{"id":571,"date":"2026-07-08T03:16:03","date_gmt":"2026-07-07T19:16:03","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=571"},"modified":"2026-07-09T11:13:33","modified_gmt":"2026-07-09T03:13:33","slug":"%e9%9b%86%e6%88%90%e6%b5%8b%e8%af%95%e4%b8%8e%e6%80%a7%e8%83%bd%e7%9c%9f%e7%9b%b8%ef%bc%9a%e4%b8%80%e4%b8%aa%e4%ba%ba%e5%81%9a%e7%9a%84%e6%a1%86%e6%9e%b6%ef%bc%8c%e5%87%ad%e4%bb%80%e4%b9%88%e8%83%bd","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/571\/","title":{"rendered":"(25) \u96c6\u6210\u6d4b\u8bd5\u4e0e\u6027\u80fd\u5bc6\u7801\uff1a\u4e00\u4e2a\u4eba\u505a\u7684\u6846\u67b6\uff0c\u51ed\u4ec0\u4e48\u80fd\u6bd4PyTorch\u5feb\uff1f"},"content":{"rendered":"\n

\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341\u4e94<\/h2>\n\n\n\n

\u4ece\u7b2c\u4e00\u7bc7\u63d0\u51fa\u201c\u4e00\u4e2a\u4eba\u7528 AI \u5199\u51fa\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u6709\u53ef\u80fd\u6bd4 PyTorch \u66f4\u5feb\u201d\u8fd9\u4e2a\u547d\u9898\u5f00\u59cb\uff0c\u6211\u4eec\u9646\u9646\u7eed\u7eed\u8bb2\u4e86\u9759\u6001\u56fe\u7f16\u8bd1\u3001MemoryPlan\u3001CUDA Graph\u3001\u6570\u636e\u683c\u5f0f\u3001\u878d\u5408\u7b97\u5b50\u3001\u591a\u6d41\u5e76\u53d1\u3001\u5206\u5e03\u5f0f\u901a\u4fe1\u3001Philox \u786e\u5b9a\u6027\u8bad\u7ec3\u2026\u2026\u628a\u8fd9\u4e9b\u90e8\u4ef6\u4e32\u8d77\u6765\uff0c\u5176\u5b9e\u90fd\u5728\u56de\u7b54\u540c\u4e00\u4e2a\u95ee\u9898\uff1a\u4e00\u4e2a\u4ece\u96f6\u5f00\u59cb\u7684\u5c0f\u56e2\u961f\u6846\u67b6\uff0c\u5230\u5e95\u80fd\u4e0d\u80fd\u5728\u771f\u5b9e\u7684\u8bad\u7ec3\u541e\u5410\u4e0a\u80dc\u8fc7\u5f53\u4eca\u6700\u4e3b\u6d41\u7684\u751f\u4ea7\u7ea7\u6846\u67b6\uff1f<\/strong><\/p>\n\n\n\n

\u8fd9\u4e2a\u95ee\u9898\u7684\u7b54\u6848\uff0c\u4e0d\u80fd\u9760\u60c5\u6000\uff0c\u4e5f\u4e0d\u80fd\u9760\u67b6\u6784\u56fe\uff0c\u53ea\u80fd\u9760\u96c6\u6210\u6d4b\u8bd5\u4e0e\u6027\u80fd\u6570\u636e<\/strong>\u3002\u6240\u4ee5\u672c\u7bc7\u8981\u628a\u524d\u9762\u6240\u6709\u7684\u8bbe\u8ba1\u644a\u5f00\uff0c\u7528\u4e24\u4e2a\u5177\u4f53\u7684\u7aef\u5230\u7aef\u793a\u4f8b\u2014\u2014MNIST \u4e0a\u7684\u7b80\u5355 MLP\uff0c\u4ee5\u53ca A100\u00d78 \u4e0a\u7684 ImageNet VGG16BN\u2014\u2014\u6765\u8bf4\u660e Tech-Renaissance \u7684\u6027\u80fd\u771f\u76f8\uff0c\u4e5f\u987a\u4fbf\u804a\u804a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u505a\u6d4b\u8bd5\u4e0e\u516c\u5e73\u5bf9\u6bd4\u8fd9\u4ef6\u4e8b\u672c\u8eab\u7684\u95e8\u9053\u3002<\/p>\n\n\n\n

\u4e00\u3001\u6d4b\u8bd5\u4e0e Benchmark\uff1a\u6df1\u5ea6\u5b66\u4e60\u7684\u201c\u5c3a\u5b50\u201d\u6709\u8bb2\u7a76<\/h2>\n\n\n\n

\u5728\u8bb2\u6570\u636e\u4e4b\u524d\uff0c\u5fc5\u987b\u5148\u8bb2\u4e00\u4e2a\u5e38\u8bc6\uff1a\u8861\u91cf\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u6027\u80fd\uff0c\u8fdc\u6bd4\u8dd1\u4e00\u4e2a\u811a\u672c\u770b\u8017\u65f6\u590d\u6742\u3002<\/strong><\/p>\n\n\n\n

\u4e00\u4e2a\u6846\u67b6\u8981\u5bf9\u5916\u58f0\u79f0\u201c\u6211\u5f88\u5feb\u201d\uff0c\u81f3\u5c11\u5f97\u8de8\u8fc7\u4e09\u9053\u95e8\u69db\u3002\u7b2c\u4e00\u9053\u95e8\u69db\u662f\u6b63\u786e\u6027<\/strong>\uff1a\u7b97\u5b50\u6570\u503c\u5fc5\u987b\u548c\u53c2\u8003\u5b9e\u73b0\u4e00\u81f4\uff0c\u5426\u5219\u8bad\u7ec3\u6536\u655b\u66f2\u7ebf\u518d\u6f02\u4eae\u4e5f\u662f\u4f2a\u547d\u9898\u3002\u7b2c\u4e8c\u9053\u95e8\u69db\u662f\u53ef\u590d\u73b0\u6027<\/strong>\uff1a\u540c\u6837\u7684\u4ee3\u7801\u3001\u540c\u6837\u7684\u79cd\u5b50\u3001\u540c\u6837\u7684\u786c\u4ef6\uff0c\u8dd1\u4e24\u6b21\u7ed3\u679c\u5e94\u8be5\u4e00\u81f4\uff0c\u5426\u5219\u6027\u80fd\u6570\u5b57\u5c31\u4e0d\u53ef\u4fe1\u3002\u7b2c\u4e09\u9053\u95e8\u69db\u624d\u662f\u516c\u5e73\u5bf9\u6bd4<\/strong>\uff1a\u4f60\u8981\u548c PyTorch \u6bd4\uff0c\u5c31\u4e0d\u80fd\u8ba9\u5b83\u88f8\u5954\uff0c\u800c\u5f97\u628a\u5b83\u4e5f\u8c03\u5230\u6700\u5f3a\u6a21\u5f0f\uff1b\u4f60\u8981\u6bd4\u8bad\u7ec3\u541e\u5410\uff0c\u5c31\u5f97\u628a\u4e24\u8fb9\u7684\u7f16\u8bd1\u65f6\u95f4\u3001\u6570\u636e\u52a0\u8f7d\u3001\u9884\u70ed\u3001\u8ba1\u65f6\u53e3\u5f84\u90fd\u5bf9\u9f50\u3002<\/p>\n\n\n\n

\u8fd9\u4e09\u9053\u95e8\u69db\u5bf9\u5e94\u4e09\u79cd\u6d4b\u8bd5\u3002<\/p>\n\n\n\n

\/\/ \u6d4b\u8bd5\u5206\u5c42\u7684\u4e00\u79cd\u6734\u7d20\u8868\u793a\nenum class TestLayer {\n    CORRECTION,   \/\/ \u5355\u5143\/\u7b97\u5b50\u6b63\u786e\u6027\uff1a\u9010\u4e2a\u7b97\u5b50 vs PyTorch \u53c2\u8003\u503c\n    OP,           \/\/ \u529f\u80fd\u4e0e\u8fb9\u754c\uff1a\u4e0d\u540c dtype\u3001shape\u3001Region \u914d\u7f6e\n    PERF,         \/\/ \u5fae\u57fa\u51c6\uff1a\u5355\u4e2a\u7b97\u5b50\u6216\u7b97\u5b50\u7ec4\u5408\u7684\u6027\u80fd\n    EXAMPLE       \/\/ \u96c6\u6210\u6d4b\u8bd5\uff1a\u7aef\u5230\u7aef\u8bad\u7ec3\u793a\u4f8b\u4e0e\u516c\u5e73\u5bf9\u6bd4\n};<\/pre>\n\n\n\n
    \n
  • \u5355\u5143\u6d4b\u8bd5 \/ \u7b97\u5b50\u6b63\u786e\u6027\u6d4b\u8bd5<\/strong>\u9a8c\u8bc1\u6bcf\u4e2a\u7b97\u5b50\u524d\u5411\u3001\u53cd\u5411\u7684\u6570\u503c\u7cbe\u5ea6\uff1b<\/li>\n\n\n\n
  • \u96c6\u6210\u6d4b\u8bd5<\/strong>\u9a8c\u8bc1\u4ece\u6570\u636e\u52a0\u8f7d\u5230\u6a21\u578b\u8bad\u7ec3\u518d\u5230\u6307\u6807\u8f93\u51fa\u7684\u5b8c\u6574\u94fe\u8def\u80fd\u8dd1\u901a\u3001\u8dd1\u5bf9\uff1b<\/li>\n\n\n\n
  • \u6027\u80fd\u6d4b\u8bd5 \/ Benchmark<\/strong>\u5219\u5728\u4fdd\u8bc1\u524d\u4e24\u8005\u6210\u7acb\u7684\u524d\u63d0\u4e0b\uff0c\u5ea6\u91cf\u7aef\u5230\u7aef\u8bad\u7ec3\u541e\u5410\u6216\u8017\u65f6\u3002<\/li>\n<\/ul>\n\n\n\n
    \u5c11\u4e86\u4efb\u4f55\u4e00\u5c42\uff0c\u6027\u80fd\u6570\u5b57\u90fd\u53ef\u80fd\u662f\u6c99\u4e0a\u5efa\u5854\u3002<\/p>\n\n\n\n
    \u4e1a\u754c\u5bf9\u6b64\u5176\u5b9e\u65e9\u6709\u5171\u8bc6\u3002MLPerf \u4e4b\u6240\u4ee5\u80fd\u6210\u4e3a\u6700\u53d7\u8ba4\u53ef\u7684 AI \u6027\u80fd\u57fa\u51c6\u4e4b\u4e00\uff0c\u6838\u5fc3\u5e76\u4e0d\u5728\u4e8e\u5b83\u9009\u4e86\u54ea\u4e9b\u6a21\u578b\uff0c\u800c\u5728\u4e8e\u5b83\u5236\u5b9a\u4e86\u4e00\u6574\u5957\u4e25\u683c\u7684\u516c\u5e73\u89c4\u5219\uff1a\u56fa\u5b9a\u6570\u636e\u96c6\u3001\u56fa\u5b9a\u8d85\u53c2\u6570\u3001\u56fa\u5b9a\u6536\u655b\u76ee\u6807\u3001\u8be6\u7ec6\u7684\u8f6f\u786c\u4ef6\u73af\u5883\u62ab\u9732\u3001\u53ef\u5ba1\u8ba1\u7684\u63d0\u4ea4\u4ee3\u7801\u3002\u6b63\u5982 MLPerf Training Benchmark \u8bba\u6587\u91cc\u5f3a\u8c03\u7684\uff0c\u8bad\u7ec3\u6027\u80fd\u5bf9\u6bd4\u5fc5\u987b\u201c\u5728\u516c\u5e73\u7684\u6761\u4ef6\u4e0b\u9f13\u52b1\u521b\u65b0\u201d\uff0c\u5426\u5219\u52a0\u901f\u6bd4\u7684\u6570\u5b57\u5c31\u6ca6\u4e3a\u5404\u8bf4\u5404\u8bdd\u7684\u5e7f\u544a\u8bcd\u3002<\/p>\n\n\n\n
    \u5b9e\u9645\u64cd\u4f5c\u4e2d\uff0c\u4e0d\u516c\u5e73\u5bf9\u6bd4\u7684\u9677\u9631\u968f\u5904\u53ef\u89c1\u3002\u6700\u5178\u578b\u7684\u51e0\u79cd\u5305\u62ec\uff1a<\/p>\n\n\n\n
      \n
    • \u7f16\u8bd1\u65f6\u95f4\u4e0d\u9694\u79bb<\/strong>\uff1a\u628a torch.compile<\/code> \u7684\u9996\u6b21\u7f16\u8bd1\u6216 Tech-Renaissance \u7684 task.compile()<\/code> \u7b97\u8fdb\u603b\u8017\u65f6\uff0c\u6216\u8005\u53cd\u8fc7\u6765\u6545\u610f\u4e0d\u7b97\uff0c\u90fd\u4f1a\u663e\u8457\u626d\u66f2\u7ed3\u679c\uff1b<\/li>\n\n\n\n
    • \u6570\u636e\u52a0\u8f7d\u4e0d\u5bf9\u7b49<\/strong>\uff1a\u4e00\u8fb9\u7528 pin_memory + persistent_workers<\/code>\uff0c\u53e6\u4e00\u8fb9\u7528\u9ed8\u8ba4\u5355\u7ebf\u7a0b\uff1b<\/li>\n\n\n\n
    • \u8d85\u53c2\u6570\u5fae\u5c0f\u5dee\u5f02<\/strong>\uff1aweight decay \u662f\u5426\u6392\u9664 bias\u3001\u5b66\u4e60\u7387 warm-up \u533a\u95f4\u3001label smoothing\u3001\u6570\u636e\u589e\u5f3a\u987a\u5e8f\uff0c\u90fd\u53ef\u80fd\u5f71\u54cd\u6536\u655b\u901f\u5ea6\uff1b<\/li>\n\n\n\n
    • \u5c3e batch \u672a\u9884\u70ed<\/strong>\uff1a\u52a8\u6001\u56fe\u7f16\u8bd1\u5668\u5728\u9047\u5230\u4e0d\u5b8c\u6574 batch \u65f6\u53ef\u80fd\u91cd\u65b0\u7f16\u8bd1\uff0c\u82e5\u53ea\u5728\u5b8c\u6574 batch \u4e0a\u9884\u70ed\uff0c\u8ba1\u65f6\u5faa\u73af\u91cc\u4f1a\u591a\u51fa\u4e00\u7b14\u5f00\u9500\uff1b<\/li>\n\n\n\n
    • \u6307\u6807\u53e3\u5f84\u4e0d\u7edf\u4e00<\/strong>\uff1a\u6bd4\u541e\u5410\u3001\u6bd4 time-to-train\u3001\u6bd4\u6bcf epoch \u8017\u65f6\uff0c\u4e0d\u540c\u53e3\u5f84\u4e0b\u7684\u7ed3\u8bba\u53ef\u80fd\u5b8c\u5168\u4e0d\u540c\u3002<\/li>\n<\/ul>\n\n\n\n
      \u8fd9\u91cc\u8fd8\u9700\u8981\u533a\u5206\u51e0\u79cd\u5e38\u7528\u6307\u6807\u3002\u8bad\u7ec3\u6846\u67b6\u901a\u5e38\u5173\u6ce8\u541e\u5410\u7387<\/strong>\uff08images\/sec \u6216 samples\/sec\uff09\uff0c\u5b83\u53cd\u6620\u5355\u4f4d\u65f6\u95f4\u5185\u80fd\u5904\u7406\u591a\u5c11\u6837\u672c\uff1btime-to-train<\/strong> \u5219\u8981\u6c42\u6a21\u578b\u8fbe\u5230\u67d0\u4e2a\u76ee\u6807\u7cbe\u5ea6\u6240\u9700\u7684\u603b\u65f6\u95f4\uff0c\u66f4\u9002\u5408\u7aef\u5230\u7aef\u6bd4\u62fc\uff1b\u6bcf epoch \u8017\u65f6<\/strong>\u6700\u76f4\u89c2\uff0c\u4f46\u4e0d\u540c\u6846\u67b6\u7684\u6570\u636e\u52a0\u8f7d\u548c\u9a8c\u8bc1\u7b56\u7565\u5dee\u5f02\u4f1a\u653e\u5927\u8fd9\u4e2a\u6570\u503c\u3002\u81f3\u4e8e\u63a8\u7406\u573a\u666f\uff0clatency \u548c throughput \u7684\u6743\u91cd\u53c8\u5b8c\u5168\u4e0d\u540c\u3002\u6211\u4eec\u5728 Tech-Renaissance \u7684\u5bf9\u6bd4\u4e2d\u4e3b\u8981\u4f7f\u7528\u541e\u5410\u7387\u548c\u6bcf epoch \u8017\u65f6\uff0c\u540c\u65f6\u62a5\u544a\u6700\u7ec8\u7cbe\u5ea6\uff0c\u786e\u4fdd\u8bfb\u8005\u65e2\u770b\u5230\u901f\u5ea6\uff0c\u4e5f\u770b\u5230\u901f\u5ea6\u6ca1\u6709\u4ee5\u727a\u7272\u6b63\u786e\u6027\u4e3a\u4ee3\u4ef7\u3002<\/p>\n\n\n\n
      \u6240\u4ee5\u6211\u4eec\u7684\u539f\u5219\u5f88\u7b80\u5355\uff1a\u5148\u4fdd\u8bc1\u6b63\u786e\uff0c\u518d\u8c08\u901f\u5ea6\uff1b\u8981\u505a\u5bf9\u6bd4\uff0c\u5c31\u628a\u5bf9\u624b\u4e5f\u6b66\u88c5\u5230\u7259\u9f7f\u3002<\/strong><\/p>\n\n\n\n
      \u4e8c\u3001Tech-Renaissance \u7684\u6d4b\u8bd5\u5206\u5c42<\/h2>\n\n\n\n
      Tech-Renaissance \u7684\u6d4b\u8bd5\u76ee\u5f55\u5e76\u4e0d\u590d\u6742\uff0c\u4f46\u5206\u5de5\u975e\u5e38\u660e\u786e\uff1a<\/p>\n\n\n\n
      \u76ee\u5f55<\/th>\u804c\u8d23<\/th>\u5178\u578b\u793a\u4f8b<\/th><\/tr><\/thead>
      tests\/correction<\/code><\/td>\u7b97\u5b50\u6570\u5b66\u6b63\u786e\u6027\u9a8c\u8bc1\uff0c\u901a\u5e38\u4e0e PyTorch \u751f\u6210\u53c2\u8003\u6570\u636e\u9010\u5143\u7d20\u5bf9\u6bd4<\/td>test_softmax_ce.cpp<\/code> \u8c03\u7528 PyTorch \u811a\u672c\u751f\u6210 logits\/labels\/loss\/\u68af\u5ea6\u53c2\u8003\u503c\uff0c\u518d\u4e0e\u672c\u6846\u67b6\u7ed3\u679c\u6bd4\u8f83 MSE<\/td><\/tr>
      tests\/perf<\/code><\/td>\u5355\u4e2a\u7b97\u5b50\u6216\u7b97\u5b50\u7ec4\u5408\u7684\u6027\u80fd\u5fae\u57fa\u51c6<\/td>test_softmax_ce_perf.cpp<\/code>\u3001perf_cbr_fwd.cpp<\/code> \u7b49<\/td><\/tr>
      tests\/op<\/code><\/td>\u7b97\u5b50\u529f\u80fd\u4e0e\u8fb9\u754c case \u6d4b\u8bd5<\/td>\u8986\u76d6\u4e0d\u540c dtype\u3001shape\u3001Region \u914d\u7f6e<\/td><\/tr>
      tests\/example<\/code><\/td>\u7aef\u5230\u7aef\u8bad\u7ec3\u793a\u4f8b\uff0c\u7528\u4e8e\u96c6\u6210\u9a8c\u8bc1\u4e0e\u516c\u5e73\u5bf9\u6bd4<\/td>mlp_mnist.cpp<\/code>\u3001test_vgg16bn.cpp<\/code><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
      \u76ee\u524d tests\/correction<\/code> \u76ee\u5f55\u8986\u76d6\u4e86 SoftmaxCE\u3001GAP\u3001Flatten+FC+ReLU+FC\u3001Range Clear\u3001D2D Copy\u3001FP32\/FP16 \u4e92\u8f6c\u3001NaN \u68c0\u6d4b\u3001\u591a\u79cd\u4f18\u5316\u5668\uff08SGD\/Momentum\/Nesterov\/Adam\/AdamW\/LARS\uff09\u7684 weight\/bias \u8def\u5f84\u7b49\u5173\u952e\u7b97\u5b50\u3002tests\/perf<\/code> \u5219\u5bf9 SoftmaxCE\u3001GAP\u3001CBR\u3001MaxPool\u3001Dropout\u3001Cast\u3001Clear \u7b49\u7b97\u5b50\u505a\u4e86\u5fae\u57fa\u51c6\uff0c\u65b9\u4fbf\u5728\u4f18\u5316\u524d\u540e\u5feb\u901f\u89c2\u5bdf\u5355\u70b9\u53d8\u5316\u3002tests\/op<\/code> \u8fdb\u4e00\u6b65\u8986\u76d6\u8fb9\u754c case\uff0c\u6bd4\u5982\u4e0d\u540c shape\u3001dtype\u3001Region \u7ec4\u5408\u4e0b\u7684\u884c\u4e3a\u3002<\/p>\n\n\n\n
      \u4ee5 tests\/correction\/test_softmax_ce.cpp<\/code> \u4e3a\u4f8b\uff0c\u5b83\u4f1a\u5148\u8c03\u7528\u4e00\u4e2a PyTorch \u811a\u672c\u751f\u6210\u53c2\u8003\u5f20\u91cf\uff1a<\/p>\n\n\n\n
      std::ostringstream py;\npy << TR_PYTHON_EXECUTABLE << \" \"\n   << TR_PROJECT_ROOT << \"\/tests\/correction\/test_softmax_ce.py\"\n   << \" --batch \" << cfg.batch\n   << \" --num_classes \" << cfg.num_classes\n   << \" --seed \" << cfg.seed\n   << \" --dtype \" << py_dtype;\n\nTR_CHECK(std::system(py.str().c_str()) == 0, RuntimeError,\n         \"Python failed. Command: \" << py.str());<\/pre>\n\n\n\n
      \u7136\u540e\u52a0\u8f7d\u53c2\u8003\u6570\u636e\uff0c\u6784\u9020\u672c\u6846\u67b6\u7684 ComputationGraph<\/code>\uff0c\u5206\u522b\u5728 CPU\u3001GPU FP32\u3001GPU AMP \u4e09\u79cd\u6a21\u5f0f\u4e0b\u8dd1\u524d\u5411\u548c\u53cd\u5411\uff0c\u6700\u540e\u4e0e\u53c2\u8003\u503c\u6bd4\u8f83 MSE\u3002\u8fd9\u79cd\u201c\u7528 PyTorch \u5f53\u88c1\u5224\u201d\u7684\u601d\u8def\u8d2f\u7a7f\u4e86\u5927\u90e8\u5206\u7b97\u5b50\u6d4b\u8bd5\uff1a\u65e2\u7136\u5927\u5bb6\u5bf9 PyTorch \u7684\u6570\u503c\u7ed3\u679c\u6709\u57fa\u672c\u4fe1\u4efb\uff0c\u90a3\u5c31\u8ba9 PyTorch \u751f\u6210\u53c2\u8003\uff0c\u672c\u6846\u67b6\u53bb\u903c\u8fd1\u5b83\u3002<\/p>\n\n\n\n
      \u96c6\u6210\u6d4b\u8bd5\u5219\u66f4\u8fdb\u4e00\u6b65\u3002tests\/example\/mlp_mnist.cpp<\/code> \u548c tests\/example\/test_vgg16bn.cpp<\/code> \u4e0d\u662f\u5355\u72ec\u6d4b\u67d0\u4e2a\u7b97\u5b50\uff0c\u800c\u662f\u628a\u6570\u636e\u52a0\u8f7d\u3001\u9884\u5904\u7406\u3001\u6a21\u578b\u3001\u635f\u5931\u51fd\u6570\u3001\u4f18\u5316\u5668\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u3001\u9a8c\u8bc1\u6307\u6807\u6574\u4e2a\u94fe\u8def\u8dd1\u901a\uff0c\u5e76\u548c\u5bf9\u5e94\u7684 PyTorch \u811a\u672c\u5728\u540c\u7b49\u6761\u4ef6\u4e0b\u5bf9\u6bd4\u8bad\u7ec3\u901f\u5ea6\u4e0e\u6700\u7ec8\u7cbe\u5ea6\u3002\u5b83\u4eec\u65e2\u662f\u793a\u4f8b\uff0c\u4e5f\u662f\u6027\u80fd\u57fa\u51c6\u3002<\/p>\n\n\n\n
      \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u672c\u6846\u67b6\u628a\u786e\u5b9a\u6027<\/strong>\u4e5f\u5f53\u4f5c\u6d4b\u8bd5\u7684\u4e00\u90e8\u5206\u3002\u5728 docs\/MNIST_BEST_ADAMW_TR4_VS_PYTORCH.md<\/code> \u4e2d\uff0c\u6211\u4eec\u8bb0\u5f55\u4e86 SoftmaxCE\u3001MaxPool \u53cd\u5411\u3001Dropout \u4e09\u4e2a\u786e\u5b9a\u6027\u4fee\u590d\uff1a<\/p>\n\n\n\n
        \n
      • SoftmaxCE \u7684\u635f\u5931\u5f52\u7ea6\u4ece atomicAdd<\/code> \u6539\u4e3a partial-sum \u7f13\u51b2\u533a + \u5355 block \u56fa\u5b9a\u987a\u5e8f\uff1b<\/li>\n\n\n\n
      • MaxPool \u53cd\u5411\u5728\u91cd\u53e0\u7a97\u53e3\u573a\u666f\u4e0b\u53cd\u8f6c\u904d\u5386\u65b9\u5411\uff0c\u907f\u514d\u591a\u7ebf\u7a0b\u7ade\u4e89\u540c\u4e00 dx<\/code> \u5730\u5740\uff1b<\/li>\n\n\n\n
      • Dropout \u7528 Xorshift64* \u5728\u8bbe\u5907\u7aef\u6bcf\u6b21\u524d\u5411\u786e\u5b9a\u6027\u65cb\u8f6c\u79cd\u5b50\u3002<\/li>\n<\/ul>\n\n\n\n
        \u8fd9\u4e9b\u4fee\u590d\u4e0d\u662f\u4e3a\u4e86\u5237\u699c\uff0c\u800c\u662f\u4e3a\u4e86\u8ba9\u201c\u8dd1\u4e24\u6b21\u7ed3\u679c\u4e00\u81f4\u201d\u8fd9\u4ef6\u672c\u5e94\u5929\u7ecf\u5730\u4e49\u7684\u4e8b\u771f\u7684\u6210\u7acb\u3002\u4e00\u4e2a\u8fde\u53ef\u590d\u73b0\u90fd\u505a\u4e0d\u5230\u7684\u6846\u67b6\uff0c\u6ca1\u6709\u8d44\u683c\u8c08\u6027\u80fd\u5bf9\u6bd4\u3002<\/p>\n\n\n\n
        \u4e09\u3001\u793a\u4f8b\u4e00\uff1aMLP MNIST\u2014\u2014\u5c0f\u4efb\u52a1\u91cc\u7684\u5927\u5dee\u8ddd<\/h2>\n\n\n\n
        \u6211\u4eec\u5148\u770b\u6700\u7b80\u5355\u7684\u7aef\u5230\u7aef\u793a\u4f8b\uff1atests\/example\/mlp_mnist.cpp<\/code>\u3002<\/p>\n\n\n\n
        \u8fd9\u4e2a\u793a\u4f8b\u8bad\u7ec3\u4e00\u4e2a 4 \u5c42 MLP\uff08784\u21921024\u2192512\u2192256\u219210\uff09\uff0c\u4f7f\u7528 AdamW + CosineAnnealingLR + Warmup\uff0c\u5728 MNIST \u4e0a\u8dd1 100 \u4e2a epoch\u3002\u5b8c\u6574\u4ee3\u7801\u53ea\u6709 64 \u884c\uff1a<\/p>\n\n\n\n
        #include <chrono>\n#include <iomanip>\n#include <iostream>\n#include <string>\n\n#include \"renaissance.h\"\n\nusing namespace tr;\n\nint main() {\n    GLOBAL_SETTING\n        .use_gpu(\"0\")\n        .amp(true)\n        .manual_seed(123)\n        .global_batch_size(200)\n        .input_resolution(28);\n\n    PREPROCESSOR_SETTING\n        .dataset(\"mnist\", std::string(TR_PROJECT_ROOT) + \"\/data\/mnist\")\n        .download(true)\n        .preprocess_workers(8)\n        .normalization(NormMode::MNIST)\n        .train_transforms(\n            Pad(2),\n            RandomCrop(28),\n            RandomRotation(20.0f, 0),\n            RandomScale(0.8f, 1.2f),\n            RandomErasing(0.5f)\n        )\n        .commit();\n\n    BluePrint mlp = seq(\n        fc(1024, true), relu(),\n        fc(512, true), relu(),\n        fc(256, true), relu(),\n        fc(10, true)\n    );\n\n    constexpr int kTotalEpochs = 100;\n\n    DeepLearningTask task;\n    task.model(mlp)\n        .loss(CrossEntropyLoss().label_smoothing(0.1f))\n        .total_epochs(kTotalEpochs)\n        .optimizer(AdamW().weight_decay(1e-4f))\n        .scheduler(CosineAnnealingLR().base_lr(0.001f).warmup(5));\n\n    task.compile();\n\n    auto t0 = std::chrono::steady_clock::now();\n    auto result = task.run();\n    auto t1 = std::chrono::steady_clock::now();\n    auto elapsed = std::chrono::duration<double>(t1 - t0).count();\n\n    std::cout << \"\\n========== TRAINING RESULT ==========\\n\"\n              << std::fixed << std::setprecision(2)\n              << \"  Best Top-1:      \" << result.best_top1 * 100.0f << \"%\\n\"\n              << \"  Best Epoch:      \" << result.best_epoch << \"\\n\"\n              << \"  Total Time:      \" << elapsed << \" s\\n\"\n              << \"  Time per Epoch:  \" << elapsed \/ kTotalEpochs << \" s\\n\"\n              << \"=====================================\\n\";\n\n    return 0;\n}<\/pre>\n\n\n\n
        \u8fd9\u6bb5\u4ee3\u7801\u91cc\u7684\u5173\u952e\u5bf9\u9f50\u70b9\u5982\u4e0b\uff1a<\/p>\n\n\n\n
        \u914d\u7f6e\u9879<\/th>Tech-Renaissance<\/th>PyTorch<\/th>\u8bf4\u660e<\/th><\/tr><\/thead>
        \u7f51\u7edc\u7ed3\u6784<\/td>784\u21921024\u2192512\u2192256\u219210\uff0cReLU\uff0cbias=True<\/td>\u5b8c\u5168\u76f8\u540c<\/td>fc(N, true)<\/code> \u8868\u793a\u5e26 bias<\/td><\/tr>
        \u521d\u59cb\u5316<\/td>Kaiming Uniform fan_in\uff0cbias=0<\/td>\u5b8c\u5168\u76f8\u540c<\/td>\u6846\u67b6\u9ed8\u8ba4\u5373 fan_in<\/td><\/tr>
        \u4f18\u5316\u5668<\/td>AdamW\uff08\u03b21=0.9, \u03b22=0.999, \u03b5=1e-8, wd=1e-4\uff09<\/td>\u5b8c\u5168\u76f8\u540c<\/td>AdamW().weight_decay(1e-4f)<\/code>\uff0cbias \u9ed8\u8ba4\u6392\u9664 wd<\/td><\/tr>
        \u5b66\u4e60\u7387<\/td>CosineAnnealing + Warmup(5)\uff0cbase_lr=0.001\uff0c\u03b7min=1e-6<\/td>\u6570\u5b66\u7b49\u4ef7<\/td>.base_lr(0.001f).warmup(5)<\/code><\/td><\/tr>
        Label Smoothing<\/td>0.1<\/td>0.1<\/td>.label_smoothing(0.1f)<\/code><\/td><\/tr>
        Batch Size<\/td>200<\/td>200<\/td>global_batch_size(200)<\/code><\/td><\/tr>
        \u6570\u636e\u589e\u5f3a<\/td>Pad(2)\u2192RandomCrop(28)\u2192RandomRotation(20\u00b0)\u2192RandomScale(0.8~1.2)\u2192RandomErasing(0.5)<\/td>\u53c2\u6570\u5bf9\u9f50<\/td>NormMode::MNIST<\/code> \u8986\u76d6 Normalize\uff1b\u65e0 RandomAutocontrast<\/code><\/td><\/tr>
        \u968f\u673a\u79cd\u5b50<\/td>123\uff0c\u786e\u5b9a\u6027\u8bad\u7ec3<\/td>123\uff08\u591a worker \u4ecd\u6709\u6ce2\u52a8\uff09<\/td>\u8bed\u4e49\u5bf9\u9f50<\/td><\/tr>
        AMP<\/td>FP16 \u81ea\u52a8\u6df7\u5408\u7cbe\u5ea6<\/td>FP16 \u81ea\u52a8\u6df7\u5408\u7cbe\u5ea6<\/td>amp(true)<\/code><\/td><\/tr>
        TF32<\/td>\u5f00\u542f<\/td>\u5f00\u542f<\/td>\u663e\u5f0f\u542f\u7528\uff08\u811a\u672c\u4e2d\u5bf9\u9f50\uff09<\/td><\/tr>
        \u56fe\u4f18\u5316<\/td>\u9759\u6001 CUDA Graph<\/td>torch.compile(mode=\"max-autotune\")<\/code><\/td>\u53cc\u65b9\u6700\u5f3a<\/td><\/tr>
        \u7f16\u8bd1\u65f6\u95f4<\/td>task.compile()<\/code> \u5728\u8ba1\u65f6\u524d\u5b8c\u6210<\/td>dummy batch \u9884\u70ed\u540e\u91cd\u65b0\u521d\u59cb\u5316<\/td>\u5747\u4e0d\u8ba1\u5165\u8bad\u7ec3\u8017\u65f6<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
        \u5bf9\u5e94\u7684 PyTorch \u5bf9\u6bd4\u811a\u672c tests\/example\/mlp_mnist_pytorch.py<\/code> \u4e5f\u505a\u4e86\u4e25\u683c\u5bf9\u9f50\uff1a\u7f51\u7edc\u7ed3\u6784\u5b8c\u5168\u4e00\u81f4\u3001Kaiming Uniform fan_in \u521d\u59cb\u5316\u3001bias \u521d\u59cb\u5316\u4e3a 0\u3001AdamW\uff08\u03b21=0.9, \u03b22=0.999, \u03b5=1e-8, wd=1e-4\uff09\u3001CosineAnnealing + Warmup(5)\u3001label smoothing=0.1\u3001batch size=200\u3001seed=123\u3001torch.compile(mode=\"max-autotune\")<\/code>\u3001AMP \u5f00\u542f\u3001\u7f16\u8bd1\u65f6\u95f4\u901a\u8fc7 dummy batch \u9694\u79bb\u3002<\/p>\n\n\n\n
        \u6ce8\uff1aPyTorch 1.12 \u4e4b\u540e\u9ed8\u8ba4\u5173\u95ed\u4e86 cuDNN \u5377\u79ef\u7684 TF32\uff0c\u4e3a\u4e86\u4fdd\u8bc1\u516c\u5e73\uff0c\u5bf9\u6bd4\u811a\u672c\u4e2d\u901a\u8fc7 torch.backends.cudnn.allow_tf32=True<\/code> \u624b\u52a8\u5f00\u542f\u3002<\/p>\n\n\n\n
        \u7ed3\u679c\u5982\u4e0b\uff08\u6307\u6807\u4e3a\u6bcf\u79d2\u8bad\u7ec3\u7684 epoch \u6570\uff0c\u57fa\u4e8e 5 \u6b21\u8fd0\u884c\u603b\u7528\u65f6\u7684\u4e2d\u4f4d\u6570\u8ba1\u7b97\uff09\uff1a<\/p>\n\n\n\n
        \u6307\u6807<\/th>RTX4060<\/th>L20<\/th>A100<\/th>A10<\/th>RTX5090<\/th>T4<\/th>RTX4090<\/th><\/tr><\/thead>
        Tech-Renaissance (epochs\/s)<\/td>3.11<\/td>2.54<\/td>2.66<\/td>2.70<\/td>2.61<\/td>2.19<\/td>2.28<\/td><\/tr>
        PyTorch (epochs\/s)<\/td>0.45<\/td>0.33<\/td>0.33<\/td>0.34<\/td>0.25<\/td>0.16<\/td>0.17<\/td><\/tr>
        \u52a0\u901f\u6bd4<\/td>6.90x<\/td>7.64x<\/td>8.02x<\/td>8.05x<\/td>10.58x<\/td>13.27x<\/td>13.58x<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
        \u5728 7 \u4e2a\u4e0d\u540c\u7684 GPU \u5e73\u53f0\u4e0a\uff0cTech-Renaissance \u8dd1\u5b8c 100 \u4e2a epoch \u7684\u4e2d\u4f4d\u603b\u8017\u65f6\u5728 39 \u79d2\u5de6\u53f3\uff08AMP\uff09\uff0c\u800c PyTorch \u5728 220\u2013625 \u79d2\u4e4b\u95f4\uff0c\u5176\u4e2d L20\u3001A100\u3001A10 \u4e09\u4e2a\u5e73\u53f0\u96c6\u4e2d\u5728 294\u2013303 \u79d2\uff0cRTX4060 \u660e\u663e\u66f4\u5feb\uff08\u7ea6 222 \u79d2\uff09\uff0cRTX5090\uff08\u7ea6 400 \u79d2\uff09\u3001RTX4090\uff08\u7ea6 588 \u79d2\uff09\u3001T4\uff08\u7ea6 625 \u79d2\uff09\u5219\u660e\u663e\u66f4\u6162\uff0c\u5e73\u53f0\u95f4\u5dee\u5f02\u8f83\u5927\u3002\u51c6\u786e\u7387\u65b9\u9762\uff0cTech-Renaissance \u7a33\u5b9a\u5728 99.54%\uff0cPyTorch \u5728 99.43\u201399.51% \u4e4b\u95f4\uff0c\u5dee\u8ddd\u4e0d\u8d85\u8fc7 0.11 \u4e2a\u767e\u5206\u70b9\uff0c\u5904\u4e8e\u540c\u4e00\u6c34\u5e73\u3002<\/p>\n\n\n\n
        \u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0cMLP \u662f\u4e00\u4e2a\u6846\u67b6\u5f00\u9500\u653e\u5927\u5668<\/strong>\uff1a\u6a21\u578b\u5c0f\u3001batch \u5c0f\u3001\u6bcf\u4e2a epoch \u7684 GPU \u8ba1\u7b97\u91cf\u4e0d\u5927\uff0cPython \u524d\u7aef\u8c03\u5ea6\u3001DataLoader \u591a\u8fdb\u7a0b\u901a\u4fe1\u3001torch.compile<\/code> \u7684\u7f16\u8bd1\u4e0e\u91cd\u7f16\u8bd1\u3001\u4f18\u5316\u5668\u9010\u53c2\u6570 launch kernel \u7b49\u5f00\u9500\u4f1a\u88ab\u663e\u8457\u653e\u5927\u3002 Tech-Renaissance \u5728\u8fd9\u4e2a\u573a\u666f\u4e0b\u80fd\u8d62\u8fd9\u4e48\u591a\uff0c\u4e3b\u8981\u5f97\u76ca\u4e8e\u9759\u6001 CUDA Graph \u6d88\u9664\u4e86 Python \u8c03\u5ea6\u4e0e kernel launch \u5f00\u9500\u3001CPVS \u9a8c\u8bc1\u96c6\u7f13\u5b58\u8df3\u8fc7\u4e86\u6bcf\u8f6e\u91cd\u590d\u7684\u9884\u5904\u7406\u3001FusedNormalization \u628a ToTensor+Normalize+RandomErasing \u5408\u5e76\u6210\u4e00\u6b21 CPU \u904d\u5386\u3001\u624b\u5199 AdamW CUDA kernel \u628a\u53c2\u6570\u66f4\u65b0\u878d\u6210\u5355\u6b21 kernel\u3002<\/p>\n\n\n\n
        \u5173\u4e8e CPVS \u9700\u8981\u591a\u8bf4\u4e00\u53e5\uff1adocs\/MNIST_BEST_ADAMW_TR4_VS_PYTORCH.md<\/code> \u660e\u786e\u6307\u51fa\uff0cCPVS\uff08Cross-Process Validation Sharing\uff09\u5728\u672c\u793a\u4f8b\u4e2d\u9ed8\u8ba4\u5f00\u542f\uff0cpreprocessor.cpp<\/code> \u4e2d using_cpvs_<\/code> \u9ed8\u8ba4\u4e3a true<\/code>\u3002\u5173\u95ed CPVS \u65f6\uff0c\u540c\u4e00\u6d4b\u8bd5\u7684\u52a0\u901f\u6bd4\u4f1a\u4ece\u7ea6 7.5\u00d7 \u964d\u5230\u7ea6 5.1\u00d7\uff1b\u5f00\u542f\u540e\u6bcf epoch \u80fd\u8282\u7701\u7ea6 0.16 \u79d2\uff0c100 \u4e2a epoch \u7d2f\u8ba1\u8282\u7701\u7ea6 15\u201316 \u79d2\u3002\u8fd9\u662f\u6846\u67b6\u539f\u751f\u80fd\u529b\u5e26\u6765\u7684\u771f\u5b9e\u5dee\u5f02\uff0c\u4e0d\u662f\u8ba1\u65f6\u53e3\u5f84\u4e0a\u7684\u201c\u4f5c\u5f0a\u201d\u3002\u5f53\u7136\uff0cCPVS \u7684\u6536\u76ca\u4e0e\u9a8c\u8bc1\u96c6\u5927\u5c0f\u548c\u9884\u5904\u7406\u590d\u6742\u5ea6\u76f8\u5173\uff0c\u4e0d\u80fd\u4e0d\u52a0\u9a8c\u8bc1\u5730\u63a8\u5e7f\u5230\u6240\u6709\u4efb\u52a1\u3002<\/p>\n\n\n\n
        \u56db\u3001\u793a\u4f8b\u4e8c\uff1aVGG16BN ImageNet\u2014\u2014\u591a\u5361\u5927\u89c4\u6a21\u4e0a\u7684\u771f\u7ae0<\/h2>\n\n\n\n
        \u5c0f\u4efb\u52a1\u8d62\u4e86\u4e0d\u591f\u6709\u8bf4\u670d\u529b\uff0c\u518d\u770b\u5927\u89c4\u6a21\u573a\u666f\uff1atests\/example\/test_vgg16bn.cpp<\/code>\u3002<\/p>\n\n\n\n
        \u8fd9\u4e2a\u793a\u4f8b\u5728 A100\u00d78 \u4e0a\u8bad\u7ec3 VGG-16-BN\uff0c\u4f7f\u7528 ImageNet-1K\uff0cglobal batch size=2048\uff08\u5355\u5361 256\uff09\uff0cSGD + momentum=0.9 + Nesterov\uff0cCosineAnnealing + Warmup(10)\uff0cpeak LR=0.36\uff0cweight decay=1e-4\uff0c\u6570\u636e\u589e\u5f3a\u5305\u62ec RandomResizedCrop(0.08~1.0)\u3001RandomHorizontalFlip\u3001ColorJitter\u3001RandomErasing\u3002\u6a21\u578b\u7684\u7f51\u7edc\u7ed3\u6784\u4e0e PyTorch \u53c2\u8003\u5b9e\u73b0\u9010\u5c42\u5bf9\u9f50\uff0c\u4ee3\u7801\u4e2d\u6bcf\u4e00\u5c42\u90fd\u6807\u6ce8\u4e86 [\u5bf9\u9f50 PyTorch]<\/code>\u3002<\/p>\n\n\n\n
        \u7531\u4e8e\u5b8c\u6574\u4ee3\u7801\u8f83\u957f\uff0c\u8fd9\u91cc\u53ea\u8d34\u51fa\u6838\u5fc3\u914d\u7f6e\u7247\u6bb5\uff1a<\/p>\n\n\n\n
        GLOBAL_SETTING\n    .manual_seed(123)\n    .global_batch_size(2048)       \/\/ local=256 @ 8 GPUs\n    .train_resolution(224)\n    .val_resolution(224)\n    .use_tf32(true);\n\nPREPROCESSOR_SETTING\n    .dataset(\"imagenet\", \"\/root\/epfs\/dataset\/imagenet\")\n    .load_workers(16)\n    .preprocess_workers(128)\n    .normalization(NormMode::IMAGENET)\n    .train_transforms(\n        RandomResizedCrop(224, 0.08f, 1.0f),\n        RandomHorizontalFlip(),\n        ColorJitter(0.2f, 0.2f, 0.2f, 0.1f),\n        RandomErasing(0.25f, {0.02f, 0.33f}, {0.3f, 3.3f})\n    )\n    .val_transforms(\n        Resize(256),\n        CenterCrop(224)\n    )\n    .commit();\n\nBluePrint vgg16bn = seq(\n    conv(64, 3, 1, 1), bn(), relu(),\n    conv(64, 3, 1, 1), bn(), relu(),\n    maxpool(2, 2, 0),\n    \/\/ ... Block 2 ~ Block 5 ...\n    flatten(),\n    fc(4096, true), relu(), dropout(0.5),\n    fc(4096, true), relu(), dropout(0.5),\n    fc(1000, true)\n);\n\nDeepLearningTask task;\ntask.model(vgg16bn)\n    .loss(CrossEntropyLoss().label_smoothing(0.1f))\n    .initializer(Initializer()\n        .conv(InitKind::KAIMING_UNIFORM)\n        .fc(InitKind::KAIMING_UNIFORM)\n        .fan(FanMode::FAN_IN))\n    .total_epochs(100)\n    .optimizer(SGD()\n        .momentum(0.9f)\n        .weight_decay(1e-4f)\n        .nesterov(true))\n    .scheduler(CosineAnnealingLR()\n        .base_lr(0.36f)\n        .warmup_start_lr(0.01f)\n        .warmup(10)\n        .eta_min(1e-6f)\n        .step_by_epoch());\n\ntask.compile(CompileInfo::ALL);\nauto result = task.run();<\/pre>\n\n\n\n
        \u5bf9\u5e94\u7684 PyTorch \u811a\u672c\u542f\u7528\u4e86 torch.compile(mode=\"max-autotune\")<\/code>\u3001DDP\u3001SyncBatchNorm\u3001channels_last<\/code>\u3001pin_memory<\/code>\u3001persistent_workers<\/code>\u3001prefetch_factor=8<\/code>\u3001TF32\uff1bTensorFlow \u811a\u672c\u542f\u7528\u4e86 XLA\u3002\u4e09\u8005\u5728\u540c\u4e00\u53f0 A100\u00d78 \u673a\u5668\u4e0a\u8fd0\u884c\uff0c\u4f7f\u7528\u76f8\u540c\u7684\u6a21\u578b\u7ed3\u6784\u3001\u8d85\u53c2\u6570\u3001\u6570\u636e\u589e\u5f3a\u548c\u9884\u5904\u7406\u7ebf\u7a0b\u6570\uff0c\u4e14\u5747\u6392\u9664\u7f16\u8bd1\u7528\u65f6\u3002<\/p>\n\n\n\n
        \u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n
        \u6846\u67b6<\/th>\u541e\u5410\u91cf (Images\/sec)<\/th>\u6bcf Epoch \u7528\u65f6 (s)<\/th>\u52a0\u901f\u6bd4 (vs PyTorch)<\/th><\/tr><\/thead>
        PyTorch + torch.compile<\/td>7,351.20<\/td>174.280<\/td>baseline<\/td><\/tr>
        TensorFlow + XLA<\/td>7,766.34<\/td>164.964<\/td>+5.65%<\/td><\/tr>
        Tech-Renaissance<\/strong><\/td>9,310.13<\/strong><\/td>137.610<\/strong><\/td>+26.65%<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
        \u8bad\u7ec3\u7cbe\u5ea6\u65b9\u9762\uff0cTech-Renaissance \u8fbe\u5230\u4e86 TOP-1 73.80% \/ TOP-5 91.71%\uff0c\u7b26\u5408 VGG16BN \u5728 ImageNet \u4e0a\u7684\u9884\u671f\u6c34\u51c6\u3002<\/p>\n\n\n\n
        \u4e3a\u4e86\u8ba9\u8fd9\u4e2a\u5bf9\u6bd4\u7ecf\u5f97\u8d77\u5ba1\u89c6\uff0c\u6211\u4eec\u8fd8\u4e13\u95e8\u5199\u4e86 docs\/VGG16BN_FAIRNESS.md<\/code> \u8fdb\u884c\u9010\u9879\u5ba1\u8ba1\u3002\u5ba1\u8ba1\u7ed3\u8bba\u662f\uff1a\u9ed8\u8ba4\u914d\u7f6e\u4e0b\u7684\u5bf9\u6bd4\u662f\u516c\u5e73\u7684\u3002<\/strong> \u53cc\u65b9\u5728\u7f51\u7edc\u7ed3\u6784\u3001BatchNorm \u53c2\u6570\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u516c\u5f0f\u3001weight decay \u6392\u9664\u7b56\u7565\u3001\u6570\u636e\u589e\u5f3a\u53c2\u6570\u3001\u9a8c\u8bc1\u9884\u5904\u7406\u3001TF32 \u8bbe\u7f6e\u3001\u521d\u59cb\u5316\u65b9\u5f0f\u3001\u8ba1\u65f6\u53e3\u5f84\u7b49\u6838\u5fc3\u7ef4\u5ea6\u4e0a\u4e25\u683c\u4e00\u81f4\uff1b\u5b58\u5728\u7684\u5dee\u5f02\u9879\uff08\u5982 FusedNormalization \u878d\u5408\u9884\u5904\u7406\u3001\u539f\u751f NHWC \u5e03\u5c40\u3001\u9759\u6001 CUDA Graph\u3001C++ \u4e24\u7ea7\u6570\u636e\u52a0\u8f7d\u7ebf\u7a0b\u6a21\u578b\uff09\u5747\u5c5e\u4e8e\u6846\u67b6\u81ea\u8eab\u67b6\u6784\u80fd\u529b\u5dee\u5f02\uff0c\u800c\u975e\u6d4b\u8bd5\u4f5c\u5f0a\u3002<\/p>\n\n\n\n
        \u4e0e MNIST MLP \u4e0d\u540c\u7684\u662f\uff0cVGG16BN \u7684\u516c\u5e73\u5ba1\u8ba1\u660e\u786e\u628a CPVS\u3001FULLY \u8bad\u7ec3\u96c6\u7f13\u5b58\u3001DTS \u538b\u7f29\u683c\u5f0f\u9ed8\u8ba4\u5173\u95ed<\/strong>\uff0c\u4ee5\u4fbf\u628a\u6ce8\u610f\u529b\u96c6\u4e2d\u5728\u8bad\u7ec3\u541e\u5410\u672c\u8eab\u3002\u5982\u679c\u628a CPVS \u6216 DTS \u6253\u5f00\uff0c\u6570\u5b57\u8fd8\u4f1a\u8fdb\u4e00\u6b65\u53d8\u5316\uff0c\u4f46\u90a3\u5c31\u5c5e\u4e8e\u201c\u6846\u67b6\u80fd\u529b\u5dee\u5f02\u201d\u4e0e\u201c\u6d4b\u8bd5\u914d\u7f6e\u5dee\u5f02\u201d\u4e4b\u95f4\u7684\u53d6\u820d\uff0c\u9700\u8981\u5728\u62a5\u544a\u91cc\u5355\u72ec\u8bf4\u660e\u3002<\/p>\n\n\n\n
        \u8fd9\u6b63\u662f\u6211\u4eec\u60f3\u8981\u4f20\u8fbe\u7684\u6001\u5ea6\uff1a\u6027\u80fd\u4f18\u52bf\u53ef\u4ee5\u6765\u81ea\u67b6\u6784\u8bbe\u8ba1\u9009\u62e9\uff0c\u4f46\u524d\u63d0\u662f\u628a\u8fd9\u79cd\u9009\u62e9\u900f\u660e\u5730\u6446\u51fa\u6765\uff0c\u8ba9\u4efb\u4f55\u4eba\u90fd\u80fd\u590d\u73b0\u548c\u5ba1\u8ba1\u3002<\/strong><\/p>\n\n\n\n
        \u4e94\u3001\u6027\u80fd\u5bc6\u7801\uff1a\u4e0d\u662f\u5355\u70b9\u5947\u6280\uff0c\u800c\u662f\u7cfb\u7edf\u5de5\u7a0b<\/h2>\n\n\n\n
        \u770b\u5230\u4e0a\u9762\u7684\u6570\u5b57\uff0c\u5f88\u591a\u4eba\u7b2c\u4e00\u53cd\u5e94\u662f\uff1a”\u4f60\u4eec\u5230\u5e95\u7528\u4e86\u4ec0\u4e48\u9ed1\u79d1\u6280\uff1f”<\/p>\n\n\n\n
        \u7b54\u6848\u662f\uff1a\u6ca1\u6709\u9ed1\u79d1\u6280\uff0c\u53ea\u6709\u628a\u8bad\u7ec3\u7ba1\u7ebf\u4ece\u78c1\u76d8\u5230 GPU \u7684\u6bcf\u4e00\u4e2a\u73af\u8282\u90fd\u91cd\u65b0\u8bbe\u8ba1\u4e86\u4e00\u904d\u3002<\/strong> Tech-Renaissance \u7684\u901f\u5ea6\u4f18\u52bf\u4e0d\u662f\u67d0\u4e00\u4e2a kernel \u5199\u5f97\u6f02\u4eae\uff0c\u800c\u662f\u6574\u6761\u94fe\u8def\u7684\u6570\u636e\u5e03\u5c40\u3001\u9884\u5904\u7406\u3001\u4f20\u8f93\u3001\u8ba1\u7b97\u3001\u901a\u4fe1\u3001\u66f4\u65b0\u2014\u2014\u51e0\u4e4e\u6bcf\u4e2a\u73af\u8282\u90fd\u88ab\u88c1\u526a\u8fc7\u3001\u878d\u5408\u8fc7\u3001\u9759\u6001\u5316\u8fc7\u3001\u6d41\u6c34\u7ebf\u5316\u8fc7\u3002\u4e0b\u9762\u628a\u8fd9\u4e9b\u8bbe\u8ba1\u62c6\u6210 11 \u6761\uff0c\u5b83\u4eec\u4e0d\u662f\u5b64\u7acb\u7684\u4f18\u5316\u70b9\uff0c\u800c\u662f\u56f4\u7ed5\u540c\u4e00\u4e2a\u76ee\u6807\u2014\u2014\u5728\u7f16\u8bd1\u671f\u5c3d\u53ef\u80fd\u591a\u5730\u786e\u5b9a\u4e00\u5207\uff0c\u5728\u8fd0\u884c\u671f\u53ea\u505a\u6700\u5c11\u7684\u4e8b<\/strong>\u2014\u2014\u9010\u5c42\u5c55\u5f00\u7684\u7cfb\u7edf\u5de5\u7a0b\u3002<\/p>\n\n\n\n
        \n\n\n\n
        1. \u9759\u6001\u663e\u5b58\u89c4\u5212\uff1a\u8fd0\u884c\u671f\u4e0d\u5206\u914d\uff0c\u4e0d\u788e\u7247\uff0c\u4e0d\u6296\u52a8<\/h3>\n\n\n\n
        \u4e3b\u6d41\u6846\u67b6\u7684\u8bad\u7ec3\u5faa\u73af\u91cc\uff0ccudaMalloc<\/code>\/cudaFree<\/code> \u6216\u7f13\u5b58\u5206\u914d\u5668\u7684\u7533\u8bf7\u4e0e\u91ca\u653e\u662f\u6301\u7eed\u53d1\u751f\u7684\u9690\u6027\u5f00\u9500\uff0c\u8fd8\u4f1a\u5e26\u6765\u5730\u5740\u4e0d\u7a33\u5b9a\u548c\u5185\u5b58\u788e\u7247\u3002Tech-Renaissance \u4ece\u8bbe\u8ba1\u4e4b\u521d\u5c31\u89c4\u5b9a\uff1a\u8fd0\u884c\u671f\u4e0d\u505a\u52a8\u6001\u663e\u5b58\u5206\u914d\u3002<\/strong><\/p>\n\n\n\n
        MemoryPlan<\/code> \u5728\u7f16\u8bd1\u671f\u628a\u6240\u6709\u5f20\u91cf\u6309\u8bed\u4e49\u5206\u533a\u6392\u5e03\u5230 68 \u4e2a\u547d\u540d Region \u4e2d\uff08\u542b\u4e00\u4e2a\u8fb9\u754c\u54e8\u5175\u5171 69 \u4e2a\u69fd\u4f4d\uff09\uff1a\u6743\u91cd\u533a\u3001\u68af\u5ea6\u533a\u3001\u52a8\u91cf\u533a\uff08\u4e00\u9636\/\u4e8c\u9636\uff09\u3001EMA \u533a\u3001AMP FP16 \u6743\u91cd\u533a\u3001\u8f93\u5165\u53cc\u7f13\u51b2\u533a\u3001\u7279\u5f81\u56fe\u533a\u3001\u7ed3\u679c\u533a\u7b49\u3002\u6bcf\u4e2a DTensor<\/code> \u53ea\u662f\u4e00\u5f20 (region, offset, shape, stride)<\/code> \u63cf\u8ff0\u7b26\uff0c\u4e0d\u6301\u6709\u5b9e\u9645\u5185\u5b58\u3002\u6240\u6709 Region \u7684\u9996\u5730\u5740\u6309 align_up_256<\/code> \u5bf9\u9f50\u5230 256 \u5b57\u8282\uff0c\u6ee1\u8db3 Tensor Core \u4e0e cuDNN \u7684\u6700\u4f18\u8bbf\u5b58\u7ea6\u675f\u3002<\/p>\n\n\n\n
        \u8fd9\u4e00\u8bbe\u8ba1\u5e26\u6765\u4e86\u4e09\u4e2a\u8fde\u9501\u7ea2\u5229\uff1a<\/p>\n\n\n\n
          \n
        • \u5730\u5740\u7a33\u5b9a\u6027\u5929\u7136\u6210\u7acb<\/strong>\u2014\u2014CUDA Graph \u91cd\u653e\u8981\u6c42\u6240\u6709\u6307\u9488\u4e0d\u53d8\uff0c\u9759\u6001\u5206\u914d\u8ba9\u8fd9\u4e2a\u8981\u6c42\u96f6\u6210\u672c\u6ee1\u8db3\uff1b<\/li>\n\n\n\n
        • \u540c\u8bed\u4e49\u5f20\u91cf\u7269\u7406\u8fde\u7eed<\/strong>\u2014\u2014\u6240\u6709\u6743\u91cd\u5728\u4e00\u8d77\u3001\u6240\u6709\u68af\u5ea6\u5728\u4e00\u8d77\u3001\u6240\u6709\u52a8\u91cf\u5728\u4e00\u8d77\uff0c\u4e3a\u6574\u533a\u6279\u91cf\u64cd\u4f5c\u5960\u5b9a\u57fa\u7840\uff1b<\/li>\n\n\n\n
        • \u591a\u5361\u95f4\u5171\u4eab\u540c\u4e00\u4efd\u5e03\u5c40\u56fe\u7eb8<\/strong>\u2014\u2014\u540c\u4e00\u5f20\u91cf\u5728\u4e0d\u540c rank \u4e0a\u7684 offset \u5b8c\u5168\u4e00\u81f4\uff0c\u5206\u5e03\u5f0f\u901a\u4fe1\u548c\u968f\u673a\u6570\u6d88\u8d39\u987a\u5e8f\u5929\u7136\u5bf9\u9f50\u3002<\/li>\n<\/ul>\n\n\n\n
          \n\n\n\n
          2. \u7b97\u5b50\u878d\u5408\uff1a\u628a\u591a\u6b21\u5185\u5b58\u904d\u5386\u538b\u6210\u4e00\u6b21<\/h3>\n\n\n\n
          \u8bad\u7ec3 CNN \u65f6\uff0c\u771f\u6b63\u62d6\u6162\u901f\u5ea6\u7684\u5f80\u5f80\u4e0d\u662f FLOPs\uff0c\u800c\u662f\u663e\u5b58\u5e26\u5bbd\u3002A100 \u4e0a FP16 Tensor Core \u7a20\u5bc6\u5cf0\u503c\u7b97\u529b\u7ea6 312 TFLOPS\uff0c\u4f46 HBM \u5e26\u5bbd\u53ea\u6709\u7ea6 2 TB\/s\u2014\u2014\u6309\u7167 roofline \u6a21\u578b\uff0c\u5f88\u591a\u9010\u5143\u7d20\u64cd\u4f5c\u90fd\u5361\u5728”\u5185\u5b58\u5899”\u4e0a\u3002\u51cf\u5c11\u4e2d\u95f4\u5f20\u91cf\u7684\u8bfb\u5199\uff0c\u662f\u6bd4\u5806\u7b97\u529b\u66f4\u76f4\u63a5\u7684\u63d0\u901f\u624b\u6bb5\u3002<\/p>\n\n\n\n
          CBR \u878d\u5408\u7b97\u5b50<\/strong>\u628a Conv + BatchNorm + ReLU \u5408\u5e76\u4e3a\u4e00\u5f20 cuDNN Frontend Graph\u3002\u8bad\u7ec3\u8def\u5f84\u867d\u7136\u4e0d\u80fd\u50cf\u63a8\u7406\u90a3\u6837\u628a BN \u6298\u53e0\u8fdb\u5377\u79ef\u6743\u91cd\uff0c\u4f46\u901a\u8fc7 conv_fprop + genstats<\/code>\u3001bn_finalize<\/code>\u3001bn_apply + ReLU<\/code> \u4e09\u6bb5\u5f0f cuDNN Graph\uff0cBN \u7684\u7edf\u8ba1\u91cf\u8ba1\u7b97\u88ab\u9690\u85cf\u5728\u5377\u79ef\u5c3e\u90e8\uff0cBN \u7684\u4eff\u5c04\u53d8\u6362\u4e0e ReLU \u6fc0\u6d3b\u88ab\u878d\u5408\u5728\u540c\u4e00\u4e2a kernel \u91cc\uff0c\u4e2d\u95f4\u7ed3\u679c\u5c3d\u91cf\u7559\u5728\u5bc4\u5b58\u5668\/\u5171\u4eab\u5185\u5b58\uff0c\u907f\u514d\u628a\u5b8c\u6574\u7279\u5f81\u56fe\u53cd\u590d\u5199\u56de HBM\u3002\u4ee5 256\u00d756\u00d756\u00d764<\/code> \u7684 FP16 \u7279\u5f81\u56fe\u4e3a\u4f8b\uff0c\u5355\u5c42\u5c31\u80fd\u7701\u6389\u7ea6 100 MB \u7ea7\u522b\u7684\u65e0\u6548\u8bfb\u5199\u3002<\/p>\n\n\n\n
          FusedNormalization<\/strong> \u5219\u628a\u6570\u636e\u589e\u5f3a\u672b\u7aef\u7684 ToTensor + RandomHorizontalFlip + Normalize + RandomErasing<\/code> \u56db\u4e2a\u539f\u672c\u72ec\u7acb\u7684\u6b65\u9aa4\uff0c\u5408\u5e76\u6210\u4e00\u6b21 CPU \u5185\u5b58\u904d\u5386\u3002\u4e00\u5f20 uint8 \u56fe\u7247\u8fdb\u53bb\uff0c\u4e00\u5f20\u5df2\u7ecf\u5f52\u4e00\u5316\u597d\u7684 FP32\/FP16 \u56fe\u7247\u51fa\u6765\uff0c\u914d\u5408 F16C \u6307\u4ee4\u5728 CPU \u7aef\u76f4\u63a5\u8f93\u51fa FP16\uff0c\u8ba9 H2D \u4f20\u8f93\u91cf\u518d\u51cf\u534a\u3002\u5728 ImageNet \u6bcf epoch 128 \u4e07\u5f20\u56fe\u7247\u7684\u8bad\u7ec3\u89c4\u6a21\u4e0b\uff0c\u8fd9\u4e00\u6b65\u7684\u5e26\u5bbd\u8282\u7701\u81f3\u5173\u91cd\u8981\u3002<\/p>\n\n\n\n
          \n\n\n\n
          3. \u6574\u533a\u6279\u91cf\u64cd\u4f5c\uff1a\u628a\u591a\u6b21 kernel launch \u538b\u6210\u4e00\u6b21<\/h3>\n\n\n\n
          \u4f20\u7edf\u6846\u67b6\u91cc\uff0c\u4f18\u5316\u5668\u9700\u8981\u904d\u5386\u6bcf\u4e2a\u53c2\u6570\uff0c\u9010\u4e2a\u8c03\u7528 kernel \u66f4\u65b0\u6743\u91cd\u548c\u52a8\u91cf\u3002\u4e00\u4e2a 100 \u5c42\u7684\u7f51\u7edc\uff0c\u5149\u662f\u4f18\u5316\u5668\u66f4\u65b0\u5c31\u53ef\u80fd\u4ea7\u751f\u6570\u767e\u6b21 kernel launch\uff0c\u6bcf\u6b21\u90fd\u6709\u56fa\u5b9a\u7684 CPU\u2192GPU \u8c03\u5ea6\u5f00\u9500\uff08\u901a\u5e38\u5728 5\u201310 \u03bcs \u91cf\u7ea7\uff09\uff0c\u79ef\u5c11\u6210\u591a\u3002<\/p>\n\n\n\n
          \u5f97\u76ca\u4e8e MemoryPlan \u628a\u540c\u7c7b\u578b\u5f20\u91cf\u7269\u7406\u8fde\u7eed\u5b58\u653e\uff0cTech-Renaissance \u5f15\u5165 RangeOp<\/strong> \u8fd9\u4e00 Region \u7ea7\u62bd\u8c61\uff0c\u7528\u4e00\u4e2a kernel \u626b\u8fc7\u4e00\u6574\u6bb5\u8fde\u7eed\u5185\u5b58\uff0c\u65e0\u89c6\u5f20\u91cf\u8fb9\u754c\uff1a<\/p>\n\n\n\n
            \n
          • \u68af\u5ea6\u6e05\u96f6<\/strong>\uff1a\u4e00\u6b21 RANGE_CLEAR<\/code> \u8986\u76d6\u6574\u4e2a\u68af\u5ea6\u533a\uff0c\u800c\u975e\u9010\u5c42\u9010\u53c2\u6570\u6e05\u96f6\uff1b<\/li>\n\n\n\n
          • \u7cbe\u5ea6\u8f6c\u6362<\/strong>\uff1aRANGE_CAST_FP32_TO_FP16<\/code> \u4e00\u6b21\u5b8c\u6210\u5168\u6a21\u578b\u6743\u91cd\u6216\u68af\u5ea6\u7684\u7c7b\u578b\u8f6c\u6362\uff1b<\/li>\n\n\n\n
          • \u878d\u5408\u4f18\u5316\u5668<\/strong>\uff1aSGD\/Momentum\/AdamW \u7684 weight \u66f4\u65b0\u548c bias \u66f4\u65b0\u5206\u522b\u7528\u4e00\u4e2a kernel \u8986\u76d6\u6574\u533a\uff0c\u65e0\u8bba\u6a21\u578b 10 \u5c42\u8fd8\u662f 100 \u5c42\uff0c\u4f18\u5316\u5668 step \u7684 kernel \u6570\u91cf\u6052\u5b9a\uff1b\u4e14 AMP \u53cd\u7f29\u653e\u4e0e\u53c2\u6570\u66f4\u65b0\u5728\u540c\u4e00\u4e2a kernel \u5faa\u73af\u4e2d\u5b8c\u6210\uff0c\u907f\u514d PyTorch \u4e2d scaler.unscale_()<\/code> \u4e0e optimizer.step()<\/code> \u4e24\u6b21\u6574\u533a\u904d\u5386\uff1b<\/li>\n\n\n\n
          • NCCL \u68af\u5ea6\u901a\u4fe1<\/strong>\uff1a\u68af\u5ea6\u5df2\u6309 G_BN_BIAS..G_FIRST_CONV<\/code> \u548c G_DEEP_CONV..R_RESULT<\/code> \u4e24\u4e2a\u8fde\u7eed Region \u6392\u5e03\uff0cAllReduce \u76f4\u63a5\u5bf9\u8fde\u7eed\u5b57\u8282\u8303\u56f4\u53d1\u8d77\uff0c\u4e24\u6876\u9759\u6001\u5206\u6876\u65e2\u51cf\u5c11\u4e86 NCCL kernel \u542f\u52a8\u6b21\u6570\uff0c\u53c8\u8ba9\u6df1\u5c42\u68af\u5ea6\u901a\u4fe1\u53ef\u88ab\u9996\u5c42\u53cd\u5411\u8ba1\u7b97\u63a9\u76d6\u3002<\/li>\n<\/ul>\n\n\n\n
            \n\n\n\n
            4. CUDA Graph \u5168\u6355\u83b7\uff1a\u628a CPU \u6d3e\u53d1\u5f00\u9500\u538b\u5230\u96f6<\/h3>\n\n\n\n
            \u5373\u4f7f kernel launch \u5df2\u7ecf\u964d\u5230\u4e2a\u4f4d\u6570\uff0c\u6bcf\u6b21 CPU \u63d0\u4ea4 GPU \u5de5\u4f5c\u4ecd\u9700\u8d70 CUDA \u9a71\u52a8\u5c42\u3002CUDA Graph \u628a\u8fd9\u4e2a\u95ee\u9898\u5f7b\u5e95\u89e3\u51b3\uff1a\u5c06\u4e00\u6574\u6bb5 GPU \u64cd\u4f5c\u5e8f\u5217\uff08kernel\u3001memcpy\u3001event\uff09\u6355\u83b7\u6210\u4e00\u5f20\u56fe\uff0c\u4e4b\u540e\u53ea\u9700\u4e00\u6b21 cudaGraphLaunch<\/code> \u5c31\u80fd\u91cd\u653e\u5168\u90e8\u64cd\u4f5c\u3002<\/p>\n\n\n\n
            Tech-Renaissance \u628a\u8bad\u7ec3\u5faa\u73af\u7684\u5168\u9636\u6bb5<\/strong>\u90fd\u6355\u83b7\u8fdb\u4e86 CUDA Graph\uff1a\u4ece H2D \u4f20\u8f93\u3001\u9996\u5c42\u524d\u5411\u3001\u6df1\u5c42\u524d\u5411+\u53cd\u5411\u3001\u68af\u5ea6 AllReduce\u3001BN \u7edf\u8ba1\u91cf\u540c\u6b65\u3001\u4f18\u5316\u5668\u66f4\u65b0\u3001EMA \u66f4\u65b0\uff0c\u5230\u6700\u540e\u7684 FP32\u2192FP16 \u6743\u91cd\u56de\u62f7\u548c A\/B \u7f13\u51b2\u5207\u6362\uff0c\u5168\u90e8\u5728\u6355\u83b7\u8303\u56f4\u5185\u3002MultiStreamCaptureState<\/code> \u5728\u7f16\u8bd1\u671f\u7ba1\u7406\u4e09\u6761\u8ba1\u7b97\u6d41\u3001\u4e00\u6761\u4f20\u8f93\u6d41\u3001\u4e00\u6761\u66f4\u65b0\u6d41\u4e4b\u95f4\u7684\u4f9d\u8d56\u548c\u4e8b\u4ef6\u540c\u6b65\u3002\u8fd0\u884c\u671f CPU \u53ea\u8d1f\u8d23\u6309\u987a\u5e8f launch \u5df2\u7ecf\u6355\u83b7\u597d\u7684\u56fe\uff0c\u8fde”\u4ec0\u4e48\u65f6\u5019\u7b49\u8c01”\u8fd9\u4ef6\u4e8b\uff0c\u90fd\u5df2\u7ecf\u88ab\u7f16\u8bd1\u671f\u51b3\u5b9a\u4e86\u3002<\/p>\n\n\n\n
            \u8fd9\u5c31\u662f Tech-Renaissance \u4e0e PyTorch \u6700\u6839\u672c\u7684\u67b6\u6784\u5dee\u5f02\uff1aPyTorch Eager \u7684\u6bcf\u4e2a\u7b97\u5b50\u90fd\u8981\u7ecf\u8fc7 Python dispatcher \u2192 Autograd \u2192 C++ kernel launch \u4e09\u5c42\u8c03\u7528\uff1btorch.compile<\/code> \u8bd5\u56fe\u5728\u52a8\u6001\u56fe\u4e4b\u4e0a\u53e0\u52a0 JIT \u7f16\u8bd1\uff0c\u4f46\u9700\u8981\u9762\u5bf9 graph break\u3001\u52a8\u6001\u5f62\u72b6\u3001\u5730\u5740\u7a33\u5b9a\u6027\u7b49\u7ed3\u6784\u6027\u6311\u6218\u3002Tech-Renaissance \u4ece\u8bbe\u8ba1\u4e4b\u521d\u5c31\u662f\u9759\u6001\u56fe\uff0cCUDA Graph \u5168\u6355\u83b7\u662f\u81ea\u7136\u7ed3\u679c\uff0c\u4e0d\u9700\u8981\u5728\u7075\u6d3b\u6027\u548c\u6027\u80fd\u4e4b\u95f4\u53cd\u590d\u62c9\u626f\u3002\u8bad\u7ec3\u8fdb\u5165\u7a33\u5b9a\u5faa\u73af\u540e\uff0c\u6ca1\u6709\u5355\u4e2a kernel launch\u3001\u6ca1\u6709\u8fd0\u884c\u671f\u663e\u5b58\u5206\u914d\u3001\u6ca1\u6709 Python \u7ea7\u8c03\u5ea6<\/strong>\u3002<\/p>\n\n\n\n
            \n\n\n\n
            5. \u591a\u6d41\u5e76\u53d1\uff1a\u8ba1\u7b97\u3001\u901a\u4fe1\u3001\u4f20\u8f93\u540c\u65f6\u8dd1<\/h3>\n\n\n\n
            Tech-Renaissance \u4e3a\u6bcf\u4e2a rank \u7ef4\u62a4\u4e94\u6761\u975e\u963b\u585e CUDA \u6d41\uff0c\u5206\u5de5\u660e\u786e\uff1a<\/p>\n\n\n\n
              \n
            • TRANS<\/strong>\uff1a\u8d1f\u8d23 H2D \u5f02\u6b65\u6570\u636e\u4f20\u8f93\uff1b<\/li>\n\n\n\n
            • COMP_1 \/ COMP_2 \/ COMP_3<\/strong>\uff1a\u4e09\u6761\u8ba1\u7b97\u6d41\uff0c\u5206\u522b\u627f\u8f7d\u4e3b\u5e72 GEMM\/\u5377\u79ef\u3001\u5f52\u7ea6\/\u6c60\u5316\/BN\u3001\u6fc0\u6d3b\/dX \u7b49\u4e0d\u540c\u7c7b\u578b\u7684\u8ba1\u7b97\uff0c\u7b97\u5b50\u5230\u6d41\u7684\u6620\u5c04\u5728 op_stream_policy.cpp<\/code> \u4e2d\u9759\u6001\u51b3\u5b9a\uff1b<\/li>\n\n\n\n
            • UPDATE<\/strong>\uff1a\u8d1f\u8d23\u68af\u5ea6\u6e05\u96f6\u3001\u7c7b\u578b\u8f6c\u6362\u3001NaN \u68c0\u6d4b\u3001\u68af\u5ea6\u7f29\u653e\u3001AllReduce\u3001\u4f18\u5316\u5668\u66f4\u65b0\u3001EMA \u66f4\u65b0\u7b49”\u8bad\u7ec3\u540e\u52e4”\u5de5\u4f5c\u3002<\/li>\n<\/ul>\n\n\n\n
              \u8fd9\u4e9b\u6d41\u4e4b\u95f4\u7684\u540c\u6b65\u5173\u7cfb\u5728\u7f16\u8bd1\u671f\u5c31\u5df2\u901a\u8fc7 CUDA event \u5199\u5165\u56fe\u4e2d\uff0c\u800c\u975e\u7c97\u7c92\u5ea6\u7684 cudaDeviceSynchronize<\/code>\u3002\u4e00\u4e2a\u5178\u578b\u7684\u8bad\u7ec3\u8fed\u4ee3\u4e2d\uff1a\u5f53 COMP_1 \u6267\u884c\u9996\u5c42\u524d\u5411\u65f6\uff0cTRANS \u6d41\u5df2\u5f00\u59cb\u642c\u8fd0\u4e0b\u4e00 batch \u7684\u6570\u636e\uff1b\u5f53 COMP_2 \u548c COMP_3 \u5e76\u884c\u6267\u884c\u6df1\u5c42\u524d\u5411\u548c\u53cd\u5411\u65f6\uff0cUPDATE \u6d41\u5728\u505a\u68af\u5ea6 AllReduce\uff1b\u5f53\u4f18\u5316\u5668\u5728 UPDATE \u6d41\u4e0a\u66f4\u65b0\u6743\u91cd\u65f6\uff0c\u4e0b\u4e00\u8f6e\u7684 H2D \u4f20\u8f93\u53c8\u5f00\u59cb\u4e86\u3002\u8ba1\u7b97\u3001\u901a\u4fe1\u3001\u6570\u636e\u642c\u8fd0\u8fd9\u4e09\u7c7b\u64cd\u4f5c\u5728\u7269\u7406\u4e0a\u91cd\u53e0\u6267\u884c\uff0cGPU \u7684 SM \u548c Copy Engine \u88ab\u540c\u65f6\u5582\u9971\u3002<\/p>\n\n\n\n
              \n\n\n\n
              6. \u53cc\u7f13\u51b2\u6d41\u6c34\u7ebf\uff1a\u8ba9 GPU \u4e0d\u518d\u7b49\u6570\u636e<\/h3>\n\n\n\n
              \u4e94\u6761\u6d41\u8981\u60f3\u771f\u7684\u91cd\u53e0\u8d77\u6765\uff0c\u6570\u636e\u4f9b\u7ed9\u5fc5\u987b\u8ddf\u5f97\u4e0a\u3002Tech-Renaissance \u5728\u6570\u636e\u7ba1\u7ebf\u7684\u4e24\u4e2a\u5173\u952e\u73af\u8282\u90fd\u91c7\u7528\u4e86\u53cc\u7f13\u51b2\uff1a<\/p>\n\n\n\n
                \n
              • TransferStation \u53cc\u7f13\u51b2<\/strong>\uff1a\u6bcf\u4e2a GPU \u5bf9\u5e94\u4e24\u5757\u9501\u9875\u5185\u5b58\uff0cCPU \u9884\u5904\u7406\u5411 A \u533a\u5199\u5165\u65f6\uff0cGPU \u8ba1\u7b97 B \u533a\u5bf9\u5e94\u7684 batch\uff1bA \u533a\u586b\u6ee1\u540e\u89e6\u53d1\u5f02\u6b65 H2D \u4f20\u8f93\uff0cCPU \u5207\u6362\u5230 B \u533a\u7ee7\u7eed\u5199\u3002CPU \u9884\u5904\u7406\u3001H2D \u4f20\u8f93\u3001GPU \u8ba1\u7b97\u4e09\u6bb5\u5f62\u6210\u6d41\u6c34\u7ebf\uff0cGPU \u4e0d\u4f1a\u56e0\u4e3a\u7b49\u6570\u636e\u800c\u7a7a\u8f6c\u3002<\/li>\n\n\n\n
              • A\/B \u53cc\u7f13\u51b2\u56fe\u6267\u884c<\/strong>\uff1aTRANSFER_A\/B<\/code>\u3001FIRST_LAYER_FWD_A\/B<\/code> \u7b49\u5b50\u56fe\u6210\u5bf9\u5b58\u5728\uff0cGraphExecutor<\/code> \u5728\u8fd0\u884c\u65f6\u4ea4\u66ff launch\uff0c\u8ba9\u4e0b\u4e00\u4e2a batch \u7684\u6570\u636e\u642c\u8fd0\u548c\u5f53\u524d batch \u7684\u8ba1\u7b97\u5b8c\u5168\u91cd\u53e0\u3002<\/li>\n<\/ul>\n\n\n\n
                \u53cc\u7f13\u51b2\u7684\u4ef7\u503c\u4e0d\u662f\u8ba9\u67d0\u4e00\u6b65\u53d8\u5feb\uff0c\u800c\u662f\u8ba9 GPU \u5c3d\u53ef\u80fd\u5c11\u5730”\u6328\u997f”\u7b49\u5f85\u6570\u636e\u3002<\/p>\n\n\n\n
                \n\n\n\n
                7. DTS \u81ea\u5b9a\u4e49\u683c\u5f0f\uff1a\u628a\u968f\u673a\u5c0f\u6587\u4ef6\u8bfb\u53d8\u6210\u987a\u5e8f\u5927\u5757\u8bfb<\/h3>\n\n\n\n
                ImageNet \u8bad\u7ec3\u96c6\u5305\u542b 128 \u4e07\u5f20 JPEG \u56fe\u7247\uff0c\u5206\u6563\u5728 1000 \u4e2a\u5b50\u76ee\u5f55\u4e2d\u3002\u5927\u91cf\u5c0f\u6587\u4ef6\u7684\u968f\u673a IO \u662f\u4f20\u7edf\u6570\u636e\u52a0\u8f7d\u7684\u7b2c\u4e00\u9053\u74f6\u9888\u2014\u2014\u6587\u4ef6\u7cfb\u7edf\u7684 inode \u67e5\u627e\u548c\u968f\u673a seek \u4f1a\u628a\u78c1\u76d8\u541e\u5410\u62d6\u5230\u8fdc\u4f4e\u4e8e\u987a\u5e8f\u8bfb\u53d6\u7684\u6c34\u5e73\u3002<\/p>\n\n\n\n
                Tech-Renaissance \u7684 DTS\uff08Deep-learning Training Store\uff09<\/strong> \u683c\u5f0f\u628a\u8bad\u7ec3\u96c6\u6253\u5305\u6210\u5355\u4e00\u4e8c\u8fdb\u5236\u6587\u4ef6\uff0c\u4ee5\u56fa\u5b9a 16 MB \u7684 Block \u4e3a\u5355\u4f4d\u7ec4\u7ec7\u6837\u672c\uff0c\u6309 First-Fit Decreasing \u88c5\u7bb1\u7b97\u6cd5\u7d27\u5bc6\u6392\u5217\uff0c\u5e73\u5747\u7a7a\u95f4\u6d6a\u8d39\u4ec5\u7ea6 0.08%\uff0813 KB \/ 16 MB\uff09\u3002\u8bad\u7ec3\u65f6\uff0cIO worker \u4ee5 Block \u4e3a\u5355\u4f4d\u987a\u5e8f\u8bfb\u53d6\uff0c\u78c1\u76d8\u5e26\u5bbd\u88ab\u5145\u5206\u5229\u7528\uff1bBlock \u5185\u90e8\u7684\u6837\u672c\u901a\u8fc7 SlotMeta \u504f\u79fb\u91cf\u96f6\u62f7\u8d1d\u5b9a\u4f4d\u3002\u6bcf\u4e2a IO worker \u9759\u6001\u9886\u53d6\u56fa\u5b9a Block\uff0c\u65e0\u9501\u3001\u65e0\u961f\u5217\u7ade\u4e89\u3002\u5728\u7b14\u8bb0\u672c\u7b49 IO \u53d7\u9650\u73af\u5883\u4e0b\uff0cDTS \u80fd\u628a ImageNet \u8bad\u7ec3\u96c6\u52a0\u8f7d\u65f6\u95f4\u4ece 1600 \u79d2\u7ea7\u538b\u7f29\u5230 37 \u79d2\u7ea7\u3002<\/p>\n\n\n\n
                \n\n\n\n
                8. \u66f4\u5feb\u7684\u7b97\u5b50\uff1a\u81ea\u5b9a\u4e49 kernel + \u7ecf\u9a8c\u641c\u7d22 + \u79bb\u7ebf\u7a77\u4e3e<\/h3>\n\n\n\n
                Tech-Renaissance \u5e76\u4e0d\u6ee1\u8db3\u4e8e”\u8c03 cuDNN \u5c31\u884c”\u3002\u5728\u5173\u952e\u8def\u5f84\u4e0a\uff0c\u6846\u67b6\u901a\u8fc7\u4e09\u5c42\u7b56\u7565\u69a8\u53d6\u66f4\u591a\u6027\u80fd\uff1a<\/p>\n\n\n\n
                  \n
                • \u79bb\u7ebf\u7a77\u4e3e + \u7ecf\u9a8c\u56fa\u5316<\/strong>\uff1a\u9488\u5bf9 A100\u3001RTX 5090 \u7b49\u76ee\u6807 GPU\uff0c\u9884\u5148\u5bf9 ResNet-50\u3001VGG16BN \u7b49\u6a21\u578b\u7684\u5377\u79ef\u5c42\u5f62\u72b6\u7ec4\u5408\u505a\u7a77\u4e3e\u5f0f\u5f15\u64ce\u57fa\u51c6\u6d4b\u8bd5\uff0c\u628a\u6700\u4f18\u5f15\u64ce tag \u4ee5 constexpr<\/code> \u6570\u7ec4\u7f16\u8bd1\u8fdb\u6846\u67b6\uff08cbr_experience_a100_fp16.hpp<\/code> \u7b49\uff09\u3002\u8fd0\u884c\u65f6\u901a\u8fc7\u4e8c\u5206\u67e5\u627e\u5339\u914d\u5f62\u72b6\uff0c\u96f6\u8fd0\u884c\u65f6\u641c\u7d22\u5f00\u9500\uff0c\u6309 winner \u2192 backup \u2192 \u542f\u53d1\u5f0f\u7684\u4f18\u5148\u7ea7\u6784\u5efa\u5f15\u64ce\u3002<\/li>\n\n\n\n
                • \u624b\u5199\u878d\u5408 CUDA kernel<\/strong>\uff1a\u4f18\u5316\u5668\uff08SGD\/AdamW\/LARS \u7684 weight\/momentum\/gradient \u66f4\u65b0\u878d\u6210\u5355\u6b21\u6574\u533a\u904d\u5386\uff09\u3001SoftmaxCE\u3001MaxPool \u786e\u5b9a\u6027\u53cd\u5411\u3001Philox \u968f\u673a\u6570\u751f\u6210\u7b49\u5173\u952e\u8def\u5f84\u90fd\u6709\u81ea\u5b9a\u4e49\u5b9e\u73b0\uff0c\u901a\u8fc7 __launch_bounds__<\/code> \u7cbe\u786e\u63a7\u5236\u5bc4\u5b58\u5668\u7528\u91cf\u4ee5\u4fdd\u8bc1\u5360\u7528\u7387\u3002<\/li>\n\n\n\n
                • LARS \u4e24\u9636\u6bb5\u5f52\u7ea6<\/strong>\uff1aPhase 1 \u542f\u52a8\u591a\u8fbe 65535 \u4e2a block \u5e76\u884c\u5f52\u7ea6\uff0cPhase 2 \u5355\u7ebf\u7a0b\u6c47\u603b\u8ba1\u7b97 trust ratio\uff0cFC\/\u9996\u5c42\u5377\u79ef\/\u6df1\u5c42\u5377\u79ef\u5206\u522b\u6620\u5c04\u5230\u4e09\u6761\u8ba1\u7b97\u6d41\u4e0a\u5e76\u884c\u3002<\/li>\n<\/ul>\n\n\n\n
                  \u8fd9\u4e9b\u4f18\u5316\u4e0d\u662f”\u66b4\u529b\u5199 kernel”\uff0c\u800c\u662f\u5148 profile\u3001\u518d\u641c\u7d22\u3001\u518d\u56fa\u5316\uff0c\u628a\u6bcf\u6b21\u8fed\u4ee3\u4e2d\u91cd\u590d\u6700\u591a\u6b21\u7684\u8def\u5f84\u538b\u69a8\u5230\u63a5\u8fd1\u786c\u4ef6\u6781\u9650\u3002<\/p>\n\n\n\n
                  \n\n\n\n
                  9. \u5185\u5b58\u7f13\u5b58\uff1a\u80fd\u4e0d\u8fdb\u78c1\u76d8\u5c31\u4e0d\u8fdb\u78c1\u76d8<\/h3>\n\n\n\n
                  \u5927\u591a\u6570\u8bad\u7ec3\u6846\u67b6\u5728\u6bcf\u4e2a epoch \u90fd\u4f1a\u91cd\u65b0\u8bfb\u53d6\u9a8c\u8bc1\u96c6\u3001\u91cd\u65b0\u505a\u9884\u5904\u7406\u3002\u5bf9\u4e8e\u9a8c\u8bc1\u96c6\u4e0d\u5927\u4f46 epoch \u6570\u5f88\u591a\u7684\u573a\u666f\uff08\u5982 MNIST \u8dd1 100 \u4e2a epoch\uff09\uff0c\u8fd9\u662f\u4e00\u7b14\u88ab\u4e25\u91cd\u4f4e\u4f30\u7684\u6d6a\u8d39\u3002<\/p>\n\n\n\n
                    \n
                  • CPVS\uff08Cached Preprocessed Validation Set\uff09<\/strong>\uff1a\u9a8c\u8bc1\u96c6\u5728\u9996\u6b21\u9884\u5904\u7406\u540e\u88ab\u7f13\u5b58\u5230\u5185\u5b58\uff0c\u540e\u7eed epoch \u76f4\u63a5 memcpy \u590d\u7528\uff0c\u5b8c\u5168\u8df3\u8fc7\u6587\u4ef6\u8bfb\u53d6\u548c\u56fe\u50cf\u89e3\u7801\u3002\u5728 MNIST MLP \u793a\u4f8b\u4e2d\uff0cCPVS \u6bcf epoch \u8282\u7701\u7ea6 0.16 \u79d2\uff0c100 \u4e2a epoch \u7d2f\u8ba1\u8282\u7701\u7ea6 15\u201316 \u79d2\u2014\u2014\u4ec5\u6b64\u4e00\u9879\uff0c\u52a0\u901f\u6bd4\u5c31\u4ece\u7ea6 5.1\u00d7 \u63d0\u5347\u5230\u7ea6 7.5\u00d7\u3002<\/li>\n\n\n\n
                  • FULLY \u6a21\u5f0f<\/strong>\uff1aMNIST\u3001CIFAR \u7b49\u5c0f\u6570\u636e\u96c6\u88ab\u5b8c\u6574\u52a0\u8f7d\u8fdb\u5185\u5b58\uff0c\u7b2c\u4e00\u4e2a epoch \u4e4b\u540e\u4e0d\u518d\u8bfb\u76d8\uff0cshuffle \u4e5f\u5728\u5185\u5b58\u4e2d\u5b8c\u6210\u3002ImageNet \u5728 512 GB \u5185\u5b58\u670d\u52a1\u5668\u4e0a\uff0cFULLY \u6a21\u5f0f\u80fd\u628a\u5b8c\u6574 epoch \u52a0\u8f7d\u8017\u65f6\u4ece RAW \u6587\u4ef6\u5939\u7684 96.9 \u79d2\u964d\u5230 DTS \u683c\u5f0f\u7684 25.8 \u79d2\u3002<\/li>\n<\/ul>\n\n\n\n
                    \u8fd9\u4e9b\u7f13\u5b58\u7b56\u7565\u5728 VGG16BN \u516c\u5e73\u5bf9\u6bd4\u4e2d\u9ed8\u8ba4\u5173\u95ed\uff0c\u4ee5\u4fbf\u805a\u7126\u8bad\u7ec3\u541e\u5410\u672c\u8eab\uff1b\u4f46\u5b83\u4eec\u4f5c\u4e3a\u6846\u67b6\u539f\u751f\u80fd\u529b\uff0c\u5728\u771f\u5b9e\u8bad\u7ec3\u573a\u666f\u4e2d\u80fd\u63d0\u4f9b\u7a33\u5b9a\u6536\u76ca\u3002<\/p>\n\n\n\n
                    \n\n\n\n
                    10. \u539f\u751f NHWC + 256 \u5b57\u8282\u5bf9\u9f50\uff1a\u6d88\u9664\u5e03\u5c40\u8f6c\u6362\u7a0e<\/h3>\n\n\n\n
                    Tech-Renaissance \u4ece\u6570\u636e\u52a0\u8f7d\u5230\u7b97\u5b50\u6267\u884c\uff0c\u5168\u7a0b\u7edf\u4e00\u4f7f\u7528 NHWC<\/strong> \u5185\u5b58\u5e03\u5c40\u3002\u5728 NCHW \u4e0b\uff0c\u540c\u4e00\u4e2a\u50cf\u7d20\u4f4d\u7f6e\u7684 C \u4e2a\u901a\u9053\u8de8\u8d8a\u4e86\u6574\u4e2a H\u00d7W \u5e73\u9762\uff0c\u5377\u79ef\u6838\u6ed1\u52a8\u65f6\u9700\u8981\u5bf9\u901a\u9053\u505a\u591a\u6b21\u5206\u6563\u8bbf\u95ee\uff1bNHWC \u4e0b\uff0c\u540c\u4e00\u50cf\u7d20\u7684\u901a\u9053\u503c\u5728\u5185\u5b58\u4e2d\u7d27\u6328\u7740\uff0c\u4e00\u6b21\u8bfb\u53d6\u5c31\u80fd\u62ff\u5230\u5168\u90e8\u901a\u9053\u3002\u5bf9\u4e8e Tensor Core \u7684\u77e9\u9635\u4e58\u6cd5\uff0cNHWC\uff08channels_last\uff09\u5929\u7136\u4e0e cuDNN \u7684\u6700\u4f18\u8def\u5f84\u4e00\u81f4\uff0c\u5f7b\u5e95\u907f\u514d\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u4e2d NCHW\u2194NHWC \u7684\u9690\u5f0f\u683c\u5f0f\u8f6c\u6362\u3002<\/p>\n\n\n\n
                    \u914d\u5408\u9996\u5730\u5740 256 \u5b57\u8282\u5bf9\u9f50\uff0ccuDNN \u80fd\u591f\u4ee5\u6700\u4f18\u6a21\u5f0f\u8bbf\u95ee\u6570\u636e\uff0c\u4e0d\u9700\u8981\u989d\u5916\u7684 padding \u6216 transpose\u3002DTensor \u8fd8\u8bbe\u8ba1\u4e86\u53cc\u8f68 stride\u2014\u2014GPU \u4fa7\u8d70\u5bf9\u9f50 stride \u4fdd\u8bc1\u6700\u4f18\u8bbf\u5b58\uff0cCPU \u4fa7\u8d70\u7d27\u51d1 stride \u51cf\u5c11\u4f20\u8f93\u91cf\u2014\u2014\u8ba9\u4e24\u79cd\u573a\u666f\u5404\u53d6\u6240\u9700\uff0c\u4e92\u4e0d\u59a5\u534f\u3002<\/p>\n\n\n\n
                    \n\n\n\n
                    11. AMP \u6df7\u5408\u7cbe\u5ea6\uff1a\u7b97\u5f97\u5c11\uff0c\u5b58\u5f97\u5c11\uff0c\u642c\u5f97\u5c11<\/h3>\n\n\n\n
                    FP16 \u7684\u663e\u5b58\u5360\u7528\u548c\u5e26\u5bbd\u6d88\u8017\u90fd\u662f FP32 \u7684\u4e00\u534a\u3002\u5728\u663e\u5b58\u5e26\u5bbd\u53d7\u9650\u7684 CNN \u8bad\u7ec3\u4e2d\uff0c\u8fd9\u610f\u5473\u7740\u7279\u5f81\u56fe\u53ef\u4ee5\u66f4\u5feb\u5730\u8bfb\u5199\uff0c\u76f4\u63a5\u63d0\u5347\u6709\u6548\u541e\u5410\u3002<\/p>\n\n\n\n
                    Tech-Renaissance \u7684 AMP \u5b9e\u73b0\u662f\u56fe\u7ea7\u3001\u9759\u6001\u3001\u663e\u5f0f\u7684<\/strong>\uff1a\u6240\u6709\u7cbe\u5ea6\u8f6c\u6362\uff08CAST_FP32_TO_FP16\u3001CAST_FP16_TO_FP32\uff09\u3001\u635f\u5931\u7f29\u653e\uff08\u56fa\u5b9a\u503c\uff0c\u53ea\u51cf\u4e0d\u589e\uff09\u3001NaN \u68c0\u6d4b\u4e0e\u68af\u5ea6\u88c1\u526a\uff0c\u90fd\u5728\u7f16\u8bd1\u671f\u5199\u5165\u8ba1\u7b97\u56fe\uff0c\u8fd0\u884c\u671f\u6309\u56fa\u5b9a\u987a\u5e8f\u542f\u52a8 CUDA Graph\u3002\u4e0e PyTorch \u7684\u52a8\u6001 AMP\uff08autocast + GradScaler\uff09\u4e0d\u540c\uff0c\u8fd9\u79cd\u9759\u6001\u65b9\u5f0f\u907f\u514d\u4e86\u8fd0\u884c\u65f6\u7684\u4e0a\u4e0b\u6587\u5207\u6362\u548c\u52a8\u6001\u7f29\u653e\u53d8\u91cf\u7684\u989d\u5916\u5f00\u9500\uff0c\u4e14\u5929\u7136\u9002\u5408 CUDA Graph \u5168\u6355\u83b7\u3002CBR \u878d\u5408\u3001FusedNormalization\u3001Range Cast \u7b49\u8bbe\u8ba1\u4e5f\u90fd\u56f4\u7ed5 AMP \u8def\u5f84\u505a\u4e86\u6df1\u5ea6\u9002\u914d\u3002<\/p>\n\n\n\n
                    \n\n\n\n
                    \u5c0f\u7ed3\uff1a\u53e0\u52a0\uff0c\u800c\u4e0d\u662f\u5355\u70b9<\/h3>\n\n\n\n
                    \u8fd9 11 \u6761\u8bbe\u8ba1\u5355\u72ec\u770b\u90fd\u4e0d\u7b97\u60ca\u5929\u52a8\u5730\uff0c\u4f46\u7ec4\u5408\u8d77\u6765\u5c31\u5f62\u6210\u4e86\u53ef\u89c2\u7684\u7efc\u5408\u6536\u76ca\u3002\u6211\u4eec\u53ef\u4ee5\u628a\u5b83\u4eec\u6620\u5c04\u5230\u8bad\u7ec3\u7ba1\u7ebf\u7684\u65f6\u95f4\u8f74\u4e0a\uff0c\u770b\u770b\u6bcf\u4e2a\u9636\u6bb5\u5206\u522b\u88ab\u54ea\u4e9b\u8bbe\u8ba1\u4f18\u5316\u8fc7\uff1a<\/p>\n\n\n\n
                    \u7ba1\u7ebf\u9636\u6bb5<\/th>\u5bf9\u5e94\u7684\u4f18\u5316\u8bbe\u8ba1<\/th><\/tr><\/thead>
                    \u6570\u636e\u5728\u78c1\u76d8\u4e0a<\/td>DTS \u81ea\u5b9a\u4e49\u683c\u5f0f\u3001FULLY \u5168\u91cf\u52a0\u8f7d<\/td><\/tr>
                    CPU \u9884\u5904\u7406<\/td>\u591a\u7ebf\u7a0b\u9759\u6001\u9886\u53d6\u3001FusedNormalization \u878d\u5408\u3001NUMA \u611f\u77e5\u9501\u9875\u5185\u5b58<\/td><\/tr>
                    CPU\u2192GPU \u4f20\u8f93<\/td>TransferStation \u53cc\u7f13\u51b2\u3001TRANS \u6d41\u5f02\u6b65\u62f7\u8d1d\u3001AMP FP16 \u51cf\u534a\u4f20\u8f93\u91cf<\/td><\/tr>
                    GPU \u524d\u5411\/\u53cd\u5411<\/td>NHWC \u539f\u751f\u5e03\u5c40\u3001AMP \u6df7\u5408\u7cbe\u5ea6\u3001CBR \u878d\u5408\u7b97\u5b50\u3001\u4e09\u8ba1\u7b97\u6d41\u5e76\u53d1<\/td><\/tr>
                    \u68af\u5ea6\u901a\u4fe1<\/td>\u4e24\u6876 Region AllReduce\u3001UPDATE \u6d41\u901a\u4fe1\u8ba1\u7b97\u91cd\u53e0<\/td><\/tr>
                    \u4f18\u5316\u5668\u66f4\u65b0<\/td>\u878d\u5408\u4f18\u5316\u5668\u6574\u533a RangeOp\u3001LARS \u4e09\u6d41\u5e76\u884c\u5f52\u7ea6<\/td><\/tr>
                    \u7cbe\u5ea6\/\u72b6\u6001\u7ef4\u62a4<\/td>FP32 \u4e3b\u6743\u91cd\u3001\u9759\u6001 loss scaling\u3001NaN \u68c0\u6d4b\u3001EMA \u66f4\u65b0<\/td><\/tr>
                    \u5168\u9636\u6bb5\u8c03\u5ea6<\/td>\u9759\u6001\u56fe\u7f16\u8bd1\u3001CUDA Graph \u5168\u6355\u83b7\u3001MemoryPlan \u5730\u5740\u7a33\u5b9a<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                    \u6240\u4ee5\uff0c\u4e0b\u4e00\u6b21\u6709\u4eba\u95ee\u4f60”\u8fd9\u4e2a\u6846\u67b6\u51ed\u4ec0\u4e48\u6bd4 PyTorch \u5feb”\u7684\u65f6\u5019\uff0c\u6700\u51c6\u786e\u7684\u56de\u7b54\u4e0d\u662f\u67d0\u4e00\u6761\u4f18\u5316\uff0c\u800c\u662f\uff1a\u56e0\u4e3a\u5b83\u628a\u4ece\u78c1\u76d8\u5230 GPU \u7684\u6574\u6761\u8bad\u7ec3\u7ba1\u7ebf\uff0c\u90fd\u5f53\u6210\u4e00\u4e2a\u7cfb\u7edf\u5de5\u7a0b\u91cd\u65b0\u8bbe\u8ba1\u4e86\u4e00\u904d\u3002<\/strong> VGG16BN \u4e0a +26.65% \u7684\u541e\u5410\u3001MNIST MLP \u4e0a 7\u201313 \u500d\u7684\u52a0\u901f\uff0c\u5c31\u662f\u8fd9\u4e2a\u7cfb\u7edf\u5de5\u7a0b\u7684\u81ea\u7136\u7ed3\u679c\u3002\u6bcf\u4e00\u70b9\u4f18\u5316\u90fd\u516c\u5f00\u5728\u6e90\u7801\u548c\u914d\u7f6e\u91cc\uff0c\u7ecf\u5f97\u8d77\u590d\u73b0\u548c\u5ba1\u8ba1\u3002<\/p>\n\n\n\n
                    \u516d\u3001\u516c\u5e73\u5bf9\u6bd4\u7684\u65b9\u6cd5\u8bba\uff1a\u8ba9\u5bf9\u624b\u4e5f\u6b66\u88c5\u5230\u7259\u9f7f<\/h2>\n\n\n\n
                    \u6211\u4eec\u5728\u5bf9\u6bd4\u4e2d\u9075\u5faa\u51e0\u4e2a\u539f\u5219\uff0c\u4e5f\u5efa\u8bae\u4efb\u4f55\u505a\u6846\u67b6 benchmark \u7684\u4eba\u9075\u5faa\uff1a<\/p>\n\n\n\n
                      \n
                    1. \u540c\u786c\u4ef6\u3001\u540c\u8f6f\u4ef6\u6808<\/strong>\uff1aCUDA\u3001cuDNN\u3001\u9a71\u52a8\u7248\u672c\u4fdd\u6301\u4e00\u81f4\uff1b<\/li>\n\n\n\n
                    2. \u6a21\u578b\u7ed3\u6784\u4e0e\u8d85\u53c2\u6570\u4e25\u683c\u5bf9\u9f50<\/strong>\uff1a\u9010\u5c42\u6838\u5bf9\u3001\u9010\u53c2\u6570\u6838\u5bf9\uff0c\u4e0d\u653e\u8fc7 weight decay \u5206\u7ec4\u3001label smoothing\u3001warm-up \u533a\u95f4\uff1b<\/li>\n\n\n\n
                    3. \u5f00\u542f\u5bf9\u624b\u7684\u6700\u5f3a\u6a21\u5f0f<\/strong>\uff1aPyTorch \u7528 torch.compile(mode=\"max-autotune\")<\/code>\uff0cTensorFlow \u5f00 XLA\uff0cDDP\/SyncBN\/channels_last\/pin_memory \u5168\u90e8\u5230\u4f4d\uff1b<\/li>\n\n\n\n
                    4. \u7f16\u8bd1\u65f6\u95f4\u9694\u79bb<\/strong>\uff1aTech-Renaissance \u7684 task.compile()<\/code> \u4e0d\u8ba1\u5165\u603b\u8017\u65f6\uff0cPyTorch \u901a\u8fc7 dummy batch \u89e6\u53d1\u7f16\u8bd1\u540e\u91cd\u65b0\u521d\u59cb\u5316\uff1b<\/li>\n\n\n\n
                    5. \u591a\u5e73\u53f0\u4ea4\u53c9\u9a8c\u8bc1<\/strong>\uff1aMLP \u5728 7 \u4e2a\u4e0d\u540c GPU \u5e73\u53f0\u4e0a\u90fd\u8dd1\u51fa\u4f18\u52bf\uff0c\u907f\u514d\u5355\u4e00\u786c\u4ef6\u7684\u5076\u7136\u6027\uff1b<\/li>\n\n\n\n
                    6. \u62ab\u9732\u5dee\u5f02\u9879\u4e0e\u5c40\u9650<\/strong>\uff1a\u4e0d\u9690\u85cf\u6846\u67b6\u80fd\u529b\u5dee\u5f02\uff0c\u4e5f\u4e0d\u628a\u7279\u5b9a\u4efb\u52a1\u7684\u7ed3\u8bba\u6cdb\u5316\u3002<\/li>\n<\/ol>\n\n\n\n
                      \u6b63\u5982 docs\/VGG16BN_FAIRNESS.md<\/code> \u91cc\u6240\u5199\u7684\uff1a\u201c\u67b6\u6784\u4f18\u52bf \u2260 \u4e0d\u516c\u5e73\u3002\u201d<\/strong> Tech-Renaissance \u7684 FusedNormalization\u3001\u539f\u751f NHWC\u3001\u9759\u6001 CUDA Graph \u90fd\u662f\u67b6\u6784\u5c42\u9762\u7684\u771f\u5b9e\u80fd\u529b\uff0cPyTorch \u7684 torch.compile + DDP + SyncBN<\/code> \u4e5f\u662f\u5b83\u7684\u6838\u5fc3\u7ade\u4e89\u529b\u3002\u6211\u4eec\u6bd4\u7684\u662f\u4e24\u4e2a\u6846\u67b6\u5728\u5404\u81ea\u6700\u4f18\u914d\u7f6e\u4e0b\u7684\u5b9e\u9645\u80fd\u529b\uff0c\u800c\u4e0d\u662f\u8ba9\u67d0\u4e00\u65b9\u88f8\u5954\u3002<\/p>\n\n\n\n
                      \u4e0b\u8868\u6982\u62ec\u4e86 docs\/VGG16BN_FAIRNESS.md<\/code> \u4e2d\u5ba1\u8ba1\u8fc7\u7684\u4e3b\u8981\u7ed3\u6784\u6027\u5dee\u5f02\uff0c\u4f9b\u8bfb\u8005\u5224\u65ad\u7ed3\u8bba\u9002\u7528\u8303\u56f4\uff1a<\/p>\n\n\n\n
                      \u5dee\u5f02\u9879<\/th>\u504f\u5411<\/th>\u5f71\u54cd\u91cf\u7ea7<\/th>\u662f\u5426\u5c5e\u4e8e\u516c\u5e73\u5dee\u5f02<\/th><\/tr><\/thead>
                      FusedNormalization \u878d\u5408\u9884\u5904\u7406<\/td>Tech-Renaissance<\/td>\u4e2d\u7b49<\/td>\u67b6\u6784\u80fd\u529b<\/td><\/tr>
                      Workers \u67b6\u6784\uff08I\/O \u4e0e\u9884\u5904\u7406\u7ebf\u7a0b\u5206\u79bb\uff09<\/td>\u7565\u5fae\u504f\u5411 Tech-Renaissance<\/td>\u5c0f<\/td>\u67b6\u6784\u80fd\u529b<\/td><\/tr>
                      \u8ba1\u7b97\u56fe\u6267\u884c\u6a21\u578b\uff08\u9759\u6001 CUDA Graph vs torch.compile\uff09<\/td>\u5404\u81ea\u6838\u5fc3\u4f18\u52bf<\/td>\u6838\u5fc3<\/td>\u8bbe\u8ba1\u8def\u7ebf\u5dee\u5f02<\/td><\/tr>
                      AMP \u7b56\u7565\uff08\u56fa\u5b9a\u7f29\u653e vs \u52a8\u6001\u7f29\u653e\uff09<\/td>\u4e92\u6709\u62b5\u6d88<\/td>\u4e2d\u7b49<\/td>\u5b9e\u73b0\u5dee\u5f02<\/td><\/tr>
                      \u591a GPU \u901a\u4fe1\uff08DDP \/ \u4e24\u6876\u5206\u6876 AllReduce\uff09<\/td>\u65e0\u663e\u8457\u504f\u5411<\/td>\u5c0f<\/td>\u5b9e\u73b0\u5dee\u5f02<\/td><\/tr>
                      \u5185\u5b58\u5e03\u5c40\uff08\u539f\u751f NHWC vs channels_last\uff09<\/td>\u7565\u5fae\u504f\u5411 Tech-Renaissance<\/td>\u5c0f<\/td>\u67b6\u6784\u80fd\u529b<\/td><\/tr>
                      \u68af\u5ea6\u6e05\u96f6\u65b9\u5f0f\uff08set_to_none\uff09<\/td>\u7565\u5fae\u504f\u5411 PyTorch<\/td>\u6781\u5c0f<\/td>\u5b9e\u73b0\u5dee\u5f02<\/td><\/tr>
                      ReLU inplace=True<\/td>\u7565\u5fae\u504f\u5411 PyTorch<\/td>\u5c0f<\/td>\u5b9e\u73b0\u5dee\u5f02<\/td><\/tr>
                      torch.compile \u5c3e batch \u9884\u70ed<\/td>\u7565\u5fae\u504f\u5411 Tech-Renaissance\uff08\u8ba1\u65f6\u53e3\u5f84\uff09<\/td>\u5c0f~\u4e2d<\/td>\u5df2\u77e5\u672a\u5b8c\u5168\u5bf9\u9f50\u9879<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                      \u8fd9\u5f20\u8868\u91cc\u6ca1\u6709\u201c\u5077\u5077\u5f00\u542f\u7684\u9690\u85cf\u5f00\u5173\u201d\uff0c\u6bcf\u4e00\u9879\u90fd\u53ef\u4ee5\u5bf9\u5e94\u5230\u516c\u5f00\u7684\u6e90\u7801\u6216\u914d\u7f6e\u53c2\u6570\u3002\u8fd9\u4e5f\u662f\u6211\u4eec\u6562\u4e8e\u628a\u6027\u80fd\u6570\u5b57\u653e\u51fa\u6765\u7684\u5e95\u6c14\u3002<\/p>\n\n\n\n
                      \u4e03\u3001\u672a\u6765\uff1a\u6027\u80fd\u53ea\u662f\u5f00\u59cb<\/h2>\n\n\n\n
                      V4.20.692 \u7248\u7684\u53d1\u5e03\uff0c\u53ea\u662f\u4e00\u4e2a\u91cc\u7a0b\u7891\uff0c\u4e0d\u662f\u7ec8\u70b9\u3002<\/p>\n\n\n\n
                      \u63a5\u4e0b\u6765 Tech-Renaissance \u8fd8\u4f1a\u7ee7\u7eed\u671d\u51e0\u4e2a\u65b9\u5411\u8d70\uff1a<\/p>\n\n\n\n
                        \n
                      • \u66f4\u6743\u5a01\u7684\u6027\u80fd\u9a8c\u8bc1<\/strong>\uff1a\u5c1d\u8bd5\u5728MLPerf\u7b49\u77e5\u540dbenchmark\u89c4\u5219\u4e0b\u8ddf\u4e3b\u6d41\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u8fdb\u884c\u66f4\u591a\u516c\u5e73\u5bf9\u6bd4\uff1b<\/li>\n\n\n\n
                      • \u66f4\u591a\u6a21\u578b\u7c7b\u578b<\/strong>\uff1a\u4ece CNN \u6269\u5c55\u5230 Transformer\u3001BERT\u3001ViT\uff0c\u751a\u81f3\u5927\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\uff1b<\/li>\n\n\n\n
                      • \u66f4\u7cfb\u7edf\u5316\u7684 profiling \u5de5\u5177\u94fe<\/strong>\uff1a\u7528 nsys \u6216\u81ea\u5b9a\u4e49 timeline \u628a\u6bcf\u4e00\u6beb\u79d2\u7684\u8017\u65f6\u6765\u6e90\u62c6\u6e05\u695a\uff1b<\/li>\n\n\n\n
                      • \u66f4\u4e30\u5bcc\u7684\u4f18\u5316\u5668\u4e0e\u8c03\u5ea6\u5668<\/strong>\uff1a\u8986\u76d6 Adam\u3001SGD\u3001LARS\u3001Shampoo \u7b49\u66f4\u591a\u8bad\u7ec3\u7b97\u6cd5\uff1b<\/li>\n\n\n\n
                      • \u66f4\u597d\u7684\u751f\u6001\u4e0e\u8fc1\u79fb\u8def\u5f84<\/strong>\uff1a\u964d\u4f4e PyTorch \u7528\u6237\u8fc1\u79fb\u5230 Tech-Renaissance \u7684\u6210\u672c\uff1b<\/li>\n\n\n\n
                      • \u6301\u7eed\u7684\u7b97\u5b50\u6b63\u786e\u6027\u5ba1\u8ba1<\/strong>\uff1a\u6bcf\u65b0\u589e\u4e00\u4e2a\u7b97\u5b50\u3001\u6bcf\u505a\u4e00\u6b21\u4f18\u5316\uff0c\u90fd\u8981\u5148\u8fc7 correction\/perf\/example \u4e09\u5c42\u6d4b\u8bd5\u3002<\/li>\n<\/ul>\n\n\n\n
                        \u6211\u4eec\u4e5f\u4e0d\u8bb3\u8a00\u81ea\u5df1\u7684\u77ed\u677f\uff1a\u751f\u6001\u3001\u8c03\u8bd5\u4f53\u9a8c\u3001\u52a8\u6001\u56fe\u7075\u6d3b\u6027\u3001\u793e\u533a\u89c4\u6a21\uff0c\u8fd9\u4e9b\u90fd\u4e0d\u662f\u4e00\u4e2a\u4eba\u6216\u5c0f\u56e2\u961f\u77ed\u671f\u5185\u80fd\u8ffd\u5e73\u7684\u3002Tech-Renaissance \u7684\u5b9a\u4f4d\u4e00\u76f4\u5f88\u6e05\u6670\u2014\u2014\u5b83\u662f\u4e00\u4e2a\u8ffd\u6c42\u6781\u81f4\u8bad\u7ec3\u541e\u5410\u91cf\u7684\u751f\u4ea7\u7ea7\u8bad\u7ec3\u6846\u67b6\uff0c\u800c\u4e0d\u662f PyTorch \u7684\u901a\u7528\u66ff\u4ee3\u54c1\u3002<\/p>\n\n\n\n
                        \u516b\u3001\u7ed3\u8bed<\/h2>\n\n\n\n
                        \u56de\u5230\u7b2c\u4e00\u7bc7\u63d0\u51fa\u7684\u95ee\u9898\uff1a\u4e00\u4e2a\u4eba\uff0c\u7528 AI\uff0c\u80fd\u4e0d\u80fd\u5199\u51fa\u6bd4 PyTorch \u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff1f<\/p>\n\n\n\n
                        \u7ecf\u8fc7\u8fd9 25 \u7bc7\u6587\u7ae0\uff0c\u6211\u60f3\u7b54\u6848\u5df2\u7ecf\u6bd4\u8f83\u6e05\u695a\u4e86\uff1a\u80fd\uff0c\u4f46\u4f60\u9700\u8981\u5728\u638c\u63e1\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u6838\u5fc3\u67b6\u6784\u7684\u57fa\u7840\u4e0a\uff0c\u9488\u5bf9\u7279\u5b9a\u9886\u57df\u6216\u5e94\u7528\u573a\u666f\u8fdb\u884c\u6781\u81f4\u4f18\u5316\uff0c\u4ece\u5934\u8bbe\u8ba1\u4e00\u6761\u56f4\u7ed5\u9759\u6001\u56fe\u3001CUDA Graph\u3001\u9759\u6001\u663e\u5b58\u89c4\u5212\u3001\u7b97\u5b50\u878d\u5408\u548c\u786e\u5b9a\u6027\u6267\u884c\u7684\u7ba1\u7ebf\u3002<\/strong><\/p>\n\n\n\n
                        \u4ece\u6280\u672f\u8def\u5f84\u7684\u89d2\u5ea6\uff0c\u8fd9\u4e2a\u9879\u76ee\u8bf4\u660e\u4e24\u70b9\uff1a\u7b2c\u4e00\uff0c\u5982\u679c\u4f60\u8ffd\u6c42\u6781\u81f4\u7684\u8bad\u7ec3\u541e\u5410\u91cf\uff0c\u613f\u610f\u727a\u7272\u4e00\u4e9b\u7075\u6d3b\u6027\uff0c\u6781\u81f4\u4f18\u5316\u7684\u9759\u6001\u56fe + \u5168 C++ + CUDA Graph \u8fd9\u6761\u8def\uff0c\u80fd\u7ed9\u4f60\u8d85\u51fa\u9884\u671f\u7684\u56de\u62a5\uff1b\u7b2c\u4e8c\uff0cAI\u7f16\u7a0b\u5927\u5e45\u63d0\u5347\u5f00\u53d1\u6548\u7387\uff0c\u4f46\u5b83\u53ea\u662f\u4e00\u4e2a\u8fdb\u5ea6\u7684\u52a0\u901f\u5668\uff0c\u8981\u771f\u6b63\u5b8c\u6210\u4e00\u4e2a\u6027\u80fd\u4f18\u826f\u7684\u9879\u76ee\u7684\u8bdd\uff0c\u79bb\u4e0d\u5f00\u9884\u5148\u51c6\u5907\u5927\u91cf\u7684\u67b6\u6784\u8bbe\u8ba1\u89c4\u5219\u548c\u63d0\u4f9b\u9002\u5f53\u7684\u4eba\u5de5\u5e72\u9884\uff0c\u5982\u679c\u4f60\u8fd8\u8981\u66f4\u52a0\u6781\u81f4\u7684\u6027\u80fd\uff0c\u90a3\u4f60\u5c31\u9700\u8981\u5c55\u5f00\u66f4\u52a0\u6df1\u5165\u800c\u5f7b\u5e95\u7684\u521b\u65b0\u3002<\/p>\n\n\n\n
                        \u5982\u679c\u4f60\u5bf9\u8fd9\u4e2a\u9879\u76ee\u611f\u5174\u8da3\uff0c\u6b22\u8fce\u53bb GitHub \u6216 Gitee clone \u4e0b\u6765\u8dd1\u4e00\u8dd1\uff0c\u8bd5\u4e00\u8bd5\uff0c\u63d0 Issue\uff0c\u63d0 PR\u3002\u8fd9\u662f\u4e00\u4e2a\u5b8c\u5168\u5f00\u6e90\u7684\u9879\u76ee\uff0c\u6b22\u8fce\u6240\u6709\u611f\u5174\u8da3\u7684\u670b\u53cb\u4e00\u8d77\u6765\u73a9\u3002<\/p>\n\n\n\n
                        <\/p>\n","protected":false},"excerpt":{"rendered":"
                        \u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341\u4e94 \u4ece\u7b2c\u4e00\u7bc7\u63d0\u51fa\u201c\u4e00\u4e2a\u4eba\u7528 A […]<\/p>\n","protected":false},"author":1,"featured_media":598,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-container-style":"default","site-container-layout":"default","site-sidebar-layout":"default","disable-article-header":"default","disable-site-header":"default","disable-site-footer":"default","disable-content-area-spacing":"default","footnotes":""},"categories":[15],"tags":[],"class_list":["post-571","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-15"],"_links":{"self":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/571","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/comments?post=571"}],"version-history":[{"count":6,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/571\/revisions"}],"predecessor-version":[{"id":733,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/571\/revisions\/733"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/598"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=571"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=571"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=571"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}