{"id":24775,"date":"2025-03-03T09:58:05","date_gmt":"2025-03-03T01:58:05","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=24775"},"modified":"2025-03-03T09:58:05","modified_gmt":"2025-03-03T01:58:05","slug":"%e5%ae%98%e6%96%b9%e8%af%a6%e8%a7%a3-deepseek-v3-r1-%e6%8e%a8%e7%90%86%e7%b3%bb%e7%bb%9f%ef%bc%9a%e7%90%86%e8%ae%ba%e5%88%a9%e6%b6%a6%e7%8e%87%e8%be%be-545","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/03\/03\/24775\/","title":{"rendered":"\u5b98\u65b9\u8be6\u89e3 DeepSeek-V3 \/ R1 \u63a8\u7406\u7cfb\u7edf\uff1a\u7406\u8bba\u5229\u6da6\u7387\u8fbe 545%"},"content":{"rendered":"
\"\"<\/div>\n
<\/div>\n
DeepSeek \u5b98\u65b9\u4eca\u65e5\u5728\u77e5\u4e4e\u53d1\u5e03\u300aDeepSeek-V3 \/ R1 \u63a8\u7406\u7cfb\u7edf\u6982\u89c8\u300b\u4e00\u6587\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528\u5927\u89c4\u6a21\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c(Expert Parallelism \/ EP)\u6765\u589e\u5927 batch size\uff0c\u5982\u4f55\u9690\u85cf\u4f20\u8f93\u7684\u8017\u65f6\uff0c\u5982\u4f55\u8fdb\u884c\u8d1f\u8f7d\u5747\u8861\u3002<\/div>\n
\n
\n

3 \u6708 1 \u65e5\u6d88\u606f\uff0cDeepSeek \u5b98\u65b9\u4eca\u65e5\u5728\u77e5\u4e4e\u53d1\u5e03\u300aDeepSeek-V3 \/ R1 \u63a8\u7406\u7cfb\u7edf\u6982\u89c8\u300b\u4e00\u6587\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528\u5927\u89c4\u6a21\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c(Expert Parallelism \/ EP)\u6765\u589e\u5927 batch size\uff0c\u5982\u4f55\u9690\u85cf\u4f20\u8f93\u7684\u8017\u65f6\uff0c\u5982\u4f55\u8fdb\u884c\u8d1f\u8f7d\u5747\u8861\u3002<\/p>\n

\u5b98\u65b9\u8868\u793a\uff0cDeepSeek-V3 \/ R1 \u63a8\u7406\u7cfb\u7edf\u7684\u4f18\u5316\u76ee\u6807\u662f\u66f4\u5927\u7684\u541e\u5410\uff0c\u66f4\u4f4e\u7684\u5ef6\u8fdf\u3002<\/p>\n

IT\u4e4b\u5bb6\u9644 DeepSeek \u63d0\u51fa\u7684\u65b9\u6848\u5982\u4e0b\uff1a<\/h3>\n

\u5927\u89c4\u6a21\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c(Expert Parallelism \/ EP)<\/h4>\n

\u7531\u4e8e DeepSeek-V3 \/ R1 \u7684\u4e13\u5bb6\u6570\u91cf\u4f17\u591a\uff0c\u5e76\u4e14\u6bcf\u5c42 256 \u4e2a\u4e13\u5bb6\u4e2d\u4ec5\u6fc0\u6d3b\u5176\u4e2d 8 \u4e2a\u3002\u6a21\u578b\u7684\u9ad8\u5ea6\u7a00\u758f\u6027\u51b3\u5b9a\u4e86 DeepSeek \u5fc5\u987b\u91c7\u7528\u5f88\u5927\u7684 overall batch size\uff0c\u624d\u80fd\u7ed9\u6bcf\u4e2a\u4e13\u5bb6\u63d0\u4f9b\u8db3\u591f\u7684 expert 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(Expert Parallelism \/ EP)\u3002<\/p>\n

DeepSeek \u91c7\u7528\u591a\u673a\u591a\u5361\u95f4\u7684\u4e13\u5bb6\u5e76\u884c\u7b56\u7565\u6765\u8fbe\u5230\u4ee5\u4e0b\u76ee\u7684\uff1a<\/p>\n

    \n
  • Prefill\uff1a\u8def\u7531\u4e13\u5bb6 EP32\u3001MLA \u548c\u5171\u4eab\u4e13\u5bb6 DP32\uff0c\u4e00\u4e2a\u90e8\u7f72\u5355\u5143\u662f 4 \u8282\u70b9\uff0c32 \u4e2a\u5197\u4f59\u8def\u7531\u4e13\u5bb6\uff0c\u6bcf\u5f20\u5361 9 \u4e2a\u8def\u7531\u4e13\u5bb6\u548c 1 \u4e2a\u5171\u4eab\u4e13\u5bb6<\/li>\n
  • Decode\uff1a\u8def\u7531\u4e13\u5bb6 EP144\u3001MLA \u548c\u5171\u4eab\u4e13\u5bb6 DP144\uff0c\u4e00\u4e2a\u90e8\u7f72\u5355\u5143\u662f 18 \u8282\u70b9\uff0c32 \u4e2a\u5197\u4f59\u8def\u7531\u4e13\u5bb6\uff0c\u6bcf\u5f20\u5361 2 \u4e2a\u8def\u7531\u4e13\u5bb6\u548c 1 \u4e2a\u5171\u4eab\u4e13\u5bb6<\/li>\n<\/ul>\n

    \u8ba1\u7b97\u901a\u4fe1\u91cd\u53e0<\/h4>\n

    \u591a\u673a\u591a\u5361\u7684\u4e13\u5bb6\u5e76\u884c\u4f1a\u5f15\u5165\u6bd4\u8f83\u5927\u7684\u901a\u4fe1\u5f00\u9500\uff0c\u56e0\u6b64\u4f7f\u7528\u53cc batch \u91cd\u53e0\u6765\u63a9\u76d6\u901a\u4fe1\u5f00\u9500\uff0c\u63d0\u9ad8\u6574\u4f53\u541e\u5410\u3002<\/p>\n

    \u5bf9\u4e8e prefill \u9636\u6bb5\uff0c\u4e24\u4e2a batch \u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u4ea4\u9519\u8fdb\u884c\uff0c\u4e00\u4e2a batch \u5728\u8fdb\u884c\u8ba1\u7b97\u7684\u65f6\u5019\u53ef\u4ee5\u53bb\u63a9\u76d6\u53e6\u4e00\u4e2a batch \u7684\u901a\u4fe1\u5f00\u9500;<\/p>\n

    <\/p>\n

    Prefill \u9636\u6bb5\u7684\u53cc batch \u91cd\u53e0<\/p>\n

    \u5bf9\u4e8e decode \u9636\u6bb5\uff0c\u4e0d\u540c\u9636\u6bb5\u7684\u6267\u884c\u65f6\u95f4\u6709\u6240\u5dee\u522b\uff0c\u56e0\u6b64\u5c06 attention \u90e8\u5206\u62c6\u6210\u4e86\u4e24\u4e2a stage\uff0c\u5171\u8ba1 5 \u4e2a stage \u7684\u6d41\u6c34\u7ebf\u6765\u5b9e\u73b0\u8ba1\u7b97\u548c\u901a\u4fe1\u7684\u91cd\u53e0\u3002<\/p>\n

    <\/p>\n

    Decode \u9636\u6bb5\u7684\u53cc batch \u91cd\u53e0<\/p>\n

    \u5173\u4e8e\u66f4\u591a\u53cc batch \u91cd\u53e0\u7684\u7ec6\u8282\uff0c\u53ef\u53c2\u8003 profiling \u6570\u636e GitHub \u4ed3\u5e93\uff1ahttps:\/\/github.com\/deepseek-ai\/profile-data\u3002<\/p>\n

    \u5c3d\u53ef\u80fd\u5730\u8d1f\u8f7d\u5747\u8861<\/h4>\n

    \u7531\u4e8e\u91c7\u7528\u4e86\u5f88\u5927\u89c4\u6a21\u7684\u5e76\u884c(\u5305\u62ec\u6570\u636e\u5e76\u884c\u548c\u4e13\u5bb6\u5e76\u884c)\uff0c\u5982\u679c\u67d0\u4e2a GPU \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;\u540c\u65f6\u5176\u4ed6 GPU \u56e0\u4e3a\u7b49\u5f85\u800c\u7a7a\u8f6c\uff0c\u9020\u6210\u6574\u4f53\u5229\u7528\u7387\u4e0b\u964d\u3002\u56e0\u6b64\u9700\u5c3d\u53ef\u80fd\u4e3a\u6bcf\u4e2a GPU \u5206\u914d\u5747\u8861\u7684\u8ba1\u7b97\u8d1f\u8f7d\u3001\u901a\u4fe1\u8d1f\u8f7d\u3002<\/p>\n

    Prefill Load Balancer<\/strong><\/p>\n

      \n
    • \u6838\u5fc3\u95ee\u9898\uff1a\u4e0d\u540c\u6570\u636e\u5e76\u884c(DP)\u5b9e\u4f8b\u4e0a\u7684\u8bf7\u6c42\u4e2a\u6570\u3001\u957f\u5ea6\u4e0d\u540c\uff0c\u5bfc\u81f4 core-attention \u8ba1\u7b97\u91cf\u3001dispatch \u53d1\u9001\u91cf\u4e5f\u4e0d\u540c<\/li>\n
    • \u4f18\u5316\u76ee\u6807\uff1a\u5404 GPU \u7684\u8ba1\u7b97\u91cf\u5c3d\u91cf\u76f8\u540c(core-attention \u8ba1\u7b97\u8d1f\u8f7d\u5747\u8861)\u3001\u8f93\u5165\u7684 token \u6570\u91cf\u4e5f\u5c3d\u91cf\u76f8\u540c(dispatch \u53d1\u9001\u91cf\u8d1f\u8f7d\u5747\u8861)\uff0c\u907f\u514d\u90e8\u5206 GPU \u5904\u7406\u65f6\u95f4\u8fc7\u957f<\/li>\n<\/ul>\n

      Decode Load Balancer<\/strong><\/p>\n

        \n
      • \u6838\u5fc3\u95ee\u9898\uff1a\u4e0d\u540c\u6570\u636e\u5e76\u884c(DP)\u5b9e\u4f8b\u4e0a\u7684\u8bf7\u6c42\u6570\u91cf\u3001\u957f\u5ea6\u4e0d\u540c\uff0c\u5bfc\u81f4 core-attention \u8ba1\u7b97\u91cf(\u4e0e KVCache \u5360\u7528\u91cf\u76f8\u5173)\u3001dispatch \u53d1\u9001\u91cf\u4e0d\u540c<\/li>\n
      • \u4f18\u5316\u76ee\u6807\uff1a\u5404 GPU \u7684 KVCache \u5360\u7528\u91cf\u5c3d\u91cf\u76f8\u540c(core-attention \u8ba1\u7b97\u8d1f\u8f7d\u5747\u8861)\u3001\u8bf7\u6c42\u6570\u91cf\u5c3d\u91cf\u76f8\u540c(dispatch \u53d1\u9001\u91cf\u8d1f\u8f7d\u5747\u8861)<\/li>\n<\/ul>\n

        Expert-Parallel Load Balancer<\/strong><\/p>\n

          \n
        • \u6838\u5fc3\u95ee\u9898\uff1a\u5bf9\u4e8e\u7ed9\u5b9a MoE \u6a21\u578b\uff0c\u5b58\u5728\u4e00\u4e9b\u5929\u7136\u7684\u9ad8\u8d1f\u8f7d\u4e13\u5bb6(expert)\uff0c\u5bfc\u81f4\u4e0d\u540c GPU \u7684\u4e13\u5bb6\u8ba1\u7b97\u8d1f\u8f7d\u4e0d\u5747\u8861<\/li>\n
        • \u4f18\u5316\u76ee\u6807\uff1a\u6bcf\u4e2a GPU \u4e0a\u7684\u4e13\u5bb6\u8ba1\u7b97\u91cf\u5747\u8861(\u5373\u6700\u5c0f\u5316\u6240\u6709 GPU \u7684 dispatch \u63a5\u6536\u91cf\u7684\u6700\u5927\u503c)<\/li>\n<\/ul>\n

          \u53c2\u8003\u67b6\u6784\u56fe<\/h4>\n

          <\/p>\n

          \u7ebf\u4e0a\u7cfb\u7edf\u7684\u5b9e\u9645\u7edf\u8ba1\u6570\u636e<\/h4>\n

          DeepSeek V3 \u548c R1 \u7684\u6240\u6709\u670d\u52a1\u5747\u4f7f\u7528 H800 GPU\uff0c\u4f7f\u7528\u548c\u8bad\u7ec3\u4e00\u81f4\u7684\u7cbe\u5ea6\uff0c\u5373\u77e9\u9635\u8ba1\u7b97\u548c dispatch \u4f20\u8f93\u91c7\u7528\u548c\u8bad\u7ec3\u4e00\u81f4\u7684 FP8 \u683c\u5f0f\uff0ccore-attention \u8ba1\u7b97\u548c combine \u4f20\u8f93\u91c7\u7528\u548c\u8bad\u7ec3\u4e00\u81f4\u7684 BF16\uff0c\u6700\u5927\u7a0b\u5ea6\u4fdd\u8bc1\u4e86\u670d\u52a1\u6548\u679c\u3002\u53e6\u5916\uff0c\u7531\u4e8e\u767d\u5929\u7684\u670d\u52a1\u8d1f\u8377\u9ad8\uff0c\u665a\u4e0a\u7684\u670d\u52a1\u8d1f\u8377\u4f4e\uff0c\u56e0\u6b64 DeepSeek \u5b9e\u73b0\u4e86\u4e00\u5957\u673a\u5236\uff1a<\/p>\n

            \n
          • \u767d\u5929\u8d1f\u8377\u9ad8\u65f6\u7528\u6240\u6709\u8282\u70b9\u90e8\u7f72\u63a8\u7406\u670d\u52a1\u3002<\/li>\n
          • \u665a\u4e0a\u8d1f\u8377\u4f4e\u65f6\uff0c\u51cf\u5c11\u63a8\u7406\u8282\u70b9\uff0c\u4ee5\u7528\u6765\u505a\u7814\u7a76\u548c\u8bad\u7ec3\u3002\u3001<\/li>\n<\/ul>\n

            \u5728\u6700\u8fd1\u7684 24 \u5c0f\u65f6\u91cc(\u5317\u4eac\u65f6\u95f4 2025\/02\/27 12:00 \u81f3 2025\/02\/28 12:00)\uff0cDeepSeek V3 \u548c R1 \u63a8\u7406\u670d\u52a1\u5360\u7528\u8282\u70b9\u603b\u548c\uff0c\u5cf0\u503c\u5360\u7528\u4e3a 278 \u4e2a\u8282\u70b9\uff0c\u5e73\u5747\u5360\u7528 226.75 \u4e2a\u8282\u70b9(\u6bcf\u4e2a\u8282\u70b9\u4e3a 8 \u4e2a H800 GPU)\u3002\u5047\u5b9a GPU \u79df\u8d41\u6210\u672c\u4e3a 2 \u7f8e\u91d1 \/ \u5c0f\u65f6\uff0c\u603b\u6210\u672c\u4e3a $87,072 \/ \u5929\u3002<\/p>\n

            <\/p>\n

            \u5728 24 \u5c0f\u65f6\u7edf\u8ba1\u65f6\u6bb5\u5185\uff0cDeepSeek V3 \u548c R1\uff1a<\/p>\n

              \n
            • \u8f93\u5165 token \u603b\u6570\u4e3a 608B\uff0c\u5176\u4e2d 342B tokens(56.3%)\u547d\u4e2d KVCache \u786c\u76d8\u7f13\u5b58\u3002<\/li>\n
            • \u8f93\u51fa token \u603b\u6570\u4e3a 168B\u3002\u5e73\u5747\u8f93\u51fa\u901f\u7387\u4e3a 20~22 tps\uff0c\u5e73\u5747\u6bcf\u8f93\u51fa\u4e00\u4e2a token \u7684 KVCache \u957f\u5ea6\u662f 4989\u3002<\/li>\n
            • \u5e73\u5747\u6bcf\u53f0 H800 \u7684\u541e\u5410\u91cf\u4e3a\uff1a\u5bf9\u4e8e prefill \u4efb\u52a1\uff0c\u8f93\u5165\u541e\u5410\u7ea6 73.7k tokens \/ s(\u542b\u7f13\u5b58\u547d\u4e2d);\u5bf9\u4e8e decode \u4efb\u52a1\uff0c\u8f93\u51fa\u541e\u5410\u7ea6 14.8k tokens \/ s\u3002\u4ee5\u4e0a\u7edf\u8ba1\u5305\u62ec\u4e86\u7f51\u9875\u3001APP \u548c API \u7684\u6240\u6709\u8d1f\u8f7d\u3002\u5982\u679c\u6240\u6709 tokens \u5168\u90e8\u6309\u7167 DeepSeek R1 \u7684\u5b9a\u4ef7\u8ba1\u7b97\uff0c\u7406\u8bba\u4e0a\u4e00\u5929\u7684\u603b\u6536\u5165\u4e3a $562,027\uff0c\u6210\u672c\u5229\u6da6\u7387 545%\u3002<\/li>\n<\/ul>\n

              \u201c\u5f53\u7136\u6211\u4eec\u5b9e\u9645\u4e0a\u6ca1\u6709\u8fd9\u4e48\u591a\u6536\u5165\uff0c\u56e0\u4e3a V3 \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\u201d<\/p>\n

              <\/p>\n

              DeepSeek R1 \u7684\u5b9a\u4ef7\uff1a$0.14 \/ \u767e\u4e07\u8f93\u5165 tokens (\u7f13\u5b58\u547d\u4e2d)\uff0c$0.55 \/ \u767e\u4e07\u8f93\u5165 tokens (\u7f13\u5b58\u672a\u547d\u4e2d)\uff0c$2.19 \/ \u767e\u4e07\u8f93\u51fa tokens\u3002<\/p>\n

              \u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n

              <\/div>\n

              <\/g><\/svg><\/i> \"Loading\"<\/p>\n

              <\/div>\n","protected":false},"excerpt":{"rendered":"

              DeepSeek \u5b98\u65b9\u4eca\u65e5\u5728\u77e5\u4e4e\u53d1\u5e03\u300aDeepSeek-V3 \/ R1 \u63a8\u7406\u7cfb\u7edf\u6982\u89c8\u300b\u4e00\u6587\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528\u5927 […]<\/p>\n

              <\/div>\n

              <\/g><\/svg><\/i> \"Loading\"<\/p>\n

              <\/div>\n","protected":false},"author":56,"featured_media":24777,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,20,80],"tags":[124,1526,113],"class_list":["post-24775","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-20","category-80","tag-ai","tag-deepseek","tag-113"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24775","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=24775"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24775\/revisions"}],"predecessor-version":[{"id":24783,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24775\/revisions\/24783"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/24777"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=24775"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=24775"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=24775"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}