{"id":715651,"date":"2026-02-04T19:07:14","date_gmt":"2026-02-04T11:07:14","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=715651"},"modified":"2026-02-04T19:07:14","modified_gmt":"2026-02-04T11:07:14","slug":"%e8%85%be%e8%ae%af%e6%b7%b7%e5%85%83ai-infra%e6%a0%b8%e5%bf%83%e6%8a%80%e6%9c%af%e5%bc%80%e6%ba%90%ef%bc%8c%e6%8e%a8%e7%90%86%e5%90%9e%e5%90%90%e6%8f%90%e5%8d%8730","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=715651","title":{"rendered":"\u817e\u8baf\u6df7\u5143AI Infra\u6838\u5fc3\u6280\u672f\u5f00\u6e90\uff0c\u63a8\u7406\u541e\u5410\u63d0\u534730%"},"content":{"rendered":"<p>IT\u4e4b\u5bb6 2 \u6708 4 \u65e5\u6d88\u606f\uff0c<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e8%85%be%e8%ae%af\" title=\"View all posts in \u817e\u8baf\" target=\"_blank\" class=\"tag_link\">\u817e\u8baf<\/a>\u6df7\u5143 AI Infra \u56e2\u961f\u4eca\u65e5\u5ba3\u5e03\u63a8\u51fa\u5f00\u6e90\u751f\u4ea7\u7ea7\u9ad8\u6027\u80fd LLM \u63a8\u7406\u6838\u5fc3\u7b97\u5b50\u5e93 <strong>HPC-Ops<\/strong>\u3002<\/p>\n<p>\u8be5\u7b97\u5b50\u5e93\u5ba3\u79f0\u57fa\u4e8e\u751f\u4ea7\u73af\u5883\u75db\u70b9\uff0c\u91c7\u7528 CUDA \u548c CuTe \u4ece\u96f6\u6784\u5efa\uff0c\u901a\u8fc7\u62bd\u8c61\u5316\u5de5\u7a0b\u67b6\u6784\u3001\u5fae\u67b6\u6784\u6df1\u5ea6\u9002\u914d\u53ca\u6307\u4ee4\u7ea7\u6781\u81f4\u4f18\u5316\u7b49\uff0c\u964d\u4f4e\u5e95\u5c42\u7b97\u5b50\u5f00\u53d1\u95e8\u69db\uff0c\u5c06\u6838\u5fc3\u7b97\u5b50\u6027\u80fd\u903c\u8fd1\u786c\u4ef6\u5cf0\u503c\uff0c\u5b9e\u73b0\u4e86\u6027\u80fd\u7a81\u7834\u3002<\/p>\n<\/p>\n<p>HPC-Ops \u7b97\u5b50\u5e93\u67b6\u6784\u56fe<\/p>\n<p>\u5728\u771f\u5b9e\u573a\u666f\u4e0b\uff0c\u57fa\u4e8e HPC-Ops\uff0c<strong>\u6df7\u5143\u6a21\u578b\u63a8\u7406 QPM \u63d0\u5347 30%\uff0cDeepSeek \u6a21\u578b QPM \u63d0\u5347 17%<\/strong>\u3002\u540c\u65f6\uff0c\u5728\u5355\u7b97\u5b50\u6027\u80fd\u65b9\u9762\uff0cHPC-Ops \u5b9e\u73b0 Attention \u76f8\u6bd4 FlashInfer \/ FlashAttention \u6700\u9ad8\u63d0\u5347 2.22 \u500d\uff1bGroupGEMM \u76f8\u6bd4 DeepGEMM \u6700\u9ad8\u63d0\u5347 1.88 \u500d\uff1bFusedMoE \u76f8\u6bd4 TensorRT-LLM \u6700\u9ad8\u63d0\u5347 1.49 \u500d\u3002<\/p>\n<p>\u5728\u672a\u6765\u7684\u53d1\u5c55\u89c4\u5212\u4e2d\uff0cHPC-Ops \u5c06\u6301\u7eed\u6df1\u8015\u5927\u6a21\u578b\u63a8\u7406\u6027\u80fd\u7684\u7a81\u7834\u65b9\u5411\uff1a<\/p>\n<p>\u4e00\u65b9\u9762\uff0c<strong>\u5c06\u91cd\u70b9\u7814\u53d1\u7a00\u758f Attention \u7b97\u5b50<\/strong>\uff0c\u9488\u5bf9\u6027\u89e3\u51b3\u957f\u4e0a\u4e0b\u6587\u5927\u6a21\u578b\u7684\u5185\u5b58\u4e0e\u7b97\u529b\u74f6\u9888\uff1b<\/p>\n<p>\u53e6\u4e00\u65b9\u9762\uff0c\u4f1a\u62d3\u5c55\u66f4\u4e30\u5bcc\u7684\u91cf\u5316<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e7%ad%96%e7%95%a5\" title=\"View all posts in \u7b56\u7565\" target=\"_blank\" class=\"tag_link\">\u7b56\u7565<\/a>\uff0c<strong>\u8986\u76d6 4bit\/8bit \u6df7\u5408\u7cbe\u5ea6\u7b49\u66f4\u591a\u91cf\u5316\u65b9\u6848<\/strong>\uff0c\u8fdb\u4e00\u6b65\u5e73\u8861\u63a8\u7406\u901f\u5ea6\u4e0e\u6a21\u578b\u7cbe\u5ea6\uff1b<\/p>\n<p>\u6b64\u5916\uff0c\u7b97\u5b50\u5e93\u8fd8\u5c06\u5e03\u5c40\u8ba1\u7b97-\u901a\u4fe1\u534f\u540c\u4f18\u5316\u7684\u5185\u6838\uff0c\u901a\u8fc7\u878d\u5408\u591a GPU \u95f4\u7684\u8ba1\u7b97\u903b\u8f91\u4e0e\u901a\u4fe1\u6d41\u7a0b\uff0c<strong>\u5927\u5e45\u964d\u4f4e\u5206\u5e03\u5f0f\u63a8\u7406\u573a\u666f\u4e0b\u7684\u901a\u4fe1\u5f00\u9500<\/strong>\uff0c\u4e3a\u8d85\u5927\u89c4\u6a21\u5927\u6a21\u578b\u7684\u9ad8\u6548\u90e8\u7f72\u63d0\u4f9b\u5e95\u5c42\u652f\u6491\u3002<\/p>\n<p>IT\u4e4b\u5bb6\u9644 HPC-Ops \u5f00\u6e90\u5730\u5740\u5982\u4e0b\uff1a<\/p>\n<p>https:\/\/github.com\/Tencent\/hpc-ops<\/p>\n<p>(\u6765\u6e90\uff1a\u65b0\u6d6a\u79d1\u6280)<\/p>\n","protected":false},"excerpt":{"rendered":"<p>IT\u4e4b\u5bb6 2 \u6708 4 \u65e5\u6d88\u606f\uff0c\u817e\u8baf\u6df7\u5143 AI Infra \u56e2\u961f\u4eca\u65e5\u5ba3\u5e03\u63a8\u51fa\u5f00\u6e90\u751f\u4ea7\u7ea7\u9ad8\u6027\u80fd LLM \u63a8\u7406\u6838\u5fc3\u7b97 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"aside","meta":[],"categories":[6],"tags":[88,119],"views":59,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/715651"}],"collection":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=715651"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/715651\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=715651"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=715651"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=715651"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}