{"id":5590,"date":"2025-01-24T14:46:15","date_gmt":"2025-01-24T06:46:15","guid":{"rendered":"https:\/\/nullthought.net\/?p=5590"},"modified":"2025-02-23T11:08:03","modified_gmt":"2025-02-23T03:08:03","slug":"gated-deltanet%e6%9e%b6%e6%9e%84%ef%bc%9a%e7%bb%93%e5%90%88%e9%97%a8%e6%8e%a7%e6%9c%ba%e5%88%b6%e5%92%8cdelta%e6%9b%b4%e6%96%b0%e8%a7%84%e5%88%99%ef%bc%8c%e6%8f%90%e5%8d%87%e7%ba%bf%e6%80%a7transforme","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5590","title":{"rendered":"Gated DeltaNet\u67b6\u6784\uff1a\u7ed3\u5408\u95e8\u63a7\u673a\u5236\u548cDelta\u66f4\u65b0\u89c4\u5219\uff0c\u63d0\u5347\u7ebf\u6027Transformer\u5728\u957f\u5e8f\u5217\u5efa\u6a21\u548c\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\u4e2d\u7684\u8868\u73b0"},"content":{"rendered":"\n

\u8bba\u6587Gated Delta Networks: Improving Mamba2 with Delta Rule<\/a><\/strong>\u63d0\u51fa\u4e86Gated DeltaNet\u67b6\u6784\uff0c\u901a\u8fc7\u7ed3\u5408\u95e8\u63a7\u673a\u5236\u548cDelta\u66f4\u65b0\u89c4\u5219\uff0c\u63d0\u5347\u7ebf\u6027Transformer\u5728\u957f\u5e8f\u5217\u5efa\u6a21\u548c\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

\u8bba\u6587\u4f5c\u8005\u4e3aSonglin Yang, Jan Kautz, Ali Hatamizadeh\uff0c\u6765\u81eaNVIDIA\u3002<\/p>\n\n\n\n

\"\"
Gated Delta Networks: Improving Mamba2 with Delta Rule<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/strong><\/p>\n\n\n\n

Transformer\u67b6\u6784\u56e0\u5176\u5353\u8d8a\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5728\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4e2d\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6210\u679c\u3002\u7136\u800c\uff0cTransformer\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u5177\u6709\u5e73\u65b9\u7ea7\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u968f\u7740\u5e8f\u5217\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u8ba1\u7b97\u5f00\u9500\u6025\u5267\u589e\u52a0\uff0c\u5bfc\u81f4\u8bad\u7ec3\u548c\u63a8\u7406\u7684\u8d44\u6e90\u9700\u6c42\u6781\u9ad8\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86\u7ebf\u6027Transformer\u6a21\u578b\uff0c\u65e8\u5728\u901a\u8fc7\u6539\u8fdb\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u65b9\u5f0f\uff0c\u964d\u4f4e\u5176\u8ba1\u7b97\u590d\u6742\u5ea6\u3002\u7ebf\u6027Transformer\u7684\u5173\u952e\u601d\u60f3\u662f\u7528\u6838\u5316\u7684\u70b9\u79ef\u6ce8\u610f\u529b\u66ff\u4ee3\u4f20\u7edf\u7684softmax\u6ce8\u610f\u529b\uff0c\u4ece\u800c\u4f7f\u5f97\u8ba1\u7b97\u590d\u6742\u5ea6\u7531\u4f20\u7edf\u7684 O(L2<\/sup>) \u964d\u4f4e\u81f3 O(L)\uff0c\u5176\u4e2dL\u662f\u5e8f\u5217\u7684\u957f\u5ea6\u3002<\/p>\n\n\n\n

\u5c3d\u7ba1\u7ebf\u6027Transformer\u6a21\u578b\u5728\u63a8\u7406\u9636\u6bb5\u5927\u5e45\u63d0\u5347\u4e86\u8ba1\u7b97\u6548\u7387\uff0c\u4f46\u5b83\u4eec\u4ecd\u7136\u5b58\u5728\u4e00\u4e9b\u5c40\u9650\u6027\uff0c\u5c24\u5176\u662f\u5728\u5904\u7406\u957f\u5e8f\u5217\u548c\u8fdb\u884c\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\u65f6\u3002\u4f8b\u5982\uff0c\u5f53\u5e8f\u5217\u957f\u5ea6\u8d85\u8fc7\u6a21\u578b\u7684\u7ef4\u5ea6\u65f6\uff0c\u7ebf\u6027Transformer\u4f1a\u9762\u4e34\u201c\u8bb0\u5fc6\u78b0\u649e\u201d\u7684\u95ee\u9898\uff0c\u5bfc\u81f4\u65e0\u6cd5\u7cbe\u786e\u68c0\u7d22\u91cd\u8981\u4fe1\u606f\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0cMamba2\u6a21\u578b\u5f15\u5165\u4e86\u95e8\u63a7\u673a\u5236\u548c\u8870\u51cf\u89c4\u5219\uff0c\u4ee5\u589e\u5f3a\u957f\u5e8f\u5217\u5efa\u6a21\u80fd\u529b\u3002\u7136\u800c\uff0cMamba2\u5728\u8bb0\u5fc6\u7ba1\u7406\u4e0a\u4ecd\u7136\u5b58\u5728\u4e0d\u8db3\uff0c\u7279\u522b\u662f\u5728\u9700\u8981\u52a8\u6001\u6e05\u9664\u8bb0\u5fc6\u65f6\u7684\u5904\u7406\u80fd\u529b\u8f83\u5f31\u3002\u4e3a\u6b64\uff0c\u672c\u6587\u63d0\u51fa\u4e86Gated DeltaNet\u67b6\u6784\uff0c\u901a\u8fc7\u7ed3\u5408\u95e8\u63a7\u673a\u5236\u548cDelta\u66f4\u65b0\u89c4\u5219\uff0c\u65e8\u5728\u63d0\u5347\u7ebf\u6027Transformer\u5728\u957f\u5e8f\u5217\u5efa\u6a21\u548c\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u4e3b\u8981\u8d21\u732e\u4e0e\u521b\u65b0\u70b9<\/strong><\/p>\n\n\n\n

    \n
  1. \u63d0\u51fa\u4e86\u95e8\u63a7Delta\u89c4\u5219<\/strong>\uff1a\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u8bb0\u5fc6\u66f4\u65b0\u673a\u5236\u2014\u2014\u95e8\u63a7Delta\u89c4\u5219\u3002\u8be5\u89c4\u5219\u901a\u8fc7\u5f15\u5165\u4e00\u4e2a\u6570\u636e\u9a71\u52a8\u7684\u95e8\u63a7\u56e0\u5b50\u6765\u52a8\u6001\u8c03\u6574\u8bb0\u5fc6\u66f4\u65b0\u8fc7\u7a0b\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5728\u9700\u8981\u65f6\u8fc5\u901f\u6e05\u9664\u4e0d\u518d\u9700\u8981\u7684\u4fe1\u606f\uff0c\u540c\u65f6\u53c8\u80fd\u591f\u9ad8\u6548\u5730\u66f4\u65b0\u91cd\u8981\u7684\u8bb0\u5fc6\u5185\u5bb9\u3002\u4e0e\u4f20\u7edf\u7684\u7ebf\u6027Transformer\u6a21\u578b\u76f8\u6bd4\uff0c\u95e8\u63a7Delta\u89c4\u5219\u5728\u8bb0\u5fc6\u7ba1\u7406\u4e0a\u66f4\u5177\u7075\u6d3b\u6027\u548c\u7cbe\u786e\u6027\u3002<\/li>\n\n\n\n
  2. \u5e76\u884c\u8bad\u7ec3\u7b97\u6cd5\u7684\u4f18\u5316<\/strong>\uff1a\u4e3a\u4e86\u9ad8\u6548\u5730\u5b9e\u73b0\u95e8\u63a7Delta\u89c4\u5219\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u786c\u4ef6\u9ad8\u6548\u7684\u5e76\u884c\u8bad\u7ec3\u7b97\u6cd5\u3002\u8be5\u7b97\u6cd5\u5229\u7528\u4e86\u6700\u65b0\u7684\u786c\u4ef6\u52a0\u901f\u6280\u672f\uff0c\u5982GPU\u7684\u5f20\u91cf\u6838\u5fc3\uff0c\u4ee5\u5b9e\u73b0\u5927\u89c4\u6a21\u5e76\u884c\u8ba1\u7b97\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u8bad\u7ec3\u7684\u901f\u5ea6\u548c\u6548\u7387\u3002<\/li>\n\n\n\n
  3. \u6df7\u5408\u6a21\u578b\u67b6\u6784<\/strong>\uff1a\u5728Gated DeltaNet\u7684\u57fa\u7840\u4e0a\uff0c\u672c\u6587\u8fd8\u63d0\u51fa\u4e86\u51e0\u79cd\u6df7\u5408\u6a21\u578b\u67b6\u6784\uff0c\u5c06Gated DeltaNet\u5c42\u4e0e\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\uff08SWA\uff09\u6216Mamba2\u5c42\u7ed3\u5408\u3002\u8fd9\u4e9b\u6df7\u5408\u6a21\u578b\u5728\u6027\u80fd\u4e0a\u8f83\u5355\u4e00\u7684Gated DeltaNet\u6a21\u578b\u6709\u663e\u8457\u7684\u63d0\u5347\uff0c\u5c24\u5176\u662f\u5728\u957f\u5e8f\u5217\u548c\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\u4e0a\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e09\u3001\u6a21\u578b\u539f\u7406\u4e0e\u8bbe\u8ba1<\/strong><\/p>\n\n\n\n

    1.\u7ebf\u6027\u6ce8\u610f\u529b\u7684\u57fa\u7840<\/strong>\uff1a
    \u7ebf\u6027Transformer\u6a21\u578b\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u4f20\u7edf\u7684softmax\u6ce8\u610f\u529b\u673a\u5236\u66ff\u6362\u4e3a\u4e00\u79cd\u57fa\u4e8e\u77e9\u9635\u8fd0\u7b97\u7684\u6838\u5316\u70b9\u79ef\u6ce8\u610f\u529b\u673a\u5236\u3002\u4f20\u7edf\u7684\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3a O(L2<\/sup>)\uff0c\u800c\u7ebf\u6027\u6ce8\u610f\u529b\u901a\u8fc7\u6539\u53d8\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u65b9\u5f0f\uff0c\u80fd\u591f\u5c06\u590d\u6742\u5ea6\u964d\u4f4e\u5230\u7ebf\u6027\u7ea7\u522b\uff0c\u5373 O(L)\u3002\u5176\u8ba1\u7b97\u8fc7\u7a0b\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u4e2a\u9012\u5f52\u7684\u7ebf\u6027\u66f4\u65b0\u8fc7\u7a0b\uff0c\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5176\u4e2d\uff0cSt<\/sub>\u662f\u5728\u7b2ct\u6b65\u7684\u72b6\u6001\uff0cvt<\/sub>\u548c kt<\/sub>\u5206\u522b\u662f\u8f93\u5165\u7684\u503c\u548c\u952e\uff0c\u77e9\u9635\u4e58\u6cd5\u7684\u7ed3\u679c\u4ee3\u8868\u4e86\u6a21\u578b\u7684\u72b6\u6001\u66f4\u65b0\u8fc7\u7a0b\u3002<\/p>\n\n\n\n

    2.Delta\u89c4\u5219\u4e0e\u95e8\u63a7\u673a\u5236\u7684\u7ed3\u5408<\/strong>\uff1a
    Delta\u89c4\u5219\u901a\u8fc7\u6bcf\u6b21\u66f4\u65b0\u65f6\u5bf9\u8bb0\u5fc6\u8fdb\u884c\u589e\u91cf\u4fee\u6539\uff0c\u4ece\u800c\u4fdd\u6301\u8bb0\u5fc6\u7684\u8fde\u7eed\u6027\u3002\u8be5\u89c4\u5219\u7684\u66f4\u65b0\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5176\u4e2d\uff0c\u03b2t<\/sub>\u662f\u66f4\u65b0\u7684\u6bd4\u4f8b\u56e0\u5b50\uff0c\u5b83\u51b3\u5b9a\u4e86\u65b0\u65e7\u4fe1\u606f\u7684\u6743\u8861\u3002\u5728Delta\u89c4\u5219\u4e2d\uff0c\u65e7\u7684\u952e\u503c\u5bf9\u88ab\u66ff\u6362\u4e3a\u65b0\u7684\u952e\u503c\u5bf9\uff0c\u4f46\u6bcf\u6b21\u66f4\u65b0\u53ea\u4f1a\u4fee\u6539\u4e00\u4e2a\u952e\u503c\u5bf9\uff0c\u56e0\u6b64\u5bf9\u4e8e\u9700\u8981\u5feb\u901f\u6e05\u9664\u65e7\u4fe1\u606f\u7684\u4efb\u52a1\uff0cDelta\u89c4\u5219\u7684\u8868\u73b0\u53ef\u80fd\u4f1a\u53d7\u5230\u9650\u5236\u3002<\/p>\n\n\n\n

    \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0cGated DeltaNet\u63d0\u51fa\u4e86\u95e8\u63a7Delta\u89c4\u5219\uff0c\u901a\u8fc7\u5728Delta\u89c4\u5219\u7684\u57fa\u7840\u4e0a\u5f15\u5165\u95e8\u63a7\u673a\u5236\u3002\u95e8\u63a7\u56e0\u5b50\u03b1t<\/sub>\u200b\u51b3\u5b9a\u4e86\u8bb0\u5fc6\u7684\u8870\u51cf\u7a0b\u5ea6\uff0c\u4ece\u800c\u80fd\u591f\u7075\u6d3b\u5730\u6e05\u9664\u4e0d\u518d\u9700\u8981\u7684\u8bb0\u5fc6\u5185\u5bb9\u3002\u95e8\u63a7Delta\u89c4\u5219\u7684\u66f4\u65b0\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6a21\u578b\u80fd\u591f\u5728\u9700\u8981\u65f6\u6e05\u9664\u4e0d\u518d\u9700\u8981\u7684\u8bb0\u5fc6\uff0c\u540c\u65f6\u66f4\u65b0\u91cd\u8981\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n

    3.\u786c\u4ef6\u9ad8\u6548\u7684\u5e76\u884c\u8bad\u7ec3<\/strong>\uff1a
    \u4e3a\u4e86\u4f7f\u95e8\u63a7Delta\u89c4\u5219\u80fd\u591f\u5728\u5927\u89c4\u6a21\u6570\u636e\u4e0a\u9ad8\u6548\u8bad\u7ec3\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u786c\u4ef6\u9ad8\u6548\u7684\u5e76\u884c\u8bad\u7ec3\u7b97\u6cd5\u3002\u901a\u8fc7\u4f7f\u7528GPU\u7684\u5f20\u91cf\u6838\u5fc3\uff08Tensor Core\uff09\u52a0\u901f\u77e9\u9635\u4e58\u6cd5\u64cd\u4f5c\uff0c\u5e76\u7ed3\u5408\u5206\u5757\u5e76\u884c\u8ba1\u7b97\uff08chunkwise parallelism\uff09\uff0c\u6a21\u578b\u80fd\u591f\u5728\u4fdd\u6301\u8bad\u7ec3\u6548\u7387\u7684\u540c\u65f6\uff0c\u6700\u5927\u5316\u786c\u4ef6\u8d44\u6e90\u7684\u5229\u7528\u7387\u3002\u8be5\u65b9\u6cd5\u7684\u5173\u952e\u5728\u4e8e\u5c06\u8f93\u5165\u5e8f\u5217\u62c6\u5206\u6210\u591a\u4e2a\u5c0f\u5757\uff0c\u5e76\u901a\u8fc7\u6bcf\u4e2a\u5c0f\u5757\u7684\u6700\u7ec8\u72b6\u6001\u6765\u8ba1\u7b97\u8f93\u51fa\u3002\u8fd9\u79cd\u5206\u5757\u5e76\u884c\u5f62\u5f0f\u4e0d\u4ec5\u53ef\u4ee5\u51cf\u5c11\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u8fd8\u80fd\u591f\u5728\u591aGPU\u73af\u5883\u4e0b\u6709\u6548\u5206\u914d\u8ba1\u7b97\u4efb\u52a1\uff0c\u4ece\u800c\u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n\n\n\n

    \u56db\u3001\u5b9e\u9a8c\u5206\u6790<\/strong><\/p>\n\n\n\n

    1.\u8bed\u8a00\u5efa\u6a21\u4e0e\u5e38\u8bc6\u63a8\u7406<\/strong>\uff1a
    \u5728\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e2d\uff0cGated DeltaNet\u76f8\u6bd4\u4f20\u7edf\u7684\u7ebf\u6027Transformer\u6a21\u578b\uff0c\u5982Mamba2\u548cDeltaNet\uff0c\u5728\u56f0\u60d1\u5ea6\uff08perplexity\uff09\u4e0a\u8868\u73b0\u51fa\u8272\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0cGated DeltaNet\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u80fd\u591f\u66f4\u597d\u5730\u7ba1\u7406\u8bb0\u5fc6\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u6a21\u578b\u7684\u56f0\u60d1\u5ea6\uff0c\u5c24\u5176\u662f\u5728\u5e38\u8bc6\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u51c6\u786e\u7387\u4e5f\u5f97\u5230\u4e86\u63d0\u5347\u3002\u6b64\u5916\uff0c\u6df7\u5408\u6a21\u578b\uff08\u5982Gated DeltaNet-H1\u548cGated DeltaNet-H2\uff09\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u6027\u80fd\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u590d\u6742\u63a8\u7406\u4efb\u52a1\u65f6\uff0c\u6df7\u5408\u6a21\u578b\u7684\u4f18\u52bf\u66f4\u4e3a\u660e\u663e\u3002<\/p>\n\n\n\n

    2.\u957f\u5e8f\u5217\u5916\u63a8\u80fd\u529b<\/strong>\uff1a
    \u5728\u5904\u7406\u8d85\u957f\u5e8f\u5217\u65f6\uff0cGated DeltaNet\u5c55\u73b0\u4e86\u8f83\u5f3a\u7684\u5916\u63a8\u80fd\u529b\u3002\u5b9e\u9a8c\u4e2d\uff0c\u6a21\u578b\u5728\u957f\u5ea6\u8fbe\u523020K\u7684\u5e8f\u5217\u65f6\u4ecd\u80fd\u4fdd\u6301\u8f83\u4f4e\u7684\u56f0\u60d1\u5ea6\uff0c\u800c\u4f20\u7edf\u7684\u7ebf\u6027Transformer\u6a21\u578b\uff08\u5982DeltaNet\uff09\u5219\u5728\u9762\u5bf9\u957f\u5e8f\u5217\u65f6\u8868\u73b0\u51fa\u8f83\u5927\u7684\u56f0\u60d1\u5ea6\u3002\u901a\u8fc7\u7ed3\u5408\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u548c\u95e8\u63a7DeltaNet\u5c42\uff0c\u6df7\u5408\u6a21\u578b\u5728\u957f\u5e8f\u5217\u63a8\u7406\u4efb\u52a1\u4e0a\u8868\u73b0\u5f97\u66f4\u4e3a\u7a33\u5b9a\u3002<\/p>\n\n\n\n

    3.\u4fe1\u606f\u68c0\u7d22\u4e0e\u8bb0\u5fc6\u7ba1\u7406<\/strong>\uff1a
    \u5728S-NIAH\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cGated DeltaNet\u5728\u5355\u9488\u68c0\u7d22\uff08S-NIAH-1\uff09\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u4f18\u4e8eMamba2\u548cDeltaNet\uff0c\u5c3d\u7ba1\u5728\u67d0\u4e9b\u66f4\u590d\u6742\u7684\u4efb\u52a1\uff08\u5982S-NIAH-2\u548cS-NIAH-3\uff09\u4e2d\uff0cGated DeltaNet\u7684\u68c0\u7d22\u51c6\u786e\u7387\u7565\u6709\u4e0b\u964d\uff0c\u4f46\u4f9d\u7136\u4f18\u4e8e\u5176\u4ed6\u6a21\u578b\u3002Gated DeltaNet\u901a\u8fc7\u5176\u95e8\u63a7Delta\u89c4\u5219\uff0c\u6709\u6548\u5730\u7ba1\u7406\u4e86\u8bb0\u5fc6\u7684\u6e05\u9664\u4e0e\u66f4\u65b0\uff0c\u907f\u514d\u4e86\u56e0\u8bb0\u5fc6\u78b0\u649e\u5bfc\u81f4\u7684\u68c0\u7d22\u5931\u8d25\u3002<\/p>\n\n\n\n

    4.\u8bad\u7ec3\u6548\u7387\u4e0e\u541e\u5410\u91cf<\/strong>\uff1a
    Gated DeltaNet\u7684\u8bad\u7ec3\u541e\u5410\u91cf\u4e0eDeltaNet\u51e0\u4e4e\u76f8\u5f53\uff0c\u4e14\u8fdc\u8d85Mamba2\u3002\u8fd9\u5f97\u76ca\u4e8e\u95e8\u63a7Delta\u89c4\u5219\u548c\u786c\u4ef6\u52a0\u901f\u8bad\u7ec3\u7b97\u6cd5\u7684\u4f18\u5316\uff0c\u7279\u522b\u662f\u5728\u5229\u7528GPU\u7684\u5f20\u91cf\u6838\u5fc3\u8fdb\u884c\u9ad8\u6548\u8ba1\u7b97\u65f6\uff0c\u6a21\u578b\u7684\u8bad\u7ec3\u901f\u5ea6\u5f97\u5230\u4e86\u663e\u8457\u63d0\u5347\u3002\u901a\u8fc7\u8fd9\u79cd\u786c\u4ef6\u9ad8\u6548\u7684\u8bad\u7ec3\u65b9\u6cd5\uff0cGated DeltaNet\u80fd\u591f\u5728\u77ed\u65f6\u95f4\u5185\u5904\u7406\u5927\u91cf\u6570\u636e\uff0c\u4e14\u4fdd\u6301\u8f83\u9ad8\u7684\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n

    \u4e94\u3001\u7ed3\u8bba\u4e0e\u672a\u6765\u65b9\u5411<\/strong><\/p>\n\n\n\n

    Gated DeltaNet\u901a\u8fc7\u7ed3\u5408\u95e8\u63a7\u673a\u5236\u4e0eDelta\u89c4\u5219\uff0c\u663e\u8457\u6539\u5584\u4e86\u7ebf\u6027Transformer\u5728\u957f\u5e8f\u5217\u5efa\u6a21\u548c\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\u4e2d\u7684\u80fd\u529b\u3002\u5b83\u4e0d\u4ec5\u5728\u591a\u4e2a\u57fa\u51c6\u4efb\u52a1\u4e2d\u5c55\u73b0\u4e86\u4f18\u5f02\u7684\u6027\u80fd\uff0c\u8fd8\u901a\u8fc7\u786c\u4ef6\u9ad8\u6548\u7684\u5e76\u884c\u8bad\u7ec3\u7b97\u6cd5\uff0c\u5927\u5e45\u63d0\u5347\u4e86\u6a21\u578b\u8bad\u7ec3\u6548\u7387\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63a2\u7d22Gated DeltaNet\u5728\u8d85\u957f\u5e8f\u5217\u4e0a\u7684\u8868\u73b0\uff0c\u5e76\u5c1d\u8bd5\u5c06\u66f4\u591a\u5148\u8fdb\u7684\u6df7\u5408\u67b6\u6784\u7ed3\u5408\u5230Gated DeltaNet\u4e2d\uff0c\u4ee5\u89e3\u51b3\u66f4\u590d\u6742\u7684\u4efb\u52a1\u3002<\/p>\n\n\n\n

    \n\n\n\n

    Gated DeltaNet on GitHub: https:\/\/github.com\/NVlabs\/GatedDeltaNet<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

    \u8bba\u6587Gated Delta Networks: Improving Mamba2 with Delta Rul […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,84,95,86],"class_list":["post-5590","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-nvidia","tag-transformer","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Gated Delta Networks: Improving Mamba2 with Delta Rul…","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5590","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5590"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5590\/revisions"}],"predecessor-version":[{"id":5595,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5590\/revisions\/5595"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5590"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5590"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5590"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}