{"id":10006491,"date":"2022-09-16T10:26:42","date_gmt":"2022-09-16T02:26:42","guid":{"rendered":"https:\/\/ityet.com\/?p=10006491"},"modified":"2022-09-16T10:26:42","modified_gmt":"2022-09-16T02:26:42","slug":"%e5%9f%ba%e4%ba%8e%e9%98%bf%e9%87%8c%e4%ba%91gpu%e4%ba%91%e6%9c%8d%e5%8a%a1%e5%99%a8%e7%9a%84aiacc%e5%8a%a9%e5%8a%9buc%e6%90%9c%e7%b4%a2%e4%b8%9a%e5%8a%a1%e6%80%a7%e8%83%bd%e6%8f%90%e6%95%88380","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10006491","title":{"rendered":"\u57fa\u4e8e\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u7684AIACC\u52a9\u529bUC\u641c\u7d22\u4e1a\u52a1\u6027\u80fd\u63d0\u6548380%\uff0c\u6bcf\u5e74\u8282\u7701\u6570\u5343\u4e07\u6210\u672c"},"content":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u7528\u963f\u91cc\u4e91GPU\u8ba1\u7b97\u5b9e\u4f8b\u6765\u6ee1\u8db3UC\u6781\u81f4\u6027\u4ef7\u6bd4\u9700\u6c42<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488916\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u6587\u4e28\u963f\u91cc\u4e91\u795e\u9f99\u8ba1\u7b97\u5e73\u53f0AI\u52a0\u901f\u56e2\u961f &#038; UC\u641c\u7d22\u67b6\u6784\u90e8\u63a8\u7406\u5f15\u64ce\u56e2\u961f\u00a0\u5bfc\u8bed\uff1a\u4f5c\u4e3a\u56fd\u4ea7\u884c\u5217\u91cc\u5360\u6709\u7387\u6392\u540d\u7b2c\u4e00\u7684\u79fb\u52a8\u6d4f\u89c8\u5668\uff0cUC\u6d4f\u89c8\u5668\u81ea\u8eab\u627f\u8f7d\u7740\u6570\u4ee5\u4ebf\u8ba1\u7684\u7528\u6237\u91cf\uff0c\u5f53\u524dUC\u6d4f\u89c8\u5668\u6bcf\u5929\u7684\u670d\u52a1\u8bf7\u6c42\u5bf9\u670d\u52a1\u5668\u7684\u7b97\u529b\u53ca\u5e26\u5bbd\u8981\u6c42\u6781\u9ad8\uff0c\u56e0\u6b64\u4e5f\u5e26\u6765\u4e86\u5de8\u989d\u7684\u8fd0\u8425\u6210\u672c\u3002\u56e0\u4e3a\u4e1a\u52a1\u662f\u52a8\u6001\u53d8\u5316\u7684\uff0cUC\u5bf9\u8ba1\u7b97\u8d44\u6e90\u4e5f\u6709\u52a8\u6001\u6269\u7f29\u5bb9\u7684\u9700\u6c42\u3002\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u662f\u63d0\u4f9bGPU\u7b97\u529b\u7684\u5f39\u6027\u8ba1\u7b97\u670d\u52a1\uff0c\u5177\u6709\u8d85\u5f3a\u7684\u8ba1\u7b97\u80fd\u529b\uff0c\u670d\u52a1\u4e8e\u6df1\u5ea6\u5b66\u4e60\u3001\u79d1\u5b66\u8ba1\u7b97\u3001\u56fe\u5f62\u53ef\u89c6\u5316\u3001\u89c6\u9891\u5904\u7406\u591a\u79cd\u5e94\u7528\u573a\u666f\uff0c\u80fd\u4e3a\u5ba2\u6237\u63d0\u4f9b\u8f6f\u4ef6\u4e0e\u786c\u4ef6\u7ed3\u5408\u7684\u5b8c\u6574\u670d\u52a1\u4f53\u7cfb\uff0c\u52a9\u529b\u5ba2\u6237\u5728\u5b9e\u9645\u4e1a\u52a1\u4e2d\u5b9e\u73b0\u8d44\u6e90\u7684\u7075\u6d3b\u5206\u914d\u3001\u5f39\u6027\u6269\u5c55\u3001\u7b97\u529b\u7684\u63d0\u5347\u4ee5\u53ca\u6210\u672c\u7684\u63a7\u5236\u3002\u800c\u57fa\u4e8e\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u7684\u795e\u9f99AI\u52a0\u901f\u5f15\u64ce\uff08AIACC\uff09\u662f\u4e3a\u4e86\u6781\u81f4\u6027\u80fd\u800c\u751f\uff0c\u65e8\u5728\u6253\u9020\u4e16\u754c\u7ea7\u65e0\u4e0e\u4f26\u6bd4\u7684AI\u6027\u80fd\u4f53\u9a8c\uff0c\u540c\u65f6\u4e3a\u4e86\u5ba2\u6237\u964d\u672c\u589e\u6548\uff0c\u5b9e\u73b0\u5171\u8d62\u3002\u636e\u6089\uff0c\u521a\u516c\u5e03\u7684\u6700\u65b0\u4e16\u754cMLPerfTM\u63a8\u7406\u699c\u5355\u4e2d\uff0c\u57fa\u4e8e\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u7684AIACC\u9996\u6b21\u7a81\u7834\u5c01\u95ed\u5f0f\u573a\u666f\u4e16\u754c\u7b2c\u4e00\u3002\u00a0\u672c\u7bc7\u6587\u7ae0\u5c06\u5e26\u5927\u5bb6\u4e86\u89e3\u963f\u91cc\u4e91AIACC\u5982\u4f55\u57fa\u4e8e\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u52a9\u529bUC\u6d4f\u89c8\u5668\u7684\u641c\u7d22\u4e1a\u52a1\u5e73\u8861\u8ba1\u7b97\u6027\u80fd\u4e0e\u8fd0\u8425\u6210\u672c\u4e4b\u95f4\u7684\u77db\u76fe\uff0c\u4f7f\u5176\u5927\u5e45\u5b9e\u73b0\u964d\u672c\u589e\u6548\uff0c\u6210\u529f\u5728\u963f\u91cc\u4e91\u7684GPU\u4e91\u670d\u52a1\u5668\u843d\u5730\u3002\u00a0\u80cc \u00a0\u666f\u00a01. \u4e1a\u52a1\u80cc\u666fUC\u641c\u7d22\u627f\u8f7d\u7740UC\u4e3b\u8981\u4e1a\u52a1\u5165\u53e3\uff0c\u573a\u666f\u5305\u62ec\uff1a\u5927\u641c\u3001\u5404\u79cd\u5782\u641c\u4e1a\u52a1\u3001\u5938\u514bapp\u7b49\u3002\u641c\u7d22\u6d41\u7a0b\u4e00\u822c\u7ecf\u8fc7\u51e0\u4e2a\u9636\u6bb5\uff1a\u53ec\u56de \u2013&gt; \u7c97\u6392 -&gt; \u7cbe\u6392(L2-&gt;L4) -&gt; \u6df7\u6392\u7b49\u3002\u67b6\u6784\u5982\u4e0b\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488917\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u5728\u4e1a\u52a1\u4e2d\uff0cL3\/L4\u6392\u5e8f\u90e8\u5206\u90fd\u4f7f\u7528\u4e86QTC\u6838\u5fc3\u6a21\u578b\u3002\u968f\u7740\u4e1a\u52a1\u6d41\u91cf\u7684\u589e\u957f\uff0c\u76ee\u524d\u7cbe\u6392\u6253\u5206\u9636\u6bb5\u9762\u4e34\u5de8\u5927\u6311\u6218\uff0c\u5ef6\u8fdf\u548c\u541e\u5410\u90fd\u8981\u517c\u5f97\u3002\u00a02. QTC\u6a21\u578b\u4e0b\u56fe\u662f\u7528TF-summary\u5bf9QTC\u6a21\u578b\u505a\u53ef\u89c6\u5316\uff0c<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488918\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0QTC\u6a21\u578b\u5c5e\u4e8e\u6392\u5e8f\u6838\u5fc3\u6a21\u578b\uff0c\u6a21\u578b\u7ed3\u6784\u5206\u4e3a3\u4e2aBERT+ \u591a\u5c42Conv + \u591a\u5c42MLP\u7b49\uff0c\u5176\u4e2dBert\u8f93\u5165seq length\u6700\u5927\u957f\u5ea6\u662f512\u3002\u6a21\u578b\u603b\u5171\u6709\u5927\u7ea64500\u4e2a\u7b97\u5b50\uff0c\u8ba1\u7b97\u91cf\u5de8\u5927\u3002\u00a03. \u539f\u59cb\u6027\u80fd\u6700\u521d\u91c7\u7528\u4e86NV\u63d0\u4f9b\u7684Faster-Transformer\u8fd9\u4e00\u5957\u8f6f\u4ef6\u6765\u4f18\u5316QTC\u6a21\u578b\u7684\u63a8\u7406\uff0c\u4f46\u7531\u4e8eFaster-Transformer\u53ea\u80fd\u52a0\u901fBERT\u7f51\u7edc\uff0c\u5bfc\u81f4\u63a8\u7406\u65f6\u9700\u8981\u505a\u591a\u4e2a\u5b50\u56fe\u5272\u88c2\u8fd0\u884c\uff0c\u52a0\u4e0a\u5176\u4f59\u7f51\u7edc\u90e8\u5206\u4e5f\u672a\u505a\u4f18\u5316\uff0c\u4ece\u800c\u6574\u4f53\u6027\u80fd\u5e76\u4e0d\u597d\u3002\u9488\u5bf9A100\u673a\u578b\uff0c\u505a\u4e86\u57fa\u4e8e\u8fd9\u5957\u65b9\u6848\u7684\u6027\u80fd\u8bc4\u4f30\uff0c\u5ef6\u8fdf\u63a7\u5236\u572830ms\u4e0b\uff0cQTC\u6a21\u578b\u53ea\u80fd\u8dd1\u5230350QPS\u3002\u00a04. \u6311\u6218\u4e0e\u76ee\u6807\u968f\u7740\u6a21\u578b\u6548\u679c\u5e26\u6765\u7684\u4e1a\u52a1\u6536\u76ca\uff0c\u4fc3\u4f7f\u4e1a\u52a1\u6d41\u91cf\u4e0d\u65ad\u589e\u52a0\uff0c\u5728\u7b97\u6cd5\u9884\u7b97\u524d\u63d0\u4e0b\uff0c\u6269\u91cf\u9762\u4e34\u6bd4\u8f83\u5927\u7684\u6311\u6218\u3002\u4e00\u65b9\u9762\u8d44\u6e90\u4e0d\u591f\uff0c\u4e00\u65b9\u9762\u8d44\u6e90\u4e5f\u6ca1\u6709\u5b8c\u5168\u5229\u7528\u8d77\u6765\u5bfc\u81f4\u7684\u541e\u5410\u4e0d\u591f\uff0c\u540c\u65f6\u89c2\u5bdf\u5206\u6790\u5728GPU\u4f7f\u7528\u7387\u8fbe\u523090%\u4ee5\u4e0a\u65f6\u5ef6\u8fdf\u4f1a\u51fa\u73b0\u98d9\u5347\u7684\u60c5\u51b5\uff0c\u8fd9\u79cd\u60c5\u51b5\u5bf9\u4e8e\u7a33\u5b9a\u6027\u5e26\u6765\u4e86\u4e0d\u5c0f\u7684\u5f71\u54cd\u3002\u5982\u4f55\u5728\u539f\u6765\u9884\u7b97\u7684\u57fa\u7840\u4e0a\u63d0\u9ad8\u541e\u5410\u3001\u964d\u4f4e\u5ef6\u8fdf\u3001\u63d0\u5347\u6027\u4ef7\u6bd4\u6700\u7ec8\u89c4\u6a21\u5316\u4e0a\u7ebf\u4e3a\u4e1a\u52a1\u8d4b\u80fd\u6210\u4e3a\u6700\u5927\u7684\u6311\u6218\u3002\u00a0\u7ed3\u5408\u641c\u7d22\u4e1a\u52a1\u7684\u7279\u70b9\uff0c\u9488\u5bf9QTC\u6a21\u578b\u5728GPU\u4e0a\u63a8\u7406\u63d0\u51fa\u4e86\u51e0\u4e2a\u65b9\u9762\u7684\u9700\u6c42\uff1a\u00a01. \u00a0\u4e3a\u4e86\u4fdd\u8bc1\u7b97\u6cd5\u7684\u6548\u679c\uff0c\u63a8\u7406\u65f6\u6a21\u578b\u8f93\u51fa\u7684\u8bef\u5dee\u63a7\u5236\u5728\u5343\u5206\u4f4d\uff1b2. \u00a0\u4e3a\u4e86\u4fdd\u8bc1\u7ec8\u7aef\u7528\u6237\u5728\u641c\u7d22\u65f6\u7684\u4f53\u9a8c\uff0c\u7cbe\u6392\u90e8\u5206\u7684\u63a8\u7406\u8017\u65f6\u572830ms\u4ee5\u5185\uff1b3. \u00a0\u5728\u6709\u9650\u7684\u9884\u7b97\u4e0b\u6ee1\u8db3\u4e1a\u52a1\u7684\u9700\u6c42\uff0c\u5373QPS\u9700\u8981\u8fdc\u9ad8\u4e8e\u539f\u6709\u89e3\u51b3\u65b9\u6848\uff1b\u00a0\u7efc\u5408\u4e1a\u52a1\u65b9\u5404\u4e2a\u65b9\u9762\u7684\u9700\u6c42\uff0cUC\u5236\u5b9a\u5982\u4e0b\u7684\u76ee\u6807\uff1a\u00a0\u5728A100\u673a\u578b\u4e0a\uff0c\u4fdd\u8bc1\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\uff0cQTC\u6a21\u578b\u8017\u65f6\u572830ms\u4ee5\u5185\uff0cQPS\u8fbe\u52301120\u4ee5\u4e0a\u3002\u00a0\u4f18 \u00a0\u5316\u00a01. \u4f18\u5316\u8def\u7ebfTensorRT\u662fNvidia\u5f00\u53d1\u7684\u9488\u5bf9\u6a21\u578b\u63a8\u7406\u7684\u9ad8\u6027\u80fd\u8f6f\u4ef6\u5e93\uff0c\u4f46QTC\u6a21\u578b\u590d\u6742\u5ea6\u9ad8\uff0c\u65e0\u6cd5\u6574\u4f53\u76f4\u63a5\u7528TensorRT load\u3002\u6b64\u5916\uff0c\u76f4\u63a5\u7528TensorRT load\u65f6\u90e8\u5206\u5c42\u4f1a\u88ab\u62c6\u6210\u7ec6\u7c92\u5ea6\u7684op\uff0c\u6027\u80fd\u8f83\u5dee\u3002\u5bf9QTC\u6a21\u578b\u505a\u6df1\u5ea6\u7684\u5206\u6790\u4e4b\u540e\uff0c\u7ed3\u8bba\u662f\u7528TensorRT\u4ee5\u53ca\u5176plugin\u652f\u6301\u7684op\u53ef\u4ee5\u8986\u76d6QTC\u6a21\u578b\u7684op\uff0c\u4f46\u662f\u73b0\u6709\u7684parser\u65e0\u6cd5\u6309\u7167\u9700\u8981\u7684\u7c92\u5ea6\u53bb\u89e3\u6790QTC\u6a21\u578b\u5e76\u6620\u5c04\u5230TensorRT\u4e0a\u3002\u00a0\u9488\u5bf9\u4e0a\u8ff0\u95ee\u9898\uff0cAIACC\u4ece\u6a21\u578b\u5b9a\u4e49\u5c42\uff0c\u76f4\u63a5\u89e3\u6790QTC\u6a21\u578b\u5e76\u6620\u5c04\u5230TensorRT\u4e0a\uff0c\u907f\u514d\u8f6cPB\u540e\u5927\u7684operator\u88ab\u62c6\u5206\u6210\u7ec6\u7c92\u5ea6\u7684operator\uff0c\u4ee5\u8fbe\u5230\u6700\u4f73\u6027\u80fd\u3002\u6b64\u5916\uff0c\u8fd9\u4e48\u505a\u4e5f\u4fbf\u4e8e\u4ee5\u4e0d\u540c\u7684\u7c92\u5ea6\u8fdb\u884c\u6a21\u578b\u7cbe\u5ea6\u7684\u5206\u6790\uff0c\u4ee5\u4e0d\u540c\u7684\u7c92\u5ea6\u5b9e\u73b0op\u7684\u878d\u5408\u4e0e\u66ff\u6362\uff0c\u65b9\u9762\u96c6\u6210AIACC\u7684\u9ad8\u6027\u80fdKernel\u8fdb\u6765\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488919\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a02. Enable TensorRT &#038; \u7cbe\u5ea6\u5bf9\u9f50\u00a02.1 \u6620\u5c04\u6b63\u786e\u6027\u9a8c\u8bc1\u00a0\u89e3\u6790QTC\u6a21\u578b\u5e76\u6620\u5c04\u5230TensorRT\u4e0a\u7684\u8fc7\u7a0b\u4e2d\uff0c\u9996\u5148\u8981\u89e3\u51b3\u7684\u95ee\u9898\u662f\u6620\u5c04\u7684\u6b63\u786e\u6027\u9a8c\u8bc1\u3002\u5bf9\u6b63\u786e\u6027\u9a8c\u8bc1\uff0c\u9996\u5148\u8981\u89e3\u51b3\u57fa\u51c6\u662f\u4ec0\u4e48\u8fd9\u4e2a\u95ee\u9898\u3002\u6211\u4eec\u62ff\u5230\u7684\u4fe1\u606f\u4e3a\u4e00\u4efd\u6a21\u578b\u7ed3\u6784\u7684\u4ee3\u7801\u4e0e\u4e00\u4e2a\u8bad\u7ec3\u4e2d\u7684checkpoint\u3002\u57fa\u4e8e\u4e1a\u52a1\u65b9\u7ed9\u7684\u6709\u9650\u7684\u4fe1\u606f\uff0c\u5b8c\u6210\u4e86\u4ece\u8bad\u7ec3\u6846\u67b6\u4fa7load checkpoint\u7684\u903b\u8f91\uff0c\u5e76\u6839\u636e\u9700\u6c42\uff0c\u7559\u4e0b\u4e86\u53ef\u4ee5\u83b7\u53d6\u4efb\u610f\u4e2d\u95f4\u7ed3\u679c\u7684\u63a5\u53e3\u3002\u5728\u9a8c\u8bc1\u6620\u5c04\u6b63\u786e\u6027\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u4ee5\u8bad\u7ec3\u6846\u67b6\u4fa7\u8fd0\u884c\u63a8\u7406\u7684\u7ed3\u679c\u4e3a\u5bf9\u6bd4\u7684baseline\uff0c\u548c\u6211\u4eec\u6784\u5efa\u7684TensorRT\u4fa7\u7684Engine\u7684\u8fd0\u884c\u7ed3\u679c\u505a\u5bf9\u6bd4\uff0c\u6765\u5224\u65ad\u662f\u5426\u6b63\u786e\u7684\u628aQTC\u6a21\u578b\u6620\u5c04\u5230TensorRT\u4e0a\u3002\u00a0\u4e0b\u9762\u4ee5\u4e00\u4e2aself-attention\u5c42\u4e3a\u4f8b\uff0c\u5c55\u5f00\u600e\u4e48\u505aQTC\u6a21\u578b\u5230TensorRT\u7684\u6620\u5c04\u5e76\u4fdd\u8bc1\u8fd0\u7b97\u7ed3\u679c\u7684\u4e00\u81f4\u3002\u00a0\u5728TensorRT\u4fa7\uff0c\u4e3a\u4e86\u6821\u9a8c\u6620\u5c04\u7684\u6b63\u786e\u6027\uff0c\u9700\u8981\u6709\u63a8\u7406\u7684\u811a\u672c\u6765\u9a8c\u8bc1build\u597d\u7684engine\u7684\u6b63\u786e\u6027\u3002\u540c\u5728\u8bad\u7ec3\u6846\u67b6\u4fa7\u6784\u5efa\u6b63\u786e\u6027\u7684baseline\u7c7b\u4f3c\uff0c\u6211\u4eec\u57fa\u4e8eTensorRT\u5f00\u53d1\u4e86\u4e00\u5957load build\u597d\u7684engine\u5e76\u505a\u63a8\u7406\u7684\u811a\u672c\uff0c\u6765\u83b7\u53d6engine\u7684\u63a8\u7406\u7ed3\u679c\uff0c\u7528\u4e8e\u548c\u8bad\u7ec3\u6846\u67b6\u4fa7\u7684baseline\u505a\u5bf9\u6bd4\u3002\u00a0TensorRT\u4e3a\u4e86\u8ffd\u6c42\u66f4\u9ad8\u7684\u6027\u80fd\uff0c\u5bf9self-attention\u5c42\u7684\u8fd0\u7b97\u505a\u4e86\u7b49\u4ef7\u53d8\u5316\uff0c\u628a\u90e8\u5206\u8f6c\u6362\u63d0\u524d\u5230\u6a21\u578bbuild\u9636\u6bb5\uff0c\u8fdb\u800c\u51cf\u5c11\u6a21\u578b\u63a8\u7406\u9700\u8981\u7684operation\u3002\u5176\u4ed6\u7684\u4e00\u4e9b\u7ec6\u8282\uff0c\u5305\u62ecmask\u5904\u7406\u3001\u6570\u636e\u7c7b\u578b\u3001hidden size\u7b49\uff0c\u6309\u7167\u7c7b\u4f3c\u903b\u8f91\uff0c\u4e00\u4e00\u5bf9\u9f50\u540e\uff0c\u6211\u4eec\u5373\u53ef\u5728TensorRT\u4fa7\u6784\u5efa\u51fa\u8fd0\u884c\u7ed3\u679c\u4e0e\u8bad\u7ec3\u6846\u67b6\u4fa7\u8fd0\u884c\u7ed3\u679c\u4e00\u81f4\u7684engine\u3002\u00a0\u5bf9\u6bd4TensorRT\u76f4\u63a5load checkpoint\u7684\u6a21\u5f0f\uff0c\u5728\u6211\u4eec\u7684\u6620\u5c04\u4e2d\uff0c\u628aself-attention\u6620\u5c04\u4e3a1\u4e2aop\uff0c\u907f\u514d\u4e86\u62c6\u6210\u591a\u4e2a\u7ec6\u788eop\u800c\u5bfc\u81f4\u7684\u6027\u80fd\u95ee\u9898\u3002\u8fd9\u79cd\u6a21\u5f0f\u4e0bself-attention\u7684\u4fe1\u606f\u5728\u4e00\u4e2a\u8282\u70b9\u4e0a\uff0c\u4e5f\u65b9\u4fbf\u6211\u4eec\u540e\u7eed\u505akernel\u7684\u4f18\u5316\uff0c\u4f8b\u5982\u878d\u5408self-attention\u5185\u90e8\u7684GEMM\u7b49\u591a\u4e2a\u64cd\u4f5c\uff0c\u4ee5\u8fbe\u5230\u66f4\u597d\u7684\u65b0\u80fd\u3002\u00a02.2 \u6570\u503c\u95ee\u9898\u89e3\u51b3\u00a0\u6570\u503c\u95ee\u9898\u662fAI\u4efb\u52a1\u4e2d\u5f88\u96be\u5b9a\u4f4d\u89e3\u51b3\u7684\u4e00\u7c7b\u95ee\u9898\uff0c\u56e0\u4e3a\u6ca1\u6709\u7f16\u8bd1\u5668\u6216\u8005\u5176\u4ed6\u7684\u62a5\u9519\u63d0\u793a\u6765\u5e2e\u52a9\u6211\u4eec\u5b9a\u4f4d\u95ee\u9898\u3002\u5728\u628aQTC\u6a21\u578b\u6620\u5c04\u5230TensorRT\u7684\u65f6\u5019\uff0c\u9047\u5230\u4e86\u6570\u503c\u95ee\u9898\uff0c\u73b0\u8c61\u4e3a\u591a\u4e2a\u76f8\u540c\u7684\u57fa\u7840\u6a21\u5757\u53e0\u52a0\u5bfc\u81f4\u4e2d\u95f4\u7684feature map\u4e2d\u6709NAN\u7684\u5f02\u5e38\u503c\uff0c\u8fdb\u800c\u5bfc\u81f4\u6700\u7ec8\u7684\u7ed3\u679c\u8bef\u5dee\u8fdc\u8fdc\u8d85\u51fa\u4e1a\u52a1\u56e2\u961f\u8981\u6c42\u7684\u5343\u5206\u4f4d\u3002\u00a0\u7b2c\u4e00\u65f6\u95f4\u628a\u8fd9\u4e2a\u95ee\u9898\u5b9a\u4e3a\u6570\u503c\u95ee\u9898\uff0c\u662f\u56e0\u4e3a\u5728QTC\u6a21\u578b\u5230TensorRT\u6620\u5c04\u8fc7\u7a0b\u4e2d\uff0c\u6bcf\u4e00\u4e2a\u5b50\u6a21\u5757\u6211\u4eec\u90fd\u505a\u4e86\u5355\u5143\u6d4b\u8bd5\u6765\u9a8c\u8bc1\u6620\u5c04\u7684\u6b63\u786e\u6027\uff0c\u4e0d\u5b58\u5728\u6620\u5c04\u9519\u8bef\u5bfc\u81f4\u7684\u5f02\u5e38\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u63a8\u7406\u7684\u6027\u80fd\uff0c\u6211\u4eec\u5f00\u542f\u4e86FP16\uff0c\u90e8\u5206\u4e2d\u95f4\u7684\u8fd0\u7b97\u662f\u7528FP16\u8fdb\u884c\uff0cFP16\u7684\u8868\u8fbe\u80fd\u529b\u76f8\u6bd4FP32\u5dee\u5f88\u591a\uff0c\u8fd9\u662f\u5f15\u5165\u6570\u503c\u95ee\u9898\u7684\u4e00\u4e2a\u98ce\u9669\u70b9\u3002\u591a\u4e2a\u76f8\u540c\u7684\u57fa\u7840\u6a21\u5757\u53e0\u52a0\u5bfc\u81f4\u6700\u7ec8\u7684\u8f93\u51fa\u4e2d\u6709NAN\u7684\u5f02\u5e38\u503c\uff0c\u5927\u6982\u7387\u662f\u56e0\u4e3a\u8ba1\u7b97\u4e2d\u51fa\u73b0\u4e86\u6781\u5927or\u6781\u5c0f\u7684\u6570\u503c\uff0c\u5bfc\u81f4\u67d0\u4e9b\u64cd\u4f5c\uff0c\u4f8b\u5982\u505aexp\/\u9664\u6cd5\u7684\u65f6\u5019\uff0c\u51fa\u73b0\u7684\u5f02\u5e38\u7684\u884c\u4e3a\u3002\u00a0\u5206\u6790\u6e05\u695a\u95ee\u9898\u540e\uff0c\u89e3\u51b3\u95ee\u9898\u7684\u65b9\u6cd5\u4e5f\u5f88\u7b80\u5355\uff0c\u5b8c\u5584TensorRT\u4e2d\u8fd9\u90e8\u5206exp\u8ba1\u7b97\u903b\u8f91\uff0c\u907f\u514dexp\u8ba1\u7b97\u4e2d\u51fa\u73b0\u5927\u7684\u8d1f\u503c\u5373\u53ef\u3002\u00a02.3 \u5176\u4ed6\u7684N\u4e2a\u95ee\u9898\u7684\u6982\u8ff0\u00a0\u5728\u628aQTC\u6a21\u578b\u6620\u5c04\u5230TensorRT\u5e76\u5bf9\u9f50\u7cbe\u5ea6\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u8fd8\u9047\u5230\u4e86\u5176\u4ed6\u4e00\u7cfb\u5217\u5927\u5927\u5c0f\u5c0f\u7684\u5176\u4ed6\u95ee\u9898\uff0c\u8fd9\u91cc\u4e0d\u4e00\u4e00\u5c55\u5f00\uff0c\u53ea\u5728\u7b80\u5355\u5217\u4e3e\u4e00\u4e0b\uff0c\u5305\u62ec\u7528shuffle\u66ff\u6362reduce\u6765\u63d0\u9ad8\u5b9e\u73b0reshape\u7684\u8fd0\u7b97\u7684\u6548\u7387\uff0c\u7528identity\u5c42\u6765\u5b9e\u73b0data type cast\u7684\u903b\u8f91\uff0cconv\u5c42\u7684\u53c2\u6570\u8f6c\u6362\uff0cpooling\u7684padding\u5904\u7406\uff0creshape\u540e\u63a5FC\u5c42\u53c2\u6570\u6392\u5e03\u7b49\u95ee\u9898\u3002\u00a03. \u5173\u952eKernel\u4f18\u5316seq_length=35\/512\u65f6\uff0cTensorRT\u672a\u5bf9Multi-Head Attention\u505a\u9488\u5bf9\u6027\u4f18\u5316\u3002seq_length\u4e3a512\u7684Bert\u8017\u65f6\u5360\u6bd4\u8f83\u5927\uff0c\u5982\u679c\u5bf9\u8fd9\u90e8\u5206\u505a\u9488\u5bf9\u6027\u4f18\u5316\uff0c\u4f1a\u83b7\u53d6\u8f83\u5927\u5e45\u5ea6\u7684\u6027\u80fd\u63d0\u5347\u3002\u5bf9\u8fd9\u90e8\u5206\u7684\u4f18\u5316\u6211\u4eec\u5206\u4e24\u6b65\u8fdb\u884c\uff0c\u7b2c\u4e00\u6b65\u4f18\u5316\u4e86seq_length=512\u7684Multi-Head Attention\u4e2d\u7684softmax\u7684\u5b9e\u73b0\uff0c\u7b2c\u4e8c\u6b65\u5219\u662f\u9488\u5bf9seq_length\u4e3a35\/512\u7684case\u505a\u4e86\u7c7b\u4f3cTensorRT\u9488\u5bf9seq_length=64\u7684\u60c5\u5f62\u4e0b\u505a\u7684\u878d\u5408\u3002\u00a0\u4e0b\u56fe\u662fseq_length\u4e3a512\u7684\u60c5\u51b5\u4e0b\uff0c\u672a\u4f18\u5316\u524d\u7684\u4e00\u4e2aTransformer\u5c42\u8c03\u7528\u7684Kernel\uff0c\u5176\u4e2d\u7eff\u6846\u4e2d\u7684kernel\u4e3a\u7b2c\u4e00\u6b65\u4f18\u5316\u7684Kernel\uff0c\u7ea2\u6846\u4e2d\u7684\u90e8\u5206\u5219\u662f\u7b2c\u4e8c\u6b65\u4f18\u5316\u4e2d\u878d\u5408\u4e3a\u4e00\u4e2a\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488920\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a04. Enable CUDA-GraphCUDA Graph\u662f\u4e00\u4e2a\u628a\u6240\u6709kernel\u7b97\u5b50\u7ed3\u5408\uff08capture\uff09\u6210Graph\uff0c\u7136\u540e\u6574\u4f53launch\u8fd9\u4e2aGraph\uff0c\u51cf\u5c11\u9891\u7e41\u7684kernel launch\u6765\u5e26\u7684\u5f00\u9500\u4ee5\u53cakernel\u4e2d\u95f4\u7684gap\u3002\u4e0b\u56fe\u662f\u666e\u901a\u7684kernel\u6267\u884c\u548cGraph\u6267\u884c\u7684\u533a\u522b\u3002\u53ef\u4ee5\u770b\u51fa\uff0c\u5728kernel\u6267\u884c\u65f6\u56e0\u4e3a\u9700\u8981CPU\u548cGPU\u5207\u6362\uff0c\u9020\u6210\u5c0f\u7b97\u5b50\u95f4\u4f1a\u6709\u6bd4\u8f83\u5927\u7684GPU idle\u65f6\u95f4\uff08gap\u5f15\u8d77\uff09\uff0c\u540c\u65f6\u5982\u679c\u5c0f\u7b97\u5b50\u6267\u884c\u7684\u65f6\u95f4\u6bd4\u8f83\u77ed\uff0c\u90a3\u4e48launch\u7684\u65f6\u95f4\u5360\u6bd4\u5c31\u6210\u4e86\u5927\u5934\uff0c\u9020\u6210GPU\u5229\u7528\u7387\u4f4e\u4e0b\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488921\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u5728CUDA11\u7248\u672c\u548cAmpere\u67b6\u6784\u4e0b\uff0cCUDA Graph\u672c\u8eab\u505a\u4e86\u5f88\u5927\u7684\u6539\u8fdb\uff0c\u5728CUDA\u5185\u90e8\u505a\u4e86\u5e76\u884c\u5316\u5904\u7406\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488922\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u4ece\u673a\u5236\u56fe\u53ef\u4ee5\u770b\u5230\uff0cCUDA\u5185\u90e8\u5728\u6267\u884cGraph\u65f6\uff0c\u67e5\u627e\u4f9d\u8d56\u5173\u7cfb\uff0c\u5982\u679c\u53d1\u73b0\u65e0\u76f4\u63a5\u4f9d\u8d56\u4e14\u76ee\u524d\u8d44\u6e90\u6ee1\u8db3\u7b97\u5b50\u6267\u884c\uff08\u6bd4\u5982\uff1astream processor\/share memory\/register\u7b49\uff09\u5219\u5e76\u884c\u6267\u884c\uff0c\u63d0\u9ad8GPU\u5229\u7528\u7387\u3002\u540c\u65f6\u6211\u4eec\u4e5f\u5bf9CUDA Graph\u4e0b\u505a\u4e86\u4e00\u4e9b\u4f18\u5316\uff08\u6bd4\u5982\u589e\u52a0memory cache\u673a\u5236\u3001graph min update\u3001multi-stream\u5904\u7406\u7b49\uff09\u66f4\u597d\u5730\u63d0\u5347\u4e86\u6027\u80fd\u3002\u00a0Enable CUDA Graph\u6709\u4e24\u79cd\u65b9\u5f0f\uff0c\u5176\u4e2d\u6d41\u6355\u83b7\u63d0\u4f9b\u4e86\u4e00\u79cd\u4ece\u73b0\u6709\u7684\u57fa\u4e8e\u6d41\u7684 API \u521b\u5efa\u56fe\u7684\u673a\u5236\u3002\u5728QTC\u6a21\u578b\u4e2d\uff0c\u901a\u8fc7\u6d41\u6355\u83b7\u7684\u65b9\u5f0f\u6765\u52a0\u901f\u57fa\u4e8eTensorRT\u6784\u5efa\u7684\u56fe\u3002Enable CUDA Graph\u540e\uff0c\u5728A100\u4e0a\u5ef6\u8fdf\u67093%\u4e0b\u964d\uff0c\u541e\u5410\u63d0\u9ad8\u4e864%\uff0c\u5728GPU\u4f7f\u7528\u7387\u9ad8\u65f6latency\u4e5f\u66f4\u7a33\u5b9a\u3002\u00a0\u4e1a \u52a1 \u7ed3 \u679c\u00a0\u4e0b\u56fe\u4e2d\uff0cFaster-Transformer\u662f\u6307UC\u539f\u6709\u7684\u57fa\u4e8eFaster-Transformer\u7684\u5728GPU\u4e0a\u7684\u89e3\u51b3\u65b9\u6848\uff1bAIACC-Optimization\u662f\u6211\u4eec\u4f18\u5316\u540e\u7684\u6027\u80fd\u6570\u636e\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042488923\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0\u5728UC\u7684\u751f\u4ea7\u73af\u5883\u4e2d\uff0cA100\u673a\u578b\u4e0aQPS\u8fbe\u5230\u4e861330\uff0c\u4e3a\u6700\u521dFaster-Transformer\u7248\u672c\u76843.8\u500d\uff0c\u8d85\u51fa\u4e86\u9884\u8bbe\u7684\u76ee\u6807\u3002\u00a0\u76ee\u524d\uff0c\u91c7\u7528AIACC\u8fd9\u4e00\u7cfb\u5217\u4f18\u5316\u7684QTC\u6a21\u578b\u5df2\u7ecf\u5728\u90e8\u5206\u641c\u7d22\u4e1a\u52a1\u4e0a\u7ebf\u4f7f\u7528\uff0c\u540e\u7eed\u5373\u5c06\u5927\u89c4\u6a21\u7ebf\u4e0a\u4f7f\u7528\u3002\u6309\u6700\u4f4e\u7684\u9884\u671f\u4f7f\u7528\u91cf\u6765\u8ba1\u7b97\uff0c\u6bcf\u5e74\u80fd\u8282\u7701\u8fbe\u6570\u5343\u4e07\u7684\u6210\u672c\u3002\u00a0\u5728\u8fd9\u6b21\u5408\u4f5c\u4e2d\uff0cUC\u641c\u7d22\u7cbe\u6392\u5e94\u7528\u5728\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u4e0a\u8fd0\u884c\u7684\u6027\u4ef7\u6bd4\u5f97\u5230\u4e86\u663e\u8457\u63d0\u5347\uff0c\u53ef\u4ee5\u628a\u4e1a\u52a1\u6279\u91cf\u90e8\u7f72\u5728\u751f\u6001\u6210\u719f\u7684\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u4e0a\u3002\u4e91\u5e73\u53f0\u5f39\u6027\u6269\u7f29\u5bb9\u7684\u4f18\u70b9\u8ba9UC\u53ef\u4ee5\u6839\u636e\u4e1a\u52a1\u9700\u6c42\uff0c\u6309\u9700\u4ece\u963f\u91cc\u4e91\u4e0a\u83b7\u53d6GPU\u4e91\u670d\u52a1\u5668\u8d44\u6e90\u3002\u6b64\u5916\uff0c\u63a8\u7406\u7528\u7684\u673a\u578b\u548c\u8bad\u7ec3\u7528\u7684\u673a\u578b\u7edf\u4e00\uff0c\u8ba9\u751f\u4ea7\u94fe\u8def\u4e0a\u4ece\u7b97\u6cd5\u5230\u4f18\u5316\u518d\u5230\u90e8\u7f72\u7684\u5de5\u7a0b\u5e08\u5408\u4f5c\u66f4\u65b9\u4fbf\uff0c\u4e5f\u65b9\u4fbf\u540e\u7eed\u5728\u7ebf\u4e0a\u5b9e\u73b0\u8bad\u7ec3\u4e0e\u63a8\u7406\u4efb\u52a1\u7684\u6df7\u5408\u90e8\u7f72\uff0c\u6700\u5927\u5316GPU\u8d44\u6e90\u7684\u5229\u7528\u3002\u00a0\u963f\u91cc\u4e91AIACC\u9488\u5bf9\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u786c\u4ef6\u505a\u4e86\u6df1\u5165\u7684\u4f18\u5316\uff0c\u8fbe\u6210\u4e86\u5728\u4e0d\u989d\u5916\u5f15\u5165\u65b0\u786c\u4ef6\u5e73\u53f0\u7684\u60c5\u51b5\u4e0b\uff0c\u7528\u963f\u91cc\u4e91GPU\u8ba1\u7b97\u5b9e\u4f8b\u6765\u6ee1\u8db3UC\u6781\u81f4\u6027\u4ef7\u6bd4\u9700\u6c42\u7684\u76ee\u6807\u3002\u6b64\u5916\uff0c\u7528UC\u771f\u5b9e\u5e94\u7528\u6253\u78e8\u7684\u4f18\u5316\u6027\u80fd\u5de5\u4f5c\u90fd\u6c89\u6dc0\u5230AIACC\u4e2d\uff0c\u540e\u7eed\u53ef\u4ee5\u89c4\u6a21\u5316\u670d\u52a1\u66f4\u591a\u7684\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u7684\u5ba2\u6237\u3002\u00a0\u70b9\u51fb\u8fd9\u91cc\uff0c\u4e86\u89e3\u963f\u91cc\u4e91GPU\u4e91\u670d\u52a1\u5668\u3002\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=7g84rDQDLVXt5jdV51cK0Q%3D%3D.tSNF8VETbu86PpxLUoNvI0buwNsBiCWVucR6uQcWxoSuUpJRrupdzzJtE9tS9ZV%2B\" rel=\"nofollow\">https:\/\/click.aliyun.com\/m\/10&#8230;<\/a>\u672c\u6587\u4e3a\u963f\u91cc\u4e91\u539f\u521b\u5185\u5bb9\uff0c\u672a\u7ecf\u5141\u8bb8\u4e0d\u5f97\u8f6c\u8f7d\u3002<\/p>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000042488914\">https:\/\/segmentfault.com\/a\/1190000042488914<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u7528\u963f\u91cc\u4e91GPU\u8ba1\u7b97\u5b9e\u4f8b\u6765\u6ee1\u8db3UC\u6781\u81f4\u6027\u4ef7\u6bd4\u9700\u6c42\u6587\u4e28\u963f\u91cc\u4e91\u795e\u9f99\u8ba1\u7b97\u5e73\u53f0AI\u52a0\u901f\u56e2\u961f &#038; UC\u641c [&hellip;]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[],"tags":[],"class_list":["post-10006491","post","type-post","status-publish","format-standard","hentry"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006491","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10006491"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006491\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10006491"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10006491"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10006491"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}