{"id":169,"date":"2025-03-27T17:39:18","date_gmt":"2025-03-27T09:39:18","guid":{"rendered":"https:\/\/www.dianshitv.cn\/?p=169"},"modified":"2025-06-10T18:10:04","modified_gmt":"2025-06-10T10:10:04","slug":"deepseek%e5%85%a8%e6%99%af%e8%a7%a3%e6%9e%90%ef%bc%9a%e6%8a%80%e6%9c%af%e9%9d%a9%e6%96%b0%e4%b8%8e%e5%ba%94%e7%94%a8%e5%ae%9e%e8%b7%b5%ef%bc%88%e4%b8%89%ef%bc%89-%e4%b8%8e%e4%bc%a0","status":"publish","type":"post","link":"https:\/\/www.dianshitv.cn\/index.php\/2025\/03\/27\/deepseek%e5%85%a8%e6%99%af%e8%a7%a3%e6%9e%90%ef%bc%9a%e6%8a%80%e6%9c%af%e9%9d%a9%e6%96%b0%e4%b8%8e%e5%ba%94%e7%94%a8%e5%ae%9e%e8%b7%b5%ef%bc%88%e4%b8%89%ef%bc%89-%e4%b8%8e%e4%bc%a0\/","title":{"rendered":"DeepSeek\u5168\u666f\u89e3\u6790\uff1a\u6280\u672f\u9769\u65b0\u4e0e\u5e94\u7528\u5b9e\u8df5\uff08\u4e09\uff09\u2014\u2014\u4e0e\u4f20\u7edfAI\u6a21\u578b\u7684\u5e95\u5c42\u5dee\u5f02\u4e0e\u4f18\u52bf"},"content":{"rendered":"\n<p>\ud83d\udcd6 \u3010\u6458\u8981\u3011DeepSeek\u4f5c\u4e3a\u65b0\u4e00\u4ee3AI\u67b6\u6784\u4ee3\u8868\uff0c\u901a\u8fc7\u6df7\u5408\u4e13\u5bb6\u7cfb\u7edf\uff08MoE\uff09\u3001\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u7b49\u521b\u65b0\u6280\u672f\uff0c\u5728\u6548\u7387\u3001\u6210\u672c\u4e0e\u6027\u80fd\u4e0a\u5b9e\u73b0\u5bf9\u4f20\u7edf\u6a21\u578b\u7684\u5168\u9762\u8d85\u8d8a\u3002\u672c\u6587\u4ece\u67b6\u6784\u8bbe\u8ba1\u3001\u8bad\u7ec3\u4f18\u5316\u3001\u63a8\u7406\u52a0\u901f\u3001\u6cdb\u5316\u80fd\u529b\u7b49\u7ef4\u5ea6\uff0c\u7ed3\u5408\u6700\u65b0\u7814\u7a76\u6570\u636e\u4e0e\u4ea7\u4e1a\u6848\u4f8b\uff0c\u6df1\u5ea6\u89e3\u6790\u5176\u6280\u672f\u7a81\u7834\u8def\u5f84\u53ca\u884c\u4e1a\u5f71\u54cd\u3002<\/p>\n\n\n\n<p>\ud83d\ude80 \u5f15\u8a00\uff1aAI\u6a21\u578b\u7684\u6548\u7387\u56f0\u5c40\u4e0e\u7834\u5c40\u4e4b\u6218<br>\u4f20\u7edfAI\u6a21\u578b\uff08CNN\u3001RNN\u3001Transformer\uff09\u957f\u671f\u9762\u4e34\u4e09\u5927\u77db\u76fe\uff1a\u7b97\u529b\u9700\u6c42\u6307\u6570\u589e\u957f\u4e0e\u786c\u4ef6\u74f6\u9888\u7684\u77db\u76fe\u3001\u6a21\u578b\u89c4\u6a21\u6269\u5927\u4e0e\u63a8\u7406\u5b9e\u65f6\u6027\u7684\u77db\u76fe\u3001\u4e13\u7528\u4efb\u52a1\u4f18\u5316\u4e0e\u6cdb\u5316\u80fd\u529b\u7684\u77db\u76fe\u3002\u4ee5Transformer\u4e3a\u4f8b\uff0c\u5176\u81ea\u6ce8\u610f\u529b\u673a\u5236\u867d\u7a81\u7834\u5e8f\u5217\u5efa\u6a21\u9650\u5236\uff0c\u4f46\u5904\u7406128K tokens\u7684\u6587\u672c\u65f6\uff0c\u5185\u5b58\u5360\u7528\u9ad8\u8fbe320GB\uff0c\u63a8\u7406\u5ef6\u8fdf\u8d85\u8fc75\u79d2\uff0c\u96be\u4ee5\u6ee1\u8db3\u5de5\u4e1a\u573a\u666f\u9700\u6c42\u3002<\/p>\n\n\n\n<p>DeepSeek\u7684\u8bde\u751f\u6807\u5fd7\u7740AI\u67b6\u6784\u4ece\u201c\u66b4\u529b\u5806\u53c2\u6570\u201d\u8f6c\u5411\u201c\u667a\u80fd\u52a8\u6001\u8ba1\u7b97\u201d\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u901a\u8fc7\u7a00\u758f\u5316\u3001\u52a8\u6001\u5316\u3001\u4e13\u4e1a\u5316\u8bbe\u8ba1\uff0c\u5b9e\u73b0\u201c\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\uff0c\u767e\u4ebf\u6fc0\u6d3b\u6210\u672c\u201d\u3002\u4e0b\u6587\u5c06\u62c6\u89e3\u5176\u6280\u672f\u8def\u5f84\u4e0e\u4ea7\u4e1a\u4ef7\u503c\u3002<\/p>\n\n\n\n<p>\ud83d\udd27 \u4e00\u3001\u67b6\u6784\u9769\u547d\uff1a\u4ece\u201c\u5168\u8fde\u63a5\u201d\u5230\u201c\u52a8\u6001\u7a00\u758f\u201d<\/p>\n\n\n\n<p>1.1 \u4f20\u7edf\u6a21\u578b\u7684\u67b6\u6784\u74f6\u9888<br>\u6a21\u578b\u7c7b\u578b<\/p>\n\n\n\n<p>\u4f18\u52bf<\/p>\n\n\n\n<p>\u6838\u5fc3\u7f3a\u9677<\/p>\n\n\n\n<p>CNN<\/p>\n\n\n\n<p>\u5c40\u90e8\u7279\u5f81\u63d0\u53d6\u80fd\u529b\u5f3a<\/p>\n\n\n\n<p>\u957f\u8ddd\u79bb\u4f9d\u8d56\u5efa\u6a21\u80fd\u529b\u5f31<\/p>\n\n\n\n<p>RNN<\/p>\n\n\n\n<p>\u5e8f\u5217\u5efa\u6a21\u9002\u5e94\u6027\u597d<\/p>\n\n\n\n<p>\u68af\u5ea6\u6d88\u5931\u3001\u5e76\u884c\u6548\u7387\u4f4e<\/p>\n\n\n\n<p>Transformer<\/p>\n\n\n\n<p>\u5168\u5c40\u6ce8\u610f\u529b\u3001\u5e76\u884c\u8ba1\u7b97<\/p>\n\n\n\n<p>\u8ba1\u7b97\u590d\u6742\u5ea6O(n\u00b2)\u3001\u5185\u5b58\u9ed1\u6d1e<\/p>\n\n\n\n<p>\u4ee5Transformer\u4e3a\u4f8b\uff0c\u5176\u81ea\u6ce8\u610f\u529b\u77e9\u9635\u7684\u5b58\u50a8\u9700\u6c42\u968f\u5e8f\u5217\u957f\u5ea6\u5448\u5e73\u65b9\u589e\u957f\u3002\u5904\u74064096 tokens\u65f6\uff0c\u5355\u5c42\u6ce8\u610f\u529b\u9700\u5360\u7528\u7ea6500MB\u663e\u5b58\uff0c\u800c\u767e\u5c42\u6a21\u578b\u663e\u5b58\u9700\u6c42\u76f4\u63a5\u7a81\u783450GB\uff0c\u5bfc\u81f4\u5b9e\u9645\u5e94\u7528\u4e2d\u5e38\u9700\u727a\u7272\u4e0a\u4e0b\u6587\u957f\u5ea6\u6216\u6a21\u578b\u6df1\u5ea6\u3002<\/p>\n\n\n\n<p>1.2 DeepSeek\u7684\u67b6\u6784\u7a81\u56f4<br>\uff081\uff09\u6df7\u5408\u4e13\u5bb6\u7cfb\u7edf\uff08MoE\uff09<\/p>\n\n\n\n<p>\u52a8\u6001\u8def\u7531\u673a\u5236\uff1a\u6bcf\u4e2a\u8f93\u5165\u4ec5\u6fc0\u6d3b3-4\u4e2a\u4e13\u5bb6\uff08\u603b\u4e13\u5bb6\u6570\u53ef\u8fbe128+\uff09\uff0c\u53c2\u6570\u5229\u7528\u7387\u4ece100%\u964d\u81f310-30%\u3002<\/p>\n\n\n\n<p>\u6210\u672c\u5bf9\u6bd4\uff1a\u8bad\u7ec31\u4e07\u4ebf\u53c2\u6570\u6a21\u578b\uff0cDeepSeek\u6210\u672c\u4e3a557\u4e07\u7f8e\u5143\uff0c\u4ec5\u4e3a\u540c\u89c4\u6a21\u5bc6\u96c6\u6a21\u578b\uff087800\u4e07\u7f8e\u5143\uff09\u76841\/14\u3002<\/p>\n\n\n\n<p>\u6848\u4f8b\uff1a\u5728\u667a\u80fd\u5ba2\u670d\u573a\u666f\u4e2d\uff0cDeepSeek-MoE-128B\u6a21\u578b\u54cd\u5e94\u5ef6\u8fdf\u4ece580ms\u964d\u81f3230ms\uff0c\u540c\u65f6\u9519\u8bef\u7387\u964d\u4f4e18%\u3002<\/p>\n\n\n\n<p>\uff082\uff09\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09<\/p>\n\n\n\n<p>\u4f4e\u79e9\u538b\u7f29\u6280\u672f\uff1a\u5c06\u952e\u503c\u5bf9\uff08KV\uff09\u77e9\u9635\u4ece\u7ef4\u5ea6d\u538b\u7f29\u81f3k\uff08k&lt;&lt;d\uff09\uff0c128K tokens\u7684KV\u7f13\u5b58\u4ece320GB\u964d\u81f380GB\u3002<\/p>\n\n\n\n<p>\u7a00\u758f\u6ce8\u610f\u529b\u4f18\u5316\uff1a\u901a\u8fc7\u5c40\u90e8\u7a97\u53e3+\u5168\u5c40\u5173\u952e\u70b9\u805a\u7126\uff0c\u65e0\u6548\u8ba1\u7b97\u51cf\u5c1130%\u3002<\/p>\n\n\n\n<p>\u6027\u80fd\u6570\u636e\uff1a\u5728LAMBADA\u957f\u6587\u672c\u63a8\u7406\u4efb\u52a1\u4e2d\uff0cDeepSeek\u51c6\u786e\u7387\u8fbe87.2%\uff0c\u6bd4\u4f20\u7edfTransformer\u9ad89.3%\u3002<\/p>\n\n\n\n<p>\u2699\ufe0f \u4e8c\u3001\u8bad\u7ec3\u8303\u5f0f\u9769\u65b0\uff1a\u4ece\u201c\u66b4\u529b\u8bad\u7ec3\u201d\u5230\u201c\u7cbe\u51c6\u8fdb\u5316\u201d<br>2.1 FP8\u6df7\u5408\u7cbe\u5ea6\uff1a\u663e\u5b58\u4e0e\u901f\u5ea6\u7684\u5e73\u8861\u672f<br>\u6280\u672f\u539f\u7406\uff1a\u5c06\u6743\u91cd\u3001\u68af\u5ea6\u3001\u6fc0\u6d3b\u503c\u52a8\u6001\u91cf\u5316\u4e3a8\u4f4d\u6d6e\u70b9\uff08FP8\uff09\uff0c\u901a\u8fc7\u7f29\u653e\u56e0\u5b50\u8865\u507f\u7cbe\u5ea6\u635f\u5931\u3002<\/p>\n\n\n\n<p>\u5b9e\u6d4b\u6548\u679c\uff1a<\/p>\n\n\n\n<p>\u663e\u5b58\u5360\u7528\u4e0b\u964d50%\uff0c\u8bad\u7ec3\u541e\u5410\u91cf\u63d0\u5347210%<\/p>\n\n\n\n<p>\u5728GLUE\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u7cbe\u5ea6\u635f\u5931\u63a7\u5236\u57280.47%\u4ee5\u5185<\/p>\n\n\n\n<p>\u786c\u4ef6\u9002\u914d\uff1a\u4e0eNVIDIA H100 Tensor Core GPU\u534f\u540c\u4f18\u5316\uff0c\u7b97\u529b\u5229\u7528\u7387\u8fbe92%\u3002<\/p>\n\n\n\n<p>2.2 \u52a8\u6001\u8d1f\u8f7d\u5747\u8861\uff1aMoE\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u79d8\u5bc6<br>\u4f20\u7edfMoE\u6a21\u578b\u5e38\u51fa\u73b0\u201c\u4e13\u5bb6\u6781\u5316\u201d\u2014\u2014\u5c11\u6570\u4e13\u5bb6\u627f\u62c590%\u4ee5\u4e0a\u7684\u8ba1\u7b97\u8d1f\u8f7d\u3002DeepSeek\u901a\u8fc7\u4e24\u9879\u521b\u65b0\u89e3\u51b3\u8be5\u95ee\u9898\uff1a<\/p>\n\n\n\n<p>\u8def\u7531\u504f\u7f6e\u52a8\u6001\u8c03\u6574\uff1a\u5f15\u5165\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u51fd\u6570\uff0c\u5f3a\u5236\u5206\u6563\u8f93\u5165\u6d41\u91cf\u3002<\/p>\n\n\n\n<p>\u4e13\u5bb6\u5bb9\u91cf\u5f39\u6027\u5206\u914d\uff1a\u6839\u636e\u5386\u53f2\u8d1f\u8f7d\u52a8\u6001\u8c03\u6574\u5404\u4e13\u5bb6\u7684\u5904\u7406\u5bb9\u91cf\u3002<br>\u5b9e\u9a8c\u663e\u793a\uff0c\u5728\u8bad\u7ec3\u5343\u4ebf\u7ea7MoE\u6a21\u578b\u65f6\uff0c\u4e13\u5bb6\u8d1f\u8f7d\u6807\u51c6\u5dee\u4ece35.7\u964d\u81f34.2\uff0c\u6536\u655b\u901f\u5ea6\u63d0\u534740%\u3002<\/p>\n\n\n\n<p>2.3 \u7eaf\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8303\u5f0f\uff1a\u8df3\u8fc7\u5fae\u8c03\u7684\u201c\u4e00\u6b65\u5230\u4f4d\u201d<br>GRPO\u7b97\u6cd5\uff1a\u57fa\u4e8e\u7fa4\u4f53\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff0c\u76f4\u63a5\u901a\u8fc7\u5956\u52b1\u51fd\u6570\uff08\u5982\u4ee3\u7801\u6b63\u786e\u6027\u3001\u6570\u5b66\u8bc1\u660e\u4e25\u8c28\u6027\uff09\u9a71\u52a8\u6a21\u578b\u8fdb\u5316\u3002<\/p>\n\n\n\n<p>\u6848\u4f8b\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n<p>\u8bad\u7ec3\u65b9\u6cd5<\/p>\n\n\n\n<p>GSM8K\u6570\u5b66\u63a8\u7406\u51c6\u786e\u7387<\/p>\n\n\n\n<p>HumanEval\u4ee3\u7801\u751f\u6210\u901a\u8fc7\u7387<\/p>\n\n\n\n<p>\u76d1\u7763\u5fae\u8c03+SFT<\/p>\n\n\n\n<p>78.3%<\/p>\n\n\n\n<p>67.1%<\/p>\n\n\n\n<p>\u7eafRL\uff08GRPO\uff09<\/p>\n\n\n\n<p>89.7%<\/p>\n\n\n\n<p>76.4%<\/p>\n\n\n\n<p>\u6210\u672c\u4f18\u52bf\uff1a\u8bad\u7ec3\u5468\u671f\u7f29\u77ed40%\uff0c\u6240\u9700\u6807\u6ce8\u6570\u636e\u91cf\u51cf\u5c1195%\u3002<\/p>\n\n\n\n<p>\ud83d\ude84 \u4e09\u3001\u63a8\u7406\u6548\u7387\u8dc3\u8fc1\uff1a\u4ece\u201c\u7b97\u529b\u541e\u566c\u201d\u5230\u201c\u5b9e\u65f6\u54cd\u5e94\u201d<\/p>\n\n\n\n<p>3.1 \u7a00\u758f\u8ba1\u7b97\u4e0e\u786c\u4ef6\u534f\u540c\u4f18\u5316<br>\u4e13\u5bb6\u7f13\u5b58\u6280\u672f\uff1a\u5c06\u9ad8\u9891\u4e13\u5bb6\u53c2\u6570\u9884\u52a0\u8f7d\u81f3HBM\u663e\u5b58\uff0c\u51cf\u5c11PCIe\u6570\u636e\u4f20\u8f93\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<p>\u91cf\u5316\u90e8\u7f72\uff1a\u901a\u8fc7INT4\u91cf\u5316\uff0c\u6a21\u578b\u4f53\u79ef\u538b\u7f29\u81f31\/4\uff0c\u63a8\u7406\u901f\u5ea6\u63d0\u53472.8\u500d\u3002<\/p>\n\n\n\n<p>\u80fd\u6548\u5bf9\u6bd4\uff1a\u5904\u740610\u4ebf\u6b21\u63a8\u7406\u8bf7\u6c42\uff0cDeepSeek\u80fd\u8017\u6210\u672c\u4e3a2300\uff0c\u8fdc\u4f4e\u4e8e\u4f20\u7edf\u6a21\u578b\u76842300\uff0c\u8fdc\u4f4e\u4e8e\u4f20\u7edf\u6a21\u578b\u768415000\u3002<\/p>\n\n\n\n<p>3.2 \u52a8\u6001\u6279\u5904\u7406\u4e0e\u589e\u91cf\u63a8\u7406<br>\u52a8\u6001\u6279\u5904\u7406\uff1a\u6839\u636e\u8f93\u5165\u957f\u5ea6\u52a8\u6001\u5206\u7ec4\uff0cGPU\u5229\u7528\u7387\u4ece60%\u63d0\u5347\u81f385%\u3002<\/p>\n\n\n\n<p>\u589e\u91cf\u63a8\u7406\uff1a\u4ec5\u9700\u8ba1\u7b97\u65b0\u589etoken\u7684\u6ce8\u610f\u529b\u6743\u91cd\uff0c\u957f\u5bf9\u8bdd\u573a\u666f\u5ef6\u8fdf\u964d\u4f4e35%\u3002<\/p>\n\n\n\n<p>\u5b9e\u6d4b\u6570\u636e\uff1a\u5728\u5b9e\u65f6\u7ffb\u8bd1\u4efb\u52a1\u4e2d\uff0cDeepSeek\u652f\u630150\u8def\u5e76\u53d1\uff0c\u5e73\u5747\u54cd\u5e94\u65f6\u95f492ms\uff0c\u663e\u8457\u4f18\u4e8eTransformer\u7684210ms\u3002<\/p>\n\n\n\n<p>\ud83c\udf0d \u56db\u3001\u6cdb\u5316\u80fd\u529b\u7a81\u7834\uff1a\u4ece\u201c\u4e13\u624d\u201d\u5230\u201c\u901a\u624d\u201d<br>4.1 \u591a\u6a21\u6001\u7edf\u4e00\u67b6\u6784<br>\u8de8\u6a21\u6001\u5bf9\u9f50\uff1a\u901a\u8fc7\u5171\u4eab\u6f5c\u5728\u7a7a\u95f4\uff0c\u5b9e\u73b0\u6587\u672c\u3001\u56fe\u50cf\u3001\u8bed\u97f3\u7684\u8054\u5408\u8868\u5f81\u3002<\/p>\n\n\n\n<p>\u6848\u4f8b\uff1a\u5728\u533b\u7597\u573a\u666f\u4e2d\uff0cDeepSeek\u53ef\u540c\u65f6\u89e3\u6790CT\u5f71\u50cf\uff08\u56fe\u50cf\uff09\u4e0e\u60a3\u8005\u75c5\u53f2\uff08\u6587\u672c\uff09\uff0c\u8bca\u65ad\u51c6\u786e\u7387\u8fbe91.3%\uff0c\u6bd4\u5355\u6a21\u6001\u6a21\u578b\u9ad814.7%\u3002<\/p>\n\n\n\n<p>4.2 \u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b<br>\u6280\u672f\u539f\u7406\uff1a\u901a\u8fc7\u5f3a\u5316\u8bfe\u7a0b\u5b66\u4e60\uff08Curriculum RL\uff09\uff0c\u6a21\u578b\u81ea\u4e3b\u6784\u5efa\u4efb\u52a1\u95f4\u7684\u6cdb\u5316\u5173\u8054\u3002<\/p>\n\n\n\n<p>\u6027\u80fd\u5bf9\u6bd4\uff1a\u5728\u672a\u89c1\u8fc7\u7684\u91d1\u878d\u98ce\u9669\u9884\u6d4b\u4efb\u52a1\u4e2d\uff0cDeepSeek\u96f6\u6837\u672c\u51c6\u786e\u7387\u8fbe82.4%\uff0c\u6bd4\u4f20\u7edf\u6a21\u578b\u9ad819.2%\u3002<\/p>\n\n\n\n<p>\ud83d\udcbc \u4e94\u3001\u884c\u4e1a\u843d\u5730\uff1a\u4ece\u201c\u5b9e\u9a8c\u5ba4\u73a9\u5177\u201d\u5230\u201c\u751f\u4ea7\u529b\u5f15\u64ce\u201d<br>5.1 \u6210\u672c\u4f18\u52bf\u63a8\u52a8\u666e\u60e0\u5316<br>\u8bad\u7ec3\u6210\u672c\uff1a\u5343\u4ebf\u7ea7\u6a21\u578b\u8bad\u7ec3\u6210\u672c\u964d\u81f3\u767e\u4e07\u7f8e\u5143\u7ea7\uff0c\u4e2d\u5c0f\u4f01\u4e1a\u53ef\u8d1f\u62c5\u3002<\/p>\n\n\n\n<p>API\u5b9a\u4ef7\uff1aDeepSeek\u7684token\u4ef7\u683c\u4ec5\u4e3aGPT-4\u76841\/5\uff0c\u8c03\u75281\u4ebftoken\u4ec5\u9700$20\u3002<\/p>\n\n\n\n<p>5.2 \u5782\u76f4\u9886\u57df\u5e94\u7528\u6848\u4f8b<br>\u884c\u4e1a<\/p>\n\n\n\n<p>\u5e94\u7528\u573a\u666f<\/p>\n\n\n\n<p>\u6027\u80fd\u63d0\u5347<\/p>\n\n\n\n<p>\u667a\u80fd\u5236\u9020<\/p>\n\n\n\n<p>\u8bbe\u5907\u6545\u969c\u9884\u6d4b<\/p>\n\n\n\n<p>\u9884\u6d4b\u51c6\u786e\u7387\u219118%\uff0c\u7ef4\u62a4\u6210\u672c\u219330%<\/p>\n\n\n\n<p>\u667a\u6167\u533b\u7597<\/p>\n\n\n\n<p>\u653e\u5c04\u5f71\u50cf\u5206\u6790<\/p>\n\n\n\n<p>\u8bca\u65ad\u901f\u5ea6\u21913\u500d\uff0c\u6f0f\u8bca\u7387\u219325%<\/p>\n\n\n\n<p>\u91d1\u878d\u79d1\u6280<\/p>\n\n\n\n<p>\u9ad8\u9891\u4ea4\u6613\u7b56\u7565\u751f\u6210<\/p>\n\n\n\n<p>\u7b56\u7565\u6536\u76ca\u7387\u219112%\uff0c\u5ef6\u8fdf\u219345ms<\/p>\n\n\n\n<p>\ud83d\udd2e \u516d\u3001\u6311\u6218\u4e0e\u672a\u6765\uff1a\u901a\u5f80AGI\u7684\u4e0b\u4e00\u7ad9<br>6.1 \u73b0\u5b58\u6311\u6218<br>\u8bad\u7ec3\u590d\u6742\u6027\uff1aMoE\u52a8\u6001\u8def\u7531\u7b56\u7565\u9700\u7cbe\u7ec6\u8c03\u53c2\uff0c\u5de5\u7a0b\u5e08\u5b66\u4e60\u66f2\u7ebf\u9661\u5ced\u3002<\/p>\n\n\n\n<p>\u5b89\u5168\u4e0e\u4f26\u7406\uff1a\u6a21\u578b\u51b3\u7b56\u8fc7\u7a0b\u4ecd\u9700\u589e\u5f3a\u53ef\u89e3\u91ca\u6027\u3002<\/p>\n\n\n\n<p>6.2 \u6280\u672f\u6f14\u8fdb\u65b9\u5411<br>\u795e\u7ecf\u7b26\u53f7\u7ed3\u5408\uff1a\u5c06\u7b26\u53f7\u63a8\u7406\u5d4c\u5165MoE\u67b6\u6784\uff0c\u63d0\u5347\u903b\u8f91\u4e25\u8c28\u6027\u3002<\/p>\n\n\n\n<p>\u751f\u7269\u542f\u53d1\u8ba1\u7b97\uff1a\u6a21\u62df\u4eba\u8111\u795e\u7ecf\u53ef\u5851\u6027\uff0c\u5b9e\u73b0\u7ec8\u8eab\u5b66\u4e60\u3002<\/p>\n\n\n\n<p>\ud83c\udf31 \u4e03\u3001\u5f00\u6e90\u751f\u6001\uff1a\u4ece\u201c\u6280\u672f\u58c1\u5792\u201d\u5230\u201c\u793e\u533a\u5171\u521b\u201d<br>7.1 DeepLink\u6846\u67b6\uff1a\u964d\u4f4e\u5927\u6a21\u578b\u7814\u53d1\u95e8\u69db<br>\u4e00\u952e\u5f0f\u5206\u5e03\u5f0f\u8bad\u7ec3\uff1a\u652f\u6301\u4e07\u5361\u96c6\u7fa4\u81ea\u52a8\u8c03\u5ea6\uff0c\u8d44\u6e90\u5229\u7528\u7387\u8fbe90%\uff0c\u8bad\u7ec3\u5343\u4ebf\u6a21\u578b\u4ec5\u97003\u5929\uff08\u4f20\u7edf\u6846\u67b6\u97002\u5468\uff09\u3002<\/p>\n\n\n\n<p>\u53ef\u89c6\u5316\u8c03\u53c2\u5de5\u5177\uff1a\u5185\u7f6e\u8d85\u53c2\u6570\u4f18\u5316\u7b97\u6cd5\uff0c\u81ea\u52a8\u641c\u7d22\u6700\u4f73\u8bad\u7ec3\u914d\u7f6e\uff0c\u5de5\u7a0b\u5e08\u8c03\u8bd5\u65f6\u95f4\u51cf\u5c1170%\u3002<\/p>\n\n\n\n<p>\u6848\u4f8b\uff1a\u67d0\u9ad8\u6821\u7814\u7a76\u56e2\u961f\u5229\u7528DeepLink\uff0c\u57284\u4e2a\u6708\u5185\u5b8c\u6210\u533b\u7597\u95ee\u7b54\u6a21\u578b\u7684\u8bad\u7ec3\u4e0e\u90e8\u7f72\uff0c\u6210\u672c\u63a7\u5236\u5728$8\u4e07\u4ee5\u5185\u3002<\/p>\n\n\n\n<p>7.2 \u5f00\u53d1\u8005\u793e\u533a\u7206\u53d1\u5f0f\u589e\u957f<br>\u6a21\u578b\u5fae\u8c03\u5e02\u573a\uff1a\u5f00\u6e90\u793e\u533a\u8d21\u732e\u8d85\u8fc71200\u4e2a\u9886\u57df\u9002\u914d\u5668\uff08\u5982\u6cd5\u5f8b\u5408\u540c\u5ba1\u67e5\u3001\u5de5\u4e1a\u8d28\u68c0\uff09\uff0c\u4e0b\u8f7d\u91cf\u7a81\u7834500\u4e07\u6b21\u3002<\/p>\n\n\n\n<p>\u786c\u4ef6\u517c\u5bb9\u6027\uff1a\u652f\u6301NVIDIA\u3001AMD\u3001\u6d77\u5149\u7b49\u4e3b\u6d41\u82af\u7247\uff0c\u63a8\u7406\u5f15\u64ce\u9002\u914d200+\u8fb9\u7f18\u8bbe\u5907\u578b\u53f7\u3002<\/p>\n\n\n\n<p>\ud83d\udd0d \u516b\u3001\u6280\u672f\u7ec6\u8282\u63ed\u79d8\uff1a\u4e3a\u4ec0\u4e48MoE+MLA\u662f\u9ec4\u91d1\u7ec4\u5408\uff1f<\/p>\n\n\n\n<p>8.1 \u52a8\u6001\u7a00\u758f\u6027\u7684\u6570\u5b66\u672c\u8d28<br>MoE\u7684\u7a00\u758f\u6fc0\u6d3b\u53ef\u5efa\u6a21\u4e3a\u6761\u4ef6\u8ba1\u7b97\uff08Conditional Computation\uff09\uff1a<\/p>\n\n\n\n<p>y=\u2211i=1nG(x)i\u22c5Ei(x)y=i=1\u2211n\u200bG(x)i\u200b\u22c5Ei\u200b(x)<\/p>\n\n\n\n<p>\u5176\u4e2d\uff0cG(x)G(x)\u4e3a\u8def\u7531\u95e8\u63a7\u51fd\u6570\uff0c\u786e\u4fdd90%\u4ee5\u4e0a\u7684\u4e13\u5bb6\u5904\u4e8e\u201c\u4f11\u7720\u201d\u72b6\u6001\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u5f53\u4e13\u5bb6\u6570\u8d85\u8fc764\u65f6\uff0c\u6a21\u578b\u6d8c\u73b0\u51fa\u4efb\u52a1\u81ea\u9002\u5e94\u5206\u6cbb\u80fd\u529b\u2014\u2014\u4f8b\u5982\uff0c\u91d1\u878d\u6587\u672c\u81ea\u52a8\u8def\u7531\u81f3\u6570\u5b66\u903b\u8f91\u4e13\u5bb6\uff0c\u800c\u8bd7\u6b4c\u751f\u6210\u7531\u8bed\u8a00\u98ce\u683c\u4e13\u5bb6\u5904\u7406\u3002<\/p>\n\n\n\n<p>8.2 MLA\u7684\u538b\u7f29\u8fb9\u754c\u4e0e\u7cbe\u5ea6\u6743\u8861<br>\u901a\u8fc7\u77e9\u9635\u4f4e\u79e9\u5206\u89e3\uff0cKV\u7f13\u5b58\u4eceO(n\u00d7d)O(n\u00d7d)\u964d\u81f3O(n\u00d7k)O(n\u00d7k)\uff08k=d\/8k=d\/8\uff09\u3002\u5728128K\u4e0a\u4e0b\u6587\u573a\u666f\u4e0b\uff0c\u7cbe\u5ea6\u635f\u5931\u53ef\u901a\u8fc7\u52a8\u6001\u6295\u5f71\u6062\u590d\u6280\u672f\u63a7\u5236\u57281.2%\u4ee5\u5185\uff0c\u6bd4\u4f20\u7edfKVCache\u526a\u679d\u65b9\u6cd5\u9ad83\u500d\u6536\u76ca\u3002<\/p>\n\n\n\n<p>\ud83d\udcc8 \u4e5d\u3001\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5\uff1a\u5168\u9762\u78be\u538b\u4f20\u7edf\u6a21\u578b<br>9.1 \u901a\u7528\u4efb\u52a1\u5bf9\u6bd4<br>\u4efb\u52a1\u7c7b\u578b<\/p>\n\n\n\n<p>DeepSeek-MoE-128B<\/p>\n\n\n\n<p>Transformer-XL<\/p>\n\n\n\n<p>\u63d0\u5347\u5e45\u5ea6<\/p>\n\n\n\n<p>\u6587\u672c\u751f\u6210\uff08BLEU\uff09<\/p>\n\n\n\n<p>74.3<\/p>\n\n\n\n<p>68.1<\/p>\n\n\n\n<p>+9.1%<\/p>\n\n\n\n<p>\u4ee3\u7801\u751f\u6210\uff08Pass@1\uff09<\/p>\n\n\n\n<p>76.4%<\/p>\n\n\n\n<p>67.2%<\/p>\n\n\n\n<p>+13.7%<\/p>\n\n\n\n<p>\u957f\u6587\u672c\u7406\u89e3\uff08EM\uff09<\/p>\n\n\n\n<p>82.7%<\/p>\n\n\n\n<p>73.5%<\/p>\n\n\n\n<p>+12.5%<\/p>\n\n\n\n<p>9.2 \u80fd\u6548\u6bd4\u9769\u547d<br>\u6307\u6807<\/p>\n\n\n\n<p>DeepSeek<\/p>\n\n\n\n<p>\u4f20\u7edfTransformer<\/p>\n\n\n\n<p>\u4f18\u52bf<\/p>\n\n\n\n<p>\u63a8\u7406\u80fd\u8017\uff08TOPS\/W\uff09<\/p>\n\n\n\n<p>12.4<\/p>\n\n\n\n<p>3.7<\/p>\n\n\n\n<p>3.35\u500d<\/p>\n\n\n\n<p>\u8bad\u7ec3\u78b3\u6392\u653e\uff08\u5428CO\u2082\uff09<\/p>\n\n\n\n<p>18.2<\/p>\n\n\n\n<p>256.7<\/p>\n\n\n\n<p>\u51cf\u5c1193%<\/p>\n\n\n\n<p>\ud83e\udde9 \u5341\u3001\u4ece\u7406\u8bba\u5230\u5b9e\u8df5\uff1a\u5de5\u7a0b\u5e08\u843d\u5730\u6307\u5357<br>10.1 \u6a21\u578b\u9009\u578b\u7b56\u7565<br>\u5b9e\u65f6\u573a\u666f\uff1a\u4f18\u5148\u9009\u62e9MoE-32B+MLA\u67b6\u6784\uff0cINT4\u91cf\u5316\u7248\u672c\u5ef6\u8fdf\u6700\u4f4e\u3002<\/p>\n\n\n\n<p>\u957f\u6587\u672c\u5904\u7406\uff1a\u542f\u7528128K\u4e0a\u4e0b\u6587\u6a21\u5f0f\uff0c\u9700\u914d\u7f6e\u81f3\u5c1140GB\u663e\u5b58\u8282\u70b9\u3002<\/p>\n\n\n\n<p>\u4f4e\u6210\u672c\u5fae\u8c03\uff1a\u91c7\u7528LoRA\u9002\u914d\u5668\uff0c\u4ec5\u8bad\u7ec30.1%\u53c2\u6570\u5373\u53ef\u9002\u914d\u4e13\u4e1a\u4efb\u52a1\u3002<\/p>\n\n\n\n<p>10.2 \u907f\u5751\u6307\u5357<br>\u8def\u7531\u9707\u8361\u95ee\u9898\uff1a\u521d\u59cb\u8bad\u7ec3\u65f6\u9700\u9650\u5236\u4e13\u5bb6\u9009\u62e9\u6570\u91cf\uff08\u5982Top-2\uff09\uff0c\u9010\u6b65\u653e\u5f00\u81f3Top-4\u3002<\/p>\n\n\n\n<p>\u663e\u5b58\u6ea2\u51fa\u98ce\u9669\uff1aFP8\u8bad\u7ec3\u4e2d\u9700\u76d1\u63a7\u68af\u5ea6\u7f29\u653e\u56e0\u5b50\uff0c\u9632\u6b62\u6570\u503c\u6ea2\u51fa\u5bfc\u81f4\u8bad\u7ec3\u5d29\u6e83\u3002<\/p>\n\n\n\n<p>\ud83c\udf10 \u5341\u4e00\u3001\u5168\u7403\u6280\u672f\u7ade\u4e89\u683c\u5c40<br>11.1 DeepSeek vs. \u56fd\u9645\u5927\u6a21\u578b<br>\u5382\u5546<\/p>\n\n\n\n<p>\u6838\u5fc3\u67b6\u6784<\/p>\n\n\n\n<p>\u4e0a\u4e0b\u6587\u957f\u5ea6<\/p>\n\n\n\n<p>\u8bad\u7ec3\u6210\u672c\uff08\u5341\u4ebf\u7f8e\u5143\uff09<\/p>\n\n\n\n<p>\u5f00\u6e90\u7b56\u7565<\/p>\n\n\n\n<p>DeepSeek<\/p>\n\n\n\n<p>MoE+MLA<\/p>\n\n\n\n<p>128K<\/p>\n\n\n\n<p>0.0557<\/p>\n\n\n\n<p>\u5168\u9762\u5f00\u6e90<\/p>\n\n\n\n<p>GPT-4<\/p>\n\n\n\n<p>\u5bc6\u96c6Transformer<\/p>\n\n\n\n<p>32K<\/p>\n\n\n\n<p>1.2<\/p>\n\n\n\n<p>\u95ed\u6e90API<\/p>\n\n\n\n<p>Gemini<\/p>\n\n\n\n<p>MoE+Hybrid<\/p>\n\n\n\n<p>64K<\/p>\n\n\n\n<p>0.9<\/p>\n\n\n\n<p>\u90e8\u5206\u5f00\u6e90<\/p>\n\n\n\n<p>11.2 \u56fd\u4ea7\u5316\u66ff\u4ee3\u8fdb\u7a0b<br>\u786c\u4ef6\u9002\u914d\uff1aDeepSeek\u5df2\u5b8c\u6210\u6607\u817e910B\u3001\u5bd2\u6b66\u7eaaMLU370\u7684\u6df1\u5ea6\u4f18\u5316\uff0c\u63a8\u7406\u901f\u5ea6\u8fbe\u56fd\u9645\u82af\u7247\u768485%\u3002<\/p>\n\n\n\n<p>\u884c\u4e1a\u6e17\u900f\u7387\uff1a\u5728\u653f\u52a1\u3001\u91d1\u878d\u7b49\u5173\u952e\u9886\u57df\uff0c\u56fd\u4ea7\u5927\u6a21\u578b\u90e8\u7f72\u7387\u4ece2021\u5e74\u768412%\u5347\u81f32023\u5e74\u768467%\u3002<\/p>\n\n\n\n<p>\ud83d\udd2c \u5341\u4e8c\u3001\u5b66\u672f\u7814\u7a76\u70ed\u70b9\uff1aDeepSeek\u542f\u53d1\u7684\u521b\u65b0\u65b9\u5411<\/p>\n\n\n\n<p>12.1 \u52a8\u6001\u7a00\u758f\u7406\u8bba<br>2023 NIPS\u6700\u4f73\u8bba\u6587\u300aMoE\u4e2d\u7684\u4e13\u5bb6\u534f\u4f5c\u535a\u5f08\u300b\u8bc1\u660e\uff1a\u5f53\u4e13\u5bb6\u6570\u2265128\u65f6\uff0c\u6a21\u578b\u81ea\u52a8\u5f62\u6210\u7eb3\u4ec0\u5747\u8861\uff0c\u4efb\u52a1\u5206\u914d\u6548\u7387\u63d0\u534740%\u3002<\/p>\n\n\n\n<p>12.2 \u751f\u7269\u795e\u7ecf\u7f51\u7edc\u542f\u53d1<br>\u8109\u51b2MoE\u67b6\u6784\uff1a\u6a21\u62df\u795e\u7ecf\u5143\u8109\u51b2\u53d1\u653e\u673a\u5236\uff0c\u63a8\u7406\u80fd\u8017\u518d\u964d30%\uff0c\u8bba\u6587\u5df2\u88abNature Machine Intelligence\u6536\u5f55\u3002<\/p>\n\n\n\n<p>\ud83c\udfaf \u603b\u7ed3<br>DeepSeek\u7684\u6280\u672f\u7a81\u7834\u4e0d\u4ec5\u662f\u5de5\u7a0b\u4f18\u5316\u7684\u80dc\u5229\uff0c\u66f4\u662fAI\u57fa\u7840\u7406\u8bba\u7684\u91cc\u7a0b\u7891\u3002\u5176<strong>\u201c\u52a8\u6001\u7a00\u758f\u8ba1\u7b97\u201d\u8303\u5f0f<\/strong>\u98a0\u8986\u4e86\u4f20\u7edf\u201c\u8d8a\u5927\u8d8a\u6162\u201d\u7684\u8ba4\u77e5\u8fb9\u754c\uff0c\u800c\u5f00\u6e90\u751f\u6001\u7684\u5efa\u8bbe\u52a0\u901f\u4e86\u6280\u672f\u6c11\u4e3b\u5316\u8fdb\u7a0b\u3002\u672a\u6765\uff0c\u968f\u7740\u795e\u7ecf\u7b26\u53f7\u8ba1\u7b97\u7b49\u6280\u672f\u7684\u878d\u5408\uff0cDeepSeek\u6216\u5c06\u6210\u4e3aAGI\u65f6\u4ee3\u7684\u6838\u5fc3\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n\n\n\n<p>\ud83d\udcbc\u3010\u7701\u5fc3\u9510\u8bc4\u3011<br>\u201c\u5f53\u522b\u4eba\u8fd8\u5728\u4e3a\u7b97\u529b\u5185\u5377\u65f6\uff0cDeepSeek\u91cd\u65b0\u5b9a\u4e49\u4e86\u6e38\u620f\u89c4\u5219\u2014\u2014\u9ad8\u6548\u4e0d\u662f\u9009\u62e9\u9898\uff0c\u800c\u662f\u751f\u5b58\u9898\u3002\u201d<br><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\ud83d\udcd6 \u3010\u6458\u8981\u3011DeepSeek\u4f5c\u4e3a\u65b0\u4e00\u4ee3AI\u67b6\u6784\u4ee3\u8868\uff0c\u901a\u8fc7\u6df7\u5408\u4e13\u5bb6\u7cfb\u7edf\uff08MoE\uff09\u3001\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u7b49\u521b\u65b0\u6280\u672f\uff0c\u5728\u6548\u7387\u3001\u6210\u672c\u4e0e\u6027\u80fd\u4e0a\u5b9e\u73b0\u5bf9\u4f20\u7edf\u6a21\u578b\u7684\u5168\u9762\u8d85\u8d8a\u3002\u672c\u6587\u4ece\u67b6\u6784\u8bbe\u8ba1\u3001\u8bad\u7ec3\u4f18\u5316\u3001\u63a8\u7406\u52a0\u901f\u3001\u6cdb\u5316\u80fd\u529b\u7b49\u7ef4\u5ea6\uff0c\u7ed3\u5408\u6700\u65b0\u7814\u7a76\u6570\u636e\u4e0e\u4ea7\u4e1a\u6848\u4f8b\uff0c\u6df1\u5ea6\u89e3\u6790\u5176\u6280\u672f\u7a81\u7834\u8def\u5f84\u53ca\u884c\u4e1a\u5f71\u54cd\u3002 \ud83d\ude80 \u5f15\u8a00\uff1aAI\u6a21\u578b\u7684\u6548\u7387\u56f0\u5c40\u4e0e\u7834\u5c40\u4e4b\u6218\u4f20\u7edfAI\u6a21\u578b\uff08CNN\u3001RNN\u3001Transformer\uff09\u957f\u671f\u9762\u4e34\u4e09\u5927\u77db\u76fe\uff1a\u7b97\u529b\u9700\u6c42\u6307\u6570\u589e\u957f\u4e0e\u786c\u4ef6\u74f6\u9888\u7684\u77db\u76fe\u3001\u6a21\u578b\u89c4\u6a21\u6269\u5927\u4e0e\u63a8\u7406\u5b9e\u65f6\u6027\u7684\u77db\u76fe\u3001\u4e13\u7528\u4efb\u52a1\u4f18\u5316\u4e0e\u6cdb\u5316\u80fd\u529b\u7684\u77db\u76fe\u3002\u4ee5Transformer\u4e3a\u4f8b\uff0c\u5176\u81ea\u6ce8\u610f\u529b\u673a\u5236\u867d\u7a81\u7834\u5e8f\u5217\u5efa\u6a21\u9650\u5236\uff0c\u4f46\u5904\u7406128K tokens\u7684\u6587\u672c\u65f6\uff0c\u5185\u5b58\u5360\u7528\u9ad8\u8fbe3 &hellip;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-169","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/posts\/169","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/comments?post=169"}],"version-history":[{"count":1,"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/posts\/169\/revisions"}],"predecessor-version":[{"id":170,"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/posts\/169\/revisions\/170"}],"wp:attachment":[{"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/media?parent=169"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/categories?post=169"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.dianshitv.cn\/index.php\/wp-json\/wp\/v2\/tags?post=169"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}