{"id":1445,"date":"2025-12-01T21:18:54","date_gmt":"2025-12-01T13:18:54","guid":{"rendered":"https:\/\/lingbo.online\/?p=1445"},"modified":"2025-12-02T22:11:00","modified_gmt":"2025-12-02T14:11:00","slug":"%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6%e5%8f%8a%e5%85%b6%e4%b8%bb%e8%a6%81%e5%8f%98%e4%bd%93%ef%bc%88mha%e3%80%81mqa%e3%80%81gqa%ef%bc%89","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6%e5%8f%8a%e5%85%b6%e4%b8%bb%e8%a6%81%e5%8f%98%e4%bd%93%ef%bc%88mha%e3%80%81mqa%e3%80%81gqa%ef%bc%89\/","title":{"rendered":"\u6ce8\u610f\u529b\u673a\u5236\u53ca\u5176\u4e3b\u8981\u53d8\u4f53\uff08MHA\u3001MQA\u3001GQA\uff09"},"content":{"rendered":"

\u7b80\u4ecb<\/h3>\n

Transformer\u7684\u6838\u5fc3\u90e8\u5206\u662fAttention\u6ce8\u610f\u529b\u673a\u5236\uff0c\u901a\u8fc7\u5f15\u5165Attention\u673a\u5236\uff0c\u6a21\u578b\u80fd\u591f\u5904\u7406\u5e8f\u5217\u6570\u636e\u7684\u65f6\u5019\u6709\u6548\u5173\u6ce8\u5e8f\u5217\u4e2d\u7684\u91cd\u70b9\u4fe1\u606f\uff0c\u800c\u4e0d\u4f9d\u8d56\u4e8e\u56fa\u5b9a\u7684\u987a\u5e8f\u6216\u8ddd\u79bb\u3002
\n\u4e0b\u56fe\u5c55\u793a\u4e86Attention\u5728NLP\u9886\u57df\u4e2d\u7684\u5177\u4f53\u542b\u4e49\uff1a\u5982\u679c\u8003\u8651\u4e00\u4e2a\u5355\u8bcdit\u7684\u7279\u5f81\uff0c\u90a3\u4e48\u5b83\u7684\u7279\u5f81\u5c06\u6839\u636e\u522b\u7684\u5355\u8bcd\u7684\u7279\u5f81\u52a0\u6743\u5f97\u5230\uff0c\u6bd4\u5982\u8bf4the animal\u548cit\u7684\u5173\u7cfb\u6bd4\u8f83\u8fd1\uff0c\u90a3\u4e48the animal\u5bf9\u5e94\u7684\u6743\u91cd\u503c\u5c31\u4f1a\u5f88\u9ad8\uff0c\u4ece\u800c\u5f71\u54cd\u4e0b\u4e00\u5c42it\u7684\u7279\u5f81\u3002
\n\"\"<\/p>\n

Self-Attention<\/h3>\n

\u81ea\u6ce8\u610f\u529b\u673a\u5236\u901a\u8fc7\u8ba1\u7b97\u8f93\u5165\u5e8f\u5217\u4e2d\u5404\u4e2a\u5143\u7d20\u4e4b\u95f4\u7684\u76f8\u5173\u6027\uff0c\u6765\u6355\u6349\u5e8f\u5217\u4e2d\u4e0d\u540c\u4f4d\u7f6e\u4e4b\u95f4\u7684\u76f8\u4e92\u4f9d\u8d56\u5173\u7cfb\u3002\u5177\u4f53\u6765\u8bf4\uff0cAttention\u673a\u5236\u5c06\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a\u5143\u7d20\u540c\u5176\u4ed6\u6240\u6709\u5143\u7d20\u8fdb\u884c\u6bd4\u8f83\uff0c\u4ece\u800c\u5224\u5b9a\u54ea\u4e9b\u5143\u7d20\u5bf9\u4e8e\u5f53\u524d\u5143\u7d20\u7684\u8f93\u51fa\u6700\u4e3a\u91cd\u8981\u3002\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u8fc7\u7a0b\u5305\u62ec\uff1a
\n\"\"<\/p>\n

1\u3001\u8ba1\u7b97Q\u3001K\u548cV\u77e9\u9635\uff1a<\/strong><\/p>\n

Q=XW^Q<\/span><\/center>
\n
K=XW^K<\/span><\/center>
\n
K=XW^V<\/span><\/center><\/p>\n

\u5176\u4e2d\uff0cX<\/span>\u662f\u8f93\u5165\u5e8f\u5217\uff0c\u7ef4\u5ea6\u4e3a\uff08N\uff0cd\uff09<\/span>\uff0c\u5176\u4e2dN<\/span>\u662f\u5e8f\u5217\u7684\u957f\u5ea6\uff0cd<\/span>\u662f\u6bcf\u4e00\u4e2a\u5e8f\u5217\u7684\u7279\u5f81\u7ef4\u5ea6\u3002W^Q<\/span>\u3001W^K<\/span>\u3001W^V<\/span>\u662f\u9700\u8981\u88ab\u8bad\u7ec3\u7684\u6743\u91cd\u77e9\u9635\uff0c\u7ef4\u5ea6\u4e3a\uff08d,d'\uff09<\/span>\u3002\u53d8\u6362\u540e\uff0c\u67e5\u8be2Q<\/span>\u3001\u952eK<\/span>\u548c\u6307V<\/span>\u7684\u7ef4\u5ea6\u90fd\u662f\uff08N,d'\uff09<\/span>\uff0c\u4e00\u822c\u6765\u8bf4\uff0cd'<\/span>\u548cd<\/span>\u53d6\u503c\u76f8\u540c\uff0c\u5747\u4e3a\u6a21\u578b\u7684\u9690\u85cf\u5c42\u7ef4\u5ea6\u3002<\/p>\n

2\u3001\u8ba1\u7b97\u81ea\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u7136\u540e\u6c42\u5173\u8054\u5ea6<\/strong>
\n\u81ea\u6ce8\u610f\u529b\u77e9\u9635\u7684\u8868\u8fbe\u5f0f\u4e3a:<\/p>\n

A=QK^T<\/span><\/center><\/p>\n

\u4f7f\u7528\u70b9\u79ef\u8ba1\u7b97\u67e5\u8be2Q<\/span>\u548c\u952eK<\/span>\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\uff0c\u56e0\u6b64\uff0cA<\/span>\u7684\u7ef4\u5ea6\u662f\uff08N,N\uff09<\/span>
\n\u5173\u8054\u5ea6\u7684\u8ba1\u7b97\u65b9\u6cd5\u4e3a\uff1a<\/p>\n

A'=softmax(\\frac{A}{\\sqrt{d_k}})<\/span><\/center><\/p>\n

\u6b64\u5904\u901a\u8fc7Softmax\u5bf9\u8ba1\u7b97\u7ed3\u679c\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u5f97\u5230\u6ce8\u610f\u529b\u6743\u91cd\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u8ba1\u7b97\u5173\u8054\u5ea6\u65f6\uff0c\u9700\u8981\u505ascale\u7684\u64cd\u4f5c\uff0c\u9664\\sqrt{d_k}<\/span>\uff0c\u4ece\u800c\u9632\u6b62\u70b9\u79ef\u8fc7\u5927\uff0c\u4fdd\u6301Softmax\u7a33\u5b9a\u3002<\/p>\n

3\u3001\u52a0\u6743\u6c42\u548c\uff0c\u5f97\u5230\u81ea\u6ce8\u610f\u529b\u8f93\u51fa<\/strong><\/p>\n

Q=A'V<\/span><\/center><\/p>\n

\u6700\u7ec8output\u7684\u8f93\u51fa\u7ef4\u5ea6\u4e3a (N,d') <\/span><\/p>\n

Multi-Head Attention<\/h3>\n

\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff08Multi-head Self-attention, MHA\uff09\u662f\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8fdb\u9636\uff0c\u540c\u65f6\u4e5f\u662ftransformer\u67b6\u6784\u7684\u6838\u5fc3\u7ec4\u4ef6\u4e4b\u4e00\u3002\u901a\u8fc7\u52a0\u5165\u6ce8\u610f\u529b\u5934\uff0c\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u5141\u8bb8\u6a21\u578b\u5728\u4e0d\u540c\u7684\u8868\u793a\u5b50\u7a7a\u95f4\u4e2d\u5e76\u884c\u5b66\u4e60\uff0c\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u5bf9\u4e8e\u4e0d\u540c\u7279\u5f81\u7684\u6355\u6349\u80fd\u529b\uff0c\u63d0\u5347\u6a21\u578b\u6027\u80fd\u3002
\n\u5728MHA\u4e2d\uff0c\u6bcf\u4e00\u4e2a\u201c\u5934\u201d\u53ef\u4ee5\u770b\u4f5c\u662f\u5728\u4e0d\u540c\u7684\u8868\u793a\u5b50\u7a7a\u95f4\u4e2d\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u4ece\u800c\u4f7f\u5f97\u6bcf\u4e2a\u5934\u80fd\u591f\u63d0\u53d6\u5230\u8f93\u5165\u5e8f\u5217\u7684\u4e0d\u540c\u7279\u5f81\u3002\u4f8b\u5982\uff0c\u5728NLP\u4efb\u52a1\u4e2d\uff0c\u67d0\u4e2a\u5934\u53ef\u80fd\u4e13\u6ce8\u4e8e\u53e5\u6cd5\u7ed3\u6784\uff0c\u53e6\u5916\u4e00\u4e2a\u5934\u6709\u53ef\u80fd\u4e13\u6ce8\u4e8e\u63d0\u53d6\u8bed\u4e49\u5185\u5bb9\u7b49\u3002
\n\"\"<\/p>\n

1\u3001\u8ba1\u7b97Q\u3001K\u548cV<\/strong>
\n\u8ba1\u7b97\u65b9\u6cd5\u4e0eSelf-Attention\u76f8\u540c\uff1a<\/p>\n

Q=XW^Q<\/span><\/center>
\n
K=XW^K<\/span><\/center>
\n
K=XW^V<\/span><\/center><\/p>\n

2\u3001\u5206\u5272\u5934<\/strong>
\n\u5c06Q<\/span>\u3001K<\/span>\u5206\u5272\u6210H<\/span>\u4e2a\u5934\uff0c\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\u662f\\frac{d}{H}<\/span>\uff0c\u901a\u8fc7\u77e9\u9635\u91cd\u6392\u5b9e\u73b0\uff1a<\/p>\n

Q_i=split(Q,\\frac{d}{H})<\/span><\/center>
\n
K_i=split(K,\\frac{d}{H})<\/span><\/center>
\n
V_i=split(V,\\frac{d}{H})<\/span><\/center><\/p>\n

\u5176\u4e2d\uff0cQ_i<\/span>\u3001K_i<\/span>\u548cV_i<\/span>
\n\u662f\u7b2ci<\/span>\u4e2ahead\u5934\u7684\u67e5\u8be2\u77e9\u9635\u3001\u952e\u77e9\u9635\u548c\u503c\u77e9\u9635\uff0c\u7ecf\u8fc7\u5206\u5272\u540e\uff0cQ<\/span>\u3001K<\/span>\u548cV<\/span>\u7684\u7ef4\u5ea6\u662f\uff08N,H,\\frac{d}{H}\uff09<\/span><\/p>\n

3\u3001\u8ba1\u7b97\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u6c42\u5173\u8054\u5ea6<\/strong>
\n\u5206\u522b\u8ba1\u7b97\u6bcf\u4e00\u4e2ahead\u7684\u6ce8\u610f\u529b\u5206\u6570\uff0c\u6c42\u5173\u8054\u5ea6\uff1a<\/p>\n

A_i = Q_{i}K_{i}^{T}<\/span><\/center>
\n
A_i' = softmax\\left( \\frac{A_i}{\\sqrt{\\frac{d}{h}}} \\right)<\/span><\/center><\/p>\n

4\u3001\u52a0\u6743\u6c42\u548c<\/strong><\/p>\n

O_i=A_i'V_i<\/span><\/center><\/p>\n

5\u3001\u62fc\u63a5\u5934\u7684\u8f93\u51fa\uff0c\u5e76\u7ecf\u8fc7\u7ebf\u6027\u5c42\u83b7\u5f97\u6700\u7ec8\u8f93\u51fa<\/strong>
\n\u5c06\u6240\u6709\u7684head\u5934\u7684\u8f93\u51fa\u8fdb\u884c\u62fc\u63a5\uff1a<\/p>\n

O=concat(O_1\uff0cO_2\uff0c\u2026\uff0cO_i,\u2026,O_h)<\/span><\/center><\/p>\n

\u6700\u540e\uff0c\u901a\u8fc7\u4e00\u4e2a\u7ebf\u6027\u5c42W_o<\/span>\u5bf9\u62fc\u63a5\u7684\u8f93\u51fa\u8fdb\u884c\u6295\u5f71\uff0c\u4ee5\u5f97\u5230\u6700\u7ec8\u7684\u591a\u5934\u6ce8\u610f\u529b\u8f93\u51fa\uff1a<\/p>\n

O'=OW^o<\/span><\/center><\/p>\n

Multi-Query Attention<\/h3>\n

\u591a\u67e5\u8be2\u6ce8\u610f\u529b\u673a\u5236\u662f\u5bf9\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u7684\u6539\u8fdb\uff0c\u65e8\u5728\u51cf\u5c11\u6a21\u578b\u7684\u53c2\u6570\u6570\u91cf\u548c\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u540c\u65f6\u4fdd\u6301\u6216\u63d0\u9ad8\u6a21\u578b\u7684\u6027\u80fd\u3002\u4e0e\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\u4e0d\u540c\uff0c\u5728MQA\u4e2d\uff0c\u6240\u6709\u7684Query\u5171\u4eab\u540c\u4e00\u7ec4\u952e\u548c\u503c\uff0c\u4f46\u6bcf\u4e2aQeury\u53ef\u4ee5\u6709\u4e0d\u540c\u7684\u6743\u91cd\u3002\u8fd9\u79cd\u65b9\u6cd5\u51cf\u5c11\u4e86\u9700\u8981\u5b58\u50a8\u7684\u952e\u548c\u503c\u7684\u8d28\u91cf\uff0c\u540c\u65f6\u4ecd\u7136\u5141\u8bb8\u901a\u8fc7\u591a\u4e2a\u67e5\u8be2\u6765\u6355\u6349\u4e0d\u540c\u5b50\u7a7a\u95f4\u7684\u7279\u5f81\u3002
\n\"\"<\/p>\n

1\u3001\u8ba1\u7b97Q\u3001K\u548cV<\/strong><\/p>\n

Q=XW^Q<\/span><\/center>
\n
K=XW^K\uff08\u5171\u4eab\uff09<\/span><\/center>
\n
V=XW^V\uff08\u5171\u4eab\uff09<\/span><\/center><\/p>\n

\u5176\u4e2dX<\/span>\u662f\u8f93\u5165\u5e8f\u5217\uff0c\u7ef4\u5ea6\u4e3a\uff08N,d\uff09<\/span>\uff0c\u5176\u4e2dN<\/span>\u662f\u5e8f\u5217\u7684\u957f\u5ea6\uff0cd<\/span>\u662f\u6bcf\u4e2a\u5143\u7d20\u7684\u7279\u5f81\u7ef4\u5ea6\u3002W^Q<\/span>\u3001W^K<\/span>\u548cW^V<\/span>\u662f\u9700\u8981\u88ab\u8bad\u7ec3\u7684\u6743\u91cd\u77e9\u9635<\/p>\n

2\u3001\u5206\u5272\u5934<\/strong>
\n\u5c06Q<\/span>\u3001K<\/span>\u548cV<\/span>\u5206\u5272\u6210h<\/span>\u4e2a\u5934\uff0c\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\u662f\\frac{d}{h}<\/span>\uff0c\u901a\u8fc7\u77e9\u9635\u91cd\u6392\u5b9e\u73b0\uff1a<\/p>\n

Q_i=split(Q,\\frac{d}{h})<\/span><\/center><\/p>\n

\u5176\u4e2d\uff0cQ_i<\/span>\u662f\u7b2ci<\/span>\u4e2ahead\u5934\u7684\u67e5\u8be2\u77e9\u9635\u3002\u6b64\u5904\u7684\u8ba1\u7b97\u6b65\u9aa4\u4e0eMHA\u7684\u8ba1\u7b97\u8fc7\u7a0b\u662f\u5b8c\u5168\u76f8\u540c\u7684\u3002
\n\u5728MQA\u4e2d\uff0c\u952e\u548c\u503c\u5f97\u77e9\u9635\u662f\u5171\u4eab\u7684\uff0c\u6a21\u578b\u53c2\u6570\u6570\u91cf\u51cf\u5c11\uff0c\u4ece\u800c\u8fbe\u5230\u51cf\u5c11\u5185\u5b58\u5360\u7528\u5e76\u52a0\u5feb\u63a8\u7406\u901f\u5ea6\u7684\u76ee\u7684\u3002\u7136\u800c\uff0c\u5171\u4eab\u952e\u548c\u503c\u77e9\u9635\u53ef\u80fd\u4f1a\u9650\u5236\u6a21\u578b\u6355\u83b7\u4e0d\u540c\u5934\u4e4b\u95f4\u4fe1\u606f\u7684\u80fd\u529b\uff0c\u56e0\u6b64\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u53ef\u80fd\u4f1a\u727a\u7272\u6027\u80fd\u3002<\/p>\n

Grouped Query Attention<\/h3>\n

\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u673a\u5236\u8fdb\u4e00\u6b65\u4f18\u5316\u4e86\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u901a\u8fc7\u5c06\u67e5\u8be2\u5411\u91cf\u5206\u7ec4\uff0c\u4f7f\u6bcf\u4e2a\u7ec4\u5171\u4eab\u4e00\u4e2a\u952e\u548c\u503c\u3002\u4e0e\u591a\u67e5\u8be2\u6ce8\u610f\u529b\u673a\u5236\u7c7b\u4f3c\uff0cGQA\u964d\u4f4e\u4e86\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u51cf\u5c11\u4e86\u5185\u5b58\u5360\u7528\uff0c\u5e76\u4e14\u901a\u8fc7\u5206\u7ec4\u65b9\u5f0f\u4fdd\u7559\u4e86\u6a21\u578b\u7684\u8868\u793a\u80fd\u529b\uff1a<\/p>\n

1\u3001\u8ba1\u7b97Q\u3001K\u548cV<\/strong><\/p>\n

Q=XW^Q<\/span><\/center>
\n
K=XW^K<\/span><\/center>
\n
V=XW^V<\/span><\/center><\/p>\n

2\u3001\u6309\u7ec4\u5206\u5272\u67e5\u8be2\u77e9\u9635Q<\/strong>
\n\u5c06\u67e5\u8be2\u77e9\u9635Q<\/span>\u5206\u5272\u6210G<\/span>\u4e2a\u7ec4\uff0c\u6bcf\u4e2a\u7ec4\u5305\u542b\\frac{H}{G}<\/span>\u4e2a\u5934\uff0c\u5f97\u5230G[katex]\u4e2a\u67e5\u8be2\u5b50\u77e9\u9635[katex]Q_1,Q_2,\u2026,Q_G[katex]<\/p>\n<p>**3\u3001\u8ba1\u7b97\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u7136\u540e\u6c42\u5173\u8054\u5ea6**\n\n\u5206\u522b\u8ba1\u7b97\u6bcf\u4e00\u4e2a\u7ec4[katex]i<\/span>\u7684\u5173\u8054\u5ea6\uff0c<\/p>\n

A_i'=softmax(\\frac{Q_iK_i^T}{\\sqrt{d}})<\/span><\/center><\/p>\n

4\u3001\u52a0\u6743\u6c42\u548c<\/strong>
\n\u5bf9\u4e8e\u7b2ci<\/span>\u4e2a\u7ec4\uff0c\u8f93\u51fa\u7ed3\u679c\u4e3a\uff1a<\/p>\n

O_i=A_i'V_i<\/span><\/center><\/p>\n

5\u3001\u62fc\u63a5\u5934\u7684\u8f93\u51fa\uff0c\u5e76\u7ecf\u8fc7\u7ebf\u6027\u5c42\u83b7\u5f97\u6700\u7ec8\u7684\u8f93\u51fa<\/strong>
\n\u5c06\u6240\u6709\u5934\u7684\u8f93\u51fa\u8fdb\u884cconcat<\/p>\n

O=concat(O_1,O_2,\u2026\uff0cO_i,\u2026,O_G)<\/span><\/center><\/p>\n

\u6700\u540e\uff0c\u901a\u8fc7\u4e00\u4e2a\u7ebf\u6027\u5c42W_o<\/span>\u5bf9\u62fc\u63a5\u7684\u8f93\u51fa\u8fdb\u884c\u6295\u5f71\uff0c\u4ee5\u5f97\u5230\u6700\u7ec8\u7684\u591a\u5934\u6ce8\u610f\u529b\u8f93\u51fa:<\/p>\n

O'=OW_O<\/span><\/center><\/p>\n

GQA\u80fd\u591f\u5728\u6548\u7387\u548c\u8d28\u91cf\u4e4b\u95f4\u5bfb\u6c42\u5e73\u8861\uff0c\u9002\u7528\u4e8e\u9700\u8981\u5904\u7406\u5927\u91cf\u6570\u636e\u548c\u4fdd\u6301\u5feb\u901f\u54cd\u5e94\u7684\u573a\u666f\u4e2d\u3002\u7136\u800c\uff0c\u4ed6\u4e5f\u8981\u5728\u5b9e\u73b0\u548c\u6027\u80fd\u4e0a\u8fdb\u884c\u4ed4\u7ec6\u7684\u6743\u8861\u3002<\/p>\n

\u603b\u7ed3<\/h3>\n\n\n\n\n\n\n\n\n
\u6ce8\u610f\u529b\u673a\u5236<\/th>\n\u6838\u5fc3\u673a\u5236<\/th>\n\u6838\u5fc3\u5dee\u5f02<\/th>\n\u4f18\u52bf<\/th>\n\u52a3\u52bf<\/th>\n\u9002\u7528\u573a\u666f<\/th>\n<\/tr>\n<\/thead>\n
MHA\uff08\u591a\u5934\u6ce8\u610f\u529b\uff09<\/td>\n\u5c06Q\u3001K\u3001V\u901a\u8fc7\u7ebf\u6027\u6620\u5c04\u5206\u89e3\u4e3a\u591a\u4e2a\u5e76\u884c\u6ce8\u610f\u529b\u5934\uff0c\u6bcf\u4e2a\u5934\u90e8\u62e5\u6709\u72ec\u7acb\u7684Q\u3001K\u3001V\u53c2\u6570<\/strong>\uff0c\u72ec\u7acb\u8ba1\u7b97\u6ce8\u610f\u529b\u5206\u6570\u540e\u62fc\u63a5\u8f93\u51fa<\/td>\n\u65e0\u53c2\u6570\u5171\u4eab\uff0c\u65e0\u4efb\u4f55\u538b\u7f29\u7b56\u7565\uff0c\u6bcf\u4e2a\u5934\u5b8c\u5168\u72ec\u7acb\u5efa\u6a21<\/td>\n1. \u591a\u5934\u72ec\u7acb\u5efa\u6a21\u80fd\u6355\u6349\u4e30\u5bcc\u7684\u8bed\u4e49\u7279\u5f81\u548c\u591a\u7ef4\u5ea6\u4f9d\u8d56\u5173\u7cfb\uff0c\u8868\u8fbe\u80fd\u529b\u6781\u5f3a\uff1b2. \u6280\u672f\u6210\u719f\u7a33\u5b9a\uff0c\u5728\u591a\u4efb\u52a1\u548c\u591a\u6a21\u6001\u573a\u666f\u4e2d\u8868\u73b0\u4f18\u5f02\uff1b3. \u65e0\u9700\u590d\u6742\u53c2\u6570\u8c03\u4f18\uff0c\u9002\u914d\u591a\u6570\u57fa\u7840Transformer\u4efb\u52a1<\/td>\n1. \u5185\u5b58\u590d\u6742\u5ea6\u8fbeO(h\u00b7d)\uff0c\u53c2\u6570\u91cf\u548c\u663e\u5b58\u5360\u7528\u6781\u9ad8\uff0c\u957f\u5e8f\u5217\u4efb\u52a1\u4e2d\u6613\u51fa\u73b0\u5185\u5b58\u74f6\u9888\uff1b2. \u8ba1\u7b97\u590d\u6742\u5ea6\u5448\u5e73\u65b9\u589e\u957f\uff0c\u63a8\u7406\u901f\u5ea6\u6162\uff0c\u90e8\u7f72\u6210\u672c\u9ad8<\/td>\n1. \u4e2d\u5c0f\u89c4\u6a21\u6a21\u578b\u7684\u9ad8\u7cbe\u5ea6\u4efb\u52a1\uff0c\u5982\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u6458\u8981\u7b49\uff1b2. \u5bf9\u6a21\u578b\u6027\u80fd\u8981\u6c42\u9ad8\u4e8e\u90e8\u7f72\u6548\u7387\u7684\u5b66\u672f\u7814\u7a76\u548c\u57fa\u7840\u6a21\u578b\u8bad\u7ec3<\/td>\n<\/tr>\n
MQA\uff08\u591a\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/td>\n\u6240\u6709\u67e5\u8be2\u5934\u4fdd\u6301\u72ec\u7acb\uff0c\u5168\u5c40\u5171\u4eab\u540c\u4e00\u7ec4K\u3001V\u53c2\u6570<\/strong>\uff0c\u4ec5\u901a\u8fc7\u72ec\u7acb\u67e5\u8be2\u5934\u6355\u6349\u90e8\u5206\u7279\u5f81\u5b50\u7a7a\u95f4\u4fe1\u606f<\/td>\nK\/V\u5168\u5c40\u5171\u4eab\u5b9e\u73b0\u53c2\u6570\u538b\u7f29\uff0c\u820d\u5f03\u5934\u95f4\u72ec\u7acb\u7684K\/V\u7279\u5f81<\/td>\n1. \u5927\u5e45\u51cf\u5c11K\/V\u53c2\u6570\u6570\u91cf\u548cKV\u7f13\u5b58\u5360\u7528\uff0c\u5185\u5b58\u590d\u6742\u5ea6\u4f4e\u81f3O(d)\uff1b2. \u63a8\u7406\u901f\u5ea6\u6781\u5feb\uff0c\u8f83MHA\u541e\u5410\u91cf\u53ef\u63d0\u534730%-50%\uff1b3. \u6a21\u578b\u7ed3\u6784\u7b80\u5316\uff0c\u90e8\u7f72\u6210\u672c\u4f4e<\/td>\n1. \u8fc7\u5ea6\u5171\u4eabK\/V\u5bfc\u81f4\u4e0a\u4e0b\u6587\u4fe1\u606f\u6355\u6349\u80fd\u529b\u4e0b\u964d\uff0c\u6a21\u578b\u7cbe\u5ea6\u6709\u660e\u663e\u635f\u5931\uff1b2. \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6a21\u578b\u7a33\u5b9a\u6027\u5dee\uff0c\u5bb9\u6613\u51fa\u73b0\u6536\u655b\u6ce2\u52a8<\/td>\n1. \u5185\u5b58\u53d7\u9650\u7684\u957f\u5e8f\u5217\u5feb\u901f\u63a8\u7406\u573a\u666f\uff0c\u5982\u5b9e\u65f6\u6587\u672c\u751f\u6210\u63a5\u53e3\uff1b2. \u5bf9\u7cbe\u5ea6\u635f\u5931\u5bb9\u5fcd\u5ea6\u8f83\u9ad8\u7684\u8f7b\u91cf\u5316\u90e8\u7f72\u573a\u666f\uff0c\u5982\u79fb\u52a8\u7aef\u7b80\u5355\u5bf9\u8bdd\u6a21\u578b<\/td>\n<\/tr>\n
GQA\uff08\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff09<\/td>\n\u628a\u67e5\u8be2\u5934\u5212\u5206\u4e3a\u82e5\u5e72\u7ec4\uff0c\u6bcf\u7ec4\u5185\u5171\u4eab\u4e00\u7ec4K\u3001V\u53c2\u6570\uff0c\u7ec4\u95f4K\u3001V\u76f8\u4e92\u72ec\u7acb<\/strong>\uff0c\u517c\u987e\u5206\u7ec4\u72ec\u7acb\u6027\u4e0e\u5171\u4eab\u6027<\/td>\nK\/V\u5206\u7ec4\u5171\u4eab\u5b8c\u6210\u53c2\u6570\u538b\u7f29\uff0c\u901a\u8fc7\u7ec4\u6570\u8c03\u8282\u5171\u4eab\u7a0b\u5ea6<\/td>\n1. \u53ef\u901a\u8fc7\u8c03\u8282\u5206\u7ec4\u6570\u91cf\u7075\u6d3b\u5e73\u8861\u6027\u80fd\u4e0e\u6548\u7387\uff0c\u7ec4\u6570\u7b49\u4e8e1\u65f6\u7b49\u6548MQA\uff0c\u7b49\u4e8e\u5934\u6570\u65f6\u7b49\u6548MHA\uff1b2. \u63a8\u7406\u901f\u5ea6\u63a5\u8fd1MQA\uff0c\u540c\u65f6\u6027\u80fd\u903c\u8fd1MHA\uff0cT5-XXL\u5b9e\u9a8c\u4e2d\u5ef6\u8fdf\u964d\u4f4e50%\u4e14\u7cbe\u5ea6\u635f\u5931\u5c0f\u4e8e1%\uff1b3. \u8bad\u7ec3\u548c\u63a8\u7406\u7a33\u5b9a\u6027\u4f18\u4e8eMQA<\/td>\n1. \u5206\u7ec4\u6570\u91cf\u4e3a\u8d85\u53c2\u6570\uff0c\u9700\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u53cd\u590d\u8c03\u8bd5\u4ee5\u9002\u914d\uff1b2. \u76f8\u6bd4MQA\uff0c\u53c2\u6570\u5171\u4eab\u7a0b\u5ea6\u964d\u4f4e\uff0c\u6781\u7aef\u573a\u666f\u4e0b\u7684\u5185\u5b58\u4f18\u52bf\u7565\u5f31<\/td>\n1. \u5927\u6a21\u578b\u7684\u5de5\u4e1a\u5316\u90e8\u7f72\uff0c\u5982LLaMA-2\u3001ChatGLM2\u7b49\u5927\u6a21\u578b\u7684\u63a8\u7406\u4f18\u5316\uff1b2. \u517c\u987e\u7528\u6237\u4f53\u9a8c\u548c\u90e8\u7f72\u6210\u672c\u7684\u573a\u666f\uff0c\u5982\u667a\u80fd\u5ba2\u670d\u3001\u5185\u5bb9\u751f\u6210\u5e73\u53f0<\/td>\n<\/tr>\n
MLA\uff08\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff09<\/td>\n\u901a\u8fc7\u4f4e\u79e9\u6295\u5f71\u5c06\u8f93\u5165\u6620\u5c04\u5230\u4f4e\u7ef4\u6f5c\u5728\u7a7a\u95f4\uff0c\u5bf9Q\u3001K\u3001V\u8fdb\u884c\u77e9\u9635\u5206\u89e3\u538b\u7f29<\/strong>\uff0c\u642d\u914d\u89e3\u8026\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\u4f18\u5316\u7279\u5f81\u6355\u6349<\/td>\n\u4f4e\u79e9\u77e9\u9635\u5206\u89e3\u5b9e\u73b0\u53c2\u6570\u538b\u7f29\uff0c\u501f\u52a9\u6f5c\u5728\u7a7a\u95f4\u964d\u4f4e\u8ba1\u7b97\u8d1f\u8377<\/td>\n1. \u4f4e\u79e9\u5206\u89e3\u53ef\u5c06\u53c2\u6570\u89c4\u6a21\u5927\u5e45\u538b\u7f29\uff0c\u5982r=64\u65f6\u53ef\u538b\u7f2916\u500d\uff0cKV\u7f13\u5b58\u91cf\u4ec5\u4e3aMHA\u76841\/10\uff1b2. \u957f\u5e8f\u5217\u5efa\u6a21\uff0810k tokens\u4ee5\u4e0a\uff09\u65f6\u8ba1\u7b97\u901f\u5ea6\u8f83MHA\u63d0\u53474-8\u500d\uff1b3. \u89e3\u8026\u4f4d\u7f6e\u7f16\u7801\u51cf\u5c11\u5934\u90e8\u95f4\u5e72\u6270\uff0c\u4fdd\u6301\u8f83\u597d\u7684\u7279\u5f81\u6355\u6349\u7cbe\u5ea6<\/td>\n1. \u8bad\u7ec3\u521d\u671f\u6536\u655b\u901f\u5ea6\u8f83\u6162\uff0c\u9700\u8981\u9002\u914d\u6f5c\u5728\u7a7a\u95f4\u7684\u53c2\u6570\u5b66\u4e60\uff1b2. \u6f5c\u5728\u7a7a\u95f4\u7684\u7ef4\u5ea6\u8bbe\u8ba1\u5bf9\u6027\u80fd\u5f71\u54cd\u6781\u5927\uff0c\u9700\u4e13\u4e1a\u8c03\u4f18\uff0c\u9002\u914d\u573a\u666f\u6709\u4e00\u5b9a\u5c40\u9650\u6027<\/td>\n1. 10k tokens\u4ee5\u4e0a\u7684\u8d85\u957f\u5e8f\u5217\u5efa\u6a21\u4efb\u52a1\uff0c\u5982\u957f\u6587\u6863\u5206\u6790\u3001\u89c6\u9891\u5e27\u5e8f\u5217\u5904\u7406\uff1b2. \u8d44\u6e90\u53d7\u9650\u7684\u8fb9\u7f18\u8bbe\u5907\u90e8\u7f72\uff0c\u5982\u5d4c\u5165\u5f0fAI\u63a8\u7406\u573a\u666f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\"\"<\/p>\n","protected":false},"excerpt":{"rendered":"

\u7b80\u4ecb Transformer\u7684\u6838\u5fc3\u90e8\u5206\u662fAttention\u6ce8\u610f\u529b\u673a\u5236\uff0c\u901a\u8fc7\u5f15\u5165Attention\u673a\u5236\uff0c\u6a21\u578b\u80fd\u591f\u5904\u7406\u5e8f\u5217\u6570\u636e\u7684\u65f6\u5019\u6709 …<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-1445","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1445","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1445"}],"version-history":[{"count":14,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1445\/revisions"}],"predecessor-version":[{"id":1464,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1445\/revisions\/1464"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1445"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1445"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1445"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}