{"id":53839,"date":"2025-02-18T08:47:46","date_gmt":"2025-02-18T00:47:46","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53839\/"},"modified":"2025-02-18T08:47:46","modified_gmt":"2025-02-18T00:47:46","slug":"%e5%be%ae%e4%bf%a1deepseek","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53839\/","title":{"rendered":"\u5fae\u4fe1+DeepSeek"},"content":{"rendered":"<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\"><span style=\"letter-spacing: 0.04em;text-indent: 2em\">ima.copilot\u521a\u51fa\u6765\u7684\u65f6\u5019, \u7528\u4e86\u4e00\u6bb5\u65f6\u95f4. \u9664\u4e86\u6df7\u5143\u81ea\u8eab\u6a21\u578b\u6709\u4e00\u4e9b\u5dee\u8ddd\u4ee5\u5916, \u5bf9\u6574\u4e2a\u516c\u4f17\u53f7\u7684\u5185\u5bb9\u751f\u6001\u4e0a\u7684\u652f\u6301\u662f\u975e\u5e38\u597d\u7684. \u524d\u6bb5\u65f6\u95f4\u4e5f\u6d4b\u8bd5\u4e86\u4e00\u4e0b\u516c\u4f17\u53f7\u540e\u53f0\u7528LLM\u81ea\u52a8\u56de\u590d,\u9664\u4e86\u4e1a\u52a1\u903b\u8f91\u4e0a\u6709\u4e9b\u4ea4\u4e92\u7684\u95ee\u9898(\u4f8b\u5982\u6b63\u5e38\u7684\u7559\u8a00\u4ea4\u6d41\u548c\u641c\u7d22\u6574\u7406\u516c\u4f17\u53f7\u5185\u5bb9\u7684\u533a\u5206),\u5176\u5b9e\u6574\u7406\u7684\u5185\u5bb9\u8fd8\u662f\u53ef\u8bfb\u6027\u5f88\u9ad8\u7684, \u5f53\u7136\u4e5f\u6709\u56e0\u4e3a\u57fa\u7840\u6a21\u578b\u7684\u95ee\u9898\u5bfc\u81f4\u7684 \u6700\u8fd1\u5fae\u4fe1\u5f00\u59cb\u7070\u5ea6DeepSeek\u4e86, \u57282C\u7684\u5e02\u573a\u4e0a\u5c06\u4f1a\u8fce\u6765\u66f4\u591a\u7684\u53d8\u6570, \u7279\u522b\u662f\u5bf9\u5b57\u8282\u8c46\u5305\u7684\u751f\u6001\u4e0a\u7684\u5f71\u54cd. \u6bd5\u7adf\u817e\u8baf\u5fae\u4fe1\u7684DAU\u662f\u63a5\u8fd110\u4ebf\u7ea7\u7684.<\/span><\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u53e6\u4e00\u65b9\u9762\u6700\u8fd1\u767e\u5ea6\u548c\u706b\u5c71\u7684\u8868\u6001(\u53c2\u8003<span style=\"font-weight: bold\">\u5feb\u79d1\u6280\u7684\u62a5\u9053<\/span><sup>[1]<\/sup>)<\/p>\n<p style=\"font-size: 14px;line-height: 1.8em;letter-spacing: 0.02em;text-indent: 0em;padding-top: 20px;padding-bottom: 8px\">\u767e\u5ea6\u667a\u80fd\u4e91\u4e8b\u4e1a\u7fa4\u603b\u88c1\u6c88\u6296\u5728\u5168\u5458\u4f1a\u4e0a\u8868\u793a\uff0c\u56fd\u5185\u5927\u6a21\u578b\u53bb\u5e74\u7684\u201c\u6076\u610f\u201d\u4ef7\u683c\u6218\uff0c\u5bfc\u81f4\u884c\u4e1a\u6574\u4f53\u521b\u6536\u76f8\u8f83\u4e8e\u56fd\u5916\u5dee\u4e86\u591a\u4e2a\u6570\u91cf\u7ea7\u3002<\/p>\n<p style=\"font-size: 14px;line-height: 1.8em;letter-spacing: 0.02em;text-indent: 0em;padding-top: 20px;padding-bottom: 8px\">\u5b57\u8282\u8df3\u52a8\u65d7\u4e0b\u706b\u5c71\u5f15\u64ce\u603b\u88c1\u8c2d\u5f85\u901a\u8fc7\u670b\u53cb\u5708\u56de\u5e94\u79f0\uff0c\u5927\u6a21\u578b\u964d\u4ef7\u662f\u901a\u8fc7\u6280\u672f\u521b\u65b0\u5b9e\u73b0\u7684\uff0c\u5927\u5bb6<code>\u5e94\u50cfDeepSeek\u4e00\u6837\u805a\u7126\u57fa\u672c\u529f\uff0c\u5c11\u65e0\u7aef\u731c\u6d4b\uff0c\u5f52\u56e0\u5916\u90e8\u3002<\/code> \u8c2d\u5f85\u6307\u51fa\uff0c\u706b\u5c71\u5f15\u64ce\u7684\u8c46\u53051.5Pro\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6210\u672c\u548c\u63a8\u7406\u6210\u672c\u5747\u4f4e\u4e8eDeepSeek V3\uff0c\u66f4\u662f\u8fdc\u4f4e\u4e8e\u56fd\u5185\u5176\u4ed6\u6a21\u578b\uff0c\u5728\u5f53\u524d\u4ef7\u683c\u4e0b\u6709\u975e\u5e38\u4e0d\u9519\u7684\u6bdb\u5229\u3002\u4ed6\u8fdb\u4e00\u6b65\u89e3\u91ca\uff1a\u201c\u56fd\u5185\u5916\u7684\u5382\u5546\u90fd\u5728\u4f9d\u9760\u6280\u672f\u521b\u65b0\uff0c\u964d\u4f4e\u6a21\u578b\u4ef7\u683c\u3002\u6211\u4eec\u4e5f\u53ea\u662f\u5b9e\u73b0\u4e86Gemini 2.0 Flash\u7684\u4ef7\u683c\u6c34\u5e73\u800c\u5df2\uff0c\u8fd9\u4e2a\u4ef7\u683c\u5b8c\u5168\u662f\u4f9d\u8d56\u6280\u672f\u8fdb\u6b65\u53ef\u4ee5\u505a\u5230\u7684\u3002\u201d<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u5176\u5b9e\u5f88\u591a\u65f6\u5019\u6210\u672c\u7684\u4f30\u8ba1\u7684\u5206\u6b67, \u672c\u8d28\u662f\u6280\u672f\u4e0a\u7684\u5dee\u8ddd. \u4f8b\u5982\u5c24\u6d0b\u8001\u5e08\u4f30\u8ba1\u7684\u6570\u636e, \u548cDeepSeek-V3\u7684\u8bba\u6587\u5b9e\u73b0\u7684PD\u5206\u79bb+EP\u5e76\u884c\u6027\u80fd\u5dee\u8ddd\u8d85\u8fc710\u500d\u4ee5\u4e0a. \u672c\u8d28\u4e0a\u7684\u6210\u672c\u5dee\u5f02\u5c31\u662f\u62ff\u4e00\u4e9b\u5f00\u6e90\u793e\u533a\u7b80\u5355\u7684TP\/PP\u5e76\u884c\u7684\u7ed3\u679c\u6765\u4f30\u8ba1\u662f\u6709\u5f88\u5927\u7684\u5dee\u8ddd\u7684. \u7279\u522b\u662fGoogle Gemini 2.0 Flash\u7684\u4ef7\u683c\u6c34\u5e73\u6765\u770b, \u6280\u672f\u4e0a\u8fd8\u6709\u66f4\u591a\u7684\u4f18\u5316\u7a7a\u95f4. \u4f8b\u5982\u6628\u5929\u63d0\u5230\u7684\u4e00\u7bc7\u6587\u7ae0\u5bf9MOE\u7684\u6027\u80fd\u5206\u6790.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u8fd9\u91cc\u7b80\u5355\u7684\u505a\u4e00\u4e2aRoofline\u7684\u5206\u6790, \u4ece\u7b97\u529b\u4e0a\u8bb2DeepSeek-V3\/R1\u6a21\u578b\u7684\u7b97\u529b\u9700\u6c42\u662f\u76f8\u5bf9\u8f83\u5c0f\u7684, \u74f6\u9888\u4e3b\u8981\u662f\u5728\u8bbf\u5b58\u548cAll2All\u7684\u901a\u4fe1\u4e0a\u4ee5\u53ca\u5982\u4f55\u89e3\u51b3\u63a8\u7406\u65f6\u7684\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861\u4e0a. \u4f8b\u5982\u534e\u4e3a\u6607\u817e\u63d0\u5230\u7684:<span style=\"text-decoration: underline\">\u201c\u901a\u8fc7EP\u6df7\u5408\u5e76\u884c\u7b97\u6cd5, \u901a\u4fe1\u4f18\u5316\u6027\u80fd\u63d0\u534730%+, \u8bbf\u5b58\u6027\u80fd\u63d0\u534720%+, \u4ece\u800c\u964d\u4f4e\u4e13\u5bb6\u4e0d\u5747\u8861\u5ea6, \u63a8\u7406\u541e\u5410\u6027\u80fd\u63d0\u534720%~35%\u201d <\/span>\u53e6\u4e00\u65b9\u9762\u4ece\u8881\u8fdb\u8f89\u8001\u5e08\u7684\u4e00\u6bb5\u8bdd\u53ef\u4ee5\u77e5\u9053, \u6881\u603b\u4e3a\u5565\u8981\u63a8\u8350\u6027\u80fd\u6700\u597d\u9700\u898180\u53f0, \u4e3b\u8981\u662f\u901a\u8fc7EP\u5e76\u884c\u83b7\u5f97\u66f4\u597d\u7684Data Locality.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u4eceNetwork-Bound\u4e0a\u6765\u770b, \u6211\u4eec\u4ee5\u5355\u4e2aToken 7168B\u6765\u8ba1\u7b97, 400Gbps\u7f51\u7edc\u673a\u95f4\u4e92\u8054\u7f51\u7edc\u537350GB\/s, \u7b80\u5355\u7684\u505a\u4e00\u4e2a\u4e0a\u9650\u4f30\u8ba1. \u6a21\u578b\u9700\u8981\u4f20\u8f9360\u5c42, \u6bcf\u4e2aToken\u9700\u89818\u4e2aRouted Expert\u548c\u4e00\u4e2aShared Expert, \u5373\u5355\u4e2aToken\u9700\u8981 7168 x 9 x60 ~ 4MB\u7684\u6570\u636e, \u518d\u7b97\u4e0aAttention block\u7684\u901a\u4fe1\u91cf, \u5373\u4e00\u79d2\u5355\u5361\u80fd\u591f\u4ea7\u751f\u7684Token\u4e3a\u8d85\u8fc76000\u4e2a, \u518d\u52a0\u4e0a\u4e00\u4e9b\u901a\u4fe1\u635f\u8017\u548cSLA\u7684\u5ef6\u8fdf\u4fdd\u969c\u7ea6\u675f, \u6309\u7167\u53ea\u670930%~50%\u7684\u6298\u6263\u4f30\u8ba1, \u57fa\u672c\u4e0a<span style=\"text-decoration: underline\">\u5355\u5361\u5728\u5e76\u884c\u7b56\u7565\u6070\u5f53\u65f6\u80fd\u591f\u505a\u52301800~3000\u4e2atokens.<\/span><\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u4eceMemory-Bound\u4e0a\u6765\u770b, \u867d\u7136\u6309\u7167DSv3\u7684\u8bba\u6587Decoding\u9636\u6bb5\u9700\u8981256\u4e2atoken\u4f5c\u4e3a\u4e00\u4e2abatch, \u5176\u6570\u636e\u91cf\u4e3a1.8MB, \u4f46\u662f\u5355\u4e2aExpert\u7684\u53c2\u6570\u6570\u636e\u91cf\u4e3a44MB. \u56e0\u6b64\u5982\u679c\u5c3d\u91cf\u7684\u628a\u4e13\u5bb6\u6253\u6563, \u7136\u540e\u4fdd\u8bc1L2 Cache\u7684hit-rate\u65f6, \u5185\u5b58\u5e26\u5bbd\u7684\u6548\u7387\u4f1a\u9ad8\u51fa\u5f88\u591a\u500d.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u800c\u5355\u673a\/\u53cc\u673a\/\u56db\u673a\u7684PP\/TP\u5e76\u884c\u5219\u5f88\u96be\u83b7\u5f97\u8fd9\u6837\u7684Data Locality\u7684\u4f18\u52bf, \u8fd9\u4e5f\u5c31\u662f\u4e3a\u4ec0\u4e48\u6881\u603b\u5efa\u8bae40\u53f0\u6216\u800580\u53f0\u505a\u66f4\u5927\u89c4\u6a21EP\u5e76\u884c\u7684\u539f\u56e0.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u53e6\u4e00\u65b9\u9762\u6211\u4eec\u8fd8\u9700\u8981\u8003\u8651\u5230DeepSeek-V3\/R1\u5bf9MTP\u7684\u652f\u6301, \u4f8b\u5982Sglang\u6700\u8fd1\u5b9e\u73b0MTP\u540e, \u6027\u80fd\u51e0\u4e4e\u53c8\u7ffb\u500d\u4e86. \u6240\u4ee5\u5728MTP\u652f\u6301\u7684\u60c5\u51b5\u4e0b, \u5355\u5361\u7684TPS\u8fd8\u53ef\u4ee5\u63a5\u8fd1\u7ffb\u500d.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u8003\u8651\u5230\u4e00\u4e9b\u989d\u5916\u7684\u5f00\u9500, \u6211\u4eec\u6309\u7167\u5355\u5361\u7684\u6027\u80fd\u4e0b\u96502000TPS\u8ba1\u7b97, \u5355\u673a8\u5361\u5c31\u662f\u5dee\u4e0d\u591a16000TPS. \u6309\u7167\u6bcf\u4e2a\u7528\u623720TPS\u7684\u901f\u7387, \u5927\u6982\u5355\u53f0H20\u53ef\u4ee5\u627f\u62c5800\u4e2a\u7528\u6237, \u8003\u8651PD\u5206\u79bb\u4e2d\u7684Prefill\u8282\u70b9\u7684\u53e6\u4e00\u4e9b\u5f00\u9500, \u4f30\u8ba1\u5355\u673aH20\u505a\u5230600\u4e2a\u7528\u6237\u6280\u672f\u4e0a\u662f\u53ef\u884c\u7684.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u90a3\u4e48\u7d27\u63a5\u7740\u9488\u5bf9\u5fae\u4fe110\u4ebfDAU, \u65e9\u4e0a7~10\u70b9\u57fa\u672c\u4e0a\u662f\u5404\u79cd\u8d44\u8baf\u7c7b\u4fe1\u606f\u7684\u516c\u4f17\u53f7\u6d88\u606f\u63a8\u9001, \u4e0b\u5348\u5927\u591a\u662f\u4e00\u4e9b\u5e7f\u544a\/\u7535\u5546, \u665a\u4e0a\u5185\u5bb9\u8fd8\u4f1a\u66f4\u4e30\u5bcc, \u57fa\u672c\u4e0a\u4e00\u5929\u5185\u90fd\u53ef\u4ee5\u7ef4\u6301\u5728\u4e00\u4e2a\u76f8\u5bf9\u8f83\u9ad8\u7684\u6c34\u5e73. \u6309\u7167\u5355\u7528\u6237\u5fae\u4fe1\u4f7f\u7528\u65f6\u957f60\u5206\u949f\u6765\u4f30\u8ba1, \u5927\u6982\u5e76\u53d1\u6d3b\u8dc3\u7528\u6237\u6570\u4ee54000\u4e07\u4f30\u8ba1, \u6309\u7167\u5355\u673a800\u4e2a\u7528\u6237, \u5927\u6982\u9700\u89815\u4e07\u53f0\u4e5f\u5c31\u662f\u8bf440\u4e07\u5361. \u4f46\u5b9e\u9645\u4e0a\u5982\u679c\u8fdb\u4e00\u6b65\u653e\u5bbd\u523010Tokens\/s\u548c\u8003\u8651\u5230\u4e00\u4e9b\u6cca\u677e\u5230\u8fbe\u548c\u7528\u6237\u4f7f\u7528\u9891\u7387\u7684\u60c5\u51b5, \u5927\u698210\u4e07\u5361~20\u4e07\u5361\u5373\u53ef, \u4e5f\u5c31\u662f\u516c\u4f17\u53f7<code>\u5171\u8bc6\u7c89\u788e\u673a<\/code>\u5728\u63d0\u5230\u7684<\/p>\n<blockquote style=\"margin-top: 20px;margin-bottom: 20px;padding: 10px 10px 10px 20px;border-style: none;border-top-width: 3px;border-bottom-width: 3px;border-right-width: 3px;border-radius: 0px;width: auto;height: auto;overflow: auto\"><p>\n <span style=\"font-size: 16px;line-height: 1.5em;letter-spacing: 0em\"><\/span> <\/p>\n<p style=\"text-indent: 2em;padding-top: 20px;padding-bottom: 8px;font-size: 14px;line-height: 1.8em;letter-spacing: 0em\">\u6211\u4eec\u5728\u4e4b\u524d\u5c31\u5df2\u7ecf\u4ece\u4f9b\u5e94\u94fe\u770b\u5230\u4e86\u817e\u8baf\u52a0\u5355\u4e8610-20\u4e07\u5f20H20\uff0c\u73b0\u5728\u770b\u5fae\u4fe1\u7248Deepseek\u5c31\u662f\u660e\u786e\u7684\u7528\u9014\u3002<\/p>\n<\/blockquote>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u518d\u6765\u8bf4\u70b9\u9898\u5916\u8bdd, \u6700\u8fd1\u4e00\u6bb5\u65f6\u95f4\u9664\u4e86\u8c03\u4f18\u63a8\u7406\u4ee5\u5916, \u8fd8\u5728\u505a\u4e00\u4e9bR1\u7684\u590d\u73b0\u7684\u5de5\u4f5c. \u6574\u4e2a\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u4f5c\u6d41\u624d\u662fDeepSeek\u7684\u4e3b\u7ebf, \u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u6765\u5b9e\u73b0AGI\/ASI\u662f\u76ee\u7684, \u81f3\u4e8eMLA\/MoE\/MTP\/FP8\u7b49\u662f\u5728\u5b9e\u73b0\u8fd9\u6761\u8def\u4e0a\u7684\u624b\u6bb5, \u5305\u62ecDeepSeek\u7684app\u672c\u8eab, \u6211\u76f8\u4fe1\u6709\u4f20\u8a00\u8bf4\u6881\u603b\u8fd8\u4e0d\u592a\u60f3\u8981\u8fd9\u51e0\u5343\u4e07\u7684DAU, \u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u771f\u5b9e\u53ef\u4fe1\u7684\u60f3\u6cd5. \u7279\u522b\u662f\u6e23B\u6700\u8fd1\u5728\u505aR1\u590d\u73b0\u7684\u65f6\u5019, \u5bf9\u63a8\u7406\u6027\u80fd\u7684\u74f6\u9888\u53c8\u4f34\u968f\u7740\u63a8\u7406\u4f18\u5316\u7684\u5de5\u4f5c, \u5bf9\u8fd9\u4e00\u70b9\u7406\u89e3\u66f4\u6df1\u523b\u4e86.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u5176\u5b9e\u8c08\u8d77\u5f3a\u5316\u5b66\u4e60\u8fd9\u4e2a\u8bdd\u9898, \u4f3c\u4e4e\u6709\u8bf4\u4e0d\u5b8c\u7684\u6545\u4e8b. \u4ece\u5dee\u4e0d\u591a\u5feb25\u5e74\u524d\u641eOI\u7ade\u8d5b\u7684\u65f6\u5019\u5199\u4e86\u5927\u91cf\u7684\u52a8\u6001\u89c4\u5212\u7b97\u6cd5, \u6700\u540e\u62ff\u5956\u4fdd\u9001\u4e0a\u6d77\u897f\u5357\u67d0\u6280\u6821, \u518d\u5230\u5dee\u4e0d\u591a20\u5e74\u524d\u6bd5\u4e1a\u8bba\u6587\u662f\u57fa\u4e8e\u5143\u80de\u7ec6\u80de\u673a\u548c\u590d\u6742\u7f51\u7edc\u5728\u535a\u5f08\u8bba\u89c6\u89d2\u4e0b\u5bf9\u91d1\u878d\u8d44\u4ea7\u4ef7\u683c\u8fdb\u884c\u5b9a\u4ef7,\u901a\u8fc7Multi-Agent\u7684\u4eff\u771f\u4ea4\u6613\u5b9e\u73b0.\u57287~8\u5e74\u524dCisco\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\u548cSegmentRouting\u6784\u5efa\u7684SDN\u548cSWAN\u7f51\u7edc\u62ff\u4e86CEO\u5927\u5956, \u5e76\u53d1\u5e03\u4e86Cisco Predicatable Network. &nbsp;\u518d\u5230\u4e24\u5e74\u524d\u7528\u4e86\u4e00\u4e2a\u975e\u5e38\u7b80\u5355\u7684\u52a8\u6001\u89c4\u5212\u7b97\u6cd5\u8bbe\u8ba1eRDMA\u7684\u62e5\u585e\u63a7\u5236\u7b97\u6cd5. \u6625\u8282\u671f\u95f4\u91cd\u65b0\u6574\u7406\u4e86\u4e00\u904d\u5f3a\u5316\u5b66\u4e60\u7684\u7b97\u6cd5, \u800c\u6700\u8fd1\u4e00\u5468\u867d\u7136\u5fd9\u7740DS-R1\u63a8\u7406\u8c03\u4f18, \u4e5f\u7a7f\u63d2\u7740\u505a\u4e00\u4e9b\u590d\u73b0R1\u7684\u6a21\u578b\u8bad\u7ec3\u5de5\u4f5c, \u7136\u540e\u8bad\u7ec3\u7684\u65f6\u5019\u53d1\u73b0\u63a8\u7406\u7684\u6548\u7387\u592a\u4f4e, \u53c8\u8981\u52a8\u624b\u53bb\u4f18\u5316trl\u548cvllm, \u8fd9\u4e2a\u94fe\u6761\u5c31\u5b8c\u5168\u7406\u89e3\u4e86.<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u5f88\u591a\u65f6\u5019, \u5927\u5bb6\u5e94\u50cfDeepSeek\u4e00\u6837\u805a\u7126\u57fa\u672c\u529f\uff0c\u5c11\u65e0\u7aef\u731c\u6d4b\uff0c\u5f52\u56e0\u5916\u90e8\u3002\u8fd9\u4e5f\u662f\u6e23B\u8fd8\u5728\u7ee7\u7eed\u505a\u4e00\u4e9b\u5173\u4e8eMoE\u7b97\u6cd5\u548c\u57fa\u7840\u8bbe\u65bd\u534f\u540c\u4f18\u5316\u7684\u7814\u7a76.&nbsp;<span style=\"letter-spacing: 0.04em;text-indent: 2em\">\u4f8b\u5982\u524d\u9762\u6240\u8ff0\u548cMoE\u76f8\u5173\u7684\u5185\u5bb9<\/span><\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u548c\u6570\u5b66\u57fa\u7840\u7b97\u6cd5\u76f8\u5173\u7684, \u662f\u5426\u5b58\u5728\u4e00\u4e9b\u975e\u7ebf\u6027\u7a7a\u95f4\u7684\u9ad8\u6027\u80fd\u5b66\u4e60\u7b97\u6cd5? \u4ee3\u6570\u7ed3\u6784\u4e0a\u7684\u4e00\u4e9b\u7ea6\u675f\u662f\u5426\u53ef\u4ee5\u6784\u6210RL\u7684\u4e00\u4e9bReward\u7279\u5f81? \u901a\u8fc7RL\u4fee\u6539\u6bcf\u4e00\u5c42attention-score\u4e2dsoftmax\u7684temprature\u662f\u5426\u4e5f\u662f\u4e00\u6761\u8def?<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u8fd8\u6709\u5c31\u662f\u7b97\u6cd5\u548c\u57fa\u7840\u8bbe\u65bd\u534f\u540c\u76f8\u5173\u7684<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">&nbsp;,<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u5f53\u7136\u8fd8\u6709\u66f4\u591a\u5bf9\u56fd\u4ea7\u7b97\u529b\u7684\u652f\u6301, \u4f8b\u5982\u82af\u7247\u7684ScaleUP\u548cScaleOut\u4e92\u8054\u548cTensor\u8fd0\u7b97\u7684\u5206\u6790\u7b49<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">&nbsp;,&nbsp;&nbsp;<\/p>\n<p style=\"line-height: 2.1em;letter-spacing: 0.04em;text-indent: 2em;padding-top: 20px;padding-bottom: 8px\">\u5728\u8fd9\u4e2a\u65f6\u4ee3, \u9700\u8981\u4e00\u7fa4\u5b89\u5fc3\u505a\u6280\u672f\u7684\u4eba, \u5927\u5bb6\u52a0\u6cb9.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>ima.copilot\u521a\u51fa\u6765\u7684\u65f6\u5019, \u7528\u4e86\u4e00\u6bb5\u65f6\u95f4. \u9664\u4e86\u6df7\u5143\u81ea\u8eab\u6a21\u578b\u6709\u4e00\u4e9b\u5dee\u8ddd\u4ee5\u5916, \u5bf9\u6574\u4e2a\u516c\u4f17\u53f7\u7684\u5185\u5bb9\u751f\u6001\u4e0a\u7684\u652f\u6301\u662f\u975e\u5e38\u597d\u7684. \u524d\u6bb5\u65f6\u95f4\u4e5f\u6d4b\u8bd5\u4e86\u4e00\u4e0b\u516c\u4f17\u53f7\u540e\u53f0\u7528LLM\u81ea\u52a8\u56de\u590d,\u9664\u4e86\u4e1a\u52a1\u903b\u8f91\u4e0a\u6709\u4e9b\u4ea4\u4e92\u7684\u95ee\u9898(\u4f8b\u5982\u6b63\u5e38\u7684\u7559\u8a00\u4ea4\u6d41\u548c\u641c\u7d22\u6574\u7406\u516c\u4f17\u53f7\u5185\u5bb9\u7684\u533a\u5206),\u5176\u5b9e\u6574\u7406\u7684\u5185\u5bb9\u8fd8\u662f\u53ef\u8bfb\u6027\u5f88\u9ad8\u7684, \u5f53\u7136\u4e5f\u6709\u56e0\u4e3a\u57fa\u7840\u6a21\u578b\u7684\u95ee\u9898\u5bfc\u81f4\u7684 \u6700\u8fd1\u5fae\u4fe1\u5f00\u59cb\u7070\u5ea6DeepSeek\u4e86, \u57282C\u7684\u5e02\u573a\u4e0a\u5c06\u4f1a\u8fce\u6765\u66f4\u591a\u7684\u53d8\u6570, \u7279\u522b\u662f\u5bf9\u5b57\u8282\u8c46\u5305\u7684\u751f\u6001\u4e0a\u7684\u5f71\u54cd. \u6bd5\u7adf\u817e\u8baf\u5fae\u4fe1\u7684DAU\u662f\u63a5\u8fd110\u4ebf\u7ea7\u7684. \u53e6\u4e00\u65b9\u9762\u6700\u8fd1\u767e\u5ea6\u548c\u706b\u5c71\u7684\u8868\u6001(\u53c2\u8003\u5feb\u79d1\u6280\u7684\u62a5\u9053[1]) \u767e\u5ea6\u667a\u80fd\u4e91\u4e8b\u4e1a\u7fa4\u603b\u88c1\u6c88\u6296\u5728\u5168\u5458\u4f1a\u4e0a\u8868\u793a\uff0c\u56fd\u5185\u5927\u6a21\u578b\u53bb\u5e74\u7684\u201c\u6076\u610f\u201d\u4ef7\u683c\u6218\uff0c\u5bfc\u81f4\u884c\u4e1a\u6574\u4f53\u521b\u6536\u76f8\u8f83\u4e8e\u56fd\u5916\u5dee\u4e86\u591a\u4e2a\u6570\u91cf\u7ea7\u3002 \u5b57\u8282\u8df3\u52a8\u65d7\u4e0b\u706b\u5c71\u5f15\u64ce\u603b\u88c1\u8c2d\u5f85\u901a\u8fc7\u670b\u53cb\u5708\u56de\u5e94\u79f0\uff0c\u5927\u6a21\u578b\u964d\u4ef7\u662f\u901a\u8fc7\u6280\u672f\u521b\u65b0\u5b9e\u73b0\u7684\uff0c\u5927\u5bb6\u5e94\u50cfDeepSeek\u4e00\u6837\u805a\u7126\u57fa\u672c\u529f\uff0c\u5c11\u65e0\u7aef\u731c\u6d4b\uff0c\u5f52\u56e0\u5916\u90e8\u3002 \u8c2d\u5f85\u6307\u51fa\uff0c\u706b\u5c71\u5f15\u64ce\u7684\u8c46\u53051.5Pro\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6210\u672c\u548c\u63a8\u7406\u6210\u672c\u5747\u4f4e\u4e8eDeepSeek V3\uff0c\u66f4\u662f\u8fdc\u4f4e\u4e8e\u56fd\u5185\u5176\u4ed6\u6a21\u578b\uff0c\u5728\u5f53\u524d\u4ef7\u683c\u4e0b\u6709\u975e\u5e38\u4e0d\u9519\u7684\u6bdb\u5229\u3002\u4ed6\u8fdb\u4e00\u6b65\u89e3\u91ca\uff1a\u201c\u56fd\u5185\u5916\u7684\u5382\u5546\u90fd\u5728\u4f9d\u9760\u6280\u672f\u521b\u65b0\uff0c\u964d\u4f4e\u6a21\u578b\u4ef7\u683c\u3002\u6211\u4eec\u4e5f\u53ea\u662f\u5b9e\u73b0\u4e86Gemini 2.0 Flash\u7684\u4ef7\u683c\u6c34\u5e73\u800c\u5df2\uff0c\u8fd9\u4e2a\u4ef7\u683c\u5b8c\u5168\u662f\u4f9d\u8d56\u6280\u672f\u8fdb\u6b65\u53ef\u4ee5\u505a\u5230\u7684\u3002\u201d \u5176\u5b9e\u5f88\u591a\u65f6\u5019\u6210\u672c\u7684\u4f30\u8ba1\u7684\u5206\u6b67, \u672c\u8d28\u662f\u6280\u672f\u4e0a\u7684\u5dee\u8ddd. \u4f8b\u5982\u5c24\u6d0b\u8001\u5e08\u4f30\u8ba1\u7684\u6570\u636e, \u548cDeepSeek-V3\u7684\u8bba\u6587\u5b9e\u73b0\u7684PD\u5206\u79bb+EP\u5e76\u884c\u6027\u80fd\u5dee\u8ddd\u8d85\u8fc710\u500d\u4ee5\u4e0a. \u672c\u8d28\u4e0a\u7684\u6210\u672c\u5dee\u5f02\u5c31\u662f\u62ff\u4e00\u4e9b\u5f00\u6e90\u793e\u533a\u7b80\u5355\u7684TP\/PP\u5e76\u884c\u7684\u7ed3\u679c\u6765\u4f30\u8ba1\u662f\u6709\u5f88\u5927\u7684\u5dee\u8ddd\u7684. \u7279\u522b\u662fGoogle Gemini 2.0 Flash\u7684\u4ef7\u683c\u6c34\u5e73\u6765\u770b, \u6280\u672f\u4e0a\u8fd8\u6709\u66f4\u591a\u7684\u4f18\u5316\u7a7a\u95f4. \u4f8b\u5982\u6628\u5929\u63d0\u5230\u7684\u4e00\u7bc7\u6587\u7ae0\u5bf9MOE\u7684\u6027\u80fd\u5206\u6790. \u8fd9\u91cc\u7b80\u5355\u7684\u505a\u4e00\u4e2aRoofline\u7684\u5206\u6790, \u4ece\u7b97\u529b\u4e0a\u8bb2DeepSeek-V3\/R1\u6a21\u578b\u7684\u7b97\u529b\u9700\u6c42\u662f\u76f8\u5bf9\u8f83\u5c0f\u7684, \u74f6\u9888\u4e3b\u8981\u662f\u5728\u8bbf\u5b58\u548cAll2All\u7684\u901a\u4fe1\u4e0a\u4ee5\u53ca\u5982\u4f55\u89e3\u51b3\u63a8\u7406\u65f6\u7684\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861\u4e0a. \u4f8b\u5982\u534e\u4e3a\u6607\u817e\u63d0\u5230\u7684:\u201c\u901a\u8fc7EP\u6df7\u5408\u5e76\u884c\u7b97\u6cd5, \u901a\u4fe1\u4f18\u5316\u6027\u80fd\u63d0\u534730%+, \u8bbf\u5b58\u6027\u80fd\u63d0\u534720%+, \u4ece\u800c\u964d\u4f4e\u4e13\u5bb6\u4e0d\u5747\u8861\u5ea6, \u63a8\u7406\u541e\u5410\u6027\u80fd\u63d0\u534720%~35%\u201d \u53e6\u4e00\u65b9\u9762\u4ece\u8881\u8fdb\u8f89\u8001\u5e08\u7684\u4e00\u6bb5\u8bdd\u53ef\u4ee5\u77e5\u9053, \u6881\u603b\u4e3a\u5565\u8981\u63a8\u8350\u6027\u80fd\u6700\u597d\u9700\u898180\u53f0, \u4e3b\u8981\u662f\u901a\u8fc7EP\u5e76\u884c\u83b7\u5f97\u66f4\u597d\u7684Data Locality. \u4eceNetwork-Bound\u4e0a\u6765\u770b, \u6211\u4eec\u4ee5\u5355\u4e2aToken 7168B\u6765\u8ba1\u7b97, 400Gbps\u7f51\u7edc\u673a\u95f4\u4e92\u8054\u7f51\u7edc\u537350GB\/s, \u7b80\u5355\u7684\u505a\u4e00\u4e2a\u4e0a\u9650\u4f30\u8ba1. \u6a21\u578b\u9700\u8981\u4f20\u8f9360\u5c42, \u6bcf\u4e2aToken\u9700\u89818\u4e2aRouted Expert\u548c\u4e00\u4e2aShared Expert, \u5373\u5355\u4e2aToken\u9700\u8981 7168 x 9 x60 ~ [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53839","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53839","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53839"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53839\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53839"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53839"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53839"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}