{"id":2824,"date":"2026-03-19T09:00:15","date_gmt":"2026-03-19T01:00:15","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/03\/19\/kimi%e6%96%b0%e6%9e%b6%e6%9e%84%e8%ae%a9%e9%a9%ac%e6%96%af%e5%85%8b%e5%8f%b9%e6%9c%8d%ef%bc%8117%e5%b2%81%e9%ab%98%e4%b8%ad%e7%94%9f%e4%bd%9c%e8%80%85%e4%b8%80%e6%88%98%e6%88%90%e5%90%8d\/"},"modified":"2026-03-19T09:00:15","modified_gmt":"2026-03-19T01:00:15","slug":"kimi%e6%96%b0%e6%9e%b6%e6%9e%84%e8%ae%a9%e9%a9%ac%e6%96%af%e5%85%8b%e5%8f%b9%e6%9c%8d%ef%bc%8117%e5%b2%81%e9%ab%98%e4%b8%ad%e7%94%9f%e4%bd%9c%e8%80%85%e4%b8%80%e6%88%98%e6%88%90%e5%90%8d","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/03\/19\/kimi%e6%96%b0%e6%9e%b6%e6%9e%84%e8%ae%a9%e9%a9%ac%e6%96%af%e5%85%8b%e5%8f%b9%e6%9c%8d%ef%bc%8117%e5%b2%81%e9%ab%98%e4%b8%ad%e7%94%9f%e4%bd%9c%e8%80%85%e4%b8%80%e6%88%98%e6%88%90%e5%90%8d\/","title":{"rendered":"Kimi\u65b0\u67b6\u6784\u8ba9\u9a6c\u65af\u514b\u53f9\u670d\uff0117\u5c81\u9ad8\u4e2d\u751f\u4f5c\u8005\u4e00\u6218\u6210\u540d"},"content":{"rendered":"
\"Kimi\u65b0\u67b6\u6784\u8ba9\u9a6c\u65af\u514b\u53f9\u670d\uff0117\u5c81\u9ad8\u4e2d\u751f\u4f5c\u8005\u4e00\u6218\u6210\u540d\"\/<\/div><\/figure>\n
\"Kimi\u65b0\u67b6\u6784\u8ba9\u9a6c\u65af\u514b\u53f9\u670d\uff0117\u5c81\u9ad8\u4e2d\u751f\u4f5c\u8005\u4e00\u6218\u6210\u540d\"\/<\/div><\/figure>\n
\"Kimi\u65b0\u67b6\u6784\u8ba9\u9a6c\u65af\u514b\u53f9\u670d\uff0117\u5c81\u9ad8\u4e2d\u751f\u4f5c\u8005\u4e00\u6218\u6210\u540d\"\/<\/div><\/figure>\n
\"Kimi\u65b0\u67b6\u6784\u8ba9\u9a6c\u65af\u514b\u53f9\u670d\uff0117\u5c81\u9ad8\u4e2d\u751f\u4f5c\u8005\u4e00\u6218\u6210\u540d\"\/<\/div><\/figure>\n

\u628a\u6ce8\u610f\u529b\u673a\u5236\u201c\u65cb\u8f6c90\u5ea6\u201d<\/p>\n

17\u5c81\u9ad8\u4e2d\u751f\uff0c\u4ee5\u4e00\u4f5c\u8eab\u4efd\uff0c\u5728Kimi\u56e2\u961f\u628aIlya\u63d0\u51fa\u7684\u8bbe\u60f3\uff0c\u53d8\u6210\u4e86\u73b0\u5b9e\u3002<\/p>\n

Ilya\u4e4b\u524d\u6709\u4e2a\u9884\u8a00\uff0c\u628a\u6309\u65f6\u95f4\u5148\u540e\u987a\u5e8f\u5904\u7406\u6570\u636e\u7684LSTM\u7f51\u7edc\u201c\u65cb\u8f6c90\u5ea6\u201d\uff0c\u4e5f\u5c31\u662f\u628a\u65f6\u95f4\u8f74\u6362\u6210\u6a21\u578b\u6df1\u5ea6\u8f74\uff0c\u5c31\u53d8\u6210\u4e86\u73b0\u5728\u7684\u6b8b\u5dee\u7f51\u7edc\u3002<\/p>\n

Kimi\u56e2\u961f\u8ba4\u4e3a\uff0c\u65e2\u7136\u65f6\u95f4\u4e0a\u7684LSTM\u80fd\u5bf9\u5e94\u6df1\u5ea6\u4e0a\u7684\u6b8b\u5dee\uff0c\u90a3\u540e\u6765\u6dd8\u6c70\u4e86LSTM\u7684\u201c\u6ce8\u610f\u529b\u673a\u5236\u201d\u81ea\u7136\u4e5f\u53ef\u4ee5\u7167\u505a\u3002<\/p>\n

\u4ed6\u4eec\u65b0\u641e\u51fa\u7684Attention Residuals\u6280\u672f\uff0c\u5c31\u76f8\u5f53\u4e8e\u628a\u6ce8\u610f\u529b\u673a\u5236\u4e5f\u201c\u65cb\u8f6c\u4e8690\u5ea6\u201d\u3002<\/p>\n

\u7528\u4e86\u8fd9\u5957\u65b0\u65b9\u6cd5\u540e\uff0c\u6a21\u578b\u5728\u8ba1\u7b97\u5f53\u524d\u5c42\u65f6\u53ef\u4ee5\u806a\u660e\u5730\u201c\u56de\u5934\u770b\u201d\uff0c\u6839\u636e\u9700\u8981\u81ea\u7531\u51b3\u5b9a\u53bb\u63d0\u53d6\u524d\u9762\u54ea\u4e00\u5c42\u7684\u4fe1\u606f\u3002<\/p>\n

\u8fd9\u7bc7\u8bba\u6587\u8ba9\u9a6c\u65af\u514b\u4e5f\u6765\u56f4\u89c2\uff0c\u8868\u793a\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u3002<\/p>\n

\u9664\u4e86\u9a6c\u65af\u514b\uff0c\u8fd9\u7bc7\u8bba\u6587\u4e5f\u5f15\u53d1\u4e86\u5927\u795eKarpathy\u7684\u601d\u8003\uff0c\u76f4\u8a00\u6211\u4eec\u5bf9Attention is All You Need\u8fd9\u7bc7Transformer\u5f00\u5c71\u4e4b\u4f5c\u7684\u7406\u89e3\u8fd8\u662f\u4e0d\u591f\u3002<\/p>\n

\u8fd9\u79cd\u65b0\u673a\u5236\u653e\u5230Kimi\u81ea\u5bb6\u7684Kimi Linear 48B\u5927\u6a21\u578b\uff083B\u6fc0\u6d3b\u53c2\u6570\uff09\u4e0a\u9a8c\u8bc1\uff0c\u8bad\u7ec3\u6548\u7387\u63d0\u534725%\uff0c\u63a8\u7406\u5ef6\u8fdf\u589e\u52a0\u4e0d\u52302%\u3002<\/p>\n

\u5148\u56de\u987e\u4e00\u4e0b\u6b8b\u5dee\u8fde\u63a5\u7684\u5de5\u4f5c\u539f\u7406\u3002<\/p>\n

\u4f20\u7edf\u505a\u6cd5\u662f\uff1a\u7b2cN\u5c42\u7684\u8f93\u51fa = \u7b2cN\u5c42\u7684\u8ba1\u7b97\u7ed3\u679c + \u7b2cN-1\u5c42\u7684\u8f93\u51fa\u3002\u8fd9\u6837\u4e00\u8def\u7d2f\u52a0\u4e0b\u53bb\uff0c\u6bcf\u4e00\u5c42\u90fd\u80fd\u201c\u8bb0\u4f4f\u201d\u524d\u9762\u6240\u6709\u5c42\u7684\u4fe1\u606f\u3002<\/p>\n

\u95ee\u9898\u6765\u4e86\uff0c\u5728\u5927\u6a21\u578bPreNorm\u4e3b\u6d41\u8303\u5f0f\u4e0b\uff0c\u6b8b\u5dee\u8fde\u63a5\u4e2d\u6240\u6709\u5c42\u7684\u8d21\u732e\u90fd\u662f\u7b49\u6743\u7d2f\u52a0\u3002<\/p>\n

\u5c31\u50cf\u4e00\u4e2a\u201c\u8bb0\u5fc6\u529b\u592a\u597d\u7684\u4eba\u201d\uff0c\u628a\u6240\u6709\u7ecf\u5386\u90fd\u4ee5\u76f8\u540c\u6743\u91cd\u5b58\u8fdb\u5927\u8111\u3002\u8d21\u732e\u88ab\u9010\u6b65\u7a00\u91ca\uff0c\u65e9\u671f\u4fe1\u606f\u96be\u4ee5\u68c0\u7d22\uff0c\u4e14\u5927\u91cf\u5c42\u53ef\u88ab\u526a\u679d\u800c\u635f\u5931\u5fae\u5c0f\uff0c\u79f0\u4e4b\u4e3a\u201cPreNorm dilution problem\u201d\u3002<\/p>\n

\u66f4\u9ebb\u70e6\u7684\u662f\uff0c\u9690\u85cf\u72b6\u6001\u7684\u8303\u6570\u4f1a\u968f\u7740\u6df1\u5ea6\u4e0d\u65ad\u589e\u957f\u3002\u7814\u7a76\u4eba\u5458\u53d1\u73b0\uff0c\u5728\u6df1\u5c42\u7f51\u7edc\u4e2d\uff0c\u8fd9\u79cdunbounded growth\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002<\/p>\n

\u6708\u4e4b\u6697\u9762\u56e2\u961f\u6362\u4e86\u4e2a\u601d\u8def\uff1a\u65e2\u7136\u95ee\u9898\u51fa\u5728\u201c\u65e0\u5dee\u522b\u7d2f\u52a0\u201d\uff0c\u90a3\u5c31\u8ba9\u7f51\u7edc\u81ea\u5df1\u51b3\u5b9a\u8be5\u56de\u5fc6\u4ec0\u4e48\u3002<\/p>\n

\u56e2\u961f\u89c2\u5bdf\u5230\u4e00\u4e2a\u6709\u8da3\u7684\u5bf9\u5076\u6027\uff1a\u7f51\u7edc\u7684\u6df1\u5ea6\u7ef4\u5ea6\u548c\u5e8f\u5217\u7684\u65f6\u95f4\u7ef4\u5ea6\uff0c\u672c\u8d28\u4e0a\u662f\u540c\u6784\u7684\u3002<\/p>\n

\u5728Transformer\u5904\u7406\u5e8f\u5217\u65f6\uff0c\u7528\u6ce8\u610f\u529b\u673a\u5236\u8ba9\u5f53\u524d\u4f4d\u7f6e\u201c\u9009\u62e9\u6027\u5173\u6ce8\u201d\u4e4b\u524d\u7684\u4f4d\u7f6e\u3002\u90a3\u4e48\u5728\u6df1\u5ea6\u7ef4\u5ea6\u4e0a\uff0c\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u8ba9\u5f53\u524d\u5c42\u201c\u9009\u62e9\u6027\u5173\u6ce8\u201d\u4e4b\u524d\u7684\u5c42\uff1f<\/p>\n

Attention Residuals\u5c31\u8fd9\u4e48\u6765\u7684\uff1a<\/p>\n

\u8fd9\u6837\u4e00\u6765\uff0c\u7f51\u7edc\u53ef\u4ee5\u5b66\u4f1a\u54ea\u4e9b\u5c42\u7684\u4fe1\u606f\u5bf9\u5f53\u524d\u8ba1\u7b97\u6700\u91cd\u8981\uff0c\u5c31\u591a\u5173\u6ce8\u4e00\u70b9\uff1b\u4e0d\u76f8\u5173\u7684\u5c42\uff0c\u6743\u91cd\u81ea\u7136\u964d\u4f4e\u3002<\/p>\n

\u4f46\u8fd9\u5e26\u6765\u4e00\u4e2a\u65b0\u95ee\u9898\uff1a\u8ba1\u7b97\u91cf\u7206\u70b8\u3002<\/p>\n

\u5982\u679c\u4e00\u4e2a100\u5c42\u7684\u7f51\u7edc\uff0c\u6bcf\u4e00\u5c42\u90fd\u8981\u5bf9\u524d\u976299\u5c42\u505afull attention residual\uff0c\u590d\u6742\u5ea6\u662fO(L\u00b2)\uff0c\u6839\u672c\u8dd1\u4e0d\u52a8\u3002<\/p>\n

\u8bba\u6587\u4e2d\u7684\u89e3\u51b3\u65b9\u6848\u662fBlock AttnRes\u3002<\/p>\n

\u6838\u5fc3\u601d\u60f3\u662f\u628a\u8fde\u7eed\u7684\u82e5\u5e72\u5c42\u6253\u5305\u6210\u4e00\u4e2ablock\uff0c\u5bf9block\u5185\u90e8\u7684\u8f93\u51fa\u505a\u538b\u7f29\uff0c\u53ea\u4fdd\u7559\u4e00\u4e2a\u201c\u6458\u8981\u5411\u91cf\u201d\u3002<\/p>\n

\u5177\u4f53\u64cd\u4f5c\u5982\u4e0b\uff1a<\/p>\n

\u8fd9\u6837\u4e00\u6765\uff0cattention\u7684\u590d\u6742\u5ea6\u4eceO(L\u00b2)\u964d\u5230\u4e86O(L\u00b7B)\uff0c\u5728\u5b9e\u8df5\u4e2dB\u53ef\u4ee5\u8bbe\u5f97\u5f88\u5c0f\uff08\u8bba\u6587\u7528\u7684\u662f8-16\uff09\u3002<\/p>\n

\u6b64\u5916\uff0c\u56e2\u961f\u8fd8\u505a\u4e86\u6570\u4e2a\u5de5\u7a0b\u4f18\u5316\uff1a\u7f13\u5b58\u5f0f\u6d41\u6c34\u7ebf\u901a\u4fe1\u3001\u5e8f\u5217\u5206\u7247\u9884\u586b\u5145\u3001KV \u7f13\u5b58\u7c92\u5ea6\u4f18\u5316\u7b49\u7b49\u3002<\/p>\n

\u7406\u8bba\u8bf4\u5f97\u901a\uff0c\u4f46\u771f\u6b63\u8ba9\u4eba\u4fe1\u670d\u7684\u662f\u5927\u89c4\u6a21\u9a8c\u8bc1\u3002<\/p>\n

\u56e2\u961f\u5728\u81ea\u5bb6\u7684Kimi Linear\u67b6\u6784\u4e0a\u505a\u4e86\u6d4b\u8bd5\u3002\u8fd9\u662f\u4e00\u4e2a\u91c7\u7528\u7ebf\u6027\u6ce8\u610f\u529b\u7684\u5927\u6a21\u578b\uff0c\u603b\u53c2\u657048B\uff0c\u6fc0\u6d3b\u53c2\u65703B\uff08MoE\u67b6\u6784\uff09\u3002<\/p>\n

\u540c\u7b49\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0cAttention Residuals\u80fd\u83b7\u5f97\u66f4\u597d\u7684\u4e0b\u6e38\u6027\u80fd\uff1b\u53cd\u8fc7\u6765\u8bf4\uff0c\u8fbe\u5230\u76f8\u540c\u6027\u80fd\u9700\u8981\u7684\u8bad\u7ec3\u8ba1\u7b97\u91cf\u51cf\u5c11\u4e86\u7ea620%\uff0c\u76f8\u5f53\u4e8e\u83b7\u5f97\u4e861.25\u500d\u7684\u6548\u7387\u4f18\u52bf\u3002<\/p>\n

\u5728\u5177\u4f53\u4efb\u52a1\u4e0a\uff0c\u6570\u5b66\u63a8\u7406\uff08MATH\u3001GSM8K\uff09\u3001\u4ee3\u7801\u751f\u6210\uff08HumanEval\u3001MBPP\uff09\u5747\u6301\u5e73\u6216\u7565\u4f18\uff0c\u591a\u8bed\u8a00\u7406\u89e3\u7684\u4e00\u81f4\u6027\u4e5f\u6709\u6240\u6539\u5584\u3002<\/p>\n

\u66f4\u91cd\u8981\u7684\u662f\uff0cAttention Residuals\u662f\u4e00\u4e2adrop-in replacement\uff0c\u4e0d\u9700\u8981\u4fee\u6539\u7f51\u7edc\u5176\u4ed6\u90e8\u5206\uff0c\u76f4\u63a5\u66ff\u6362\u6b8b\u5dee\u8fde\u63a5\u5373\u53ef\u3002<\/p>\n

\u8bba\u6587\u91cc\u8fd8\u8bb2\u5230\u4e00\u4e2a\u6709\u610f\u601d\u7684\u89c6\u89d2\u3002<\/p>\n

\u56e2\u961f\u628a\u8fd9\u9879\u5de5\u4f5c\u79f0\u4e3a\u201c\u65f6\u95f4-\u6df1\u5ea6\u5bf9\u5076\u6027\u201d\uff08time-depth duality\uff09\u7684\u5e94\u7528\u3002<\/p>\n

\u5728\u4ed6\u4eec\u770b\u6765\uff0c\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u7684\u201c\u5c42\u201d\u548c\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u201c\u65f6\u95f4\u6b65\u201d\uff0c\u672c\u8d28\u4e0a\u662f\u90fd\u662f\u5bf9\u4fe1\u606f\u7684\u8fed\u4ee3\u5904\u7406\u3002<\/p>\n

Transformer\u4e4b\u6240\u4ee5\u6210\u529f\uff0c\u662f\u56e0\u4e3a\u7528attention\u66ff\u4ee3\u4e86RNN\u4e2d\u56fa\u5b9a\u7684recurrence\u3002<\/p>\n

\u90a3\u4e48\u5728\u6df1\u5ea6\u7ef4\u5ea6\u4e0a\uff0c\u662f\u4e0d\u662f\u4e5f\u8be5\u7528attention\u66ff\u4ee3\u56fa\u5b9a\u7684residual\uff1f<\/p>\n

\u66f4\u6709\u610f\u601d\u7684\u662f\uff0c\u8fd9\u7bc7\u8ba9\u9a6c\u65af\u514b\u3001Karpathy\u7b49\u4eba\u90fd\u4e3a\u4e4b\u4e00\u9707\u7684\u8bba\u6587\uff0c\u5171\u540c\u4e00\u4f5c\u4e4b\u4e00\u662f\u4e00\u540d\u5e74\u4ec517\u5c81\u7684\u9ad8\u4e2d\u751f\u2014\u2014\u9648\u5e7f\u5b87\uff08Nathan\uff09\u3002<\/p>\n

\u53e6\u5916\u4e24\u540d\u5171\u540c\u4e00\u4f5c\uff0c\u5206\u522b\u662fKimi\u7684\u5173\u952e\u4eba\u7269\u4e4b\u4e00\u3001RoPE\uff08\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff09\u7684\u63d0\u51fa\u8005\u82cf\u795e\uff08\u82cf\u5251\u6797\uff09\uff0c\u4ee5\u53caKimi Linear\u7684\u7b2c\u4e00\u4f5c\u8005\u5f20\u5b87\u3002<\/p>\n

\u8bda\u7136Attention Residuals\u662f\u56e2\u961f\u534f\u4f5c\u53d6\u5f97\u7684\u6210\u679c\uff0c\u4f46\u4e00\u540d\u9ad8\u4e2d\u751f\u51fa\u73b0\u5728\u8fd9\u6837\u7684\u56e2\u961f\u4e4b\u4e2d\uff0c\u8fd8\u4e0e\u4e24\u4f4d\u5927\u795e\u5171\u5217\u4e00\u4f5c\uff0c\u5df2\u7ecf\u8db3\u591f\u9707\u64bc\u3002<\/p>\n

a16z\u521b\u59cb\u4ebaMarc Andreessen\u3001Thinking Machines\u7684\u8054\u521b\u7b49\u4eba\u90fd\u5173\u6ce8\u4e86\u4ed6\u7684X\u8d26\u53f7\u3002<\/p>\n

\u4e00\u5e74\u524d\u624d\u521a\u521a\u5f00\u59cb\u4e86\u89e3\u5927\u6a21\u578b\u7684\u9648\u5e7f\u5b87\uff0c\u662f\u4ece\u5317\u4eac\u7684\u4e00\u573a\u9ed1\u5ba2\u677e\u5f00\u59cb\uff0c\u4e00\u8def\u8d70\u5411\u7845\u8c37\u7684\u3002<\/p>\n

\u540e\u6765\u56de\u56fd\u65f6\uff0c\u4ed6\u9009\u62e9\u52a0\u5165\u4e86Kimi\u3002<\/p>\n

\u7ecf\u624b\u8fc7\u6708\u4e4b\u6697\u9762\u6295\u8d44\u7684\u5947\u7ee9\u521b\u575b\uff08\u539fYC\u4e2d\u56fd\uff09\u521b\u59cb\u6210\u5458\u8463\u79d1\u542b\uff0c\u4e5f\u66fe\u5728\u5176\u4e2a\u4eba\u516c\u4f17\u53f7\u4e0a\u520a\u8f7d\u8fc7\u9648\u5e7f\u5b87\u7684\u4e00\u4efd\u81ea\u4f20\u3002<\/p>\n

\u53bb\u5e74\u4e8c\u6708\uff0c\u5317\u4eac\u7684\u4e00\u573a\u4e2d\u5b66\u751f\u9ed1\u5ba2\u677e\u4e0a\uff0c\u9648\u5e7f\u5b87\u5c55\u793a\u4e86\u4e00\u4e2a\u5173\u4e8e\u201c\u4eba\u7c7b\u7b2c\u4e09\u53ea\u673a\u68b0\u8f85\u52a9\u624b\u201d\u7684\u521b\u65b0\u6784\u60f3\u2014\u2014ThirdArm\u3002<\/p>\n

\u4e5f\u6b63\u662f\u8fd9\u4e2a\u9879\u76ee\uff0c\u8ba9\u4ed6\u7ed3\u8bc6\u4e86\u9ed1\u5ba2\u677e\u8bc4\u59d4\u8463\u79d1\u542b\uff0c\u540e\u8005\u4e5f\u6210\u4e3a\u4e86\u4ed6\u7684\u521b\u4e1a\u5bfc\u5e08\u3002<\/p>\n

\u5f53\u65f6\uff0c\u8463\u79d1\u542b\u8ffd\u95ee\u4ed6\uff0c\u672a\u6765\u662f\u5426\u4f1a\u6df1\u8015\u8fd9\u9879\u6280\u672f\uff0c\u8fd9\u4fc3\u4f7f\u4ed6\u5f00\u59cb\u91cd\u65b0\u5ba1\u89c6\u81ea\u5df1\u7684\u804c\u4e1a\u65b9\u5411\u3002<\/p>\n

\u968f\u540e\u4ed6\u5165\u9009\u4e86\u8463\u79d1\u542b\u53d1\u8d77\u7684\u53ea\u6709\u6781\u5c11\u6570\u4eba\u5165\u9009\u7684\u9752\u5e74\u8ba1\u5212\uff0c\u5f00\u59cb\u63a5\u89e6IOI\uff08\u56fd\u9645\u4fe1\u606f\u5b66\u5965\u6797\u5339\u514b\uff09\u91d1\u724c\u5f97\u4e3b\u53ca\u8d44\u6df1\u79d1\u7814\u4eba\u5458\u3002<\/p>\n

\u6b64\u524d\u4ed6\u66fe\u5c1d\u8bd5\u7ecf\u8425Shopify\u8de8\u5883\u7535\u5546\u5e97\u94fa\u3001\u8fd0\u8425\u77ed\u89c6\u9891\u8d26\u53f7\uff0c\u4f46\u7ecf\u8fc7\u8463\u79d1\u542b\u7684\u5efa\u8bae\uff0c\u4ed6\u51b3\u5b9a\u8f6c\u5411\u7406\u89e3\u65f6\u4ee3\u7684\u5e95\u5c42\u6280\u672f\u3002<\/p>\n

\u5f53\u65f6\u8fd8\u4e0d\u77e5\u9053Transformer\u662f\u4ec0\u4e48\u7684\u4ed6\uff0c\u5728DeepSeek\u7814\u7a76\u5458\u8881\u5883\u9633\u7684\u6307\u5bfc\u4e0b\uff0c\u5229\u7528Gemini\u4f5c\u4e3a\u8f85\u52a9\u5de5\u5177\uff0c\u901a\u8fc7\u7814\u8bfb\u7ecf\u5178\u8bba\u6587\u3001\u8ffd\u8e2aGitHub\u5f00\u6e90\u9879\u76ee\u7b49\u65b9\u5f0f\u9010\u6b65\u5efa\u7acb\u8ba4\u77e5\u3002<\/p>\n

\u6709\u4e00\u6b21\u4ed6\u5728\u63a8\u7279\u4e0a\u5206\u4eab\u4e86\u5bf9\u4e00\u7bc7\u535a\u5ba2\u7684\u53cd\u601d\u540e\uff0c\u83b7\u5f97\u4e86\u4f5c\u8005\u7684\u56de\u590d\uff0c\u8fd9\u7bc7\u5e16\u5b50\u4e5f\u56e0\u6b64\u5f15\u8d77\u4e86\u4e00\u5bb6\u7845\u8c37AI\u521d\u521b\u516c\u53f8CEO\u7684\u5173\u6ce8\u3002<\/p>\n

\u8be5\u516c\u53f8\u4e8e2024\u5e74\u5e95\u6210\u7acb\uff0c2025\u5e74\u521d\u5b8c\u6210\u4e86800\u4e07\u7f8e\u5143\u79cd\u5b50\u8f6e\u878d\u8d44\uff0c\u8d44\u65b9\u80cc\u666f\u6d89\u53caOpenAI\u4e0eAnthropic\u3002<\/p>\n

\u5728\u901a\u8fc7\u4e00\u9879\u9650\u65f6\u901a\u5bb5\u5b8c\u6210\u7684\u5b9e\u9a8c\u6d4b\u8bd5\u540e\uff0c\u4ed6\u62ff\u5230\u4e86\u5bf9\u65b9\u7684\u5f55\u7528\u901a\u77e5\u3002<\/p>\n

\u6691\u5047\u671f\u95f4\uff0c\u4ed6\u524d\u5f80\u65e7\u91d1\u5c71\u5f00\u542f\u4e86\u4e3a\u671f\u4e03\u5468\u7684\u5b9e\u4e60\u3002\u5176\u4e2d\u524d\u4e24\u5468\uff0c\u4ed6\u8d1f\u8d23\u5b9a\u4e49\u5e76\u63a8\u8fdb\u4e00\u4e2a\u6d89\u53ca144\u5f20H100\u663e\u5361\u7684\u63a2\u7d22\u6027\u9879\u76ee\u3002<\/p>\n

\u5728CEO\u76f4\u63a5\u6307\u5bfc\u4e0b\uff0c\u4ed6\u7684\u5de5\u4f5c\u5ef6\u4f38\u81f3\u8fd0\u8425\u5c42\u9762\uff0c\u53c2\u4e0e\u4e86\u62db\u8058\u7cfb\u7edf\u642d\u5efa\u3001\u6280\u672f\u5185\u5bb9\u8f93\u51fa\u53ca\u878d\u8d44\u7b56\u7565\u8ba8\u8bba\uff0c\u5e76\u83b7\u5f97\u4e0e\u65e9\u671f\u6295\u8d44\u8005Vinod Khosla\u4ea4\u6d41\u7684\u673a\u4f1a\u3002<\/p>\n

\u5728\u7845\u8c37\u671f\u95f4\uff0c\u4ed6\u7ef4\u6301\u7740\u9ad8\u5f3a\u5ea6\u5de5\u4f5c\u8282\u594f\uff0c\u901a\u8fc7\u5496\u5561\u793e\u4ea4\u4e0e\u82f1\u4f1f\u8fbe\u5de5\u7a0b\u5e08\u53ca\u521d\u521b\u521b\u59cb\u4eba\u5efa\u7acb\u8054\u7cfb\u3002\u8fd9\u6b21\u7ecf\u5386\u8ba9\u4ed6\u5c06\u79d1\u7814\u89c6\u4e3a\u4e00\u79cd\u652f\u6491\u521b\u9020\u7684\u5e95\u5c42\u80fd\u529b\u3002<\/p>\n

\u5b9e\u4e60\u7ed3\u675f\u540e\uff0c\u9648\u5e7f\u5b87\u56de\u5230\u56fd\u5185\uff0c\u5e76\u4e8e\u53bb\u5e7411\u6708\u52a0\u5165\u6708\u4e4b\u6697\u9762\u3002<\/p>\n

\u628a\u4ed6\u5438\u5f15\u8fdb\u53bb\u7684\uff0c\u6b63\u662fKimi\u4e00\u76f4\u505a\u7684Flash Linear Attention\u8fd9\u4e00\u7c7b\u9ad8\u6548attention\u5de5\u4f5c\u3002<\/p>\n

\u5b9e\u9645\u4e0a\uff0c\u6b63\u662fGitHub\u4e0a\u7684FLA\u9879\u76ee\uff0c\u5438\u5f15\u4e86\u4ed6\u5bf9\u673a\u5668\u5b66\u4e60\u7684\u5174\u8da3\u5e76\u88ab\u9080\u8bf7\u52a0\u5165Kimi\u56e2\u961f\u3002<\/p>\n

\u4e5f\u6b63\u662f\u987a\u7740\u8fd9\u6761\u7ebf\uff0c\u4ed6\u5f00\u59cb\u4e00\u8def\u5f80\u66f4\u5e95\u5c42\u94bb\uff0c\u4ece\u8bfb\u8bba\u6587\u3001\u770b\u5b9e\u73b0\uff0c\u5230\u7814\u7a76 Triton kernel\u3001\u7406\u89e3attention\u4e3a\u4ec0\u4e48\u80fd\u88ab\u8fd9\u6837\u91cd\u5199\u3001\u8fd9\u6837\u52a0\u901f\u3002<\/p>\n

\u5230\u4e86\u6708\u4e4b\u6697\u9762\uff0c\u8fd9\u6761\u8def\u4e5f\u7b97\u662f\u7ed5\u4e86\u4e00\u5708\u53c8\u843d\u56de\u539f\u70b9\u2014\u2014<\/p>\n

\u4ed6\u6700\u521d\u662f\u88ab\u5e95\u5c42\u6280\u672f\u5438\u5f15\uff0c\u6700\u540e\u505a\u7684\u4e5f\u6b63\u662f\u6700\u5e95\u5c42\u3001\u6700\u6838\u5fc3\u7684\u90a3\u90e8\u5206\u4e8b\u3002<\/p>\n

\u76f8\u6bd4\u4e8e\u8bb2\u4e00\u4e2a\u201c\u5c11\u5e74\u5929\u624d\u4e00\u8def\u5f00\u6302\u201d\u7684\u6545\u4e8b\uff0c\u9648\u5e7f\u5b87\u7684\u7ecf\u5386\u66f4\u50cf\u662f\u53e6\u4e00\u79cd\u6210\u957f\u8def\u5f84\u2014\u2014<\/p>\n

\u5148\u88ab\u65f6\u4ee3\u6700\u524d\u6cbf\u7684\u6280\u672f\u51fb\u4e2d\uff0c\u518d\u4e00\u6b65\u6b65\u628a\u5174\u8da3\u78e8\u6210\u80fd\u529b\uff0c\u628a\u80fd\u529b\u5e26\u5230\u771f\u6b63\u7684\u5927\u6a21\u578b\u7814\u53d1\u73b0\u573a\u91cc\u3002<\/p>\n

\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/github.com\/MoonshotAI\/Attention-Residuals\/\u53c2\u8003\u94fe\u63a5\uff1a[1]https:\/\/mp.weixin.qq.com\/s\/gRR99pEDWb5qsk2a2hwe2w[2]https:\/\/nathanchen.me\/public\/About%20me.html<\/p>\n

\n

\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n

\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n

\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n

\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

\u628a\u6ce8\u610f\u529b\u673a\u5236\u201c\u65cb\u8f6c90\u5ea6\u201d<\/p>\n

17\u5c81\u9ad8\u4e2d\u751f\uff0c\u4ee5\u4e00\u4f5c\u8eab\u4efd\uff0c\u5728Kimi\u56e2\u961f\u628aIlya\u63d0\u51fa\u7684\u8bbe\u60f3\uff0c\u53d8\u6210\u4e86\u73b0\u5b9e\u3002<\/p>\n

Ilya\u4e4b\u524d\u6709\u4e2a\u9884\u8a00\uff0c\u628a\u6309\u65f6\u95f4\u5148\u540e\u987a\u5e8f\u5904\u7406\u6570\u636e\u7684LSTM\u7f51\u7edc\u201c\u65cb\u8f6c90\u5ea6\u201d\uff0c\u4e5f\u5c31\u662f\u628a\u65f6\u95f4\u8f74\u6362\u6210\u6a21\u578b\u6df1\u5ea6\u8f74\uff0c\u5c31\u53d8\u6210\u4e86\u73b0\u5728\u7684\u6b8b\u5dee\u7f51\u7edc\u3002<\/p>\n

Kimi\u56e2\u961f\u8ba4\u4e3a\uff0c\u65e2\u7136\u65f6\u95f4\u4e0a\u7684LSTM\u80fd\u5bf9\u5e94\u6df1\u5ea6\u4e0a\u7684\u6b8b\u5dee\uff0c\u90a3\u540e\u6765\u6dd8\u6c70\u4e86LSTM\u7684\u201c\u6ce8\u610f\u529b\u673a\u5236\u201d\u81ea\u7136\u4e5f\u53ef\u4ee5\u7167\u505a\u3002<\/p>\n

\u4ed6\u4eec\u65b0\u641e\u51fa\u7684Attention Residu<\/p>\n","protected":false},"author":1,"featured_media":2820,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[354],"class_list":["post-2824","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-kimi"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/2824","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=2824"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/2824\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/2820"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=2824"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=2824"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=2824"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}