{"id":3489,"date":"2025-03-03T15:36:24","date_gmt":"2025-03-03T07:36:24","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=3489"},"modified":"2025-03-03T15:37:54","modified_gmt":"2025-03-03T07:37:54","slug":"deepseek%e7%aa%81%e8%a2%ad%e5%85%ac%e5%b8%83%e6%88%90%e6%9c%ac%e5%88%a9%e6%b6%a6%e7%8e%87%ef%bc%9a545","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2025\/03\/03\/deepseek%e7%aa%81%e8%a2%ad%e5%85%ac%e5%b8%83%e6%88%90%e6%9c%ac%e5%88%a9%e6%b6%a6%e7%8e%87%ef%bc%9a545\/","title":{"rendered":"DeepSeek\u7a81\u88ad\u516c\u5e03\u6210\u672c\u5229\u6da6\u7387\uff1a545%"},"content":{"rendered":"\t\t<div data-elementor-type=\"wp-post\" data-elementor-id=\"3489\" class=\"elementor elementor-3489\">\n\t\t\t\t<div class=\"elementor-element elementor-element-c7b071d e-flex e-con-boxed e-con e-parent\" data-id=\"c7b071d\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t<div class=\"elementor-element elementor-element-fcf1fbf e-flex e-con-boxed e-con e-parent\" data-id=\"fcf1fbf\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-9cccd7e elementor-widget elementor-widget-text-editor\" data-id=\"9cccd7e\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<p>\u4e94\u8fde\u5f00\u6e90\u540e\uff0cDeepSeek\u8fd8\u6709One More Thing\uff01<\/p><p>\u5c31\u5728\u521a\u521a\uff0cDeepSeek\u5b98\u65b9\u4eb2\u81ea\u63ed\u79d8\u4e86\u00a0<strong>DeepSeek-V3\/R1\u63a8\u7406\u7cfb\u7edf<\/strong>\u3002<\/p><p><img fetchpriority=\"high\" decoding=\"async\" src=\"https:\/\/q1.itc.cn\/q_70\/images03\/20250301\/7131dea5fa664497ba0655d44503541e.png\" width=\"640\" height=\"577\" \/><\/p><p>\u91cd\u70b9\u5305\u62ec\uff0c\u4f18\u5316\u541e\u5410\u91cf\u548c\u5ef6\u8fdf\u7684\u65b9\u6cd5\uff1a<\/p><ul><li><p>\u8de8\u8282\u70b9EP\u9a71\u52a8\u7684\u6279\u91cf\u6269\u5c55<\/p><\/li><li><p>\u8ba1\u7b97\u4e0e\u901a\u4fe1\u91cd\u53e0<\/p><\/li><li><p>\u8d1f\u8f7d\u5747\u8861<\/p><\/li><\/ul><p>\u8fd8\u516c\u5e03\u4e86DeepSeek\u7684\u5728\u7ebf\u670d\u52a1\u6570\u636e\u7edf\u8ba1\uff1a<\/p><ul><li><p>\u6bcf\u4e2aH800\u8282\u70b9\u6bcf\u79d2\u670973.7k\/14.8k\u4e2a\u8f93\u5165\/\u8f93\u51fatoken<\/p><\/li><li><p>\u6210\u672c\u5229\u6da6\u7387545%<\/p><\/li><\/ul><p><img decoding=\"async\" src=\"https:\/\/q2.itc.cn\/q_70\/images03\/20250301\/5794ea71ce094a17862c6e3d4c9ecae1.gif\" width=\"360\" height=\"202\" \/><\/p><p>\u66f4\u591a\u7ec6\u8282\uff0c\u4e00\u8d77\u6765\u770b\u5b98\u65b9\u539f\u6587\u2193<\/p><p>\u66f4\u5927\u7684\u541e\u5410\uff0c\u66f4\u4f4e\u7684\u5ef6\u8fdf<\/p><p>DeepSeek-V3\/R1\u63a8\u7406\u7cfb\u7edf\u7684\u4f18\u5316\u76ee\u6807\u662f\uff1a\u66f4\u5927\u7684\u541e\u5410\uff0c\u66f4\u4f4e\u7684\u5ef6\u8fdf\u3002<\/p><p>\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e24\u4e2a\u76ee\u6807\uff0c\u6211\u4eec\u7684\u65b9\u6848\u662f\u4f7f\u7528\u5927\u89c4\u6a21\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\uff08ExpertParallelism\/EP\uff09\u3002<\/p><p>\u9996\u5148EP\u4f7f\u5f97batch size\u5927\u5927\u589e\u52a0\uff0c\u4ece\u800c\u63d0\u9ad8GPU\u77e9\u9635\u4e58\u6cd5\u7684\u6548\u7387\uff0c\u63d0\u9ad8\u541e\u5410\u3002\u5176\u6b21EP\u4f7f\u5f97\u4e13\u5bb6\u5206\u6563\u5728\u4e0d\u540c\u7684GPU\u4e0a\uff0c\u6bcf\u4e2aGPU\u53ea\u9700\u8981\u8ba1\u7b97\u5f88\u5c11\u7684\u4e13\u5bb6\uff08\u56e0\u6b64\u66f4\u5c11\u7684\u8bbf\u5b58\u9700\u6c42\uff09\uff0c\u4ece\u800c\u964d\u4f4e\u5ef6\u8fdf\u3002<\/p><p>\u4f46EP\u540c\u65f6\u4e5f\u589e\u52a0\u4e86\u7cfb\u7edf\u7684\u590d\u6742\u6027\u3002\u590d\u6742\u6027\u4e3b\u8981\u4f53\u73b0\u5728\u4e24\u4e2a\u65b9\u9762\uff1a<\/p><ul><li><p>EP\u5f15\u5165\u8de8\u8282\u70b9\u7684\u4f20\u8f93\u3002\u4e3a\u4e86\u4f18\u5316\u541e\u5410\uff0c\u9700\u8981\u8bbe\u8ba1\u5408\u9002\u7684\u8ba1\u7b97\u6d41\u7a0b\u4f7f\u5f97\u4f20\u8f93\u548c\u8ba1\u7b97\u53ef\u4ee5\u540c\u6b65\u8fdb\u884c\u3002<\/p><\/li><li><p>EP\u6d89\u53ca\u591a\u4e2a\u8282\u70b9\uff0c\u56e0\u6b64\u5929\u7136\u9700\u8981Data Parallelism\uff08DP\uff09\uff0c\u4e0d\u540c\u7684DP\u4e4b\u95f4\u9700\u8981\u8fdb\u884c\u8d1f\u8f7d\u5747\u8861\u3002<\/p><\/li><\/ul><p>\u56e0\u6b64\uff0c\u672c\u6587\u7684\u4e3b\u8981\u5185\u5bb9\u662f\u5982\u4f55\u4f7f\u7528EP\u589e\u5927batch size\uff0c\u5982\u4f55\u9690\u85cf\u4f20\u8f93\u7684\u8017\u65f6\uff0c\u5982\u4f55\u8fdb\u884c\u8d1f\u8f7d\u5747\u8861\u3002<\/p><p>\u5927\u89c4\u6a21\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallelism\/EP\uff09<\/p><p>\u7531\u4e8eDeepSeek-V3\/R1\u7684\u4e13\u5bb6\u6570\u91cf\u4f17\u591a\uff0c\u5e76\u4e14\u6bcf\u5c42256\u4e2a\u4e13\u5bb6\u4e2d\u4ec5\u6fc0\u6d3b\u5176\u4e2d8\u4e2a\u3002\u6a21\u578b\u7684\u9ad8\u5ea6\u7a00\u758f\u6027\u51b3\u5b9a\u4e86\u6211\u4eec\u5fc5\u987b\u91c7\u7528\u5f88\u5927\u7684overall batch size\uff0c\u624d\u80fd\u7ed9\u6bcf\u4e2a\u4e13\u5bb6\u63d0\u4f9b\u8db3\u591f\u7684expert batch size\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u5927\u7684\u541e\u5410\u3001\u66f4\u4f4e\u7684\u5ef6\u65f6\u3002\u9700\u8981\u5927\u89c4\u6a21\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallelism\/EP\uff09\u3002<\/p><p>\u6211\u4eec\u91c7\u7528\u591a\u673a\u591a\u5361\u95f4\u7684\u4e13\u5bb6\u5e76\u884c\u7b56\u7565\u6765\u8fbe\u5230\u4ee5\u4e0b\u76ee\u7684\uff1a<\/p><ul><li><p>Prefill\uff1a\u8def\u7531\u4e13\u5bb6EP32\u3001MLA\u548c\u5171\u4eab\u4e13\u5bb6DP32\uff0c\u4e00\u4e2a\u90e8\u7f72\u5355\u5143\u662f4\u8282\u70b9\uff0c32\u4e2a\u5197\u4f59\u8def\u7531\u4e13\u5bb6\uff0c\u6bcf\u5f20\u53619\u4e2a\u8def\u7531\u4e13\u5bb6\u548c1\u4e2a\u5171\u4eab\u4e13\u5bb6<\/p><\/li><li><p>Decode\uff1a\u8def\u7531\u4e13\u5bb6EP144\u3001MLA\u548c\u5171\u4eab\u4e13\u5bb6DP144\uff0c\u4e00\u4e2a\u90e8\u7f72\u5355\u5143\u662f18\u8282\u70b9\uff0c32\u4e2a\u5197\u4f59\u8def\u7531\u4e13\u5bb6\uff0c\u6bcf\u5f20\u53612\u4e2a\u8def\u7531\u4e13\u5bb6\u548c1\u4e2a\u5171\u4eab\u4e13\u5bb6<\/p><\/li><\/ul><p>\u8ba1\u7b97\u901a\u4fe1\u91cd\u53e0<\/p><p>\u591a\u673a\u591a\u5361\u7684\u4e13\u5bb6\u5e76\u884c\u4f1a\u5f15\u5165\u6bd4\u8f83\u5927\u7684\u901a\u4fe1\u5f00\u9500\uff0c\u6240\u4ee5\u6211\u4eec\u4f7f\u7528\u4e86\u53ccbatch\u91cd\u53e0\u6765\u63a9\u76d6\u901a\u4fe1\u5f00\u9500\uff0c\u63d0\u9ad8\u6574\u4f53\u541e\u5410\u3002<\/p><p>\u5bf9\u4e8eprefill\u9636\u6bb5\uff0c\u4e24\u4e2abatch\u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u4ea4\u9519\u8fdb\u884c\uff0c\u4e00\u4e2abatch\u5728\u8fdb\u884c\u8ba1\u7b97\u7684\u65f6\u5019\u53ef\u4ee5\u53bb\u63a9\u76d6\u53e6\u4e00\u4e2abatch\u7684\u901a\u4fe1\u5f00\u9500\uff1b<\/p><p><strong>\u25b3<\/strong>Prefill\u9636\u6bb5\u7684\u53ccbatch\u91cd\u53e0<\/p><p>\u5bf9\u4e8edecode\u9636\u6bb5\uff0c\u4e0d\u540c\u9636\u6bb5\u7684\u6267\u884c\u65f6\u95f4\u6709\u6240\u5dee\u522b\uff0c\u6240\u4ee5\u6211\u4eec\u628aattention\u90e8\u5206\u62c6\u6210\u4e86\u4e24\u4e2astage\uff0c\u5171\u8ba15\u4e2astage\u7684\u6d41\u6c34\u7ebf\u6765\u5b9e\u73b0\u8ba1\u7b97\u548c\u901a\u4fe1\u7684\u91cd\u53e0\u3002<\/p><p><strong>\u25b3<\/strong>Decode\u9636\u6bb5\u7684\u53ccbatch\u91cd\u53e0<\/p><p>\u5173\u4e8e\u66f4\u591a\u53ccbatch\u91cd\u53e0\u7684\u7ec6\u8282\uff0c\u53ef\u4ee5\u53c2\u8003\u6211\u4eec\u7684profiling\u6570\u636e\u7684GitHub\u4ed3\u5e93\uff1ahttps:\/\/github.com\/deepseek-ai\/profile-data\u3002<\/p><p>\u5c3d\u53ef\u80fd\u5730\u8d1f\u8f7d\u5747\u8861<\/p><p>\u7531\u4e8e\u91c7\u7528\u4e86\u5f88\u5927\u89c4\u6a21\u7684\u5e76\u884c\uff08\u5305\u62ec\u6570\u636e\u5e76\u884c\u548c\u4e13\u5bb6\u5e76\u884c\uff09\uff0c\u5982\u679c\u67d0\u4e2aGPU\u7684\u8ba1\u7b97\u6216\u901a\u4fe1\u8d1f\u8f7d\u8fc7\u91cd\uff0c\u5c06\u6210\u4e3a\u6027\u80fd\u74f6\u9888\uff0c\u62d6\u6162\u6574\u4e2a\u7cfb\u7edf\uff1b\u540c\u65f6\u5176\u4ed6GPU\u56e0\u4e3a\u7b49\u5f85\u800c\u7a7a\u8f6c\uff0c\u9020\u6210\u6574\u4f53\u5229\u7528\u7387\u4e0b\u964d\u3002\u56e0\u6b64\u6211\u4eec\u9700\u8981\u5c3d\u53ef\u80fd\u5730\u4e3a\u6bcf\u4e2aGPU\u5206\u914d\u5747\u8861\u7684\u8ba1\u7b97\u8d1f\u8f7d\u3001\u901a\u4fe1\u8d1f\u8f7d\u3002<\/p><ul><li><p>Prefill Load Balancer<\/p><ul><li><p>\u6838\u5fc3\u95ee\u9898\uff1a\u4e0d\u540c\u6570\u636e\u5e76\u884c\uff08DP\uff09\u5b9e\u4f8b\u4e0a\u7684\u8bf7\u6c42\u4e2a\u6570\u3001\u957f\u5ea6\u4e0d\u540c\uff0c\u5bfc\u81f4core-attention\u8ba1\u7b97\u91cf\u3001dispatch\u53d1\u9001\u91cf\u4e5f\u4e0d\u540c<\/p><\/li><li><p>\u4f18\u5316\u76ee\u6807\uff1a\u5404GPU\u7684\u8ba1\u7b97\u91cf\u5c3d\u91cf\u76f8\u540c\uff08core-attention\u8ba1\u7b97\u8d1f\u8f7d\u5747\u8861\uff09\u3001\u8f93\u5165\u7684token\u6570\u91cf\u4e5f\u5c3d\u91cf\u76f8\u540c\uff08dispatch\u53d1\u9001\u91cf\u8d1f\u8f7d\u5747\u8861\uff09\uff0c\u907f\u514d\u90e8\u5206GPU\u5904\u7406\u65f6\u95f4\u8fc7\u957f<\/p><\/li><\/ul><\/li><li><p>Decode Load Balancer<\/p><ul><li><p>\u6838\u5fc3\u95ee\u9898\uff1a\u4e0d\u540c\u6570\u636e\u5e76\u884c\uff08DP\uff09\u5b9e\u4f8b\u4e0a\u7684\u8bf7\u6c42\u6570\u91cf\u3001\u957f\u5ea6\u4e0d\u540c\uff0c\u5bfc\u81f4core-attention\u8ba1\u7b97\u91cf\uff08\u4e0eKVCache\u5360\u7528\u91cf\u76f8\u5173\uff09\u3001dispatch\u53d1\u9001\u91cf\u4e0d\u540c<\/p><\/li><li><p>\u4f18\u5316\u76ee\u6807\uff1a\u5404GPU\u7684KVCache\u5360\u7528\u91cf\u5c3d\u91cf\u76f8\u540c\uff08core-attention\u8ba1\u7b97\u8d1f\u8f7d\u5747\u8861\uff09\u3001\u8bf7\u6c42\u6570\u91cf\u5c3d\u91cf\u76f8\u540c\uff08dispatch\u53d1\u9001\u91cf\u8d1f\u8f7d\u5747\u8861\uff09<\/p><\/li><\/ul><\/li><li><p>Expert-Parallel Load Balancer<\/p><ul><li><p>\u6838\u5fc3\u95ee\u9898\uff1a\u5bf9\u4e8e\u7ed9\u5b9aMoE\u6a21\u578b\uff0c\u5b58\u5728\u4e00\u4e9b\u5929\u7136\u7684\u9ad8\u8d1f\u8f7d\u4e13\u5bb6\uff08expert\uff09\uff0c\u5bfc\u81f4\u4e0d\u540cGPU\u7684\u4e13\u5bb6\u8ba1\u7b97\u8d1f\u8f7d\u4e0d\u5747\u8861<\/p><\/li><li><p>\u4f18\u5316\u76ee\u6807\uff1a\u6bcf\u4e2aGPU\u4e0a\u7684\u4e13\u5bb6\u8ba1\u7b97\u91cf\u5747\u8861\uff08\u5373\u6700\u5c0f\u5316\u6240\u6709GPU\u7684dispatch\u63a5\u6536\u91cf\u7684\u6700\u5927\u503c\uff09<\/p><\/li><\/ul><\/li><\/ul><p>\u53c2\u8003\u67b6\u6784\u56fe<\/p><p><img decoding=\"async\" src=\"https:\/\/q0.itc.cn\/q_70\/images03\/20250301\/09eb640df4c24dcb845a2e407c6311e0.jpeg\" width=\"640\" height=\"286\" \/><\/p><p>\u7ebf\u4e0a\u7cfb\u7edf\u7684\u5b9e\u9645\u7edf\u8ba1\u6570\u636e<\/p><p>DeepSeekV3\u548cR1\u7684\u6240\u6709\u670d\u52a1\u5747\u4f7f\u7528H800 GPU\uff0c\u4f7f\u7528\u548c\u8bad\u7ec3\u4e00\u81f4\u7684\u7cbe\u5ea6\uff0c\u5373\u77e9\u9635\u8ba1\u7b97\u548cdispatch\u4f20\u8f93\u91c7\u7528\u548c\u8bad\u7ec3\u4e00\u81f4\u7684FP8\u683c\u5f0f\uff0ccore-attention\u8ba1\u7b97\u548ccombine\u4f20\u8f93\u91c7\u7528\u548c\u8bad\u7ec3\u4e00\u81f4\u7684BF16\uff0c\u6700\u5927\u7a0b\u5ea6\u4fdd\u8bc1\u4e86\u670d\u52a1\u6548\u679c\u3002<\/p><p><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/q5.itc.cn\/q_70\/images03\/20250301\/14aa55825c4d4bbb82bb04bba9d23a59.jpeg\" width=\"640\" height=\"220\" \/><\/p><p>\u572824\u5c0f\u65f6\u7edf\u8ba1\u65f6\u6bb5\u5185\uff0cDeepSeekV3\u548cR1\uff1a<\/p><p>\u8f93\u5165token\u603b\u6570\u4e3a608B\uff0c\u5176\u4e2d342B tokens\uff0856.3%\uff09\u547d\u4e2dKVCache\u786c\u76d8\u7f13\u5b58\u3002<\/p><p>\u8f93\u51fatoken\u603b\u6570\u4e3a168B\u3002\u5e73\u5747\u8f93\u51fa\u901f\u7387\u4e3a20~22tps\uff0c\u5e73\u5747\u6bcf\u8f93\u51fa\u4e00\u4e2atoken\u7684KVCache\u957f\u5ea6\u662f4989\u3002<\/p><p>\u5e73\u5747\u6bcf\u53f0H800\u7684\u541e\u5410\u91cf\u4e3a\uff1a\u5bf9\u4e8eprefill\u4efb\u52a1\uff0c\u8f93\u5165\u541e\u5410\u7ea673.7k tokens\/s\uff08\u542b\u7f13\u5b58\u547d\u4e2d\uff09\uff1b\u5bf9\u4e8edecode\u4efb\u52a1\uff0c\u8f93\u51fa\u541e\u5410\u7ea614.8k tokens\/s\u3002<\/p><blockquote><p>\u5f53\u7136\u6211\u4eec\u5b9e\u9645\u4e0a\u6ca1\u6709\u8fd9\u4e48\u591a\u6536\u5165\uff0c\u56e0\u4e3aV3\u7684\u5b9a\u4ef7\u66f4\u4f4e\uff0c\u540c\u65f6\u6536\u8d39\u670d\u52a1\u53ea\u5360\u4e86\u4e00\u90e8\u5206\uff0c\u53e6\u5916\u591c\u95f4\u8fd8\u4f1a\u6709\u6298\u6263\u3002<\/p><\/blockquote>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"<p>\u9c7c\u7f8a \u7f16\u8f91\u6574\u7406<\/p>\n","protected":false},"author":1,"featured_media":3491,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3489","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3489","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=3489"}],"version-history":[{"count":7,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3489\/revisions"}],"predecessor-version":[{"id":3497,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3489\/revisions\/3497"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/3491"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=3489"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=3489"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=3489"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}