{"id":4596,"date":"2025-10-29T14:50:52","date_gmt":"2025-10-29T06:50:52","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=4596"},"modified":"2025-10-29T14:50:53","modified_gmt":"2025-10-29T06:50:53","slug":"%e3%80%90%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%e3%80%91%e4%b8%8b%e4%b8%80%e4%bb%a3%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%ba%94%e8%af%a5%e6%98%af%e4%bb%80%e4%b9%88%e6%9e%b6%e6%9e%84%ef%bc%9f","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/note-deep-learning\/4596\/","title":{"rendered":"\u3010\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\u3011\u4e0b\u4e00\u4ee3\u5927\u6a21\u578b\u5e94\u8be5\u662f\u4ec0\u4e48\u67b6\u6784\uff1f"},"content":{"rendered":"\n<p id=\"uff3276b8\">\u7531\u4e8eself-Attention\u592a\u8d35\uff0c\u51fa\u73b0\u4e862\u4e2a\u65b9\u6848\uff1a<\/p>\n\n\n\n<p id=\"u5aab1196\">\u8fd9\u4e24\u4e2a\u65b9\u6848\u4e00\u4e2a\u662flinear-Attention\uff0c\u4ece\u590d\u6742\u7684\u6838\u51fd\u6570\u5165\u624b\uff0c\u8bd5\u56fe\u5c06self-Attention\u4ece\u4e8c\u6b21\u590d\u6742\u5ea6\u8f6c\u6362\u4e3a\u7ebf\u6027\u590d\u6742\u5ea6\u3002\u4f8b\u5982Minimax\u3001Qwen3-Next\u90fd\u662f\u8fd9\u6761\u8def\u7ebf\u3002<\/p>\n\n\n\n<p id=\"uf51780dd\">\u53e6\u4e00\u4e2a\u65b9\u6848\u662fNative Sparse Attention\uff08NSA\uff09\uff0c\u5c06\u5b8c\u6574\u7684\u6ce8\u610f\u529b\u8fdb\u884c\u7a00\u758f\u5316\uff0c\u4f7f\u5176\u53ef\u4ee5\u7528\u66f4\u5c11\u7684\u8ba1\u7b97\u5b8c\u6210\u5e38\u89c4self-Attention\u8fd1\u4f3c\u7684\u6548\u679c\u3002\u4e00\u4e2a\u5178\u578b\u7684\u4f7f\u7528\u5c31\u662fDeepSeek V3.2-Exp\u3002<\/p>\n\n\n\n<p id=\"ua2463c5d\">\u4e8c\u8005\u901a\u8fc7\u4e0d\u540c\u7684\u65b9\u6848\uff0c\u6700\u7ec8\u76ee\u7684\u90fd\u662f\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6or\u964d\u4f4ekv cache\u7684\u4f7f\u7528\u3002<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"zRB95\">DSA\uff08NSA\uff09<\/h1>\n\n\n\n<p id=\"ubbefcc71\">DeepSeek V3.2-Exp\u4f7f\u7528\u7684\u7684DeepSeek(Dynamic?) Sparse Attention[1,2]\u3002\u8fd9\u4e2a\u65b9\u6848\u662f\u7531DeepSeek\u7684ACL 2025 best paper\uff08NSA\uff09\u53d1\u5c55\u800c\u6765\uff0c\u8be5\u65b9\u6848\u5f53\u65f6\u88abcursor\u7684CEO\u770b\u597d\uff0c\u4e5f\u662f\u6700\u8fd1\u88ab\u70ed\u8bae\u7684\u4e0b\u4e00\u4ee3llm\u7684\u65b9\u5411\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p id=\"u2c07a804\">Aman Sanger\uff08\u8054\u5408\u521b\u59cb\u4eba\uff09\uff1a\u5e0c\u671b\u4ed6\u4eec\u4e0b\u4e00\u4e2a\u6a21\u578b\u80fd\u7528\u4e0a\u3002\u8fd9\u79cd\u6ce8\u610f\u529b\u673a\u5236\u7684\u53ef\u6269\u5c55\u6027\u5f88\u597d\uff0c\u636e\u8bf4\u6548\u679c\u6bd4\u4f20\u7edf\u6ce8\u610f\u529b\u66f4\u597d\u3002\u6838\u5fc3\u539f\u7406\u662f\u628a\u6ce8\u610f\u529b\u5206\u6210\u4e09\u90e8\u5206\u3002\u4e00\u90e8\u5206\u662f\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\uff0c\u4e3b\u8981\u5173\u6ce8\u6700\u8fd1\u53d1\u751f\u7684\u5185\u5bb9\uff0c\u6bd4\u5982\u6700\u8fd1 4000 \u4e2a Tokens\u3002\u5269\u4e0b\u4e24\u90e8\u5206\u5219\u662f\u5206\u5757\u6ce8\u610f\u529b\u3002\u5b83\u4f1a\u6bcf\u9694\u4e00\u6bb5 Token\uff0c\u628a\u90a3\u6bb5\u5185\u5bb9\u4f5c\u4e3a\u4e00\u4e2a \u201c\u5757\u201d \u7684 key \u548c value \u4fdd\u5b58\u4e0b\u6765\uff0c\u7136\u540e query \u4f1a\u53bb\u5173\u6ce8\u8fd9\u4e9b\u5757\u3002\u4e4b\u540e\u518d\u4ece\u8fd9\u4e9b\u5757\u91cc\u9009\u51fa\u6700\u76f8\u5173\u7684\u524d K \u4e2a\uff0c\u6700\u540e\u5bf9\u8fd9\u4e9b\u5757\u505a\u5168\u5c40\u6ce8\u610f\u529b\u3002\u6211\u89c9\u5f97\u8fd9\u79cd\u505a\u6cd5\u5f88\u9177\uff0c\u56e0\u4e3a\u5b83\u786e\u5b9e\u80fd\u5f88\u597d\u5730\u5728\u5f88\u957f\u7684\u4e0a\u4e0b\u6587\u91cc\u505a\u68c0\u7d22\u3002<\/p>\n\n\n\n<p id=\"ub62f78d2\">NSA\u5176\u5b9e\u5c31\u662f\u591a\u7ec4\u6ce8\u610f\u529b\u7684\u96c6\u5408\uff0c\u4ed6\u4eec\u80fd\u505a\u5230\u66f4\u7a00\u758f\uff08\u4e5f\u5c31\u66f4\u5feb\u3001\u66f4\u8282\u7701kv cache\uff0cV3.2\u7684kv cache\u5927\u6982\u53ef\u4ee5\u8282\u7ea680%\uff08\u76f8\u8f83\u4e8ev3)\uff09\uff0c\u540c\u65f6\u5728\u6027\u80fd\u4e5f\u53ef\u4ee5\u4fdd\u6301\u3002<br><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"306\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-8-1024x306.png\" alt=\"\" class=\"wp-image-4597\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-8-1024x306.png 1024w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-8-300x90.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-8-768x229.png 768w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-8-1536x459.png 1536w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-8.png 1677w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p id=\"u4a1ad353\">\u800cDSA\u5728NSA\u7684\u57fa\u7840\u4e0a\u505a\u4e86\u4e00\u5b9a\u7684\u53d6\u820d\uff0c\u53bb\u6389\u4e86compression\u5206\u652f\uff0c\u4e5f\u53d6\u6d88\u4e86\u6ed1\u7a97\u3002\u76f4\u63a5\u5728V3\u7684\u57fa\u7840\u4e0a\u65b0\u589e\u4e86\u4e00\u4e2aIndexer\uff08Lightning Indexer\uff09\u3002\u4e3b\u8981\u52a8\u673a\u5176\u5b9e\u662f\u4e3a\u4e86\u65b9\u4fbf\u8fdb\u884ccontinue pretraining\u3002<\/p>\n\n\n\n<p id=\"ufa5a6285\">\u5728\u4ecb\u7ecdV3.2\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u56de\u770b\u4e00\u4e0bV3\u7684\u7ed3\u6784\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"epftP\">Deepseek<\/h2>\n\n\n\n<p id=\"u17592ec2\">DeepSeek V3\u4f7f\u7528\u4e86MLA\u5e76\u572814.8T token\u8fdb\u884c\u8bad\u7ec3\uff0c\u540c\u65f6\uff0c\u4e5f\u4f7f\u7528\u521b\u65b0\u7684\u4f7f\u7528\u4e86MTP\u6765\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"826\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-9-1024x826.png\" alt=\"\" class=\"wp-image-4598\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-9-1024x826.png 1024w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-9-300x242.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-9-768x619.png 768w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-9.png 1482w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u89c2\u5bdfMLA\u7684\u5177\u4f53\u5b9e\u73b0\uff0c\u7531\u4e8eRoPE\u7684\u590d\u6742\u5ea6\u4e5f\u6bd4\u8f83\u9ad8\uff0c\u56e0\u6b64\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u9996\u5148\u5bf9q\u3001kv\u8fdb\u884c\u4e862\u6b21lora\u53d8\u6362\uff08\u7ea2\u6846\u5185\uff09\uff0c\u4e3b\u8981\u76ee\u7684\u662f\u5206\u79bb\u51fa\u4e24\u7ec4\u53c2\u6570\uff0c\u53ea\u5bf9\u5176\u4e2d\u4e00\u7ec4\u8f83\u5c0f\u7684q\/k\u8fdb\u884cRoPE\u64cd\u4f5c\uff0c\u56e0\u6b64\u53ef\u4ee5\u518d\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u52a0\u5feb\u8fd0\u7b97\u901f\u5ea6\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"897\" height=\"1024\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-10-897x1024.png\" alt=\"\" class=\"wp-image-4599\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-10-897x1024.png 897w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-10-263x300.png 263w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-10-768x876.png 768w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-10.png 1042w\" sizes=\"(max-width: 897px) 100vw, 897px\" \/><\/figure><\/div>\n\n\n<p id=\"u7eb76480\">\u800cv 3.2\uff0c\u8fdb\u884c\u4e862\u5904\u66f4\u6539\u3002\u53ef\u80fd\u662f<strong>\u8003\u8651\u5230continue training<\/strong>\u7684\u9700\u6c42\uff0c\u5b9e\u9645\u4e0aV3.2\u5e76\u6ca1\u6709\u4f7f\u7528\u4e00\u5f00\u59cb\u63d0\u5230\u7684NSA\uff0c\u800c\u662f\u4f7f\u7528\u4e86\u4e00\u4e2aDSA\uff0c\u4ed6\u662f\u5728V3\u7684\u57fa\u7840\u4e0a\u8fdb\u884ccontinue training\u5b8c\u6210\u3002<\/p>\n\n\n\n<p id=\"u55355bea\">\u6539\u52a8\u4e3b\u8981\u6709\u4ee5\u4e0b\u4e24\u70b9\uff1a<\/p>\n\n\n\n<ol>\n<li>\u901a\u8fc7\u6570\u5b66\u6052\u7b49\u53d8\u6362\uff0c\u5c06kv_cache\u524d\u79fb\u3002\u76f8\u6bd4v3\u7684\u5b8c\u6574cache\uff0cv3.2\u4f7f\u7528\u4e86\u4f4e\u79e9cache\uff0c\u8fdb\u4e00\u6b65\u8282\u7701\u4e86kv cache\u5f00\u9500\u3002\u5728\u89e3\u7801\u65f6\uff0c\u901a\u8fc7\u6052\u7b49\u53d8\u6362\uff0c\u4f7f\u7528\u4e24\u4e2a\u5c0fcache\u5373\u53ef\u5b8c\u6210\u63a8\u7406\uff08\u8ba1\u7b97\u6362\u5b58\u50a8\uff09\u3002<\/li>\n\n\n\n<li>\u5f15\u5165NSA\u7684Indexer\u548ctop k select\uff0c\u9009\u62e9\u91cd\u8981\u7684topk token\u8fdb\u884c\u8fd0\u7b97\uff0cmask\u6389\u5176\u4f59\u4f4d\u7f6e\u7684Attention\u3002<\/li>\n<\/ol>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"988\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-11-1024x988.png\" alt=\"\" class=\"wp-image-4600\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-11-1024x988.png 1024w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-11-300x290.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-11-768x741.png 768w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-11.png 1233w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p id=\"uf9695809\">v3.2\u76f8\u6bd4v3\uff0c\u63a8\u7406\u6548\u7387\u8fdb\u4e00\u6b65\u63d0\u5347\u3002\u4f46\u662f\u4e3b\u8981\u63d0\u5347\u7684\u70b9\u662fcache\u7684\u8282\u7701\u3002<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"dD8K7\">Linear Attention<\/h1>\n\n\n\n<p id=\"ue710c0f7\">Linear\u7684\u6838\u5fc3\u601d\u8def\u662f\u901a\u8fc7\u6539\u53d8\u77e9\u9635\u7684\u60e9\u7f5a\u987a\u5e8f\uff0c\u5c06 <code>(QK\u1d40)V<\/code> \u8f6c\u6362\u4e3a <code>Q(K\u1d40V)<\/code>\uff0c\u5b9e\u73b0\u590d\u6742\u5ea6\u4ece <strong>O(N\u00b2d)<\/strong> \u5230 <strong>O(Nd\u00b2)<\/strong> \u7684\u6839\u672c\u6027\u7a81\u7834\u3002<\/p>\n\n\n\n<p id=\"u379f0f36\">\u8bf4\u5230Linear Attention\u5c31\u4e0d\u53ef\u907f\u514d\u7684\u5fc5\u987b\u63d0\u5230Linear RNN\u3002Linear RNN\u91cc\u7684\u7ebf\u6027\u6307\u7684\u662f\u53bb\u6389\u4f20\u7edfRNN\u4e2d\u7684\u975e\u7ebf\u6027\u4f9d\u8d56\uff0c\u4ece\u800c\u652f\u6301\u5e76\u884c\u8ba1\u7b97\uff08\u5377\u79ef\u6216\u8005parallel scan\uff09\uff1b\u800c\u7ebf\u6027attention\u91cc\u7684\u7ebf\u6027\u6307\u7684\u662f\u53bb\u6389softmax\uff0c\u4ece\u800c\u964d\u4f4e\u539f\u6765\u5e8f\u5217\u957f\u5ea6\u5e73\u65b9\u7684\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n\n\n\n<p id=\"u2d7d5b22\">\u5c3d\u7ba1\u4e24\u8005\u7684\u51fa\u53d1\u70b9\u4e0d\u540c\uff0c\u5728decoder-only\u67b6\u6784\u7684\u80cc\u666f\u4e0b\uff0cLinear Attention + causal mask\u53ef\u4ee5\u8868\u8fbe\u6210Linear RNN\u7684\u5f62\u5f0f\uff0c\u53ea\u662f\u72b6\u6001\u7a7a\u95f4\u4ece\u4ee5\u5f80\u7684\u5411\u91cf\u53d8\u6210\u4e86\u77e9\u9635\uff0c\u53ef\u4ee5\u770b\u6210\u662f\u4e00\u79cd\u6269\u5927\u72b6\u6001\u7a7a\u95f4\u7684\u505a\u6cd5\u3002\u7ebf\u6027\u6ce8\u610f\u529b\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a<strong>\u5177\u6709\u77e9\u9635\u503c\u72b6\u6001\u7684\u7ebf\u6027 RNN\u3002<\/strong><\/p>\n\n\n\n<p id=\"u54e624c2\">\u6211\u4eec\u77e5\u9053\u5f88\u591a\u4f18\u79c0\u7684Linear\u6a21\u578b\u4f8b\u5982Mamba\u3001Based\u7b49\uff0c\u6211\u4eec\u5c06\u4ee5RWKV\uff08\u7ebf\u6027rnn\uff09\u548c DeltaNet\uff08\u7ebf\u6027attention\uff09\u4e3a\u4f8b\u7ee7\u7eed\u4ecb\u7ecd\uff0c\u4e8c\u8005\u90fd\u662fLinear Attention\u4e2d\u8f83\u4e3a\u5178\u578b\u4e14\u4f18\u79c0\u7684\u4ee3\u8868\u3002<\/p>\n\n\n\n<p id=\"u6cc50d49\">\u6b63\u5982\u6211\u4eec\u4e4b\u524d\u5bf9Linear Attention\u7684\u521d\u6b65\u4ecb\u7ecd\u4e2d\uff08<a href=\"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/paper-deep-learning\/4581\/\" data-type=\"post\" data-id=\"4581\">\u3010\u8bba\u6587\u3011\u4e00\u4e2a\u57fa\u4e8e\u72b6\u6001\u8f6c\u79fb\u7684\u9ad8\u6548\u63a8\u7406\u6846\u67b6<\/a>\uff09\u63d0\u5230\u7684\uff0cLinear Attention\u4e4b\u6240\u4ee5\u6ca1\u6709\u53d6\u4ee3self Attention\uff0c\u7f57\u5217\u9664\u4e86\u51e0\u4e2a\u53ef\u80fd\u7684\u539f\u56e0\uff0c\u73b0\u5728\u6211\u4eec\u91cd\u65b0\u6574\u7406Linear Attention\u7684\u4f18\u7f3a\u70b9\uff1a<\/p>\n\n\n\n<p id=\"ube5238d7\">\u4f18\u70b9\uff1a<\/p>\n\n\n\n<ol>\n<li>\u4f4e\u65f6\u95f4\u590d\u6742\u5ea6\u3002Linear\u987e\u540d\u601d\u4e49\uff0c\u5c06\u590d\u6742\u5ea6\u7531\u5e73\u65b9\u590d\u6742\u5ea6\u964d\u4e3a\u7ebf\u6027\u590d\u6742\u5ea6\uff0c\u8282\u7701\u4e86\u5b9d\u8d35\u7684\u8ba1\u7b97\u65f6\u95f4\u3002<\/li>\n\n\n\n<li>\u4f4e\u7a7a\u95f4\u590d\u6742\u5ea6\u3002\u5728\u8282\u7701\u65f6\u95f4\u7684\u540c\u65f6\uff0cLinear\u7531\u4e8e\u4e0d\u4f1a\u5bf9\u6bcf\u4e2atoken\u90fd\u8fdb\u884c\u4e00\u4e9b\u7f13\u5b58\uff0c\u7a7a\u95f4\u590d\u6742\u5ea6O(n)\u3002Linear\u53ea\u4f1a\u6709\u4e00\u4e9b\u56fa\u5b9a\u7684\u7f13\u5b58\u7a7a\u95f4\uff0c\u65e2\u5e38\u6570\u9879O(1)\u7684\u7a7a\u95f4\u590d\u6742\u5ea6\u3002<\/li>\n\n\n\n<li>\u957f\u5e8f\u5217\u4efb\u52a1\u3002Linear\u7684\u957f\u6587\u672c\u80fd\u529b\u8f83\u4e3a\u4f18\u79c0\u76f8\u6bd4self-attention\u66f4\u5bb9\u6613\u505a\u523032k\u3001100k\u751a\u81f3\u66f4\u957f\u7684\u957f\u5ea6\u3002\u8fd9\u4f7f\u5f97Linear Attention\u76f8\u6bd4self Attention\u5c31\u5177\u6709\u66f4\u591a\u7684\u53ef\u80fd\u6027\u3002\u4f8b\u5982\u9ad8\u5206\u8fa8\u7387\u56fe\u7247\u89e3\u6790&amp;\u751f\u6210\u3001\u9ad8\u91c7\u6837\u7387\u7684\u89c6\u9891\u7406\u89e3\u548c\u751f\u6210\u3001\u957f\u5e8f\u5217TTS\u3001\u957f\u65f6\u95f4\u7528\u6237\u5386\u53f2\u884c\u4e3a\u7b49\u7684\u60f3\u8c61\u7a7a\u95f4\u66f4\u5927\u3002<\/li>\n\n\n\n<li>\u5177\u5907TTT\u7279\u6027\uff08Test Time Training\uff09\u3002\u7531\u4e8e\u6240\u6709\u7684\u72b6\u6001\u90fd\u88ab\u5b58\u50a8\u5728\u4e00\u4e2a\u72b6\u6001\u77e9\u9635\u4e2d\uff0c\u56e0\u6b64\u53ef\u4ee5\u6a21\u62df\u68af\u5ea6\u4e0b\u964d\u7684\u7b56\u7565\u6765\u66f4\u65b0\u8fd9\u4e2a\u72b6\u6001\u77e9\u9635\uff0c\u4e5f\u5c31\u505a\u5230\u4e86\u5728\u63a8\u7406\u4e2d\u8bad\u7ec3\u3002\u589e\u5f3a\u4e86\u957f\u5e8f\u5217\u4efb\u52a1\u7684\u53ef\u9760\u6027\u3002<\/li>\n<\/ol>\n\n\n\n<p id=\"u0bd09f39\">\u4f18\u70b9\u663e\u8457\uff0c\u7f3a\u70b9\u540c\u6837\u663e\u8457\u3002\u4e3b\u8981\u6709\uff1a<\/p>\n\n\n\n<ol>\n<li>\u4fe1\u606f\u6709\u635f\u538b\u7f29\u3002\u7531\u4e8e\u6240\u6709\u7684\u72b6\u6001\u90fd\u88ab\u5b58\u50a8\u5728\u4e00\u4e2a\u72b6\u6001\u77e9\u9635\u4e2d\uff0c\u4fe1\u606f\u6709\u635f\u538b\u7f29\uff0c\u5f53\u5e8f\u5217\u957f\u5ea6\u8fdc\u8d85\u9690\u85cf\u72b6\u6001\u7684\u5bb9\u91cf\u65f6\uff0c\u8fd9\u79cd\u538b\u7f29\u5c31\u4f1a\u4ea7\u751f\u4fe1\u606f\u4e22\u5931\uff0c\u8fd9\u79cd\u4e22\u5931\u5bf9\u4e8e\u9700\u8981\u7cbe\u786e\u53ec\u56de\u7684\u7279\u5b9a\u4efb\u52a1\uff08\u5982\u68c0\u7d22\uff09\u662f\u81f4\u547d\u7684\u3002<\/li>\n\n\n\n<li>\u5bf9prompt\u654f\u611f\uff0c\u6ca1\u6709<strong>\u56de\u987e\u673a\u5236<\/strong>\u3002\u540c\u6837\u662f\u7531\u4fe1\u606f\u538b\u7f29\u5bfc\u81f4\u7684\uff0c\u5047\u8bbe\u4e00\u4e2a\u8f93\u5165\u5e8f\u5217 [text,question]\uff0c\u7531\u4e8e\u5728\u65e9\u671f\u4e0d\u786e\u5b9a\u4efb\u52a1\u6027\u8d28\uff0c\u56e0\u6b64\u6a21\u578b\u5f80\u5f80\u4e0d\u77e5\u9053\u4ec0\u4e48\u4fe1\u606f\u9700\u8981\u88ab\u538b\u7f29\uff0c\u4ec0\u4e48\uff08\u4f8b\u5982\u8f93\u5165\u4e00\u4e2a\u9605\u8bfb\u7406\u89e3\u9898\uff0c\u4f46\u662f\u5b9e\u9645\u95ee\u6a21\u578b\u6709\u591a\u5c11\u4e2a\u6570\u5b57\uff09\uff0c\u4f1a\u5bfc\u81f4\u65e0\u6548\u4fe1\u606f\u88ab\u538b\u7f29\uff0c\u9020\u6210\u6027\u80fd\u4e0b\u964d\u3002\u56e0\u6b64\u7eafLinear\u6a21\u578b\u5f80\u5f80\u5bf9prompt\u5f88\u654f\u611f\uff0c\u5fc5\u987b\u8981\u95ee\u9898\u524d\u7f6e\u3002<\/li>\n\n\n\n<li>\u6ce8\u5165\u6027\u3002Linear Attention\u53ef\u80fd\u5177\u6709\u6ce8\u5165\u6027\u3002\u7b80\u5355\u8bf4\u5c31\u662f\u7531\u4e8e\u53d6\u6d88\u4e86Softmax\uff0c\u91c7\u7528 ReLU\uff0c\u56e0\u6b64\u5bf9\u4e8e\u4e0d\u540c\u7684x\u53ef\u80fd\u4f1a\u6709\u540c\u4e00\u4e2ay\u7684\u60c5\u51b5\u3002\u5982\u4e0b\u56fe\u5bf9\u4e8e\u4e0d\u540c\u7684q1\/q2\/q3\uff0c\u90fd\u6709\u76f8\u540c\u7684\u5411\u91cf\u8868\u793a\u3002\u8fd9\u4f1a\u5bf9\u6a21\u578b\u7684\u6027\u80fd\u9020\u6210\u8d1f\u5f71\u54cd\u3002<\/li>\n<\/ol>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"441\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-12-1024x441.png\" alt=\"\" class=\"wp-image-4601\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-12-1024x441.png 1024w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-12-300x129.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-12-768x331.png 768w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-12.png 1229w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ol start=\"4\">\n<li>\u6ce8\u610f\u529b\u5206\u6563\u3002Linear Attention\u5f88\u96be\u505a\u5230\u50cfself Attention\u4e00\u6837\u5c06\u5927\u90e8\u5206\u6ce8\u610f\u529b\u96c6\u4e2d\u5728\u4e00\u4e2a\u90e8\u4f4d\uff08\u6bd4\u5982\u51e0\u4e2atoken\u4e0a\uff09\uff0c\u4e3b\u8981\u539f\u56e0\u4f9d\u7136\u662f\u7531\u4e8esoftmax\u5bfc\u81f4\uff0c\u7531\u4e8e\u6ca1\u6709\u6620\u5c04\uff0cAttention score\u5fc5\u987b\u4f9d\u9760\u795e\u7ecf\u5143\u76f4\u63a5\u8f93\u51fa\uff0c\u56e0\u6b64\u5f88\u96be\u51fa\u73b0\u6781\u9ad8\u503c\u3002\u8fd9\u5bfc\u81f4\u6574\u4f53\u6ce8\u610f\u529b\u6bd4\u8f83\u5206\u6563\u3002<\/li>\n\n\n\n<li>\u8bad\u7ec3\u6548\u7387\u95ee\u9898\u3002\u867d\u7136Linear Attention\u7684\u65f6\u95f4\u7a7a\u95f4\u590d\u6742\u5ea6\u90fd\u6bd4\u8f83\u4f4e\uff0c\u4f46\u662f\u7531\u4e8e\u77e9\u9635\u8fd0\u7b97\u5177\u6709\u7279\u6b8a\u6027\uff0c\u73b0\u5728GPU\u5bf9\u4e8e\u77e9\u9635\u8ba1\u7b97\u7684\u4f18\u5316\u5df2\u7ecf\u5341\u5206\u4f18\u79c0\u3002\u540c\u65f6\u5728causal\u7684\u60c5\u51b5\u4e0b\uff0c\u5c06Attention\u8ba1\u7b97\u65b9\u5f0f\u7531\u5de6\u4e58\u6539\u6210\u4e86\u53f3\u4e58\u4f1a\u51fa\u73b0\u4e0d\u80fd\u7528\u77e9\u9635\u8ba1\u7b97\u7684\u95ee\u9898\uff0c\u5fc5\u987b\u9002\u5e94\u7d2f\u52a0\u6c42\u548c\uff0c\u8fd9\u5c31\u5bfc\u81f4\u4e86\u5fc5\u987b\u4f7f\u7528for\u5faa\u73af\u800c\u4e0d\u80fd\u4f7f\u7528\u9ad8\u6548\u7684\u77e9\u9635\u8fd0\u7b97\uff0c\u8fd9\u9650\u5236\u4e86\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u7684\u6548\u7387\u3002<\/li>\n<\/ol>\n\n\n\n<p id=\"ud0aed40b\">\u4ee5\u4e0a\u8fd9\u4e9b\u7f3a\u70b9\u662f\u539f\u751fLinear Attention\u5177\u5907\u7684\u95ee\u9898\uff0c\u73b0\u5728\u8d8a\u6765\u8d8a\u591a\u7684\u7814\u7a76\u90fd\u5728\u8bd5\u56fe\u89e3\u51b3\u8fd9\u4e9b\u7f3a\u70b9\uff0c\u4e0d\u8fc7\u4f8b\u5982\u5728\u4e4b\u524dpaper weekly\u4e2d\u4ecb\u7ecd\u8fc7\u7684Lightning Linear Attention\u4ee5\u53ca\u6211\u4eec\u63a5\u4e0b\u6765\u8981\u4ecb\u7ecd\u7684RWKVs\u3001DeltaNet\u7b49\uff0c\u5df2\u7ecf\u5728\u9010\u6b65\u89e3\u51b3\u8fd9\u4e9b\u7f3a\u70b9\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"REQtX\">Receptance Weighted Key Value\uff08RNNs\uff09<\/h2>\n\n\n\n<p id=\"udf780cd3\">RWKV\u662f\u4e00\u4e2a\u5178\u578b\u7684\u7ebf\u6027RNN\u6a21\u578b\uff0c\u5b83\u548c\u4f20\u7edf\u7684Transformer\u6a21\u578b\u7684\u4e3b\u8981\u533a\u522b\u5728\u4e8e\u6ce8\u610f\u529b\u5c42\u3002<\/p>\n\n\n\n<p id=\"u74302133\">\u5728RWKV\u4e2d\uff0c\u81ea\u6ce8\u610f\u529b\uff08self attention\uff09\u88ab\u66ff\u6362\u4e3a\u4f4d\u7f6e\u7f16\u7801\uff08Position Encoding\uff09\u548cTimeMix\uff0c\u800c\u524d\u9988\u7f51\u7edc\uff08Feed-Forward Network\uff0cFFN\uff09\u88ab\u66ff\u6362\u4e3aChannelMix\u3002\u8fd9\u4e9b\u6539\u52a8\u4f7f\u5f97RWKV\u5728\u4fdd\u6301Transformer\u9ad8\u6548\u5e76\u884c\u8bad\u7ec3\u7684\u540c\u65f6\uff0c\u4e5f\u80fd\u591f\u5728\u63a8\u7406\u65f6\u4fdd\u6301\u8f83\u4f4e\u7684\u8ba1\u7b97\u548c\u5185\u5b58\u590d\u6742\u5ea6\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"635\" height=\"649\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-13.png\" alt=\"\" class=\"wp-image-4602\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-13.png 635w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-13-294x300.png 294w\" sizes=\"(max-width: 635px) 100vw, 635px\" \/><\/figure><\/div>\n\n\n<p>\u5176\u4e2d\uff0cTime-Mixing\u548cChannel-Mixing\u7684\u72b6\u6001\u8f6c\u79fb\u65b9\u7a0b\u5982\u4e0b\uff0c<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"347\" height=\"340\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-14.png\" alt=\"\" class=\"wp-image-4603\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-14.png 347w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-14-300x294.png 300w\" sizes=\"(max-width: 347px) 100vw, 347px\" \/><\/figure><\/div>\n\n\n<p id=\"u4c12b342\">\u8ba9\u6211\u4eec\u629b\u5f03\u6570\u5b66\u8bed\u8a00\uff0c\u4eba\u8bdd\u63cf\u8ff0\u4e00\u4e0b\u8fd9\u4e2a\u7b97\u5b50\uff1a\u4ed6\u5e72\u7684\u4e8b\u662f<strong>\u628a\u524d\u6587\u548c\u5b83\u6709\u8bed\u4e49\u76f8\u5173\u6027\u7684\u6240\u6709token\u62ff\u8fc7\u6765\uff0c\u52a0\u6743\u6df7\u5408\u4ed6\u4eec\u7684embed\u4fe1\u606f\uff0c\u8d21\u732e\u7ed9\u4e0b\u4e00\u4e2atoken\u3002<\/strong><\/p>\n\n\n\n<p id=\"u7135191d\">\u4e0d\u8fc7RWKV\u7ed3\u6784\u76f8\u5bf9\u590d\u6742\uff0c\u4e5f\u4e00\u76f4\u5728\u8fdb\u884c\u8fed\u4ee3\uff0c\u5728\u6700\u8fd1\u7684RWKV-8\u4e2d\uff0c\u8fd8\u6dfb\u52a0\u4e86\u4e00\u4e2a\u56de\u987e\u673a\u5236\uff0c\u7528\u6765\u89e3\u51b3\u4e0a\u6587\u6211\u4eec\u63d0\u5230\u7684\u5bf9prompt\u654f\u611f\u8fd9\u4e00\u95ee\u9898\u3002\u56de\u987e\u673a\u5236\u662f\u6307\u5728\u4e00\u6b21\u524d\u5411\u4f20\u64ad\u4e2d\uff0c\u663e\u5f0f\u5730\u5904\u7406\u5386\u53f2\u72b6\u6001\uff08\u6216\u5386\u53f2\u8f93\u5165\uff09\u5e76\u5c06\u5176\u878d\u5165\u5f53\u524d\u72b6\u6001\u8ba1\u7b97\u7684\u8fc7\u7a0b\u3002<\/p>\n\n\n\n<p id=\"ucf0c981e\">\u56de\u987e\u673a\u5236\u662f\u4e00\u4e2aO(n)\u590d\u6742\u5ea6\u7684\u540e\u7f00\u81ea\u52a8\u673a\uff0c\u901a\u8fc7\u8fd9\u4e2a\u81ea\u52a8\u673a\u6a21\u578b\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u53ef\u4ee5\u56de\u770b\uff0c\u4f46\u662f\u6574\u4f53\u80fd\u529b\u4f9d\u7136\u6709\u9650\uff0c\u4e0d\u8fc7\u8fd9\u4e5f\u63d0\u4f9b\u4e86\u4e00\u4e2a\u9ad8\u6548\u4f18\u96c5\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4e3a\u540e\u7eed\u7684\u5de5\u4f5c\u4e86\u63d0\u4f9b\u4e86\u4e00\u4e2a\u5207\u5b9e\u53ef\u884c\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<p id=\"u6bbfc1d2\">\u4e0d\u8fc7RWKV\u5e76\u6ca1\u6709\u89e3\u51b3\u8bad\u7ec3\u6548\u7387\u95ee\u9898\uff0c\u4ee5\u53ca\u590d\u6742\u7684\u4f7f\u7528\u6210\u672c\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"lrbQJ\">DeltaNet<\/h2>\n\n\n\n<p id=\"uca54a078\">\u4f20\u7edf\u7684 Linear-Attention \u6a21\u578b\u901a\u5e38\u91c7\u7528\u7b80\u5355\u7684\u7d2f\u52a0\u6216\u5e26\u8870\u51cf\u7684\u7d2f\u52a0\u65b9\u5f0f\u6765\u66f4\u65b0\u9690\u85cf\u72b6\u6001\u3002\u8fd9\u79cd\u65b9\u5f0f\u867d\u7136\u7b80\u5355\u9ad8\u6548\uff0c\u4f46\u53ef\u80fd\u4e0d\u662f\u6700\u4f18\u7684\u4fe1\u606f\u878d\u5408\u7b56\u7565\u3002<\/p>\n\n\n\n<p id=\"u1db2d170\"><strong>DeltaNet<\/strong> \u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u66f4\u65b0\u89c4\u5219\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u9690\u85cf\u72b6\u6001\u7684\u66f4\u65b0\u5e94\u8be5\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u91cd\u6784\u6216\u9884\u6d4b\u672a\u6765\u7684\u4fe1\u606f\u3002\u5177\u4f53\u6765\u8bf4\uff0cDeltaNet \u7684\u66f4\u65b0\u89c4\u5219\u53ef\u4ee5\u770b\u4f5c\u662f\u5728\u6bcf\u4e00\u6b65\u90fd\u6700\u5c0f\u5316\u4e00\u4e2a\u5c40\u90e8\u7684<strong>\u5747\u65b9\u8bef\u5dee\uff08MSE\uff09<\/strong>\u635f\u5931\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"340\" height=\"73\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-15.png\" alt=\"\" class=\"wp-image-4604\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-15.png 340w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-15-300x64.png 300w\" sizes=\"(max-width: 340px) 100vw, 340px\" \/><\/figure><\/div>\n\n\n<p id=\"u3d75fa54\">\u53ef\u4ee5\u7406\u89e3\u4e3a\uff1a\u65b0\u72b6\u6001=\u65e7\u72b6\u6001+\u8bef\u5dee\u3002\u6bd4\u5982\u4f60\u5728\u7784\u51c6\uff0c\u73b0\u5728\u504f\u5de6\uff0c\u90a3\u4e48\u65b0\u72b6\u6001\u5176\u5b9e\u5c31\u662f\u5f53\u524d\u72b6\u6001\u7684\u504f\u5de6+\u5411\u53f3\u7684\u8bef\u5dee\u3002\u56e0\u6b64\uff0cSt-1 \u76f8\u5f53\u4e8e\u201c\u504f\u5de6\u201d\uff0cv_t\u662ftarget\uff0c\u90a3\u4e48\u4e8c\u8005\u5dee\u503c\u5c31\u662f\u201c\u5411\u53f3\u8c03\u6574\u201d\uff0c\u5728\u7ed9\u4e88\u4e00\u4e2a\u9002\u5f53\u7684\u5b66\u4e60\u7387\uff0c\u6784\u6210\u4e0b\u4e00\u4e2a\u72b6\u6001\u3002\u56e0\u4e3a\u6709\u8fd9\u4e2a\u81ea\u9002\u5e94\u7684\u5b66\u4e60\u7387\u5b58\u5728\uff0c\u5b9e\u9645\u4e0a\u8fd8\u53ef\u4ee5\u505a\u5230\u72b6\u6001\u7684\u52a8\u6001\u9009\u62e9\u3002<\/p>\n\n\n\n<p id=\"ub8209f2e\">\u56e0\u6b64\uff0cDeltaNet\u7684\u4f18\u52bf\u4e3b\u8981\u4e3a\uff1a<\/p>\n\n\n\n<ul>\n<li>\u91c7\u7528 MSE \u635f\u5931\u800c\u975e\u7ebf\u6027\u635f\u5931\uff0c\u63d0\u4f9b\u66f4\u5f3a\u7684\u8bef\u5dee\u6821\u6b63\u80fd\u529b<\/li>\n\n\n\n<li>\u52a8\u6001\u5b66\u4e60\u7387 \u03b2_t \u5b9e\u73b0\u81ea\u9002\u5e94\u8bb0\u5fc6\u66f4\u65b0<\/li>\n\n\n\n<li>\u7406\u8bba\u4e0a\u7b49\u4ef7\u4e8e Test-Time-Training (TTT) \u7684\u7279\u6b8a\u5f62\u5f0f<\/li>\n<\/ul>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"683\" height=\"476\" src=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-16.png\" alt=\"\" class=\"wp-image-4605\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-16.png 683w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2025\/10\/image-16-300x209.png 300w\" sizes=\"(max-width: 683px) 100vw, 683px\" \/><\/figure><\/div>\n\n\n<p id=\"u368f6383\">\u6211\u4eec\u6ce8\u610f\u5230\uff0c\u8fd9\u4e2a\u6a21\u578b\u8fd8\u662f\u7528\u4e86\u4e00\u4e2a\u77ed\u5377\u79ef\uff0c\u8fd9\u4e3b\u8981\u662f\u7528\u6765\u6355\u83b7\u5c40\u90e8\u4f9d\u8d56\uff0c\u5f25\u8865\u6a21\u578b\u5168\u5c40\u6ce8\u610f\u529b\u7684\u4e0d\u8db3\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"Q30N0\">Qwen3-Next<\/h2>\n\n\n\n<p id=\"u7c5c7ebe\">\u4ecb\u7ecd\u5b8c\u4e86DeltaNet\uff0c\u4ecb\u7ecd\u6211\u4eec\u7684\u91cd\u5934\u620fQwen3-Next\uff0c\u4e00\u4e2a80A3B\u7684\u6781\u81f4\u7a00\u758f\u7684\u6a21\u578b\uff0c\u8bbe\u7f6e\u4e86512\u4e2a\u4e13\u5bb6\uff0c\u5728\u63a8\u7406\u65f6\u4f7f\u75281\u4e2a\u5171\u4eab\u4e13\u5bb6\u548c10\u4e2a\u6fc0\u6d3b\u4e13\u5bb6\u3002\u4ec5\u5bf9\u524d25%\u7684\u4f4d\u7f6e\u7ef4\u5ea6\u6dfb\u52a0RoPE\u3001\u65f6\u4f7f\u7528MTP\uff08multi-token Prediction\uff09\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6a21\u578b\u7684\u8bad\u7ec3\u6548\u7387\u4e0e\u63a8\u7406\u6027\u80fd\u3002<\/p>\n\n\n\n<p id=\"ufa05270d\">\u8bad\u7ec3\u6570\u636e\u7531Qwen3\u768436T\u8bad\u7ec3\u6570\u636e\u5747\u5300\u91c7\u6837\u7684\u4e00\u4e2a15T\u5b50\u96c6\u6784\u6210\u3002\u5c3d\u7ba1\u5728model card\u4e2d\u63d0\u5230\uff0cQwen3-Next\u7684\u541e\u5410\u80fd\u529b\u662fQwen3\u76847~10\u500d\uff0c\u4f46\u662f\u7531\u4e8e\u73b0\u6709infra\u652f\u6301\u5f88\u5f31\uff0c\u5b9e\u9645\u4e0aQwen3-Next\u76ee\u524d\u7684\u6210\u672c\u662f\u663e\u8457\u5927\u4e8e\u540c\u7b49\u5c3a\u5bf8\u7684Dense\u6a21\u578b\u7684\u3002\u8fd9\u4e3b\u8981\u662f\u56e0\u4e3a\u76ee\u524dLinear\u7684\u57fa\u5efa\u76f8\u5bf9\u8f83\u5f31\uff0c\u8fd8\u9700\u8981\u793e\u533a\u8ddf\u8fdb\u76f8\u5173\u7b97\u5b50\u7684\u4f18\u5316\u3002<\/p>\n\n\n\n<p id=\"u146252ec\">Qwen3-Next\u5728\u7ed3\u6784\u8bbe\u8ba1\u4e0a\u6bd4\u8f83\u4fdd\u5b88\uff0c\u4e0d\u50cfMinimax-01\u572887.5%\u7684\u5c42\u4e2d\u4f7f\u7528Linear Attention\uff0cQwen3-Next\u572875%\u4f7f\u7528\u4e86Gated DeltaNet\uff08GDN\uff0c\u5728DeltaNet\u7ee7\u7eed\u6dfb\u52a0\u4e00\u4e2aGate\uff0c\u7528\u6765\u6d88\u9664Attention sink\uff09\u6765\u8fdb\u884c\u957f\u8ddd\u79bb\u5efa\u6a21\uff0c25%\u4f7f\u7528\u4e86\u6807\u51c6self-Attention\u6765\u4fdd\u7559\u4f8b\u5982\u68c0\u7d22\u7b49\u5173\u952e\u80fd\u529b\uff0c\u63d0\u5347\u5173\u952etoken\u7684\u6355\u6349\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/intranetproxy.alipay.com\/skylark\/lark\/0\/2025\/png\/147556570\/1760585977520-00616664-495c-47e4-a0ae-4a4eebbe4b5b.png\" alt=\"\" title=\"\"\/><\/figure>\n\n\n\n<p id=\"u994bc279\">\u8fd9\u4e2a\u6a21\u578b\u7531\u4e8e\u4f7f\u7528\u4e86\u5f88\u591a\u7684\u65b0\u6280\u672f\u4f8b\u5982Gate Attention\u3001MTP\u3001Zero-Centered RMSNorm\u3001\u878d\u5408\u67b6\u6784\u4ee5\u53ca\u9ad8\u7a00\u758fMoE\u7b49\uff0c\u6027\u80fd\u5f97\u5230\u4e86\u663e\u8457\u63d0\u5347\uff0c\u5728\u6fc0\u6d3b3B\u7684\u60c5\u51b5\u4e0b\uff0c\u5728\u90e8\u5206\u6838\u5fc3\u6307\u6807\u4e0a\u5ab2\u7f8eDense 32B\u3002<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"tlXhi\">\u603b\u7ed3<\/h1>\n\n\n\n<p id=\"u51a680aa\">\u672c\u6587\u4ecb\u7ecd\u4e86NSA\u548cLinear(RNN)\u4e24\u4e2a\u65b9\u6848\uff0c<\/p>\n\n\n\n<p id=\"u993afdd4\">Linear\u6d41\u6d3e\uff0c\u5047\u8bbe\u524d\u6587\u6240\u6709\u4fe1\u606f\u90fd\u53ef\u4ee5\u901a\u8fc7\u53ea\u770b\u4e00\u904d\u5c31\u88ab\u538b\u5230\u4e00\u4e2a\u56fa\u5b9astate\u91cc\u3002\u56e0\u6b64\u957f\u6587\u4e0bstate\u80fd\u529b\u4e0d\u884c\uff0c\u8bb0\u4e0d\u4f4f\u4fe1\u606f\uff0c\u6027\u80fd\u6349\u6025\u3002<\/p>\n\n\n\n<p id=\"ua29f261c\">sparse\u6d41\u6d3e\uff0c\u4f8b\u5982nsa\uff0c\u5047\u8bbe\u524d\u6587\u6709\u7528\u7684token\u81f3\u591a\u53ea\u6709k\u4e2a\uff0ck\u662f\u8fdc\u5c0f\u4e8en\u7684\u5e38\u6570\u3002\u5047\u8bbe\u592a\u5f3a\uff0c\u66f4\u591a\u7684\u662f\u57fa\u4e8e\u7ecf\u9a8c\uff0c\u5982\u679c\u67d0\u65f6\u523b\u5b9e\u9645\u9700\u8981\u7684token\u6bd4k\u591a\uff0c\u90a3\u5c31\u4e0d\u884c\u4e86\u3002<\/p>\n\n\n\n<p id=\"u43c26d5a\">\u540c\u65f6\uff0c\u4e8c\u8005\u90fd\u6ca1\u6709\u201c\u56de\u987e\u524d\u6587\u6240\u6709\u8bed\u4e49\u76f8\u4f3c\u7684token\u201d\u529f\u80fd\uff0c\u4e5f\u5c31\u662f\u8bf4\u90fd\u53ef\u80fd\u9700\u8981\u4e00\u4e2a\u7c7b\u4f3c\u4e8eRWKV-8\u7684\u56de\u987e\u673a\u5236\u6765\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\u3002<\/p>\n\n\n\n<p id=\"u3c9be253\">\u4e0d\u8fc7\u8fd9\u4e00\u9650\u5236\u4f9d\u7136\u662f\u5bf9\u4e8e\u7eafLinear\/sparse\u6a21\u578b\u6765\u8bf4\u3002\u76ee\u524d\u7684Linear\u6a21\u578b\u4e5f\u4f1a\u5728\u4e2d\u95f4\u7a7f\u63d2\u6807\u51c6self Attention\u6765\u6062\u590d\u4e00\u90e8\u5206\u7d22\u5f15\u6027\u80fd\uff0c\u800cV3.2\u4e5f\u5e76\u4e0d\u662f\u4f20\u7edf\u6ed1\u7a97\u4e00\u7c7b\u7684\u7684NSA\uff0cV3.2\u901a\u8fc7Indexer\u6765\u6784\u9020Attention\u7684\u7a00\u758f\u6027\uff0c\u7528\u8fc7Indexer\u6765\u9009\u62e9\u91cd\u8981token\u6765\u4e8c\u6b21\u6784\u9020\u7a00\u758f\u6027\uff0c\u4e0d\u8fc7\u8fd9\u4e2a\u7a00\u758f\u6027\u5e76\u6ca1\u6709\u6539\u53d8\u5b9e\u9645\u53c2\u4e0eAttention\u8fd0\u7b97\u7684\u77e9\u9635\u5927\u5c0f\u3002<\/p>\n\n\n\n<p id=\"udbe58e4a\">\u672a\u6765\u7684\u5f62\u6001\uff0c\u4e5f\u53ef\u80fd\u4f9d\u7136\u662fself Attention\u7684\u67d0\u79cd\u6539\u8fdb\u3002\u5c31\u50cf It is easy to make a smart model cheap,but is hard to make a cheap model smart\uff0c\u76ee\u524dLinear\u548cNSA\u76f8\u6bd4self attention\u4f9d\u7136\u662f\u6709\u4e00\u5b9a\u7684\u6027\u80fd\u52a3\u52bf\uff0c\u4f46\u662f\u968f\u7740\u6a21\u578b\u7684\u53d1\u5c55\uff0c\u4e5f\u4ec5\u4ec5\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u5b58\u5728\u4e00\u4e9b\u52a3\u52bf\u3002\u4f46\u662f\u66f4\u53ef\u80fd\u662f\u4e8c\u8005\u7684\u67d0\u79cd\u7ed3\u5408\uff0c\u6bd5\u7adfLinear\u7684\u60f3\u8c61\u7a7a\u95f4\u8db3\u591f\u5927\uff0c\u6027\u80fd\u4f18\u5316\u7684\u4e5f\u5dee\u4e0d\u591a\u4e86\uff0c\u5982\u679c\u771f\u7684\u6709\u65e0\u9650\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u4e00\u5929\uff0c\u53ef\u80fd\u6a21\u578b\u5fae\u8c03\u5c31\u6d88\u5931\u4e86\uff0c\u540c\u65f6\u4e00\u4e9b\u9ad8\u6d88\u8017\u7684\u591a\u6a21\u6001\u4efb\u52a1\u4e5f\u66f4\u5bb9\u6613\u843d\u5730\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7531\u4e8eself-Attention\u592a\u8d35\uff0c\u51fa\u73b0\u4e862\u4e2a\u65b9\u6848\uff1a \u8fd9\u4e24\u4e2a\u65b9\u6848\u4e00\u4e2a\u662flinear-Attenti [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_mi_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[9],"tags":[],"views":1199,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/4596"}],"collection":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/comments?post=4596"}],"version-history":[{"count":1,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/4596\/revisions"}],"predecessor-version":[{"id":4606,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/4596\/revisions\/4606"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=4596"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=4596"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=4596"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}