0c\u4f46\u5bf9\u4e8e\u8ffd\u6c42\u6781\u81f4\u6027\u80fd\u7684\u5927\u6a21\u578b\u6765\u8bf4\uff0c\u8fd9\u4e2a\u635f\u5931\u4e0d\u592a\u80fd\u63a5\u53d7\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 MQA \u7684\u5e94\u7528<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">MQA \u7684\u4ee3\u8868\u6027\u5e94\u7528\u662f Google \u7684 <strong>PaLM<\/strong> \u6a21\u578b\uff0c\u4ee5\u53ca\u4e00\u4e9b\u6ce8\u91cd\u6781\u81f4\u63a8\u7406\u6548\u7387\u7684\u8f7b\u91cf\u7ea7\u6a21\u578b\u3002\u4f46\u5728\u8ffd\u6c42\u80fd\u529b\u4e0a\u9650\u7684\u5927\u89c4\u6a21\u5f00\u6e90\u6a21\u578b\u4e2d\uff0cMQA \u9010\u6e10\u88ab\u4e00\u4e2a\u201c\u6298\u4e2d\u65b9\u6848\u201d\u53d6\u4ee3\u2014\u2014GQA\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001GQA\uff1a\u6548\u7387\u4e0e\u6548\u679c\u7684\u9ec4\u91d1\u5e73\u8861<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 GQA \u7684\u65b9\u6848\uff1a\u5206\u7ec4\u5171\u4eab<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GQA\uff08Grouped Query Attention\uff0c\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/strong> \u662f MHA \u548c MQA \u4e4b\u95f4\u7684\u201c\u751c\u871c\u70b9\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u505a\u6cd5\u662f\uff1a\u5c06\u6240\u6709\u7684 Query \u5934\u5206\u6210\u82e5\u5e72\u7ec4\uff08\u6bd4\u5982 8 \u4e2a\u5934\u5206\u6210 2 \u7ec4\u6216 4 \u7ec4\uff09\uff0c<strong>\u540c\u4e00\u7ec4\u5185\u7684 Query \u5934\u5171\u4eab\u540c\u4e00\u4efd Key \u548c Value<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5047\u8bbe\u6a21\u578b\u6709 32 \u4e2a\u5934\uff0c\u6211\u4eec\u5c06\u5176\u5206\u4e3a 8 \u7ec4\uff08\u6bcf\u7ec4 4 \u4e2a\u5934\uff09\u3002\u90a3\u4e48\u5bf9\u4e8e\u6bcf\u4e2a Token\uff0c\u53ea\u9700\u8981\u7f13\u5b58 8 \u7ec4 K \u548c V\u3002\u4e0e MHA\uff0832 \u7ec4 KV\uff09\u76f8\u6bd4\uff0cKV Cache \u7f29\u5c0f\u4e3a 1\/4\uff1b\u4e0e MQA\uff081 \u7ec4 KV\uff09\u76f8\u6bd4\uff0c\u53c8\u4fdd\u7559\u4e86\u66f4\u591a\u7684\u8868\u8fbe\u80fd\u529b\uff088 \u79cd\u4e0d\u540c\u7684 K \u548c V\uff09\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 GQA \u7684\u7531\u6765\u4e0e\u63a8\u5bfc<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">GQA \u5e76\u975e\u51ed\u7a7a\u51fa\u73b0\uff0c\u5b83\u7684\u63d0\u51fa\u52a8\u673a\u975e\u5e38\u76f4\u767d\uff1a<strong>\u5728\u6a21\u578b\u89c4\u6a21\u548c\u5e8f\u5217\u957f\u5ea6\u90fd\u4e0d\u65ad\u589e\u957f\u7684\u4eca\u5929\uff0cMHA \u7684 KV Cache \u5df2\u7ecf\u4e0d\u582a\u91cd\u8d1f\uff1bMQA \u7684\u538b\u7f29\u53c8\u8ba9\u6a21\u578b\u80fd\u529b\u4e0b\u964d\u592a\u591a\u3002\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u53ef\u4ee5\u8c03\u8282\u538b\u7f29\u6bd4\u7684\u65b9\u6848\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GQA \u7684\u201c\u7ec4\u6570\u201d\u6210\u4e86\u4e00\u4e2a\u53ef\u8c03\u7684\u8d85\u53c2\u6570\u3002\u7ec4\u6570\u8d8a\u5c0f\uff08\u751a\u81f3\u7b49\u4e8e 1\uff0c\u5373 MQA\uff09\uff0c\u538b\u7f29\u7387\u8d8a\u9ad8\uff0c\u63a8\u7406\u8d8a\u5feb\uff1b\u7ec4\u6570\u8d8a\u5927\uff08\u751a\u81f3\u7b49\u4e8e\u5934\u6570\uff0c\u5373 MHA\uff09\uff0c\u8868\u8fbe\u80fd\u529b\u8d8a\u5f3a\uff0c\u4f46\u663e\u5b58\u5360\u7528\u4e5f\u8d8a\u5927\u3002LLaMA-2 70B \u91c7\u7528\u4e86 8 \u7ec4 GQA\uff08\u5934\u6570 64\uff0c\u5206\u7ec4\u6570 8\uff09\uff0c\u5728\u51e0\u4e4e\u4e0d\u635f\u5931\u7cbe\u5ea6\u7684\u60c5\u51b5\u4e0b\uff0c\u6781\u5927\u5730\u63d0\u5347\u4e86\u63a8\u7406\u6548\u7387\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 GQA \u4e3a\u4ec0\u4e48\u6210\u4e3a\u201c\u6807\u914d\u201d\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u5b83\u5728\u4ee5\u4e0b\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\u90fd\u53d6\u5f97\u4e86\u5f88\u597d\u7684\u5e73\u8861\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u8bad\u7ec3\u7a33\u5b9a\u6027<\/strong>\uff1aGQA \u7684\u8bad\u7ec3\u66f2\u7ebf\u6bd4 MQA \u66f4\u7a33\u5b9a\uff0c\u4e0d\u6613\u51fa\u73b0\u80fd\u529b\u9000\u5316\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u6548\u7387<\/strong>\uff1a\u76f8\u6bd4 MHA\uff0cKV Cache \u663e\u8457\u7f29\u51cf\uff0c\u957f\u6587\u672c\u548c\u6279\u5904\u7406\u573a\u666f\u4e0b\u663e\u5b58\u4f18\u52bf\u660e\u663e\u3002<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u6027\u80fd<\/strong>\uff1a\u5728\u540c\u7b49\u53c2\u6570\u91cf\u4e0b\uff0cGQA \u7684\u6027\u80fd\u975e\u5e38\u63a5\u8fd1 MHA\uff0c\u8fdc\u8d85 MQA\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u4e8e\u8fd9\u4e9b\u539f\u56e0\uff0c<strong>GQA \u8fc5\u901f\u6210\u4e3a\u73b0\u4ee3 LLM \u7684\u4e8b\u5b9e\u6807\u51c6<\/strong>\uff1aLLaMA-2\/3 \u5728\u5927\u5c3a\u5bf8\u6a21\u578b\u4e0a\u91c7\u7528 GQA\uff0cQwen\u3001Mistral\u3001Gemma \u7b49\u6a21\u578b\u4e5f\u7eb7\u7eb7\u8ddf\u8fdb\uff0c\u751a\u81f3\u5728\u8f83\u5c0f\u7684\u6a21\u578b\u4e0a\u4e5f\u5f15\u5165\u4e86 GQA\u3002\u53ef\u4ee5\u8bf4\uff0c\u5728 2024 \u5230 2025 \u5e74\u95f4\u53d1\u5e03\u7684\u5546\u4e1a\u7ea7\u5927\u6a21\u578b\u4e2d\uff0cGQA \u5df2\u7ecf\u662f\u201c\u9ed8\u8ba4\u914d\u7f6e\u201d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001MHA\u3001MQA\u3001GQA \u6a2a\u5411\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u8ba9\u4f60\u5bf9\u8fd9\u4e09\u79cd\u65b9\u6848\u7684\u5dee\u5f02\u6709\u4e00\u4e2a\u6700\u76f4\u89c2\u7684\u611f\u53d7\uff0c\u6211\u4eec\u628a\u5b83\u4eec\u7684\u5173\u952e\u6570\u636e\u653e\u5728\u4e00\u5f20\u8868\u91cc\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7279\u6027<\/th><th>MHA<\/th><th>MQA<\/th><th>GQA<\/th><\/tr><\/thead><tbody><tr><td><strong>KV \u4efd\u6570<\/strong><\/td><td>h\uff08\u5934\u6570\uff09<\/td><td>1<\/td><td>g\uff08\u7ec4\u6570\uff0c1 &lt; g &lt; h\uff09<\/td><\/tr><tr><td><strong>KV Cache \u5927\u5c0f<\/strong><\/td><td>\u6700\u5927<\/td><td>\u6700\u5c0f\uff081\/h\uff09<\/td><td>\u5c45\u4e2d\uff08g\/h\uff09<\/td><\/tr><tr><td><strong>\u63a8\u7406\u663e\u5b58\u5360\u7528<\/strong><\/td><td>\u9ad8<\/td><td>\u6781\u4f4e<\/td><td>\u4e2d\u4f4e<\/td><\/tr><tr><td><strong>\u8868\u8fbe\u80fd\u529b<\/strong><\/td><td>\u6700\u5f3a<\/td><td>\u6700\u5f31<\/td><td>\u63a5\u8fd1 MHA<\/td><\/tr><tr><td><strong>\u8bad\u7ec3\u96be\u5ea6<\/strong><\/td><td>\u4e2d\u7b49<\/td><td>\u7565\u9ad8<\/td><td>\u4e2d\u7b49<\/td><\/tr><tr><td><strong>\u4ee3\u8868\u6a21\u578b<\/strong><\/td><td>\u539f\u7248 Transformer\u3001GPT-3<\/td><td>PaLM\u3001\u90e8\u5206\u5c0f\u6a21\u578b<\/td><td>LLaMA-2\/3\u3001Qwen\u3001Mistral<\/td><\/tr><tr><td><strong>\u9002\u7528\u573a\u666f<\/strong><\/td><td>\u79d1\u7814\u3001\u5c0f\u89c4\u6a21\u63a8\u7406<\/td><td>\u6781\u81f4\u8d44\u6e90\u53d7\u9650<\/td><td>\u901a\u7528\u751f\u4ea7\u73af\u5883<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9009\u62e9\u5efa\u8bae<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u505a\u5b66\u672f\u7814\u7a76\u3001\u8ffd\u6c42\u6781\u81f4\u7684\u6a21\u578b\u80fd\u529b\u4e0a\u9650\uff0c\u4e14\u63a8\u7406\u8d44\u6e90\u5145\u8db3 \u2192 <strong>MHA<\/strong>\uff08\u6216\u8f83\u9ad8\u7ec4\u6570\u7684 GQA\uff09<\/li>\n\n\n\n<li>\u60f3\u7528\u5927\u6a21\u578b\u505a API \u670d\u52a1\uff0c\u9700\u8981\u9ad8\u5e76\u53d1\u548c\u957f\u4e0a\u4e0b\u6587 \u2192 <strong>GQA<\/strong>\uff08\u9ed8\u8ba4\u9009\u62e9\uff09<\/li>\n\n\n\n<li>\u69a8\u5e72\u8fb9\u7f18\u8bbe\u5907\uff08\u624b\u673a\u3001\u5d4c\u5165\u5f0f\uff09\u7684\u6bcf\u4e00\u6ef4\u6027\u80fd \u2192 <strong>MQA<\/strong>\uff08\u6216 GQA \u7ec4\u6570\u4e3a 1\uff09<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u524d\u6cbf\u63a2\u7d22\uff1a\u4ece GQA \u5230 MLA \u518d\u5230\u672a\u6765<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6545\u4e8b\u5230\u4e86 GQA\uff0cAttention \u7684\u67b6\u6784\u4f3c\u4e4e\u5df2\u7ecf\u8d8b\u4e8e\u5b8c\u7f8e\u3002\u4f46\u7814\u7a76\u8005\u4eec\u5e76\u6ca1\u6709\u505c\u4e0b\u811a\u6b65\u2014\u2014\u968f\u7740\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3\u7a81\u7834 100 \u4e07 Token\uff0c\u54ea\u6015\u662f GQA \u4e5f\u5f00\u59cb\u663e\u5f97\u201c\u81c3\u80bf\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.1 DeepSeek MLA\uff1a\u4f4e\u79e9\u538b\u7f29\u7684\u5947\u8ff9<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4 MQA \u662f\u7528\u201c\u5171\u4eab\u201d\u6765\u538b\u7f29\uff0c\u90a3\u4e48 <strong>DeepSeek \u63d0\u51fa\u7684 MLA\uff08Multi-head Latent Attention\uff0c\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff09<\/strong> \u5c31\u662f\u7528\u201c\u964d\u7ef4\u201d\u6765\u538b\u7f29\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">MLA \u7684\u6838\u5fc3\u6d1e\u5bdf\u662f\uff1a<strong>Key \u548c Value \u77e9\u9635\u4e2d\u5176\u5b9e\u6709\u5927\u91cf\u7684\u5197\u4f59\u4fe1\u606f\uff0c\u6211\u4eec\u53ef\u4ee5\u7528\u4e00\u4e2a\u4f4e\u79e9\u7684\u201c\u9690\u53d8\u91cf\uff08Latent Vector\uff09\u201d\u6765\u538b\u7f29\u5b83\u4eec\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6765\u8bf4\uff0cMLA \u4e0d\u518d\u4e3a\u6bcf\u4e2a Token \u5b58\u50a8\u5b8c\u6574\u7684 K \u548c V \u5411\u91cf\u3002\u5b83\u5728\u8ba1\u7b97 K \u548c V \u65f6\uff0c\u5148\u5c06\u8f93\u5165\u6295\u5f71\u5230\u4e00\u4e2a\u7ef4\u5ea6\u975e\u5e38\u5c0f\u7684\u201c\u6f5c\u5728\u7a7a\u95f4\u201d\uff08\u6bd4\u5982 64 \u7ef4\uff09\uff0c\u7136\u540e\u7528\u8fd9\u4e2a\u5c0f\u5411\u91cf\u5728\u6bcf\u6b21\u4f7f\u7528\u65f6\u52a8\u6001\u89e3\u538b\u51fa\u5b8c\u6574\u7684 K \u548c V\u3002\u7531\u4e8e\u53ea\u9700\u8981\u7f13\u5b58\u90a3\u4e2a\u5c0f\u5c0f\u7684\u6f5c\u5728\u5411\u91cf\uff0c<strong>KV Cache \u7684\u5927\u5c0f\u53c8\u88ab\u538b\u7f29\u4e86\u4e00\u4e2a\u6570\u91cf\u7ea7<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u66f4\u5173\u952e\u7684\u662f\uff0cMLA \u7684\u89e3\u538b\u8fc7\u7a0b\u53ef\u4ee5\u4e0e RoPE \u5de7\u5999\u7ed3\u5408\u2014\u2014DeepSeek \u56e2\u961f\u8bbe\u8ba1\u4e86\u4e00\u79cd\u89e3\u8026\u7684\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\u7b56\u7565\uff0c\u786e\u4fdd\u5728\u4f4e\u79e9\u538b\u7f29\u7684\u540c\u65f6\uff0c\u4f4d\u7f6e\u4fe1\u606f\u4f9d\u7136\u88ab\u51c6\u786e\u6ce8\u5165\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ed3\u679c\u662f\u4ec0\u4e48\uff1f<strong>DeepSeek-V2 \u7528 MLA \u5b9e\u73b0\u4e86\u540c\u65f6\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u7684\u5927\u5e45\u6548\u7387\u63d0\u5347<\/strong>\uff0cKV Cache \u4ec5\u4e3a\u4f20\u7edf MHA \u7684 5%~10%\u3002\u8fd9\u4e5f\u662f DeepSeek-V3 \u80fd\u591f\u4ee5 37B \u6fc0\u6d3b\u53c2\u6570\u5339\u654c\u66f4\u5927\u89c4\u6a21\u7a20\u5bc6\u6a21\u578b\u7684\u79d8\u5bc6\u6b66\u5668\u4e4b\u4e00\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.2 \u5176\u4ed6\u524d\u6cbf\u65b9\u6848\u4e00\u89c8<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u9664\u4e86 MLA\uff0c2025 \u5e74\u5230 2026 \u5e74\u95f4\u8fd8\u6d8c\u73b0\u4e86\u8bb8\u591a\u65e8\u5728\u8fdb\u4e00\u6b65\u4f18\u5316 Attention \u7684\u5de5\u4f5c\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\uff081\uff09Lightning Attention\uff1a\u878d\u5408 FlashAttention \u548c\u8ba1\u7b97\u91cd\u5851<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531 RWKV \u56e2\u961f\u63d0\u51fa\uff0cLightning Attention \u5728\u5355\u5185\u6838\u4e2d\u9ad8\u6548\u5904\u7406 Attention \u77e9\u9635\u7684\u6240\u6709\u5143\u7d20\uff0c\u7279\u522b\u4f18\u5316\u4e86\u591a GPU \u73af\u5883\u4e0b\u7684\u901a\u4fe1\u6a21\u5f0f\u3002\u5728 3B \u89c4\u6a21\u7684\u5bc6\u96c6\u6a21\u578b\u4e0a\u63a8\u7406\u901f\u5ea6\u6bd4 FlashAttention-3 \u5feb\u4e24\u500d\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\uff082\uff09Beyond KV Caching\uff1a\u629b\u5f03\u7f13\u5b58\u7684\u6fc0\u8fdb\u5c1d\u8bd5<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSeek \u7684 <strong>RetrievalAttention<\/strong> \u548c <strong>ContextParallel<\/strong> \u5c1d\u8bd5\u7528\u7a00\u758f\u68c0\u7d22\u548c\u5e76\u884c\u4e0a\u4e0b\u6587\u5206\u7247\u6765\u53d6\u4ee3\u4f20\u7edf\u7684\u7a20\u5bc6 KV Cache\u3002\u800c LLaMA \u793e\u533a\u7684 <strong>RocketKV<\/strong> \u5219\u5728\u4e0d\u7528\u95ea\u5b58\u5378\u8f7d\u7684\u60c5\u51b5\u4e0b\uff0c\u8ba9 7B \u6a21\u578b\u5728\u5355\u5f20 4090 \u4e0a\u5c31\u80fd\u5904\u7406\u767e\u4e07 Token \u7684\u957f\u4e0a\u4e0b\u6587\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\uff083\uff09MagicPIG\uff1a\u7528 LSH \u505a\u8fd1\u4f3c\u68c0\u7d22<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">MagicPIG \u7684\u7075\u611f\u6765\u6e90\u4e8e\u8ba1\u7b97\u673a\u79d1\u5b66\u4e2d\u7ecf\u5178\u7684\u5c40\u90e8\u654f\u611f\u54c8\u5e0c\uff08LSH\uff09\u2014\u2014\u628a Key \u5411\u91cf\u54c8\u5e0c\u5230\u6876\u4e2d\uff0c\u751f\u6210\u6bcf\u4e2a Token \u65f6\u53ea\u53bb\u76f8\u5173\u7684\u6876\u91cc\u627e Key\uff0c\u800c\u4e0d\u662f\u7528 Q \u53bb\u548c\u6240\u6709\u7f13\u5b58\u7684 K \u505a\u70b9\u79ef\u3002\u8fd9\u79cd\u65b9\u6cd5\u628a Attention \u7684 O(n\u00b2) \u590d\u6742\u6027\u964d\u5230\u4e86 O(n log n)\uff0c\u7279\u522b\u9002\u7528\u4e8e\u8d85\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\uff084\uff09\u786c\u4ef6\u4e0e\u7b97\u6cd5\u534f\u540c\u8bbe\u8ba1<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740<strong>\u5b58\u5185\u8ba1\u7b97\uff08PIM\uff0cProcessing-in-Memory\uff09<\/strong> \u82af\u7247\u7684\u6210\u719f\uff0c\u4e00\u4e9b\u7814\u7a76\u8005\u5f00\u59cb\u63a2\u7d22\u5c06 Attention \u8ba1\u7b97\u5b8c\u5168\u5378\u8f7d\u5230\u65b0\u7684\u786c\u4ef6\u67b6\u6784\u4e0a\u30022025 \u5e74\u7684\u4e00\u7bc7\u8bba\u6587\u7efc\u8ff0\u4e86\u5982\u4f55\u5229\u7528 PIM \u6253\u7834 Attention \u7684\u201c\u5185\u5b58\u5899\u201d\u74f6\u9888\uff0c\u4ee3\u8868\u4e86\u8f6f\u786c\u7ed3\u5408\u7684\u65b0\u65b9\u5411\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001FlashAttention\uff1a\u4e0d\u6539\u53d8\u67b6\u6784\uff0c\u5374\u6539\u53d8\u4e00\u5207\u7684\u7b97\u6cd5<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u7ed3\u675f Attention \u7684\u8bdd\u9898\u4e4b\u524d\uff0c\u5fc5\u987b\u82b1\u4e00\u70b9\u7bc7\u5e45\u81f4\u656c\u53e6\u4e00\u4e2a\u201c\u5e55\u540e\u82f1\u96c4\u201d\u2014\u2014<strong>FlashAttention<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u524d\u9762\u8ba8\u8bba\u7684 MQA\u3001GQA\u3001MLA\uff0c\u90fd\u662f\u4ece<strong>\u67b6\u6784\u5c42\u9762<\/strong>\u538b\u7f29 KV Cache \u6216\u53c2\u6570\u91cf\u3002\u800c <strong>FlashAttention \u8d70\u7684\u662f\u53e6\u4e00\u6761\u8def\uff1a\u4e0d\u6539\u53d8 Attention \u7684\u6570\u5b66\u5b9a\u4e49\uff0c\u53ea\u6539\u53d8\u5b83\u5728 GPU \u4e0a\u7684\u8ba1\u7b97\u65b9\u5f0f\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8.1 \u95ee\u9898\u7684\u6839\u6e90\uff1a\u5185\u5b58\u5899<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u516c\u5f0f\u4e2d\u6709\u4e00\u4e2a\u201c\u5669\u68a6\u7ea7\u201d\u7684\u64cd\u4f5c\uff1a<code>Q \u00b7 K^T<\/code> \u4f1a\u4ea7\u751f\u4e00\u4e2a <code>[seq_len, seq_len]<\/code> \u7684\u77e9\u9635\u3002\u5f53\u5e8f\u5217\u957f\u5ea6\u8fbe\u5230 128K \u65f6\uff0c\u8fd9\u4e2a\u77e9\u9635\u6709\u591a\u5927\uff1f<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5047\u8bbe\u7528 FP16 \u5b58\u50a8\uff1a128K \u00d7 128K \u00d7 2 \u5b57\u8282 \u2248 <strong>32 GB<\/strong><\/li>\n\n\n\n<li>\u5373\u4f7f\u662f H100 \u7684 80GB \u663e\u5b58\uff0c\u5355\u662f\u5b58\u8fd9\u4e2a\u77e9\u9635\u5c31\u5df2\u7ecf\u6349\u895f\u89c1\u8098\u3002\u800c\u5b9e\u9645\u7684\u63a8\u7406\u4e2d\uff0c\u8fd8\u6709\u6a21\u578b\u53c2\u6570\u3001KV Cache\u3001\u4e2d\u95f4\u6fc0\u6d3b\u503c\u8981\u653e\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf\u5b9e\u73b0\u4f1a\u5c06\u6574\u4e2a Attention \u77e9\u9635\u5199\u51fa\u5230 GPU \u7684 HBM\uff08\u9ad8\u5e26\u5bbd\u663e\u5b58\uff09\u4e2d\uff0c\u8ba1\u7b97\u5b8c Softmax \u540e\u518d\u8bfb\u56de\u6765\u3002\u8fd9\u79cd\u9891\u7e41\u7684\u201c\u5199\u5165-\u8bfb\u53d6-\u5199\u5165\u201d\u9020\u6210\u4e86\u63a8\u7406\u7684\u4e25\u91cd\u74f6\u9888\u2014\u2014<strong>\u7b97\u529b\u5728\u7b49\u5f85\u6570\u636e\u642c\u8fd0\uff0cGPU \u6838\u5fc3\u5927\u91cf\u95f2\u7f6e<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8.2 FlashAttention \u7684\u89e3\u6cd5\uff1a\u5206\u5757\u8ba1\u7b97 + \u5728\u7ebf Softmax<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">FlashAttention \u505a\u4e86\u4e24\u4ef6\u975e\u5e38\u806a\u660e\u7684\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5206\u5757\u8ba1\u7b97\uff08Tiling\uff09<\/strong>\uff1a\u4e0d\u4e00\u6b21\u6027\u751f\u6210\u6574\u4e2a Attention \u77e9\u9635\uff0c\u800c\u662f\u5c06 Q \u548c K \u5206\u6210\u5c0f\u5757\uff0c\u4e00\u5757\u4e00\u5757\u5730\u8ba1\u7b97\uff0c\u7528 SRAM\uff08\u7247\u4e0a\u7f13\u5b58\uff09\u4e34\u65f6\u5b58\u50a8\u4e2d\u95f4\u7ed3\u679c\uff0c\u7b97\u5b8c\u5c31\u6254\uff0c\u7edd\u4e0d\u8ba9\u5927\u77e9\u9635\u5b8c\u6574\u5730\u51fa\u73b0\u5728 HBM \u4e2d\u3002<\/li>\n\n\n\n<li><strong>\u5728\u7ebf Softmax<\/strong>\uff1a\u63a8\u5bfc\u51fa\u4e00\u79cd\u6570\u5b66\u4e0a\u7b49\u4ef7\u7684 Softmax \u9012\u63a8\u516c\u5f0f\uff0c\u4f7f\u5f97\u6bcf\u4e00\u5757\u7684\u8ba1\u7b97\u53ef\u4ee5\u72ec\u7acb\u8fdb\u884c\uff0c\u4e0d\u9700\u8981\u7b49\u5f85\u5168\u90e8\u6570\u636e\u5c31\u4f4d\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u6548\u679c\u600e\u6837\uff1f<strong>\u663e\u5b58\u5360\u7528\u4ece O(n\u00b2) \u964d\u4f4e\u5230 O(n)<\/strong>\u3002\u5728 128K \u957f\u5e8f\u5217\u4e0a\uff0cFlashAttention \u76f8\u6bd4\u6807\u51c6\u5b9e\u73b0\uff0c\u663e\u5b58\u5360\u7528\u51cf\u5c11 80% \u4ee5\u4e0a\uff0c\u901f\u5ea6\u63d0\u5347 5-8 \u500d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u4eca\uff0c<strong>FlashAttention \u5df2\u7ecf\u662f\u6240\u6709\u4e3b\u6d41\u6846\u67b6\uff08PyTorch\u3001vLLM\u3001Hugging Face Transformers\uff09\u7684\u9ed8\u8ba4\u540e\u7aef<\/strong>\u3002\u7531 FlashAttention \u539f\u4f5c\u8005\u521b\u529e\u7684 <strong>FlexAI\uff08\u540e\u66f4\u540d\u4e3a Aither\uff09<\/strong> \u7b49\u516c\u53f8\uff0c\u751a\u81f3\u5f00\u59cb\u5c06\u8fd9\u79cd\u201c\u5206\u5757 + \u878d\u5408 Kernel\u201d\u7684\u601d\u60f3\u505a\u6210\u4e13\u7528\u7684 AI \u4e91\u670d\u52a1\uff0c\u8fdb\u4e00\u6b65\u63a8\u52a8\u4e86\u6574\u4e2a\u884c\u4e1a\u7684\u6548\u7387\u9769\u547d\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 Attention \u6f14\u8fdb\u8def\u7ebf\u5168\u666f\u56fe<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba9\u6211\u4eec\u7528\u4e00\u5f20\u65f6\u95f4\u7ebf\u8868\u683c\u6765\u56de\u987e Attention \u67b6\u6784\u7684\u6f14\u5316\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u65f6\u95f4<\/th><th>\u65b9\u6848<\/th><th>\u6838\u5fc3\u521b\u65b0<\/th><th>\u5bf9\u63a8\u7406\u7684\u5f71\u54cd<\/th><\/tr><\/thead><tbody><tr><td>2017<\/td><td>MHA<\/td><td>\u591a\u5934\u5e76\u884c\uff0c\u591a\u89d2\u5ea6\u5173\u6ce8<\/td><td>\u6784\u5efa\u4e86 Transformer \u57fa\u7840\uff0c\u4f46 KV Cache \u5de8\u5927<\/td><\/tr><tr><td>2019<\/td><td>MQA<\/td><td>\u6240\u6709\u5934\u5171\u4eab KV<\/td><td>KV Cache \u964d\u81f3 1\/h\uff0c\u4f46\u80fd\u529b\u6709\u635f<\/td><\/tr><tr><td>2022<\/td><td>FlashAttention<\/td><td>\u5206\u5757\u8ba1\u7b97 + IO \u4f18\u5316<\/td><td>\u4e0d\u6539\u53d8\u67b6\u6784\uff0c\u4f46\u8bad\u7ec3\u63a8\u7406\u901f\u5ea6\u6570\u500d\u63d0\u5347<\/td><\/tr><tr><td>2023<\/td><td>GQA<\/td><td>\u5206\u7ec4\u5171\u4eab KV<\/td><td>\u7075\u6d3b\u8c03\u8282\u538b\u7f29\u6bd4\uff0c\u80fd\u529b\u63a5\u8fd1 MHA\uff0c\u6210\u4e3a\u6807\u914d<\/td><\/tr><tr><td>2024<\/td><td>MLA<\/td><td>\u4f4e\u79e9\u538b\u7f29 KV\uff0c\u7ed3\u5408 RoPE<\/td><td>KV Cache \u518d\u538b\u7f29\u4e00\u4e2a\u6570\u91cf\u7ea7\uff0c\u8fc8\u5411\u8d85\u957f\u4e0a\u4e0b\u6587<\/td><\/tr><tr><td>2025+<\/td><td>\u4e13\u7528\u786c\u4ef6\u534f\u540c<\/td><td>\u5b58\u5185\u8ba1\u7b97\uff08PIM\uff09\u7b49<\/td><td>\u8f6f\u786c\u7ed3\u5408\u6253\u7834\u201c\u5185\u5b58\u5899\u201d\u7684\u7ec8\u6781\u65b9\u6848<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">9.2 \u4e32\u8054\u7cfb\u5217\u77e5\u8bc6\uff1aAttention \u5728 LLM \u4e2d\u7684\u4f4d\u7f6e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u6211\u4eec\u4e94\u7bc7\u6587\u7ae0\u6784\u5efa\u7684\u5b8c\u6574\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 \u5411\u91cf\u52a0\u5165\u76f8\u5bf9\u4f4d\u7f6e\u65cb\u8f6c<\/li>\n\n\n\n<li><strong>Pre-Norm\uff08RMSNorm\uff09<\/strong> \u2192 \u7a33\u5b9a\u6570\u503c\u5206\u5e03<\/li>\n\n\n\n<li><strong>Attention\uff08GQA\/MLA\uff09<\/strong> \u2192 <strong>Token \u95f4\u7684\u4fe1\u606f\u4ea4\u4e92\u3001\u4e0a\u4e0b\u6587\u611f\u77e5<\/strong> \u2190 \u6211\u4eec\u5728\u672c\u7bc7\u5b66\u5230\u7684\u5185\u5bb9<\/li>\n\n\n\n<li><strong>\u6b8b\u5dee\u8fde\u63a5<\/strong> \u2192 \u4fdd\u7559\u539f\u59cb\u4fe1\u606f<\/li>\n\n\n\n<li><strong>Pre-Norm\uff08RMSNorm\uff09<\/strong> \u2192 \u518d\u6b21\u5f52\u4e00\u5316<\/li>\n\n\n\n<li><strong>FFN\uff08SwiGLU \/ MoE\uff09<\/strong> \u2192 \u77e5\u8bc6\u5b58\u50a8\u4e0e\u975e\u7ebf\u6027\u53d8\u6362<\/li>\n\n\n\n<li><strong>\u6b8b\u5dee\u8fde\u63a5<\/strong> \u2192 \u4fdd\u7559\u4fe1\u606f<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u81f3\u6b64\uff0c\u4e00\u4e2a\u5b8c\u6574\u7684 Transformer Block \u7684\u6240\u6709\u7ec4\u4ef6\u90fd\u5df2\u8bb2\u89e3\u5b8c\u6bd5\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>\u8d85\u957f\u4e0a\u4e0b\u6587\u5e38\u6001\u5316<\/strong>\uff1a\u4ece 128K \u5230 1M \u751a\u81f3\u65e0\u9650\u957f\u5ea6\uff0cAttention \u7684\u7a00\u758f\u5316\u548c\u538b\u7f29\u673a\u5236\u5c06\u6301\u7eed\u8fdb\u5316\u3002<\/li>\n\n\n\n<li><strong>MLA \u53ca\u66f4\u9ad8\u6548\u7684\u538b\u7f29\u65b9\u6848<\/strong>\uff1aDeepSeek \u5df2\u7ecf\u8bc1\u660e\u4f4e\u79e9\u538b\u7f29\u7684\u5de8\u5927\u6f5c\u529b\uff0c\u540e\u7eed\u5c06\u6709\u66f4\u591a\u7c7b\u4f3c\u601d\u8def\u51fa\u73b0\u3002<\/li>\n\n\n\n<li><strong>\u5b58\u5185\u8ba1\u7b97\uff08PIM\uff09\u843d\u5730<\/strong>\uff1a\u5f53\u7b97\u6cd5\u4f18\u5316\u63a5\u8fd1\u6781\u9650\uff0c\u771f\u6b63\u7684\u4e0b\u4e00\u6b21\u98de\u8dc3\u53ef\u80fd\u6765\u81ea\u786c\u4ef6\u9769\u547d\u3002<\/li>\n\n\n\n<li><strong>\u52a8\u6001\u8def\u7531\u7684 Attention<\/strong>\uff1a\u4e0d\u56fa\u5b9a\u4f7f\u7528\u6240\u6709\u5934\uff0c\u800c\u662f\u6839\u636e\u8f93\u5165\u52a8\u6001\u6fc0\u6d3b\u90e8\u5206\u5934\u2014\u2014Attention \u7248\u7684\u201cMoE\u201d\u3002<\/li>\n\n\n\n<li><strong>\u591a\u6a21\u6001\u6ce8\u610f\u529b\u7edf\u4e00\u5316<\/strong>\uff1a\u6587\u672c\u3001\u56fe\u50cf\u3001\u8bed\u97f3\u3001\u89c6\u9891\u7684 Attention \u673a\u5236\u9010\u6e10\u6536\u655b\u5230\u540c\u4e00\u5957\u6846\u67b6\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 \u7ec8\u4e8e\uff0c\u6211\u4eec\u6765\u5230\u4e86 Transformer \u6700\u6838\u5fc3\u3001\u6700\u201c\u7075\u9b42\u201d\u7684\u90e8\u5206\u2014\u2014Attention\uff08\u6ce8\u610f\u529b\u673a\u5236 [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":133,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[42,37],"class_list":["post-132","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-attention","tag-llm"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/132","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=132"}],"version-history":[{"count":3,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/132\/revisions"}],"predecessor-version":[{"id":136,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/132\/revisions\/136"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/133"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=132"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=132"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=132"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}