{"id":559,"date":"2026-07-08T03:13:56","date_gmt":"2026-07-07T19:13:56","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=559"},"modified":"2026-07-08T17:53:38","modified_gmt":"2026-07-08T09:53:38","slug":"philox%e5%8f%af%e5%a4%8d%e7%8e%b0%e9%9a%8f%e6%9c%ba%e6%95%b0%e4%b8%8e%e7%a1%ae%e5%ae%9a%e6%80%a7%e8%ae%ad%e7%bb%83","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/559\/","title":{"rendered":"(23) Philox\u53ef\u590d\u73b0\u968f\u673a\u6570\u4e0e\u786e\u5b9a\u6027\u8bad\u7ec3"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\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\u4e09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u65f6\uff0c\u968f\u673a\u6570\u770b\u8d77\u6765\u53ea\u662f\u80cc\u666f\u566a\u97f3\uff1a\u6743\u91cd\u521d\u59cb\u5316\u3001\u6570\u636e\u589e\u5f3a\u3001Dropout \u63a9\u7801\u3001\u6570\u636e\u96c6\u6d17\u724c\u3001BatchNorm \u7684\u67d0\u4e9b\u5b9e\u73b0\u2026\u2026\u5b83\u4eec\u65e0\u5904\u4e0d\u5728\uff0c\u5374\u5f88\u5c11\u88ab\u653e\u5728\u805a\u5149\u706f\u4e0b\u3002\u7136\u800c\uff0c\u5f53\u4f60\u771f\u6b63\u9700\u8981\u4e00\u4e2a<strong>\u53ef\u590d\u73b0<\/strong>\u7684\u8bad\u7ec3\u7ed3\u679c\u65f6\u2014\u2014\u6bd4\u5982\u8c03\u8bd5\u4e00\u4e2a\u8be1\u5f02\u7684\u6536\u655b\u95ee\u9898\u3001\u505a\u6d88\u878d\u5b9e\u9a8c\u3001\u6216\u8005\u5411\u5ba1\u7a3f\u4eba\u8bc1\u660e\u201c\u8fd9\u4e2a\u6539\u52a8\u786e\u5b9e\u6709\u6548\u201d\u2014\u2014\u968f\u673a\u6570\u7acb\u523b\u4ece\u5e55\u540e\u8d70\u5230\u53f0\u524d\uff0c\u6210\u4e3a\u51b3\u5b9a\u6210\u8d25\u7684\u5173\u952e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u590d\u73b0\u6027\u8fd9\u4ef6\u4e8b\uff0c\u8bf4\u8d77\u6765\u7b80\u5355\uff1a\u56fa\u5b9a\u79cd\u5b50\uff0c\u4e0d\u5c31\u80fd\u5f97\u5230\u56fa\u5b9a\u7ed3\u679c\u5417\uff1f\u4f46\u771f\u6b63\u5728\u751f\u4ea7\u7ea7\u6846\u67b6\u91cc\u628a\u5b83\u505a\u7a33\uff0c\u8fdc\u6bd4\u60f3\u8c61\u4e2d\u590d\u6742\u3002\u591a\u7ebf\u7a0b\u6570\u636e\u52a0\u8f7d\u4f1a\u6253\u4e71\u91c7\u6837\u987a\u5e8f\uff0c\u591a\u5361\u5e76\u884c\u7684\u539f\u5b50\u52a0\u6cd5\u4f1a\u6539\u53d8\u6d6e\u70b9\u6c42\u548c\u987a\u5e8f\uff0cCUDA Graph \u91cc\u7684\u968f\u673a\u72b6\u6001\u4e0d\u80fd\u6bcf\u6b21\u4ece CPU \u91cd\u65b0\u5582\uff0c\u800c\u4f20\u7edf\u57fa\u4e8e\u72b6\u6001\u9012\u63a8\u7684\u4f2a\u968f\u673a\u6570\u751f\u6210\u5668\uff08\u5982 Mersenne Twister\uff09\u4e00\u65e6\u9047\u5230\u5e76\u884c\u6267\u884c\u987a\u5e8f\u53d8\u5316\uff0c\u6574\u4e2a\u5e8f\u5217\u5c31\u4f1a\u9519\u4f4d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7bc7\u6587\u7ae0\u5c31\u6765\u804a\u804a Tech-Renaissance \u5982\u4f55\u628a<strong>Philox \u8ba1\u6570\u5668\u578b\u968f\u673a\u6570\u751f\u6210\u5668<\/strong>\u8d2f\u7a7f\u5230\u8bad\u7ec3\u5168\u6d41\u7a0b\uff0c\u5e76\u4ee5\u6b64\u4e3a\u57fa\u7840\u6784\u5efa\u7aef\u5230\u7aef\u7684\u786e\u5b9a\u6027\u8bad\u7ec3\u80fd\u529b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u4e3a\u4ec0\u4e48\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u7684\u53ef\u590d\u73b0\u6027\u8fd9\u4e48\u96be\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u673a\u6027\u5728\u6df1\u5ea6\u5b66\u4e60\u91cc\u5927\u81f4\u6709\u4e09\u7c7b\u7528\u9014\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u521d\u59cb\u5316<\/strong>\uff1a\u6743\u91cd\u3001\u504f\u7f6e\u7684\u968f\u673a\u521d\u59cb\u5316\u51b3\u5b9a\u4f18\u5316\u7684\u8d77\u70b9\uff1b<\/li>\n\n\n\n<li><strong>\u6570\u636e\u4fa7<\/strong>\uff1a\u6570\u636e\u96c6\u6d17\u724c\u3001\u968f\u673a\u88c1\u526a\u3001\u968f\u673a\u7ffb\u8f6c\u3001\u989c\u8272\u6296\u52a8\u3001Random Erasing \u7b49\u589e\u5f3a\uff1b<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u4fa7<\/strong>\uff1aDropout\u3001\u67d0\u4e9b\u91c7\u6837\u6216\u6270\u52a8\u64cd\u4f5c\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8981\u8ba9\u8fd9\u4e9b\u73af\u8282\u5168\u90e8\u53ef\u590d\u73b0\uff0c\u6838\u5fc3\u4e0d\u662f\u201c\u6709\u6ca1\u6709\u968f\u673a\u6570\u201d\uff0c\u800c\u662f<strong>\u968f\u673a\u6570\u7684\u6d88\u8d39\u987a\u5e8f\u5fc5\u987b\u5b8c\u5168\u4e00\u81f4<\/strong>\u3002\u4f20\u7edf PRNG\uff08\u5982 C++ <code>std::mt19937<\/code>\u3001NumPy \u9ed8\u8ba4\u7684 PCG64\uff09\u90fd\u662f\u201c\u72b6\u6001\u673a\u201d\uff1a\u6bcf\u751f\u6210\u4e00\u4e2a\u6570\u5c31\u63a8\u8fdb\u4e00\u6b21\u5185\u90e8\u72b6\u6001\u3002\u5982\u679c\u4e24\u4e2a\u7ebf\u7a0b\u540c\u65f6\u8bf7\u6c42\u968f\u673a\u6570\uff0c\u8c01\u5148\u8c01\u540e\u4f1a\u6539\u53d8\u540e\u7eed\u6240\u6709\u4eba\u62ff\u5230\u7684\u5e8f\u5217\uff1b\u5982\u679c\u7b2c 3 \u4e2a epoch \u7684 worker \u6570\u91cf\u53d8\u4e86\uff0c\u5168\u5c40\u72b6\u6001\u4e5f\u4f1a\u4e0d\u540c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u72b6\u6001\u673a\u5f0f RNG \u7684\u7ea7\u8054\u6545\u969c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u72b6\u6001\u673a\u5f0f RNG \u6700\u5927\u7684\u9ebb\u70e6\u5728\u4e8e\u201c\u7ea7\u8054\u504f\u79fb\u201d\u3002\u5047\u8bbe\u6709\u4e24\u4e2a\u9884\u5904\u7406 worker\uff0c\u5171\u4eab\u540c\u4e00\u4e2a Mersenne Twister \u751f\u6210\u5668\u3002worker A \u5728\u67d0\u6b21\u8fed\u4ee3\u91cc\u6bd4 worker B \u591a\u6d88\u8d39\u4e86\u4e00\u4e2a\u968f\u673a\u6570\u2014\u2014\u53ef\u80fd\u662f\u56e0\u4e3a JPEG \u89e3\u7801\u901f\u5ea6\u4e0d\u540c\uff0c\u4e5f\u53ef\u80fd\u662f\u56e0\u4e3a CPU \u8c03\u5ea6\u6296\u52a8\u2014\u2014\u90a3\u4e48\u4ece\u8fd9\u4e00\u523b\u5f00\u59cb\uff0c\u4e24\u4e2a worker \u62ff\u5230\u7684\u6240\u6709\u540e\u7eed\u968f\u673a\u6570\u90fd\u4f1a\u53d1\u751f\u9519\u4f4d\u3002\u6570\u636e\u589e\u5f3a\u7684\u968f\u673a\u88c1\u526a\u4f4d\u7f6e\u53d8\u4e86\uff0cRandom Erasing \u7684\u533a\u57df\u53d8\u4e86\uff0c\u5982\u679c\u5b83\u4eec\u8fd8\u5171\u4eab\u540c\u4e00 RNG \u505a\u5176\u4ed6\u4e8b\uff0c\u6743\u91cd\u521d\u59cb\u5316\u4e5f\u4f1a\u8ddf\u7740\u6f02\u79fb\u3002\u6700\u7ec8\u4f60\u770b\u5230\u7684\u7ed3\u679c\u4e0d\u662f\u201c\u67d0\u4e2a\u6837\u672c\u589e\u5f3a\u4e0d\u540c\u201d\uff0c\u800c\u662f\u6574\u6761\u8bad\u7ec3\u8f68\u8ff9\u7684\u5206\u53c9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u66f4\u68d8\u624b\u7684\u662f\uff0c\u8fd9\u79cd\u95ee\u9898\u901a\u5e38\u4e0d\u4f1a\u5d29\u6e83\uff0c\u800c\u662f\u201c\u6084\u6084\u5730\u9519\u201d\u3002\u6a21\u578b\u7167\u6837\u6536\u655b\uff0closs \u66f2\u7ebf\u770b\u8d77\u6765\u6b63\u5e38\uff0c\u4f46\u4e24\u6b21\u8fd0\u884c\u7684\u6743\u91cd\u3001\u7cbe\u5ea6\u3001\u751a\u81f3\u8d85\u53c2\u6570\u654f\u611f\u6027\u90fd\u4e0d\u540c\u3002\u8c03\u8bd5\u65f6\u4f60\u6839\u672c\u65e0\u6cd5\u5224\u65ad\u8fd9\u662f\u7b97\u6cd5\u672c\u8eab\u7684\u7279\u6027\uff0c\u8fd8\u662f\u968f\u673a\u6570\u987a\u5e8f\u88ab\u6270\u52a8\u5bfc\u81f4\u7684\u566a\u58f0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3b\u6d41\u6846\u67b6\u7684\u505a\u6cd5\u662f\u201c\u5206\u800c\u6cbb\u4e4b\u201d\uff1aPyTorch \u4e3a CPU \u548c CUDA \u5206\u522b\u7ef4\u62a4 RNG \u72b6\u6001\uff0cCPU \u7528 Mersenne Twister\uff0cCUDA \u7528 Philox\uff1b\u7528\u6237\u9700\u8981\u624b\u52a8\u8bbe\u7f6e <code>torch.manual_seed<\/code>\u3001<code>torch.cuda.manual_seed_all<\/code>\uff0c\u4e3a\u6bcf\u4e2a DataLoader worker \u5199 <code>worker_init_fn<\/code>\uff0c\u518d\u5173\u6389 <code>cudnn.benchmark<\/code>\u3001\u6253\u5f00 <code>cudnn.deterministic<\/code>\uff0c\u751a\u81f3\u542f\u7528 <code>torch.use_deterministic_algorithms(True)<\/code>\u3002\u5373\u4fbf\u5982\u6b64\uff0c\u5b98\u65b9\u6587\u6863\u4e5f\u5766\u627f\uff1a\u8de8 PyTorch \u7248\u672c\u3001\u8de8\u5e73\u53f0\u3001CPU \u4e0e GPU \u4e4b\u95f4\u4ecd\u65e0\u6cd5\u4fdd\u8bc1\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6839\u672c\u539f\u56e0\u5728\u4e8e\uff0c\u72b6\u6001\u673a\u5f0f RNG \u7684\u201c\u53ef\u590d\u73b0\u201d\u662f\u5efa\u7acb\u5728<strong>\u6267\u884c\u987a\u5e8f\u4e25\u683c\u4e0d\u53d8<\/strong>\u7684\u524d\u63d0\u4e0a\u7684\uff0c\u800c\u73b0\u4ee3\u8bad\u7ec3\u7ba1\u7ebf\u5929\u7136\u5c31\u662f\u4e71\u5e8f\u5e76\u884c\u7684\u3002\u8981\u8df3\u51fa\u8fd9\u4e2a\u56f0\u5883\uff0c\u6700\u597d\u662f\u8ba9\u968f\u673a\u6570\u751f\u6210\u672c\u8eab\u4e0d\u518d\u4f9d\u8d56\u6267\u884c\u987a\u5e8f\u2014\u2014\u8fd9\u5c31\u662f<strong>\u8ba1\u6570\u5668\u578b RNG\uff08Counter-Based RNG\uff09<\/strong>\u7684\u51fa\u53d1\u70b9\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Philox\uff1a\u628a\u201c\u7b2c n \u4e2a\u968f\u673a\u6570\u201d\u53d8\u6210\u7eaf\u51fd\u6570<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u7531 Salmon \u7b49\u4eba\u4e8e 2011 \u5e74\u63d0\u51fa\uff0c\u662f Random123 \u5e93\u7684\u4ee3\u8868\u6027\u7b97\u6cd5\u4e4b\u4e00\uff0c\u4e5f\u662f PyTorch CUDA\u3001TensorFlow\u3001NVIDIA cuRAND \u7b49\u5e7f\u6cdb\u91c7\u7528\u7684\u5e76\u884c\u968f\u673a\u6570\u65b9\u6848\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u975e\u5e38\u4f18\u96c5\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u7ed9\u5b9a\u4e00\u4e2a\u79cd\u5b50 <code>seed<\/code> \u548c\u4e00\u4e2a\u8ba1\u6570\u5668 <code>offset<\/code>\uff0c\u8f93\u51fa <code>f(seed, offset)<\/code> \u5c31\u662f\u786e\u5b9a\u6027\u7684\u968f\u673a\u6570\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6362\u53e5\u8bdd\u8bf4\uff0c\u7b2c n \u4e2a\u968f\u673a\u6570\u4e0d\u518d\u7531\u7b2c n-1 \u4e2a\u72b6\u6001\u9012\u63a8\u800c\u6765\uff0c\u800c\u662f\u53ef\u4ee5\u76f4\u63a5\u8ba1\u7b97\u3002\u8fd9\u4e2a\u6027\u8d28\u5e26\u6765\u4e86\u51e0\u4e2a\u76f4\u63a5\u597d\u5904\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5e76\u884c\u53cb\u597d<\/strong>\uff1a\u6bcf\u4e2a\u7ebf\u7a0b\u7528\u4e0d\u540c\u7684 <code>offset<\/code> \u72ec\u7acb\u751f\u6210\uff0c\u65e0\u9700\u7ade\u4e89\u5171\u4eab\u72b6\u6001\uff1b<\/li>\n\n\n\n<li><strong>\u53ef\u8df3\u8f6c<\/strong>\uff1a\u60f3\u8981\u7b2c 100 \u4ebf\u4e2a\u968f\u673a\u6570\uff0c\u4e0d\u9700\u8981\u5148\u8dd1 100 \u4ebf\u6b65\uff1b<\/li>\n\n\n\n<li><strong>\u987a\u5e8f\u65e0\u5173<\/strong>\uff1a\u53ea\u8981 <code>(seed, offset)<\/code> \u6620\u5c04\u4e0d\u53d8\uff0c\u65e0\u8bba\u54ea\u4e2a\u7ebf\u7a0b\u5148\u7b97\u3001\u540e\u7b97\uff0c\u7ed3\u679c\u96c6\u5408\u90fd\u4e0d\u53d8\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Philox4x32-10 \u7684\u547d\u540d\u5df2\u7ecf\u8bf4\u660e\u4e86\u7ed3\u6784\uff1a4 \u4e2a 32 \u4f4d\u8ba1\u6570\u5668\u30012 \u4e2a 32 \u4f4d\u5bc6\u94a5\uff0c\u7ecf\u8fc7 10 \u8f6e\u201c\u4e58\u9ad8-\u4e58\u4f4e-XOR\u201d\u7684 Feistel \u7ed3\u6784\u53d8\u6362\u3002\u6bcf\u4e00\u8f6e\u90fd\u4f1a\u7528 Weyl \u5e8f\u5217\u5e38\u91cf <code>0x9E3779B9<\/code>\u3001<code>0xBB67AE85<\/code> \u66f4\u65b0\u5bc6\u94a5\uff0c\u4fdd\u8bc1\u5145\u5206\u6df7\u6dc6\u3002\u5b83\u901a\u8fc7 TestU01 \u7b49\u7edf\u8ba1\u68c0\u9a8c\uff0c\u5468\u671f\u957f\u8fbe <code>2^128<\/code>\uff0c\u5bf9\u4efb\u4f55\u5b9e\u9645\u8bad\u7ec3\u4efb\u52a1\u90fd\u7ef0\u7ef0\u6709\u4f59\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5355\u8f6e\u53d8\u6362\u91cc\uff0cPhilox \u628a\u4e24\u4e2a 64 \u4f4d\u4e58\u79ef\u62c6\u6210\u9ad8 32 \u4f4d\u548c\u4f4e 32 \u4f4d\uff0c\u518d\u4e0e\u5bc6\u94a5\u548c\u53e6\u4e00\u534a\u8ba1\u6570\u5668\u505a XOR\u3002\u8fd9\u79cd\u201c\u4e58\u9ad8-\u4e58\u4f4e\u201d\u64cd\u4f5c\u662f\u73b0\u4ee3 CPU \u548c GPU \u90fd\u539f\u751f\u652f\u6301\u7684\u4f4d\u8fd0\u7b97\u4e0e\u6574\u6570\u4e58\u6cd5\uff0c\u56e0\u6b64 Philox \u65e2\u5feb\u53c8\u5bb9\u6613\u5411\u91cf\u5316\u300210 \u8f6e\u8fed\u4ee3\u662f\u4e00\u4e2a\u7ecf\u9a8c\u4e0a\u8db3\u591f\u5b89\u5168\u7684\u53c2\u6570\uff1a\u518d\u591a\u51e0\u8f6e\u53ea\u4f1a\u589e\u52a0\u8ba1\u7b97\u91cf\uff0c\u800c 10 \u8f6e\u5df2\u7ecf\u80fd\u8ba9\u6bcf\u4e2a\u8f93\u51fa\u4f4d\u90fd\u5145\u5206\u4f9d\u8d56\u6bcf\u4e2a\u8f93\u5165\u4f4d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u6ca1\u6709\u4f9d\u8d56\u5916\u90e8\u5e93\uff0c\u800c\u662f\u5728 <code>include\/renaissance\/core\/philox.h<\/code> \u4e2d\u5b9e\u73b0\u4e86\u4e00\u4efd<strong>CPU\/GPU \u901a\u7528<\/strong>\u7684 Philox4x32-10 \u5185\u6838\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">constexpr uint32_t PHILOX_M4x32_0 = 0xD2511F53u;\nconstexpr uint32_t PHILOX_M4x32_1 = 0xCD9E8D57u;\nconstexpr uint32_t PHILOX_W32_0   = 0x9E3779B9u;  \/\/ golden ratio\nconstexpr uint32_t PHILOX_W32_1   = 0xBB67AE85u;  \/\/ sqrt(3)-1\n\nTR_HOST_DEVICE TR_FORCEINLINE\nvoid philox4x32_round(\n    uint32_t* ctr0, uint32_t* ctr1, uint32_t* ctr2, uint32_t* ctr3,\n    uint32_t key0, uint32_t key1)\n{\n    uint32_t lo0, lo1;\n    uint32_t hi0 = mulhilo32(PHILOX_M4x32_0, *ctr0, &amp;lo0);\n    uint32_t hi1 = mulhilo32(PHILOX_M4x32_1, *ctr2, &amp;lo1);\n\n    uint32_t new_ctr0 = hi1 ^ *ctr1 ^ key0;\n    uint32_t new_ctr1 = lo1;\n    uint32_t new_ctr2 = hi0 ^ *ctr3 ^ key1;\n    uint32_t new_ctr3 = lo0;\n\n    *ctr0 = new_ctr0;\n    *ctr1 = new_ctr1;\n    *ctr2 = new_ctr2;\n    *ctr3 = new_ctr3;\n}\n\nTR_HOST_DEVICE TR_FORCEINLINE\nvoid philox4x32_10(\n    uint32_t* ctr0, uint32_t* ctr1, uint32_t* ctr2, uint32_t* ctr3,\n    uint32_t key0, uint32_t key1)\n{\n    for (int round = 0; round &lt; 10; ++round) {\n        philox4x32_round(ctr0, ctr1, ctr2, ctr3, key0, key1);\n        key0 += PHILOX_W32_0;\n        key1 += PHILOX_W32_1;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u53d8\u91cf\u8bf4\u660e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>ctr0..ctr3<\/code>\uff1a4 \u4e2a 32 \u4f4d\u8ba1\u6570\u5668\uff0c\u8f93\u5165\u7684\u4f4e 64 \u4f4d\u7531 <code>offset<\/code> \u586b\u5145\uff0c\u9ad8 64 \u4f4d\u8865 0\uff1b<\/li>\n\n\n\n<li><code>key0, key1<\/code>\uff1a2 \u4e2a 32 \u4f4d\u5bc6\u94a5\uff0c\u7531 64 \u4f4d <code>seed<\/code> \u62c6\u5206\u5f97\u5230\uff1b<\/li>\n\n\n\n<li><code>PHILOX_M4x32_*<\/code>\uff1a\u4e24\u4e2a\u56fa\u5b9a\u4e58\u6570\uff0c\u8d1f\u8d23\u628a\u8ba1\u6570\u5668\u4fe1\u606f\u6269\u6563\u5230\u9ad8 32 \u4f4d\uff1b<\/li>\n\n\n\n<li><code>PHILOX_W32_*<\/code>\uff1aWeyl \u5e8f\u5217\u5e38\u91cf\uff0c\u6bcf\u8f6e\u52a0\u5230\u5bc6\u94a5\u4e0a\uff0c\u8ba9\u6bcf\u8f6e\u4f7f\u7528\u7684\u5bc6\u94a5\u4e0d\u540c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684 <code>TR_HOST_DEVICE<\/code> \u5b8f\u5728 CUDA\/MUSA \u4e0b\u5c55\u5f00\u4e3a <code>__host__ __device__<\/code>\uff0c\u610f\u5473\u7740\u540c\u4e00\u4efd\u7b97\u6cd5\u4ee3\u7801\u65e2\u80fd\u5728 CPU \u4e0a\u8fd0\u884c\uff0c\u4e5f\u80fd\u88ab\u7f16\u8bd1\u8fdb GPU kernel\u3002\u8fd9\u662f\u4fdd\u8bc1<strong>\u8de8\u8bbe\u5907\u6570\u503c\u4e00\u81f4<\/strong>\u7684\u7b2c\u4e00\u6b65\u3002\u5982\u679c CPU \u548c GPU \u4f7f\u7528\u4e0d\u540c\u7b97\u6cd5\u751f\u6210\u540c\u4e00\u7ec4\u201c\u6b63\u6001\u5206\u5e03\u201d\u968f\u673a\u6570\uff0c\u5373\u4fbf\u79cd\u5b50\u76f8\u540c\uff0c\u521d\u59cb\u5316\u51fa\u6765\u7684\u6743\u91cd\u4e5f\u4f1a\u4e0d\u540c\uff0c\u540e\u7eed\u8bad\u7ec3\u5c31\u65e0\u4ece\u8c08\u8d77\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 \u4e3a\u4ec0\u4e48 mulhilo \u662f Philox \u7684\u6838\u5fc3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u7684\u5b89\u5168\u6027\u5e76\u4e0d\u6765\u81ea\u590d\u6742\u7684\u67e5\u8868\u6216 S-box\uff0c\u800c\u662f\u6765\u81ea<strong>\u4e58\u6cd5\u7684\u9ad8\u4f4d\u6269\u6563<\/strong>\u300232 \u4f4d \u00d7 32 \u4f4d = 64 \u4f4d\uff0c\u9ad8 32 \u4f4d\u5305\u542b\u4e86\u4f4e 32 \u4f4d\u7ecf\u8fc7\u4e58\u6cd5\u540e\u7684\u201c\u6df7\u5408\u4fe1\u606f\u201d\u3002\u628a\u9ad8 32 \u4f4d\u548c\u4f4e 32 \u4f4d\u62c6\u5f00\uff0c\u5206\u522b\u8d4b\u7ed9\u4e0d\u540c\u7684\u8ba1\u6570\u5668\u5206\u91cf\uff0c\u518d\u7528 XOR \u4e0e\u5bc6\u94a5\u7ed3\u5408\uff0c\u5c31\u5b8c\u6210\u4e86\u5355\u8f6e\u7684\u6df7\u6dc6\u3002<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">TR_HOST_DEVICE TR_FORCEINLINE\nuint32_t mulhilo32(uint32_t a, uint32_t b, uint32_t* lo) {\n#ifdef __CUDA_ARCH__\n    uint32_t hi;\n    asm(\"mul.hi.u32 %0, %1, %2;\" : \"=r\"(hi) : \"r\"(a), \"r\"(b));\n    *lo = a * b;\n    return hi;\n#else\n    uint64_t product = static_cast&lt;uint64_t>(a) * static_cast&lt;uint64_t>(b);\n    *lo = static_cast&lt;uint32_t>(product);\n    return static_cast&lt;uint32_t>(product >> 32);\n#endif\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 GPU \u4e0a\uff0c\u6211\u4eec\u76f4\u63a5\u7528 PTX \u6307\u4ee4 <code>mul.hi.u32<\/code> \u53d6\u9ad8 32 \u4f4d\uff0c\u907f\u514d 64 \u4f4d\u4e58\u6cd5\u7684\u989d\u5916\u5bc4\u5b58\u5668\u6d88\u8017\uff1b\u5728 CPU \u4e0a\uff0c\u76f4\u63a5\u7528 64 \u4f4d\u6574\u6570\u4e58\u6cd5\u4e00\u6b21\u62ff\u5230\u9ad8\u3001\u4f4e\u4e24\u534a\u3002\u4e24\u79cd\u8def\u5f84\u7684\u8f93\u51fa\u5fc5\u987b\u9010\u4f4d\u4e00\u81f4\uff0c\u5426\u5219 CPU \u548c GPU \u751f\u6210\u7684\u968f\u673a\u6570\u5c31\u4f1a\u5206\u53c9\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u4ece\u6574\u6570\u5230\u6d6e\u70b9\uff1a\u5747\u5300\u5206\u5e03\u4e0e\u6b63\u6001\u5206\u5e03<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u4e00\u8f6e\u8f93\u51fa 4 \u4e2a 32 \u4f4d\u65e0\u7b26\u53f7\u6574\u6570\u3002\u8981\u5f97\u5230\u8bad\u7ec3\u5e38\u7528\u7684\u6d6e\u70b9\u5206\u5e03\uff0c\u9700\u8981\u505a\u4e00\u6b21\u8f6c\u6362\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5747\u5300\u5206\u5e03 <code>[0, 1)<\/code> \u91c7\u7528\u7ecf\u5178\u7684\u9ad8\u4f4d\u6620\u5c04\u6cd5\uff1a\u53d6\u7b2c\u4e00\u4e2a 32 \u4f4d\u6574\u6570\u7684\u9ad8 24 \u4f4d\uff0c\u4e58\u4ee5 <code>2^-24<\/code>\uff0c\u907f\u514d\u6574\u6570\u9664\u6cd5\u5e26\u6765\u7684\u4e0d\u5747\u5300\u6027\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">TR_HOST_DEVICE TR_FORCEINLINE\nfloat philox_uniform_float(uint64_t seed, uint64_t offset) {\n    uint32_t r[4];\n    philox_generate_4x32(seed, offset, r);\n\n    constexpr float scale = 1.0f \/ 16777216.0f;  \/\/ 2^-24\n    return static_cast&lt;float>(r[0] >> 8) * scale;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b63\u6001\u5206\u5e03\u5219\u901a\u8fc7 Box-Muller \u53d8\u6362\uff0c\u628a\u4e24\u4e2a\u5747\u5300\u968f\u673a\u6570\u8f6c\u6362\u6210\u4e00\u5bf9\u6807\u51c6\u6b63\u6001\u968f\u673a\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">TR_HOST_DEVICE TR_FORCEINLINE\nvoid philox_normal_pair(uint64_t seed, uint64_t offset, float* out0, float* out1) {\n    uint32_t r[4];\n    philox_generate_4x32(seed, offset, r);\n\n    constexpr float scale = 1.0f \/ 16777216.0f;\n    float u1 = (static_cast&lt;float>((r[0] >> 8) | 1)) * scale;  \/\/ (0, 1]\n    float u2 = static_cast&lt;float>(r[1] >> 8) * scale;          \/\/ [0, 1)\n\n    constexpr float two_pi = 6.283185307179586f;\n\n#ifdef __CUDA_ARCH__\n    float radius = sqrtf(-2.0f * logf(u1));\n    float theta = two_pi * u2;\n    float sin_theta, cos_theta;\n    sincosf(theta, &amp;sin_theta, &amp;cos_theta);\n    *out0 = radius * cos_theta;\n    *out1 = radius * sin_theta;\n#else\n    float radius = std::sqrt(-2.0f * std::log(u1));\n    float theta = two_pi * u2;\n    *out0 = radius * std::cos(theta);\n    *out1 = radius * std::sin(theta);\n#endif\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a <code>philox_normal_pair<\/code> \u4e00\u6b21\u6d88\u8017\u4e00\u4e2a <code>offset<\/code> \u5374\u4ea7\u51fa\u4e24\u4e2a\u6b63\u6001\u968f\u673a\u6570\uff0c\u6240\u4ee5 CPU \u6b63\u6001\u5206\u5e03\u751f\u6210\u5668\u5728\u8c03\u7528 <code>gen.next_offset()<\/code> \u65f6\uff0c\u9884\u7559\u7684\u662f <code>(count + 1) \/ 2<\/code> \u4e2a offset\uff0c\u800c\u4e0d\u662f <code>count<\/code> \u4e2a\u3002\u8fd9\u662f\u4e00\u4e2a\u5b9e\u73b0\u7ec6\u8282\uff0c\u4f46\u5bf9\u53ef\u590d\u73b0\u6027\u81f3\u5173\u91cd\u8981\uff1a\u5982\u679c\u751f\u6210\u5668\u6d88\u8d39 offset \u7684\u65b9\u5f0f\u53d1\u751f\u53d8\u5316\uff0c\u5373\u4f7f seed \u76f8\u540c\uff0c\u8f93\u51fa\u5e8f\u5217\u4e5f\u4f1a\u53d8\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 \u5468\u671f\u4e0e\u7edf\u8ba1\u8d28\u91cf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Philox4x32-10 \u7684\u5468\u671f\u662f <code>2^128<\/code>\uff084 \u4e2a 32 \u4f4d\u8ba1\u6570\u5668\u603b\u5171 128 \u4f4d\uff09\uff0c\u5bf9\u4e8e\u4efb\u4f55\u5b9e\u9645\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u6765\u8bf4\u90fd\u7ef0\u7ef0\u6709\u4f59\u3002\u5373\u4f7f\u6bcf\u5929\u751f\u6210 10^12 \u4e2a\u968f\u673a\u6570\uff0c\u4e5f\u8981\u8d85\u8fc7 10^19 \u5e74\u624d\u80fd\u8017\u5c3d\u5468\u671f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u7edf\u8ba1\u68c0\u9a8c\u65b9\u9762\uff0cPhilox \u901a\u8fc7\u4e86 TestU01 \u7684 SmallCrush \/ Crush \/ BigCrush \u7b49\u4e3b\u6d41\u968f\u673a\u6570\u7edf\u8ba1\u68c0\u9a8c\uff0c\u5176\u5747\u5300\u5206\u5e03\u3001\u6b63\u6001\u5206\u5e03\u3001\u4f4d\u7ea7\u72ec\u7acb\u6027\u90fd\u4e0e MT19937 \u5904\u4e8e\u540c\u4e00\u6c34\u5e73\u3002\u5bf9\u4e8e\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u800c\u8a00\uff0c\u6211\u4eec\u66f4\u5173\u5fc3\u7684\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7ed9\u5b9a\u76f8\u540c seed\uff0c\u4e0d\u540c\u8fd0\u884c\u662f\u5426\u4ea7\u751f\u5b8c\u5168\u76f8\u540c\u7684\u5e8f\u5217\uff1b<\/li>\n\n\n\n<li>\u7ed9\u5b9a\u76f8\u540c seed\uff0cCPU \u548c GPU \u662f\u5426\u4ea7\u751f\u5b8c\u5168\u76f8\u540c\u7684\u5e8f\u5217\uff1b<\/li>\n\n\n\n<li>\u591a\u7ebf\u7a0b\u5e76\u53d1\u65f6\u662f\u5426\u4e0d\u5f15\u5165\u989d\u5916\u7684\u4e0d\u786e\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u5728\u8fd9\u4e09\u70b9\u4e0a\u90fd\u8868\u73b0\u5f97\u975e\u5e38\u5e72\u51c0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001Generator\uff1a\u4e00\u4e2a\u79cd\u5b50 + \u4e00\u4e2a\u539f\u5b50\u8ba1\u6570\u5668<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Philox \u4e4b\u4e0a\uff0cTech-Renaissance \u5c01\u88c5\u4e86 <code>include\/renaissance\/core\/rng.h<\/code> \u4e2d\u7684 <code>Generator<\/code> \u7c7b\u3002\u5b83\u7684\u72b6\u6001\u6781\u7b80\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">class Generator {\npublic:\n    explicit Generator(uint64_t seed = 0) noexcept;\n    ~Generator();\n    Generator(const Generator&amp;) = delete;\n    Generator&amp; operator=(const Generator&amp;) = delete;\n    Generator(Generator&amp;&amp; other) noexcept;\n    Generator&amp; operator=(Generator&amp;&amp; other) noexcept;\n\n    void set_seed(uint64_t seed);\n    uint64_t seed() const noexcept;\n\n    std::pair&lt;uint64_t, uint64_t> get_state() const;\n    void set_state(uint64_t seed, uint64_t offset);\n\n    uint64_t next_offset(uint64_t count);\n    uint64_t current_offset() const noexcept;\n\n    int random_int(int low, int high);\n\nprivate:\n    class Impl;\n    std::unique_ptr&lt;Impl> impl_;\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Impl<\/code> \u91cc\u53ea\u6709\u4e09\u4e2a\u6210\u5458\uff1a<code>uint64_t seed_<\/code>\u3001<code>std::atomic&lt;uint64_t&gt; offset_<\/code> \u548c\u4e00\u4e2a\u4fdd\u62a4 seed\/offset \u8bfb\u5199\u7684 <code>std::mutex<\/code>\u3002\u6240\u6709\u5e76\u53d1\u8bf7\u6c42\u90fd\u901a\u8fc7 <code>next_offset(count)<\/code> \u539f\u5b50\u5730\u9884\u7559\u4e00\u6bb5\u8fde\u7eed\u7684 <code>offset<\/code> \u533a\u95f4\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">uint64_t Generator::next_offset(uint64_t count) {\n    return impl_->offset_.fetch_add(count, std::memory_order_relaxed);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd4\u56de\u7684\u662f\u672c\u6b21\u8c03\u7528\u7684\u8d77\u59cb <code>offset<\/code>\uff0c\u8c03\u7528\u8005\u53ef\u4ee5\u5b89\u5168\u5730\u5360\u7528 <code>[offset, offset + count)<\/code> \u8fd9\u6bb5\u7a7a\u95f4\u3002\u5373\u4fbf\u591a\u4e2a\u7ebf\u7a0b\u540c\u65f6\u7533\u8bf7\uff0c\u7531\u4e8e <code>fetch_add<\/code> \u7684\u539f\u5b50\u6027\uff0c\u6bcf\u4e2a\u4eba\u62ff\u5230\u7684\u533a\u95f4\u4e92\u4e0d\u91cd\u53e0\uff1b\u800c\u53ea\u8981\u7533\u8bf7\u987a\u5e8f\u76f8\u540c\uff0c\u6700\u7ec8\u751f\u6210\u7684\u968f\u673a\u6570\u5e8f\u5217\u5c31\u5b8c\u5168\u76f8\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>get_state<\/code> \u548c <code>set_state<\/code> \u8ba9\u8bad\u7ec3\u4e2d\u65ad\u540e\u53ef\u4ee5\u4ece\u68c0\u67e5\u70b9\u7cbe\u786e\u6062\u590d\u968f\u673a\u72b6\u6001\uff0c\u5bf9\u957f\u5468\u671f\u8bad\u7ec3\u5c24\u4e3a\u91cd\u8981\u3002seed \u4e0e offset \u7684\u4e8c\u5143\u7ec4\u8db3\u591f\u5c0f\uff0c\u53ef\u4ee5\u76f4\u63a5\u5e8f\u5217\u5316\u5230 checkpoint \u6587\u4ef6\u91cc\uff0c\u4e0d\u9700\u8981\u50cf Mersenne Twister \u90a3\u6837\u4fdd\u5b58 2.5 KB \u7684\u5185\u90e8\u72b6\u6001\u6570\u7ec4\u3002\u8fd9\u610f\u5473\u7740 checkpoint \u66f4\u5c0f\uff0c\u6062\u590d\u903b\u8f91\u4e5f\u66f4\u6e05\u6670\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u4e00\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u8bbe\u8ba1\u7ec6\u8282\uff1a\u5934\u6587\u4ef6\u4f7f\u7528 <strong>Pimpl \u6a21\u5f0f<\/strong> \u9690\u85cf <code>std::atomic<\/code>\u3002\u539f\u56e0\u662f MUSA SDK \u7684 <code>musa\/std\/atomic<\/code> \u4e0e\u6807\u51c6\u5e93 <code>&lt;atomic&gt;<\/code> \u5b58\u5728\u547d\u540d\u7a7a\u95f4\u51b2\u7a81\uff0c\u5982\u679c\u5728\u516c\u5171\u5934\u6587\u4ef6\u91cc\u76f4\u63a5\u5305\u542b <code>&lt;atomic&gt;<\/code>\uff0c\u4f1a\u5728\u67d0\u4e9b\u56fd\u4ea7 GPU \u7f16\u8bd1\u73af\u5883\u4e0b\u89e6\u53d1\u5927\u91cf\u9519\u8bef\u3002\u628a\u539f\u5b50\u6210\u5458\u653e\u5230 <code>.cpp<\/code> \u7684\u5b9e\u73b0\u7c7b\u4e2d\uff0c\u65e2\u89e3\u51b3\u4e86\u517c\u5bb9\u6027\u95ee\u9898\uff0c\u4e5f\u6ca1\u6709\u727a\u7272 <code>next_offset<\/code> \u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u4fdd\u5b58\u4e0e\u6062\u590d RNG \u72b6\u6001<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u957f\u5468\u671f\u8bad\u7ec3\uff0ccheckpoint \u4e0d\u4ec5\u8981\u4fdd\u5b58\u6743\u91cd\uff0c\u8fd8\u8981\u4fdd\u5b58\u968f\u673a\u72b6\u6001\uff0c\u5426\u5219\u6062\u590d\u8bad\u7ec3\u540e\u968f\u673a\u5e8f\u5217\u4f1a\u4ece\u5934\u5f00\u59cb\uff0c\u5bfc\u81f4\u7ed3\u679c\u4e0e\u4e0d\u95f4\u65ad\u8bad\u7ec3\u4e0d\u540c\u3002<code>Generator<\/code> \u63d0\u4f9b\u4e86\u975e\u5e38\u8f7b\u91cf\u7684\u72b6\u6001\u63a5\u53e3\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u8bad\u7ec3\u524d\u4fdd\u5b58\nauto [seed, offset] = gen.get_state();\nsave_checkpoint({seed, offset, model_weights});\n\n\/\/ \u6062\u590d\u8bad\u7ec3\u65f6\u52a0\u8f7d\ngen.set_state(loaded_seed, loaded_offset);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a\u72b6\u6001\u53ea\u6709\u4e24\u4e2a 64 \u4f4d\u6574\u6570\uff0c\u76f8\u6bd4 Mersenne Twister \u9700\u8981\u4fdd\u5b58 624 \u4e2a 32 \u4f4d\u6574\u6570\u7684\u72b6\u6001\u6570\u7ec4\uff0cPhilox \u7684 checkpoint \u8d1f\u62c5\u51e0\u4e4e\u4e3a\u96f6\u3002\u8fd9\u4e5f\u662f\u8ba1\u6570\u5668\u578b RNG \u5728\u5de5\u7a0b\u4e0a\u7684\u53e6\u4e00\u4e2a\u4f18\u52bf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 \u7528\u6237\u4fa7\u79cd\u5b50\u8bbe\u7f6e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u6237\u4fa7\u7684\u79cd\u5b50\u8bbe\u7f6e\u901a\u8fc7 <code>GlobalRegistry::manual_seed(seed)<\/code> \u5b8c\u6210\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">GlobalRegistry&amp; GlobalRegistry::manual_seed(uint64_t seed) {\n    \/\/ \u53ea\u8981\u8bbe\u5b9a\u79cd\u5b50\uff0c\u5c31\u5fc5\u5b9a\u662f\u60f3\u8981\u53ef\u590d\u73b0\n    \/\/ \u56e0\u6b64\u4f1a\u81ea\u52a8\u4f7f\u7528\u53ef\u590d\u73b0\u6a21\u5f0f\uff0c\u4ee5\u5fae\u5c0f\u7684\u6027\u80fd\u4ee3\u4ef7\u6362\u6765\u968f\u673a\u53ef\u590d\u73b0\u6027\n    reproducible();\n    rng_set_seed(seed);\n    return *this;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ea\u8981\u8c03\u7528 <code>manual_seed<\/code>\uff0c\u6846\u67b6\u5c31\u4f1a\u81ea\u52a8\u5f00\u542f\u201c\u53ef\u590d\u73b0\u6027\u4fdd\u9669\u201d\uff0c\u628a <code>TransferStation<\/code>\u3001\u6570\u636e\u6d17\u724c\u3001\u968f\u673a\u589e\u5f3a\u7b49\u5168\u90e8\u5207\u6362\u5230\u786e\u5b9a\u6027\u6a21\u5f0f\u3002\u8fd9\u662f Tech-Renaissance \u7684\u4e00\u4e2a\u9ad8\u5c42\u7ea6\u5b9a\uff1a<strong>\u7528\u6237\u53ea\u9700\u8981\u8bf4\u4e00\u6b21\u201c\u6211\u8981\u56fa\u5b9a\u79cd\u5b50\u201d\uff0c\u5269\u4e0b\u7684\u4e8b\u60c5\u6846\u67b6\u8d1f\u8d23<\/strong>\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cPyTorch \u7528\u6237\u901a\u5e38\u9700\u8981\u4e3a\u4e00\u5806\u72ec\u7acb\u7684 RNG \u5206\u522b\u8bbe\u79cd\u5b50\uff0c\u5e76\u624b\u52a8\u7ba1\u7406\u5404\u79cd\u540e\u7aef\u5f00\u5173\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u4e00\u4e2a\u6700\u5c0f\u7684\u53ef\u590d\u73b0\u793a\u4f8b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5b9e\u9645\u4ee3\u7801\u91cc\uff0c\u7528\u6237\u4e0d\u9700\u8981\u5173\u5fc3 Philox \u7684\u7ec6\u8282\u3002\u4ee5 <code>tests\/example\/mlp_mnist.cpp<\/code> \u4e3a\u4f8b\uff0c\u5f00\u542f\u53ef\u590d\u73b0\u8bad\u7ec3\u53ea\u9700\u8981\u5728\u914d\u7f6e\u94fe\u91cc\u52a0\u4e00\u884c\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">GLOBAL_SETTING\n    .use_gpu(\"0\")\n    .amp(true)\n    .manual_seed(123)      \/\/ \u56fa\u5b9a\u968f\u673a\u79cd\u5b50\uff0c\u4fdd\u8bc1\u7ed3\u679c\u53ef\u590d\u73b0\n    .global_batch_size(200)\n    .input_resolution(28);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>manual_seed(123)<\/code> \u80cc\u540e\u53d1\u751f\u7684\u4e8b\u60c5\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bbe\u7f6e\u5168\u5c40 <code>Generator<\/code> \u7684 seed \u4e3a 123\uff0coffset \u5f52\u96f6\uff1b<\/li>\n\n\n\n<li>\u5f00\u542f <code>reproducibility_insurance_<\/code>\uff0c\u8ba9 <code>TransferStation<\/code> \u8fdb\u5165\u53ef\u590d\u73b0\u6a21\u5f0f\uff1b<\/li>\n\n\n\n<li>\u6bcf\u4e2a <code>PreprocessWorker<\/code> \u7528 <code>global_seed ^ (worker_id &lt;&lt; 32)<\/code> \u6d3e\u751f\u81ea\u5df1\u7684\u521d\u59cb\u79cd\u5b50\uff1b<\/li>\n\n\n\n<li>\u6bcf\u4e2a rank \u7684 Dropout seed \u901a\u8fc7 SplitMix64 \u54c8\u5e0c\u4ece 123 \u6d3e\u751f\uff1b<\/li>\n\n\n\n<li>\u6743\u91cd\u521d\u59cb\u5316\u3001\u6570\u636e\u589e\u5f3a\u3001\u6570\u636e\u96c6\u6d17\u724c\u5168\u90e8\u63a5\u5165\u540c\u4e00\u5957 Philox \u4f53\u7cfb\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\uff0c\u4e00\u884c\u8c03\u7528\u5c31\u66ff\u4ee3\u4e86 PyTorch \u91cc\u5e38\u89c1\u7684\u201cseed everything\u201d\u6a21\u677f\u51fd\u6570\uff0c\u628a\u79cd\u5b50\u7ba1\u7406\u4ece\u7528\u6237\u4fa7\u8f6c\u79fb\u5230\u6846\u67b6\u4fa7\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.4 \u4e3a\u4ec0\u4e48\u4e0d\u7528 Mersenne Twister<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Mersenne Twister\uff08MT19937\uff09\u662f C++ \u6807\u51c6\u5e93\u548c\u8bb8\u591a\u79d1\u5b66\u8ba1\u7b97\u5e93\u7684\u9ed8\u8ba4\u9009\u62e9\uff0c\u5b83\u5468\u671f\u6781\u957f\uff08<code>2^19937 - 1<\/code>\uff09\u3001\u7edf\u8ba1\u8d28\u91cf\u597d\uff0c\u4f46\u6709\u4e00\u4e2a\u6839\u672c\u95ee\u9898\uff1a<strong>\u5b83\u662f\u4e00\u4e2a\u72b6\u6001\u673a<\/strong>\u3002\u8981\u751f\u6210\u7b2c n \u4e2a\u968f\u673a\u6570\uff0c\u5fc5\u987b\u5148\u751f\u6210\u524d n-1 \u4e2a\u3002\u8fd9\u4e2a\u7279\u6027\u5bfc\u81f4\u5b83\u5728\u5e76\u53d1\u573a\u666f\u4e0b\u975e\u5e38\u522b\u626d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u591a\u7ebf\u7a0b\u5171\u4eab\u4e00\u4e2a\u751f\u6210\u5668\u5fc5\u987b\u52a0\u9501\uff0c\u9501\u7ade\u4e89\u548c\u8c03\u5ea6\u987a\u5e8f\u90fd\u4f1a\u7834\u574f\u53ef\u590d\u73b0\u6027\uff1b<\/li>\n\n\n\n<li>\u6bcf\u4e2a\u7ebf\u7a0b\u4e00\u4e2a\u751f\u6210\u5668\uff0c\u7ebf\u7a0b\u521b\u5efa\u987a\u5e8f\u53d8\u5316\u4f1a\u5bfc\u81f4\u79cd\u5b50\u5206\u914d\u987a\u5e8f\u53d8\u5316\uff1b<\/li>\n\n\n\n<li>\u72b6\u6001\u6570\u7ec4 2.5 KB\uff0c\u4fdd\u5b58\u5230 checkpoint \u6216\u4f20\u9012\u7ed9 GPU \u90fd\u4e0d\u65b9\u4fbf\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u7528\u8ba1\u6570\u5668\u66ff\u4ee3\u72b6\u6001\u673a\u540e\uff0c\u8fd9\u4e9b\u95ee\u9898\u8fce\u5203\u800c\u89e3\u3002\u5b83\u7684\u72b6\u6001\u53ea\u6709\u4e24\u4e2a 64 \u4f4d\u6574\u6570\uff0c\u6bcf\u4e2a\u7ebf\u7a0b\u53ef\u4ee5\u72ec\u7acb\u8ba1\u7b97\u81ea\u5df1\u90a3\u4e00\u4efd\uff0c\u4e0d\u9700\u8981\u9501\uff0c\u4e5f\u4e0d\u9700\u8981\u7ef4\u62a4\u5386\u53f2\u72b6\u6001\u3002\u5728 GPU \u4e0a\uff0c\u8fd9\u610f\u5473\u7740\u6210\u5343\u4e0a\u4e07\u4e2a\u7ebf\u7a0b\u53ef\u4ee5\u540c\u65f6\u751f\u6210\u968f\u673a\u6570\uff0c\u800c\u4e0d\u4f1a\u6709\u4efb\u4f55\u7ade\u4e89\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.5 \u7edf\u8ba1\u7279\u6027\uff1aPhilox \u4e0e MT19937 \u7684\u5bf9\u6bd4<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u5728 <code>tests\/correction\/rng_comparison_test.cpp<\/code> \u4e2d\u5b9e\u73b0\u4e86\u4e00\u4e2a\u7b80\u5355\u7684\u7edf\u8ba1\u5bf9\u6bd4\u7a0b\u5e8f\uff0c\u7528\u540c\u6837\u7684\u79cd\u5b50\u5206\u522b\u751f\u6210 Philox \u548c MT19937 \u7684\u5747\u5300\u5206\u5e03\u3001\u6b63\u6001\u5206\u5e03\u6837\u672c\uff0c\u7136\u540e\u6bd4\u8f83\u5747\u503c\u3001\u6807\u51c6\u5dee\u3001\u6700\u5c0f\u503c\u3001\u6700\u5927\u503c\u3002Philox \u4e0e MT19937 \u7684\u7edf\u8ba1\u7279\u6027\u5728\u540c\u4e00\u91cf\u7ea7\uff0c\u5b8c\u5168\u6ee1\u8db3\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u7684\u9700\u6c42\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e2a\u5bf9\u6bd4\u7a0b\u5e8f\u76ee\u524d\u5c1a\u672a\u63a5\u5165 <code>tests\/correction\/CMakeLists.txt<\/code>\uff0c\u5b83\u66f4\u50cf\u662f\u4e00\u4efd\u968f\u7528\u968f\u53d6\u7684\u9a8c\u8bc1\u811a\u672c\uff0c\u800c\u4e0d\u662f\u6bcf\u65e5\u6301\u7eed\u96c6\u6210\u7684\u4e00\u90e8\u5206\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u4ece\u6743\u91cd\u521d\u59cb\u5316\u5230\u6570\u636e\u589e\u5f3a\uff1aPhilox \u8d2f\u7a7f\u5168\u6d41\u7a0b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u53c2\u6570\u521d\u59cb\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>src\/tensor\/tensor.cpp<\/code> \u4e2d\u7684 <code>Tensor::normal<\/code>\u3001<code>Tensor::uniform<\/code> \u7b49\u65b9\u6cd5\u76f4\u63a5\u8c03\u7528 <code>cpu_rand_*<\/code> \u7cfb\u5217\u51fd\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">void Tensor::normal(float mean, float stddev) {\n    const int64_t total_elems = numel();\n    float* data = static_cast&lt;float*>(ptr_);\n    cpu_rand_normal_float(data, total_elems, mean, stddev);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u51fd\u6570\u7684\u5b9e\u73b0\uff08<code>src\/core\/rng.cpp<\/code>\uff09\u7edf\u4e00\u9075\u5faa\u540c\u4e00\u4e2a\u6a21\u5f0f\uff1a\u5148 <code>next_offset<\/code> \u9884\u7559\u6574\u4e2a\u5f20\u91cf\u6240\u9700\u7684 offset\uff0c\u518d\u7528 OpenMP \u5e76\u884c\u751f\u6210\uff0c\u6bcf\u4e2a\u7ebf\u7a0b\u7528 <code>base_offset + i<\/code> \u4f5c\u4e3a Philox \u8ba1\u6570\u5668\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">void cpu_rand_normal_float(float* ptr, size_t count, float mean, float std, Generator&amp; gen) {\n    \/\/ Box-Muller \u6bcf\u6b21\u751f\u6210 2 \u4e2a\u6570\uff0c\u6240\u4ee5 offset \u6d88\u8017\u91cf\u4e3a (count + 1) \/ 2\n    uint64_t pairs_needed = (count + 1) \/ 2;\n    uint64_t base_offset = gen.next_offset(pairs_needed);\n    uint64_t seed = gen.seed();\n\n    int num_threads = get_num_threads(count);\n    int64_t pair_count = static_cast&lt;int64_t>(count \/ 2);\n\n#if TR_USE_OPENMP\n    #pragma omp parallel for num_threads(num_threads) schedule(static)\n#endif\n    for (int64_t i = 0; i &lt; pair_count; ++i) {\n        float n0, n1;\n        detail::philox_normal_pair(seed, base_offset + i, &amp;n0, &amp;n1);\n        ptr[i * 2]     = mean + std * n0;\n        ptr[i * 2 + 1] = mean + std * n1;\n    }\n\n    \/\/ \u5904\u7406\u5947\u6570\u5c3e\u90e8\uff08\u5982\u679c\u6709\uff09\n    if (count % 2 == 1) {\n        float n0, n1;\n        detail::philox_normal_pair(seed, base_offset + pair_count, &amp;n0, &amp;n1);\n        ptr[count - 1] = mean + std * n0;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f <code>schedule(static)<\/code> \u7684\u91cd\u8981\u6027\uff1a\u5b83\u786e\u4fdd\u7ebf\u7a0b i \u6c38\u8fdc\u5904\u7406\u540c\u4e00\u6bb5 offset\uff0c\u5373\u4fbf\u7ebf\u7a0b\u8c03\u5ea6\u6709\u6296\u52a8\uff0c\u53ea\u8981 <code>base_offset<\/code> \u548c <code>count<\/code> \u76f8\u540c\uff0c\u8f93\u51fa\u5c31\u76f8\u540c\u3002\u5982\u679c\u6362\u6210 <code>schedule(dynamic)<\/code>\uff0c\u968f\u673a\u6570\u4e0e\u7ebf\u7a0b\u7684\u7ed1\u5b9a\u5173\u7cfb\u4f1a\u968f\u8fd0\u884c\u65f6\u53d8\u5316\uff0c\u53ef\u590d\u73b0\u6027\u5c31\u4f1a\u88ab\u7834\u574f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e DTensor \u5728 GPU \u4e0a\u7684\u521d\u59cb\u5316\uff0c<code>src\/backend\/infra_kernels.cu<\/code> \u63d0\u4f9b\u4e86\u4e00\u4e2a Philox \u6b63\u6001\u5206\u5e03 kernel\uff0c\u88ab <code>TaskBase::randn<\/code> \u8c03\u7528\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">__global__ void tr_philox_normal_float_kernel(\n    int n, uint64_t seed, uint64_t base_offset,\n    float mean, float std, float* out)\n{\n    int idx = blockIdx.x * blockDim.x + threadIdx.x;\n    if (idx >= n) return;\n\n    float n0, n1;\n    philox_normal_pair(seed, base_offset + idx \/ 2, &amp;n0, &amp;n1);\n\n    if (idx % 2 == 0) {\n        out[idx] = mean + std * n0;\n    } else {\n        out[idx] = mean + std * n1;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CPU \u4e0e GPU \u7528\u7684\u662f\u540c\u4e00\u4e2a <code>philox_normal_pair<\/code>\uff0c\u53ea\u662f\u627f\u8f7d\u5e76\u884c\u7c92\u5ea6\u7684\u8f7d\u4f53\u4e0d\u540c\uff1a\u4e00\u4e2a\u662f OpenMP \u7ebf\u7a0b\uff0c\u4e00\u4e2a\u662f CUDA \u7ebf\u7a0b\u3002\u8fd9\u907f\u514d\u4e86\u201cCPU \u521d\u59cb\u5316\u4e00\u5957\u3001GPU \u521d\u59cb\u5316\u4e00\u5957\u201d\u5e26\u6765\u7684\u6570\u503c\u5206\u53c9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6846\u67b6\u521d\u59cb\u5316\u6d41\u7a0b\u4e2d\uff0c<code>TaskBase::init_all()<\/code> \u4f1a\u904d\u5386 MemoryPlan \u4e2d\u7684\u6bcf\u4e2a DTensor\uff0c\u6839\u636e <code>InitConfig<\/code> \u51b3\u5b9a\u8c03\u7528\u54ea\u79cd\u521d\u59cb\u5316\u65b9\u5f0f\u3002\u5bf9\u4e8e Kaiming\/Xavier \u7b49\u5e38\u89c1\u521d\u59cb\u5316\uff0c\u6700\u7ec8\u90fd\u4f1a\u843d\u5230 <code>Tensor::normal<\/code> \u6216 <code>Tensor::uniform<\/code>\uff08CPU \u8def\u5f84\uff09\uff0c\u6216 <code>TaskBase::randn<\/code>\uff08GPU \u8def\u5f84\uff09\u3002\u8fd9\u4e9b\u5165\u53e3\u7edf\u4e00\u4f7f\u7528\u5168\u5c40 <code>Generator<\/code>\uff0c\u6240\u4ee5\u53ea\u8981 <code>manual_seed<\/code> \u76f8\u540c\uff0c\u6240\u6709\u5c42\u7684\u521d\u59cb\u5316\u7ed3\u679c\u90fd\u5b8c\u5168\u786e\u5b9a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u6570\u636e\u589e\u5f3a\u4e0e\u6d17\u724c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u9884\u5904\u7406\u662f\u968f\u673a\u6027\u7684\u53e6\u4e00\u4e2a\u91cd\u707e\u533a\u3002Tech-Renaissance \u7684\u6bcf\u4e2a <code>PreprocessWorker<\/code> \u90fd\u6301\u6709\u81ea\u5df1\u72ec\u7acb\u7684 <code>Generator rng_<\/code>\uff0c\u5176\u79cd\u5b50\u5728 <code>ensure_rng_initialized()<\/code> \u4e2d\u6309 worker \u548c phase \u884d\u751f\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">void PreprocessWorker::ensure_rng_initialized() {\n    if (!rng_initialized_) {\n        uint64_t base_seed = initial_seed_;\n        uint64_t worker_seed = base_seed ^ (static_cast&lt;uint64_t>(param_.phase_id) &lt;&lt; 16);\n        rng_.set_seed(worker_seed);\n        rng_initialized_ = true;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d <code>initial_seed_ = global_seed ^ (worker_id &lt;&lt; 32)<\/code>\uff0c\u5728\u6784\u9020\u65f6\u5c31\u56fa\u5b9a\u4e0b\u6765\u3002\u8fd9\u610f\u5473\u7740\u4e0d\u540c worker \u7684\u79cd\u5b50\u4e0d\u540c\uff0c\u540c\u4e00 worker \u4e0d\u540c epoch \u7684\u79cd\u5b50\u4e5f\u4e0d\u540c\uff0c\u4f46\u5168\u90e8\u7531\u5168\u5c40\u79cd\u5b50\u786e\u5b9a\u6027\u6d3e\u751f\u3002worker \u4e4b\u95f4\u4e0d\u4f1a\u56e0\u4e3a\u5171\u4eab\u5168\u5c40 RNG \u800c\u4e92\u76f8\u201c\u5403\u6389\u201d\u5bf9\u65b9\u7684\u968f\u673a\u6570\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u6bcf\u4e2a epoch \u7684\u6570\u636e\u96c6\u6d17\u724c\uff0c<code>src\/data\/preprocess_worker.cpp<\/code> \u4e2d\u7684 <code>shuffle_s_indices()<\/code> \u6ca1\u6709\u4f7f\u7528 <code>Generator::random_int()<\/code>\u2014\u2014\u56e0\u4e3a\u90a3\u4f1a\u4fee\u6539 Generator \u72b6\u6001\uff0c\u53ef\u80fd\u88ab\u5176\u4ed6\u968f\u673a\u64cd\u4f5c\u6c61\u67d3\u2014\u2014\u800c\u662f\u76f4\u63a5\u8c03\u7528 Philox \u539f\u8bed\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">for (int i = n - 1; i > 0; --i) {\n    uint32_t r[4];\n    detail::philox_generate_4x32(shuffle_seed, static_cast&lt;uint64_t>(i), r);\n    uint32_t j = r[0] % (i + 1);\n    std::swap(s_shuffled_indices_[i], s_shuffled_indices_[j]);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>shuffle_seed<\/code> \u7531 <code>initial_seed_ ^ (phase_id &lt;&lt; 32)<\/code> \u5f97\u5230\uff0c<code>offset<\/code> \u76f4\u63a5\u53d6\u5faa\u73af\u53d8\u91cf <code>i<\/code>\u3002\u8fd9\u662f\u4e00\u4e2a\u5b8c\u5168\u72ec\u7acb\u7684\u968f\u673a\u6d41\uff0c\u4e0d\u4e0e\u5176\u4ed6\u4efb\u4f55\u968f\u673a\u64cd\u4f5c\u5171\u4eab\u8ba1\u6570\u5668\uff0c\u56e0\u6b64\u65e0\u8bba\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u53c2\u6570\u521d\u59cb\u5316\u3001Dropout\u3001Random Erasing \u6d88\u8017\u4e86\u591a\u5c11\u968f\u673a\u6570\uff0c\u6d17\u724c\u7ed3\u679c\u90fd\u4e0d\u4f1a\u53d7\u5f71\u54cd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>src\/data\/fused_normalization.cpp<\/code> \u91cc\uff0cRandom Erasing\u3001\u968f\u673a\u6c34\u5e73\u7ffb\u8f6c\u7b49\u589e\u5f3a\u540c\u6837\u901a\u8fc7 <code>Generator*<\/code> \u5b8c\u6210\u3002<code>FusedNormalization::uniform<\/code> \u548c <code>randint<\/code> \u90fd\u662f\u57fa\u4e8e Philox \u539f\u8bed\u7684\u8584\u5c01\u88c5\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">float FusedNormalization::uniform(float min_val, float max_val, Generator* rng) const {\n    uint64_t offset = rng->next_offset(1);\n    float rand_val = detail::philox_uniform_float(rng->seed(), offset);\n    return min_val + rand_val * (max_val - min_val);\n}\n\nint FusedNormalization::randint(int min_val, int max_val, Generator* rng) const {\n    return rng->random_int(min_val, max_val);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u786e\u4fdd\u8fd9\u4e9b\u589e\u5f3a\u4e0e\u53c2\u6570\u521d\u59cb\u5316\u3001Dropout \u4f7f\u7528\u540c\u4e00\u5957\u968f\u673a\u6570\u4f53\u7cfb\uff0c\u800c\u4e0d\u662f\u4ece\u67d0\u4e2a\u72ec\u7acb\u5e93\u91cc\u518d\u62c9\u4e00\u6761 RNG \u8fdb\u6765\u3002\u5bf9\u4e8e Random Erasing\uff0c\u5b83\u4f1a\u6839\u636e <code>erase_p_<\/code> \u51b3\u5b9a\u662f\u5426\u751f\u6210\u64e6\u9664\u77e9\u5f62\uff0c\u518d\u4f9d\u6b21\u7528 <code>uniform<\/code> \u91c7\u6837\u9762\u79ef\u6bd4\u4f8b\u3001\u957f\u5bbd\u6bd4\u3001\u4ee5\u53ca\u77e9\u5f62\u5de6\u4e0a\u89d2\u4f4d\u7f6e\uff1b\u6240\u6709\u8fd9\u4e9b\u968f\u673a\u6d88\u8d39\u90fd\u901a\u8fc7\u5f53\u524d worker \u7684 <code>Generator<\/code> \u5b8c\u6210\uff0c\u5e76\u4e14\u6d88\u8d39\u987a\u5e8f\u56fa\u5b9a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 \u53ef\u590d\u73b0\u7684\u6570\u636e\u642c\u8fd0\uff1aTransferStation \u53cc\u6a21\u5f0f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5373\u4f7f\u968f\u673a\u6570\u672c\u8eab\u786e\u5b9a\u4e86\uff0c\u5982\u679c\u591a\u4e2a worker \u5f80 GPU \u642c\u8fd0\u6570\u636e\u65f6\u7684<strong>\u5199\u5165\u987a\u5e8f<\/strong>\u4e0d\u786e\u5b9a\uff0c\u6700\u7ec8 batch \u5185\u7684\u6837\u672c\u6392\u5217\u4ecd\u53ef\u80fd\u53d8\u5316\u3002\u4e3a\u6b64\uff0c<code>src\/data\/transfer_station.cpp<\/code> \u8bbe\u8ba1\u4e86\u4e24\u79cd\u6a21\u5f0f\uff0c\u7531 <code>require_reproducibility_<\/code> \u6807\u5fd7\u63a7\u5236\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u53ef\u590d\u73b0\u6a21\u5f0f<\/strong>\uff1a\u6bcf\u4e2a worker \u6309\u56fa\u5b9a\u516c\u5f0f\u8ba1\u7b97\u81ea\u5df1\u5728 batch \u4e2d\u7684 <code>position<\/code> \u548c <code>batch_id<\/code>\uff0c\u5feb worker \u5fc5\u987b\u7b49\u6162 worker \u5b8c\u6210\u5f53\u524d batch \u624d\u80fd\u7ee7\u7eed\uff0c\u4f20\u8f93\u65f6\u673a\u5b8c\u5168\u786e\u5b9a\uff1b<\/li>\n\n\n\n<li><strong>\u975e\u53ef\u590d\u73b0\u6a21\u5f0f<\/strong>\uff1a\u7528\u539f\u5b50\u8ba1\u6570\u5668\u52a8\u6001\u5206\u914d slot\uff0c\u6027\u80fd\u66f4\u9ad8\uff0c\u4f46\u6bcf\u6b21\u8fd0\u884c\u987a\u5e8f\u53ef\u80fd\u4e0d\u540c\u3002<\/li>\n<\/ul>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">if (require_reproducibility_) {\n    std::unique_lock&lt;std::mutex> lock(mutex_);\n    cv_batch_ready_.wait(lock, [this, batch_id]() {\n        return current_batch_id_.load() >= batch_id || finished_.load();\n    });\n\n    int buf_id = current_buffer_.load(std::memory_order_acquire);\n    buffer_labels_[buf_id][position] = label;\n    size_t offset = position * snapshot_sample_bytes;\n    return buffer_data_[buf_id] + offset;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f Tech-Renaissance \u5bf9\u201c\u53ef\u590d\u73b0\u201d\u505a\u51fa\u7684\u4e00\u4e2a\u660e\u786e trade-off\uff1a\u5f00\u542f\u53ef\u590d\u73b0\u6a21\u5f0f\u540e\uff0c\u6570\u636e\u7ba1\u7ebf\u7684\u541e\u5410\u91cf\u4f1a\u7565\u6709\u4e0b\u964d\uff0c\u56e0\u4e3a\u5feb worker \u4f1a\u88ab\u6162 worker \u62d6\u4f4f\uff1b\u4f46\u6362\u6765\u7684\u662f\u5b57\u8282\u7ea7\u4e00\u81f4\u7684\u6837\u672c\u987a\u5e8f\u3002\u5bf9\u4e8e\u8c03\u8bd5\u548c\u79d1\u5b66\u5b9e\u9a8c\uff0c\u8fd9\u662f\u503c\u5f97\u7684\u3002\u5bf9\u4e8e\u8ffd\u6c42\u6781\u9650\u6027\u80fd\u7684\u751f\u4ea7\u8dd1\u5206\uff0c\u53ef\u4ee5\u5173\u95ed\u53ef\u590d\u73b0\u6a21\u5f0f\uff0c\u8ba9 TransferStation \u8d70\u65e0\u9501\u5feb\u901f\u8def\u5f84\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 Dropout\uff1a\u5728 CUDA Graph \u5185\u81ea\u52a9\u65cb\u8f6c\u79cd\u5b50<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Dropout \u7684\u96be\u70b9\u5728\u4e8e\u5b83\u53d1\u751f\u5728 GPU kernel \u91cc\uff0c\u800c Tech-Renaissance \u7684\u8bad\u7ec3\u5faa\u73af\u88ab\u5b8c\u6574\u6355\u83b7\u4e3a CUDA Graph\uff0cCPU \u4e0d\u80fd\u5728\u6bcf\u6b21 forward \u65f6\u63d2\u8fdb\u53bb\u6539\u79cd\u5b50\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u89e3\u51b3\u65b9\u6848\u662f\u5728 <code>MemoryPlan<\/code> \u91cc\u4e3a\u6bcf\u4e2a rank \u5206\u914d\u4e00\u4e2a <code>dropout_seed<\/code> \u5f20\u91cf\uff08\u5f62\u72b6 <code>{1,1,1,2}<\/code>\uff0cINT32\uff09\uff0c\u653e\u5728 <code>S_SCALAR_INT32<\/code> \u533a\u57df\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">Shape seed_shape{1, 1, 1, 2};\nauto seed_dt = alloc_impl(seed_shape, DType::INT32, Region::S_SCALAR_INT32);\nset_init_config(seed_dt.id, InitConfig{0.0f, InitKind::NONE, FanMode::FAN_IN});\nbaseline_.dropout_seed = seed_dt.id;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u521d\u59cb\u5316\u65f6\uff0c<code>TaskBase::init_all()<\/code> \u7528\u5168\u5c40\u79cd\u5b50\u901a\u8fc7\u4e00\u4e2a SplitMix64 \u98ce\u683c\u7684\u54c8\u5e0c\u4e3a\u6bcf\u4e2a rank \u751f\u6210\u72ec\u7acb\u79cd\u5b50\uff0c\u5e76 H2D \u62f7\u8d1d\u5230\u8be5\u5f20\u91cf\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">if (active_memory_plan_->baseline().dropout_seed >= 0) {\n    uint64_t global_seed = get_default_generator().seed();\n    for (int rank = 0; rank &lt; num_gpus_; ++rank) {\n        uint64_t z = global_seed + static_cast&lt;uint64_t>(rank) + 0x9e3779b97f4a7c15ULL;\n        z = (z ^ (z >> 30)) * 0xbf58476d1ce4e5b9ULL;\n        z = (z ^ (z >> 27)) * 0x94d049bb133111ebULL;\n        uint64_t rank_seed = z ^ (z >> 31);\n\n        int32_t seed_data[2] = {\n            static_cast&lt;int32_t>(rank_seed &amp; 0xFFFFFFFFULL),\n            static_cast&lt;int32_t>(rank_seed >> 32)\n        };\n\n        Tensor host_seed(Shape{1, 1, 1, 2}, DType::INT32);\n        memcpy(host_seed.data(), seed_data, sizeof(seed_data));\n        transfer_to_rank(host_seed, active_memory_plan_->get_dtensor(\n            active_memory_plan_->baseline().dropout_seed), rank);\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6bcf\u6b21 Dropout forward \u7684 CUDA Graph \u8282\u70b9\u91cc\uff0ckernel \u5148\u8c03\u7528 <code>rotate_dropout_seed_kernel<\/code> \u7528 Xorshift64* \u539f\u5730\u65cb\u8f6c\u79cd\u5b50\uff0c\u518d\u4ece\u8fd9\u4e2a\u8bbe\u5907\u4e0a\u7684\u79cd\u5b50\u751f\u6210 Philox \u6837\u672c\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">__global__ void rotate_dropout_seed_kernel(int32_t* seed_ptr) {\n    uint64_t s = (static_cast&lt;uint64_t>(static_cast&lt;uint32_t>(seed_ptr[1])) &lt;&lt; 32)\n               | static_cast&lt;uint32_t>(seed_ptr[0]);\n    if (s == 0) s = 0x9e3779b97f4a7c15ULL;  \/\/ \u9632\u5fa1 stuck\n    s ^= s >> 12;\n    s ^= s &lt;&lt; 25;\n    s ^= s >> 27;\n    s *= 0x2545F4914F6CDD1DULL;\n    seed_ptr[0] = static_cast&lt;int32_t>(s &amp; 0xFFFFFFFFULL);\n    seed_ptr[1] = static_cast&lt;int32_t>(s >> 32);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CPU \u8def\u5f84\u5728 <code>dropout_op.cpp<\/code> \u4e2d\u505a\u4e86\u9010\u4f4d\u4e00\u81f4\u7684\u5b9e\u73b0\u3002\u8fd9\u6837\u65e0\u8bba\u8d70 GPU CUDA Graph \u8fd8\u662f CPU \u56de\u9000\uff0c\u540c\u4e00\u4e2a\u5168\u5c40\u79cd\u5b50\u90fd\u4f1a\u5f97\u5230\u76f8\u540c\u7684 Dropout \u63a9\u7801\u5e8f\u5217\u3002\u5173\u952e\u5728\u4e8e\uff0c\u79cd\u5b50\u7684\u65cb\u8f6c\u5b8c\u5168\u5728\u8bbe\u5907\u7aef\u5b8c\u6210\uff0cCUDA Graph \u6355\u83b7\u4e00\u6b21\u540e\u5c31\u53ef\u4ee5\u53cd\u590d\u91cd\u653e\uff0c\u4e0d\u9700\u8981 CPU \u4ecb\u5165\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cDropout kernel \u5185\u90e8\u4f7f\u7528\u7684\u662f Philox 2&#215;64 \u7b80\u5316\u7248\uff0c\u800c\u4e0d\u662f <code>philox.h<\/code> \u4e2d\u7684 4&#215;32-10\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">__device__ float philox_sample(uint64_t seed_lo, uint64_t seed_hi, uint64_t offset) {\n    uint64_t counter = seed_lo + offset;\n    uint64_t key0 = seed_hi;\n    uint64_t key1 = seed_lo ^ 0x9e3779b97f4a7c15ULL;\n\n    for (int i = 0; i &lt; 10; ++i) {\n        \/\/ mulhilo\uff1a\u7528 64 \u4f4d\u4e58\u6cd5\u628a\u8ba1\u6570\u5668\u6269\u6563\u5230\u9ad8 64 \u4f4d\n        uint64_t lo = 0xD2B74407B1CE6E93ULL * counter;\n        uint64_t hi = __umul64hi(0xD2B74407B1CE6E93ULL, counter);\n        uint64_t new_counter = hi ^ key0 ^ key1;\n        uint64_t new_key = lo;\n        counter = new_counter;\n        key0 = new_key;\n        key1 = new_key ^ key1;\n    }\n\n    uint32_t upper = static_cast&lt;uint32_t>(counter >> 32);\n    return __uint2float_rn(upper) * (1.0f \/ 4294967296.0f);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u53d8\u91cf\u8bf4\u660e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>seed_lo<\/code> \/ <code>seed_hi<\/code>\uff1a\u7531 per-rank dropout seed \u62c6\u5206\u51fa\u7684\u4e24\u4e2a 64 \u4f4d\u91cf\uff1b<\/li>\n\n\n\n<li><code>offset<\/code>\uff1a\u6bcf\u4e2a\u8f93\u51fa\u5143\u7d20\u7684\u5168\u5c40\u6241\u5e73\u7d22\u5f15\uff1b<\/li>\n\n\n\n<li><code>counter<\/code> \/ <code>key0<\/code> \/ <code>key1<\/code>\uff1aPhilox 2&#215;64 \u7684\u6838\u5fc3\u72b6\u6001\uff0c\u6bcf\u8f6e\u901a\u8fc7 64 \u4f4d\u4e58\u9ad8-\u4e58\u4f4e\u6269\u6563\uff1b<\/li>\n\n\n\n<li>\u6700\u7ec8\u53d6 <code>counter<\/code> \u7684\u9ad8 32 \u4f4d\u6620\u5c04\u5230 <code>[0, 1)<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd 2&#215;64 \u53d8\u4f53\u6bd4 4&#215;32-10 \u66f4\u9002\u5408 Dropout\uff1a\u6bcf\u4e2a\u7ebf\u7a0b\u53ea\u9700\u8981\u4e00\u4e2a float\uff0c\u65e0\u9700\u50cf 4&#215;32 \u90a3\u6837\u4e00\u6b21\u751f\u6210 4 \u4e2a\u6574\u6570\u518d\u6d6a\u8d39 3 \u4e2a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0cDropout \u7684 CPU \u8def\u5f84\u548c GPU \u8def\u5f84\u867d\u7136\u5728\u4e0d\u540c\u6587\u4ef6\u4e2d\u5b9e\u73b0\uff0c\u4f46\u7b97\u6cd5\u662f\u9010\u4f4d\u5bf9\u9f50\u7684\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e24\u8005\u90fd\u7528 SplitMix64 \u4ece\u5168\u5c40\u79cd\u5b50\u6d3e\u751f per-rank seed\uff1b<\/li>\n\n\n\n<li>\u4e24\u8005\u90fd\u7528 Xorshift64* \u5bf9 seed \u505a\u540c\u6784\u65cb\u8f6c\uff1b<\/li>\n\n\n\n<li>\u4e24\u8005\u90fd\u7528 Philox 2&#215;64 \u91c7\u6837 <code>offset = \u5143\u7d20\u7d22\u5f15<\/code>\uff1b<\/li>\n\n\n\n<li>\u4e24\u8005\u90fd\u628a\u9ad8 32 \u4f4d\u6620\u5c04\u5230 <code>[0, 1)<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c\u5373\u4f7f\u67d0\u6b21\u8fd0\u884c\u56e0\u4e3a\u73af\u5883\u6ca1\u6709 GPU \u800c\u56de\u9000\u5230 CPU\uff0c\u53ea\u8981\u5168\u5c40\u79cd\u5b50\u76f8\u540c\uff0cDropout \u63a9\u7801\u5e8f\u5217\u4ecd\u7136\u4e00\u81f4\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u968f\u673a\u6570\u4e4b\u5916\uff1a\u786e\u5b9a\u6027\u8bad\u7ec3\u8fd8\u9700\u8981\u4ec0\u4e48\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a\u968f\u673a\u6570\u6362\u6210 Philox \u53ea\u662f\u5fc5\u8981\u6761\u4ef6\uff0c\u4e0d\u662f\u5145\u5206\u6761\u4ef6\u3002Tech-Renaissance \u4e3a\u4e86\u7aef\u5230\u7aef\u53ef\u590d\u73b0\uff0c\u8fd8\u505a\u4e86\u51e0\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u9759\u6001\u56fe + CUDA Graph \u5168\u6355\u83b7<\/strong>\uff1a\u7b97\u5b50\u5e8f\u5217\u3001\u5185\u5b58\u5730\u5740\u3001\u8c03\u5ea6\u987a\u5e8f\u5728\u7f16\u8bd1\u671f\u56fa\u5b9a\uff0c\u6d88\u9664\u4e86\u8fd0\u884c\u65f6\u8c03\u5ea6\u548c\u5730\u5740\u53d8\u5316\u5e26\u6765\u7684\u4e0d\u786e\u5b9a\u6027\u3002\u52a8\u6001\u56fe\u6846\u67b6\u91cc\u5e38\u89c1\u7684 graph break\u3001\u52a8\u6001\u5f62\u72b6\u3001\u8fd0\u884c\u65f6\u5206\u914d\uff0c\u90fd\u4f1a\u6210\u4e3a\u4e0d\u53ef\u590d\u73b0\u7684\u6765\u6e90\u3002Tech-Renaissance \u7684 MemoryPlan \u5728\u7f16\u8bd1\u671f\u5c31\u4e3a\u6bcf\u4e2a DTensor \u8ba1\u7b97\u51fa\u56fa\u5b9a\u7684 offset\uff0c\u8fd0\u884c\u671f\u6240\u6709\u5f20\u91cf\u90fd\u4ece\u8fd9\u4e2a\u9759\u6001\u6c60\u91cc\u6309\u504f\u79fb\u8bbf\u95ee\u3002\u8fd9\u610f\u5473\u7740\u540c\u4e00\u6b21\u8bad\u7ec3\u3001\u540c\u4e00\u4e2a\u6a21\u578b\u3001\u540c\u4e00\u4e2a batch\uff0c\u6bcf\u6b21 forward\/backward \u8bbf\u95ee\u7684\u663e\u5b58\u5730\u5740\u5b8c\u5168\u76f8\u540c\u3002CUDA Graph \u6355\u83b7\u4e00\u6b21\u540e\u53cd\u590d\u91cd\u653e\uff0cCPU \u4e0d\u518d\u53c2\u4e0e\u6bcf\u6b21\u8fed\u4ee3\u7684 kernel \u542f\u52a8\uff0c\u81ea\u7136\u4e5f\u4e0d\u4f1a\u56e0\u4e3a CPU \u8c03\u5ea6\u6296\u52a8\u800c\u6539\u53d8 GPU \u4e0a\u7684\u6267\u884c\u65f6\u5e8f\u3002<\/li>\n\n\n\n<li><strong>\u907f\u514d\u6c42\u548c\u987a\u5e8f\u53d8\u5316\u7684\u5f52\u7ea6 kernel<\/strong>\uff1a\u6d6e\u70b9\u52a0\u6cd5\u4e0d\u6ee1\u8db3\u7ed3\u5408\u5f8b\uff0c\u5206\u6563\u7684 <code>atomicAdd<\/code> \u4f1a\u5bfc\u81f4\u7ed3\u679c\u968f\u7ebf\u7a0b\u8c03\u5ea6\u6ce2\u52a8\u3002\u4f8b\u5982\uff0c\u591a\u4e2a CUDA block \u540c\u65f6\u5f80\u540c\u4e00\u4e2a\u5730\u5740\u7d2f\u52a0\u68af\u5ea6\u65f6\uff0cblock \u7684\u5230\u8fbe\u987a\u5e8f\u4e0d\u540c\uff0c\u6700\u7ec8\u820d\u5165\u7ed3\u679c\u5c31\u4f1a\u4e0d\u540c\u3002\u6846\u67b6\u5728\u5173\u952e\u8def\u5f84\u4e0a\u5c3d\u91cf\u4f7f\u7528\u7a33\u5b9a\u7684\u5f52\u7ea6\u7b56\u7565\uff0c\u51cf\u5c11\u591a\u5361\u4e4b\u95f4\u7684\u6570\u503c\u6f02\u79fb\u3002\u4e3e\u4e2a\u5177\u4f53\u4f8b\u5b50\uff1a\u5047\u8bbe\u4e09\u4e2a\u68af\u5ea6\u503c\u5206\u522b\u662f <code>1.0f<\/code>\u3001<code>1e-8f<\/code>\u3001<code>1e-8f<\/code>\uff0c\u5982\u679c\u5148\u7528\u5927\u6570\u52a0\u5c0f\u6570\uff1a<code>float a = (1.0f + 1e-8f) + 1e-8f; \u00a0\/\/ \u7ed3\u679c\u4ecd\u662f 1.0f\uff08\u7b2c\u4e00\u6b21\u52a0\u5df2\u88ab\u820d\u5165\uff09<br>float b = 1e-8f + 1e-8f + 1.0f; \u00a0 \u00a0\/\/ \u7ed3\u679c\u53ef\u80fd\u662f 1.0000002f<\/code>\u5728\u5355\u7cbe\u5ea6\u4e0b\uff0c<code>1.0f + 1e-8f == 1.0f<\/code>\uff0c\u56e0\u4e3a\u5c0f\u6570\u88ab\u5927\u6570\u7684\u5c3e\u6570\u201c\u5403\u6389\u201d\u4e86\u3002\u4e0d\u540c\u7684\u7d2f\u52a0\u987a\u5e8f\u4f1a\u5bfc\u81f4\u6700\u7ec8\u7ed3\u679c\u4e0d\u540c\u3002\u5982\u679c AllReduce \u7684 block \u5230\u8fbe\u987a\u5e8f\u6bcf\u6b21\u8fd0\u884c\u90fd\u4e0d\u4e00\u6837\uff0c\u90a3\u4e48\u5373\u4f7f\u68af\u5ea6\u672c\u8eab\u5b8c\u5168\u76f8\u540c\uff0c\u805a\u5408\u540e\u7684\u6743\u91cd\u66f4\u65b0\u4e5f\u4f1a\u51fa\u73b0\u5fae\u5c0f\u5dee\u5f02\uff0c\u5e76\u5728\u540e\u7eed\u4f18\u5316\u6b65\u9aa4\u4e2d\u88ab\u653e\u5927\u3002<\/li>\n\n\n\n<li><strong>\u786e\u5b9a\u6027\u7684 MaxPool \u53cd\u5411\u5b9e\u73b0<\/strong>\uff1aCUDA \u8def\u5f84\u4e2d\uff0c<code>src\/backend\/ops\/dtensor\/maxpool_op.cu<\/code> \u5b9e\u73b0\u4e86 <code>maxpool_bwd_deterministic_kernel<\/code>\uff0c\u901a\u8fc7&#8221;\u8f93\u5165\u904d\u5386 + \u6536\u96c6\u7d2f\u52a0&#8221;\u7684\u65b9\u5f0f\u907f\u514d\u4f20\u7edf <code>atomicAdd<\/code> \u5e26\u6765\u7684\u975e\u786e\u5b9a\u6027\uff1b\u5982\u679c\u8d70 cuDNN pooling \u8def\u5f84\uff0c\u5219\u663e\u5f0f\u9009\u62e9 <code>CUDNN_POOLING_MAX_DETERMINISTIC<\/code>\u3002CPU \u8def\u5f84\u5219\u4f7f\u7528\u56fa\u5b9a\u904d\u5386\u987a\u5e8f\u7684\u5b9e\u73b0\uff0c\u4fdd\u8bc1\u540c\u4e00\u8f93\u5165\u4e0b\u7ed3\u679c\u7a33\u5b9a\u3002<\/li>\n\n\n\n<li><strong>\u7981\u7528 TF32<\/strong>\uff1a<code>GlobalRegistry::use_tf32(false)<\/code> \u901a\u8fc7\u8bbe\u7f6e\u73af\u5883\u53d8\u91cf <code>NVIDIA_TF32_OVERRIDE=0<\/code>\uff0c\u5f3a\u5236 Ampere \u53ca\u66f4\u65b0\u67b6\u6784\u7684 GPU \u4f7f\u7528\u7eaf FP32 \u8ba1\u7b97\uff0c\u907f\u514d TF32 \u4f4e\u7cbe\u5ea6\u5e26\u6765\u7684\u8fd0\u884c\u95f4\u5dee\u5f02\u3002<\/li>\n\n\n\n<li><strong>\u4e00\u81f4\u7684 Region \u5e03\u5c40<\/strong>\uff1a\u6240\u6709 rank \u5171\u4eab\u540c\u4e00\u4efd MemoryPlan\uff0cDTensor \u504f\u79fb\u8de8\u5361\u4e00\u81f4\uff0c\u4fdd\u8bc1\u591a\u5361\u968f\u673a\u6570\u6d88\u8d39\u987a\u5e8f\u4e00\u81f4\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5373\u4fbf\u5982\u6b64\uff0c\u6211\u4eec\u4e5f\u5fc5\u987b\u8bda\u5b9e\u5730\u8bf4\uff1a<strong>\u7edd\u5bf9\u7684\u8de8\u5e73\u53f0\u3001\u8de8\u786c\u4ef6\u53ef\u590d\u73b0\u6027\u5728\u5de5\u7a0b\u4e0a\u5f88\u96be\u627f\u8bfa<\/strong>\u3002GPU \u9a71\u52a8\u7248\u672c\u3001cuDNN\/cuBLAS \u7248\u672c\u3001\u4e0d\u540c GPU \u7684 Tensor Core \u884c\u4e3a\u90fd\u53ef\u80fd\u5f15\u5165\u5fae\u5c0f\u5dee\u5f02\u3002Tech-Renaissance \u80fd\u4fdd\u8bc1\u7684\u662f\uff1a\u5728\u76f8\u540c\u4ee3\u7801\u3001\u76f8\u540c\u786c\u4ef6\u3001\u76f8\u540c\u79cd\u5b50\u3001\u5f00\u542f\u53ef\u590d\u73b0\u6a21\u5f0f\u7684\u524d\u63d0\u4e0b\uff0c\u591a\u6b21\u8fd0\u884c\u5f97\u5230\u4e00\u81f4\u7ed3\u679c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u4e00\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u7ec6\u8282\uff1aLARS \u4fe1\u4efb\u7cfb\u6570\u7684\u786e\u5b9a\u6027<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 B20\u300a\u878d\u5408\u4f18\u5316\u5668\u300b\u4e2d\u6211\u4eec\u63d0\u5230\u8fc7\uff0cLARS \u9700\u8981\u5bf9\u6bcf\u4e2a\u53ef\u8bad\u7ec3\u5c42\u5206\u522b\u8ba1\u7b97 <code>||w||^2<\/code> \u548c <code>||g||^2<\/code>\uff0c\u8fd9\u662f\u5178\u578b\u7684\u5f52\u7ea6\u64cd\u4f5c\u3002\u4e3a\u4e86\u4fdd\u8bc1\u53ef\u590d\u73b0\uff0cLARS \u7684\u4fe1\u4efb\u7cfb\u6570\u8ba1\u7b97\u91c7\u7528\u4e24\u9636\u6bb5\u56fa\u5b9a\u987a\u5e8f\u5f52\u7ea6\uff1a\u7b2c\u4e00\u9636\u6bb5\u6bcf\u4e2a block \u8ba1\u7b97\u90e8\u5206\u548c\u5e76\u5199\u5165\u56fa\u5b9a\u7684\u90e8\u5206\u548c\u7f13\u51b2\u533a\uff0c\u7b2c\u4e8c\u9636\u6bb5\u6309\u56fa\u5b9a\u987a\u5e8f\u7d2f\u52a0\u8fd9\u4e9b\u90e8\u5206\u548c\u3002\u53ea\u8981 block \u7f51\u683c\u5927\u5c0f\u4e0d\u53d8\uff0c\u6c42\u548c\u987a\u5e8f\u5c31\u4e0d\u53d8\uff0c\u7ed3\u679c\u4e5f\u5c31\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u4e0e PyTorch \u8def\u5f84\u7684\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u7684 RNG \u8bbe\u8ba1\u662f\u201c\u5206\u5c42\u7684\u201d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>CPU \u7528 Mersenne Twister\uff0c\u5fc5\u987b\u987a\u5e8f\u6d88\u8d39\uff1b<\/li>\n\n\n\n<li>CUDA \u7528 Philox\uff0c\u4f46\u6bcf\u4e2a\u5f20\u91cf\u64cd\u4f5c\u6709\u81ea\u5df1\u7684 offset \u8ba1\u6570\uff1b<\/li>\n\n\n\n<li>DataLoader \u9700\u8981\u5355\u72ec\u4e3a\u6bcf\u4e2a worker \u8bbe\u79cd\u5b50\uff0c\u5178\u578b\u5199\u6cd5\u5982\u4e0b\uff1a<\/li>\n<\/ul>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"python\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">def seed_worker(worker_id):\n    worker_seed = torch.initial_seed() % 2**32\n    numpy.random.seed(worker_seed)\n    random.seed(worker_seed)\n\ng = torch.Generator()\ng.manual_seed(0)\nDataLoader(train_dataset, batch_size=..., num_workers=...,\n           worker_init_fn=seed_worker, generator=g)<\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7528\u6237\u8fd8\u9700\u8981\u624b\u52a8\u5904\u7406 <code>cudnn.deterministic<\/code>\u3001<code>cudnn.benchmark<\/code>\u3001<code>use_deterministic_algorithms<\/code>\u3001<code>CUBLAS_WORKSPACE_CONFIG<\/code> \u7b49\u5f00\u5173\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u505a\u6cd5\u662f\u201c\u7edf\u4e00\u7684\u201d\uff1a\u4ece\u6743\u91cd\u521d\u59cb\u5316\u3001\u6570\u636e\u589e\u5f3a\u3001\u6570\u636e\u96c6\u6d17\u724c\u5230 Dropout\uff0c\u5168\u90e8\u57fa\u4e8e\u540c\u4e00\u4efd Philox \u539f\u8bed\uff1b\u901a\u8fc7 <code>manual_seed<\/code> \u4e00\u4e2a\u5165\u53e3\uff0c\u628a <code>Generator<\/code>\u3001worker \u884d\u751f\u79cd\u5b50\u3001Dropout per-rank seed\u3001TransferStation \u53ef\u590d\u73b0\u6a21\u5f0f\u5168\u90e8\u4e32\u8d77\u6765\u3002\u5b83\u727a\u7272\u4e86 DataLoader \u5728\u53ef\u590d\u73b0\u6a21\u5f0f\u4e0b\u7684\u6781\u9650\u541e\u5410\uff0c\u4f46\u6362\u6765\u7684\u662f\u201c\u4e00\u952e\u53ef\u590d\u73b0\u201d\u7684\u4f53\u9a8c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u5dee\u5f02\u4e5f\u53cd\u6620\u5728\u8bbe\u8ba1\u54f2\u5b66\u4e0a\u3002PyTorch \u503e\u5411\u4e8e\u628a\u63a7\u5236\u6743\u4ea4\u7ed9\u7528\u6237\uff1a\u4f60\u53ef\u4ee5\u4e3a CPU\u3001CUDA\u3001\u6bcf\u4e2a DataLoader worker\u3001\u6bcf\u4e2a\u7b2c\u4e09\u65b9\u5e93\u5206\u522b\u8bbe\u79cd\u5b50\uff0c\u7075\u6d3b\u4f46\u7e41\u7410\u3002Tech-Renaissance \u503e\u5411\u4e8e\u628a\u63a7\u5236\u6743\u6536\u56de\u5230\u6846\u67b6\uff1a\u53ea\u8981\u7528\u6237\u8868\u8fbe\u201c\u6211\u8981\u53ef\u590d\u73b0\u201d\uff0c\u6846\u67b6\u5c31\u81ea\u52a8\u628a\u6240\u6709\u968f\u673a\u6d88\u8d39\u70b9\u4e32\u5230\u540c\u4e00\u6761\u786e\u5b9a\u6027\u8f68\u9053\u4e0a\u3002\u8fd9\u4e0d\u662f\u8bf4 PyTorch \u505a\u4e0d\u5230\u53ef\u590d\u73b0\uff0c\u800c\u662f\u8bf4\u5728 Tech-Renaissance \u91cc\uff0c\u505a\u5bf9\u7684\u6210\u672c\u66f4\u4f4e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u6027\u80fd\u4e0e trade-off<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Philox \u7b97\u6cd5\u672c\u8eab\u7684\u8ba1\u7b97\u5f00\u9500\u5f88\u4f4e\uff1a10 \u8f6e\u6574\u6570\u4e58\u6cd5\u548c XOR\uff0c\u73b0\u4ee3 CPU \u548c GPU \u90fd\u80fd\u9ad8\u6548\u6267\u884c\u3002\u5728\u6211\u4eec\u7684\u5b9e\u73b0\u91cc\uff0cDropout kernel \u4e2d\u968f\u673a\u6570\u751f\u6210\u53ea\u5360\u6574\u4e2a kernel \u5f88\u5c0f\u4e00\u90e8\u5206\u65f6\u95f4\uff0c\u4e3b\u8981\u5f00\u9500\u4ecd\u7136\u662f\u663e\u5b58\u8bfb\u5199\uff1b\u53c2\u6570\u521d\u59cb\u5316\u51e0\u767e\u4e07\u4e2a float \u4e5f\u53ea\u9700\u8981\u5fae\u79d2\u7ea7\u3002\u56e0\u6b64\uff0c<strong>\u968f\u673a\u6570\u751f\u6210\u672c\u8eab\u5e76\u4e0d\u662f\u53ef\u590d\u73b0\u6027\u7684\u6027\u80fd\u74f6\u9888<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f8\u6bd4 Mersenne Twister\uff0cPhilox \u5728\u5e76\u884c\u573a\u666f\u4e0b\u751a\u81f3\u53ef\u80fd\u66f4\u5feb\uff1aMT19937 \u9700\u8981\u7ef4\u62a4\u4e00\u4e2a 624 \u5b57\u7684\u72b6\u6001\u6570\u7ec4\uff0c\u9891\u7e41\u8bbf\u95ee\u4f1a\u5e26\u6765\u7f13\u5b58\u538b\u529b\uff1b\u800c Philox \u6bcf\u6b21\u751f\u6210\u53ea\u9700\u8981\u5c11\u91cf\u5bc4\u5b58\u5668\uff0c\u6ca1\u6709\u72b6\u6001\u6570\u7ec4\uff0c\u7f13\u5b58\u53cb\u597d\u3002\u5728 CPU \u591a\u7ebf\u7a0b\u573a\u666f\u4e0b\uff0cPhilox \u4e0d\u9700\u8981\u9501\uff0c\u591a\u4e2a\u7ebf\u7a0b\u53ef\u4ee5\u65e0\u7ade\u4e89\u5730\u540c\u65f6\u751f\u6210\u968f\u673a\u6570\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u771f\u6b63\u7684\u6027\u80fd\u4ee3\u4ef7\u6765\u81ea<strong>\u6570\u636e\u9884\u5904\u7406\u9636\u6bb5\u7684\u540c\u6b65<\/strong>\u3002\u5728\u53ef\u590d\u73b0\u6a21\u5f0f\u4e0b\uff0cTransferStation \u5fc5\u987b\u201c\u5feb worker \u7b49\u6162 worker\u201d\uff0c\u4fdd\u8bc1\u6bcf\u4e2a\u6837\u672c\u5199\u5165\u9884\u5b9a\u4f4d\u7f6e\u3002\u5982\u679c worker \u4e4b\u95f4\u5904\u7406\u901f\u5ea6\u5dee\u5f02\u5f88\u5927\uff0c\u5feb worker \u4f1a\u88ab\u6162 worker \u62d6\u6162\uff0c\u6574\u4f53\u541e\u5410\u91cf\u4f1a\u6709\u6240\u4e0b\u964d\u3002\u8fd9\u4e2a\u5f00\u9500\u4e0e\u6570\u636e\u96c6\u3001worker \u6570\u91cf\u3001\u9884\u5904\u7406\u590d\u6742\u5ea6\u90fd\u76f8\u5173\uff0c\u65e0\u6cd5\u7ed9\u51fa\u4e00\u4e2a\u653e\u4e4b\u56db\u6d77\u800c\u7686\u51c6\u7684\u6570\u5b57\uff1b\u7528\u6237\u53ef\u4ee5\u6839\u636e\u573a\u666f\u5728\u201c\u53ef\u590d\u73b0\u201d\u548c\u201c\u6781\u9650\u541e\u5410\u201d\u4e4b\u95f4\u5207\u6362\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u4e2a\u6f5c\u5728\u7684\u4ee3\u4ef7\u662f\u67d0\u4e9b\u7b97\u5b50\u5fc5\u987b\u9009\u62e9\u786e\u5b9a\u6027\u5b9e\u73b0\u3002\u4f8b\u5982 MaxPool \u7684\u53cd\u5411\u4f20\u64ad\uff0c\u5982\u679c\u4f7f\u7528\u4f20\u7edf\u7684\u201c\u8f93\u51fa\u904d\u5386 + atomicAdd\u201d\u7b56\u7565\uff0c\u4e0d\u540c block \u7684\u5230\u8fbe\u987a\u5e8f\u4f1a\u5bfc\u81f4\u7ed3\u679c\u6ce2\u52a8\uff1b\u6211\u4eec\u7684\u786e\u5b9a\u6027\u5b9e\u73b0\u91c7\u7528\u201c\u8f93\u5165\u904d\u5386 + \u6536\u96c6\u7d2f\u52a0\u201d\uff0c\u907f\u514d\u4e86\u539f\u5b50\u64cd\u4f5c\uff0c\u4f46\u53ef\u80fd\u727a\u7272\u4e00\u70b9\u70b9\u5cf0\u503c\u6027\u80fd\u3002\u8fd9\u662f\u4e00\u79cd\u53ef\u9884\u6d4b\u3001\u53ef\u91cf\u5316\u7684\u53d6\u820d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u573a\u666f<\/th><th>\u662f\u5426\u5efa\u8bae\u5f00\u542f\u53ef\u590d\u73b0<\/th><th>\u7406\u7531<\/th><\/tr><\/thead><tbody><tr><td>\u8c03\u8bd5 bug<\/td><td>\u2705<\/td><td>\u6bcf\u6b21\u8fd0\u884c\u4ee3\u7801\u8def\u5f84\u76f8\u540c\uff0c\u65b9\u4fbf\u5b9a\u4f4d<\/td><\/tr><tr><td>\u8bba\u6587\u5b9e\u9a8c<\/td><td>\u2705<\/td><td>\u7ed3\u679c\u53ef\u590d\u73b0\u662f\u79d1\u5b66\u7814\u7a76\u7684\u57fa\u672c\u8981\u6c42<\/td><\/tr><tr><td>\u8d85\u53c2\u6570\u641c\u7d22<\/td><td>\u2705<\/td><td>\u6392\u9664\u968f\u673a\u6027\u5e72\u6270\uff0c\u51c6\u786e\u6bd4\u8f83\u8d85\u53c2\u6570\u6548\u679c<\/td><\/tr><tr><td>\u751f\u4ea7\u73af\u5883\u5927\u89c4\u6a21\u8dd1\u5206<\/td><td>\u274c<\/td><td>\u6027\u80fd\u4f18\u5148\uff0c\u5fae\u5c0f\u5dee\u5f02\u4e0d\u5f71\u54cd\u6700\u7ec8\u7ed3\u679c<\/td><\/tr><tr><td>ablation study<\/td><td>\u2705<\/td><td>\u51c6\u786e\u6bd4\u8f83\u4e0d\u540c\u7ec4\u4ef6\u7684\u8d21\u732e<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4e0d\u8c03\u7528 <code>manual_seed<\/code>\uff0c\u6846\u67b6\u4f1a\u9ed8\u8ba4\u4f7f\u7528 <code>auto_seed()<\/code> \u53d6\u65f6\u95f4\u6233\u4f5c\u4e3a\u79cd\u5b50\uff0c\u5e76\u5173\u95ed\u53ef\u590d\u73b0\u6027\u4fdd\u9669\uff0c\u6b64\u65f6 <code>TransferStation<\/code> \u8d70\u65e0\u9501\u5feb\u901f\u8def\u5f84\uff0c\u6570\u636e\u7ba1\u7ebf\u53ef\u4ee5\u8fbe\u5230\u6700\u9ad8\u541e\u5410\u3002\u8fd9\u79cd\u201c\u8981\u53ef\u590d\u73b0\u5c31\u4e00\u884c\u5f00\u542f\uff0c\u8981\u6027\u80fd\u5c31\u4e0d\u989d\u5916\u540c\u6b65\u201d\u7684\u8bbe\u8ba1\uff0c\u8ba9\u540c\u4e00\u4e2a\u6846\u67b6\u65e2\u80fd\u670d\u52a1\u79d1\u7814\u8c03\u8bd5\uff0c\u4e5f\u80fd\u670d\u52a1\u751f\u4ea7\u8dd1\u5206\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">7.1 \u53ef\u590d\u73b0\u6027\u7684\u8fb9\u754c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u867d\u7136\u6211\u4eec\u628a\u786e\u5b9a\u6027\u8bad\u7ec3\u505a\u5230\u4e86\u6846\u67b6\u5c42\u9762\u7684\u201c\u4e00\u952e\u5f00\u542f\u201d\uff0c\u4f46\u4ecd\u6709\u4e00\u4e9b\u8fb9\u754c\u9700\u8981\u8bf4\u660e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8de8 GPU \u67b6\u6784<\/strong>\uff1a\u4e0d\u540c\u67b6\u6784\u7684 Tensor Core\u3001FP32 \u5355\u5143\u5b9e\u73b0\u53ef\u80fd\u6709\u5fae\u5c0f\u5dee\u5f02\uff0c\u4e0d\u80fd\u4fdd\u8bc1 A100 \u548c RTX 4090 \u8dd1\u51fa\u5b8c\u5168\u4e00\u81f4\u7684 loss \u66f2\u7ebf\uff1b<\/li>\n\n\n\n<li><strong>\u8de8\u9a71\u52a8 \/ \u5e93\u7248\u672c<\/strong>\uff1acuDNN\u3001cuBLAS \u7684\u5c0f\u7248\u672c\u66f4\u65b0\u6709\u65f6\u4f1a\u6539\u53d8\u67d0\u4e9b kernel \u7684\u53d6\u820d\u6216\u820d\u5165\u884c\u4e3a\uff1b<\/li>\n\n\n\n<li><strong>\u8de8\u7f16\u8bd1\u5668 \/ \u4f18\u5316\u7ea7\u522b<\/strong>\uff1aCPU \u8def\u5f84\u7684 OpenMP \u5212\u5206\u3001\u5411\u91cf\u5316\u53ef\u80fd\u5f71\u54cd\u6d6e\u70b9\u7d2f\u52a0\u987a\u5e8f\uff1b<\/li>\n\n\n\n<li><strong>\u5206\u5e03\u5f0f\u5f02\u6b65\u901a\u4fe1<\/strong>\uff1a\u867d\u7136\u6211\u4eec\u5c3d\u91cf\u8ba9\u901a\u4fe1\u4e0e\u8ba1\u7b97\u91cd\u53e0\u5e76\u4fdd\u6301\u786e\u5b9a\u6027\uff0c\u4f46\u6781\u7aef\u60c5\u51b5\u4e0b\u7f51\u7edc\u5ef6\u8fdf\u6ce2\u52a8\u4ecd\u53ef\u80fd\u5f71\u54cd\u67d0\u4e9b\u65f6\u5e8f\u654f\u611f\u7684\u64cd\u4f5c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c\u6211\u4eec\u4e0d\u8bf4\u201c\u8de8\u5e73\u53f0\u5b57\u8282\u7ea7\u4e00\u81f4\u201d\uff0c\u800c\u662f\u8bf4\u201c\u5728\u76f8\u540c\u4ee3\u7801\u3001\u76f8\u540c\u786c\u4ef6\u3001\u76f8\u540c\u8f6f\u4ef6\u6808\u3001\u76f8\u540c\u79cd\u5b50\u4e0b\u53ef\u590d\u73b0\u201d\u3002\u8fd9\u662f\u4e00\u4e2a\u66f4\u8bda\u5b9e\u3001\u4e5f\u66f4\u6709\u5de5\u7a0b\u4ef7\u503c\u7684\u627f\u8bfa\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u673a\u6570\u4e0d\u662f\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u88c5\u9970\u54c1\uff0c\u800c\u662f\u51b3\u5b9a\u8bad\u7ec3\u80fd\u5426\u88ab\u4fe1\u4efb\u7684\u57fa\u7840\u8bbe\u65bd\u4e4b\u4e00\u3002Philox \u7684\u8ba1\u6570\u5668\u578b\u8bbe\u8ba1\u628a\u201c\u7b2c n \u4e2a\u968f\u673a\u6570\u201d\u53d8\u6210\u7eaf\u51fd\u6570\uff0c\u4ece\u6839\u672c\u4e0a\u89e3\u9664\u4e86\u5e76\u884c\u6267\u884c\u987a\u5e8f\u5bf9\u968f\u673a\u5e8f\u5217\u7684\u7ed1\u5b9a\u3002Tech-Renaissance \u56f4\u7ed5\u8fd9\u4e00\u601d\u60f3\uff0c\u6253\u9020\u4e86\u8986\u76d6\u521d\u59cb\u5316\u3001\u6570\u636e\u589e\u5f3a\u3001\u6d17\u724c\u3001Dropout \u7684\u786e\u5b9a\u6027\u7ba1\u7ebf\uff0c\u5e76\u901a\u8fc7 <code>GlobalRegistry::manual_seed<\/code> \u548c\u53ef\u590d\u73b0\u6027\u4fdd\u9669\u628a\u5b83\u66b4\u9732\u4e3a\u4e00\u884c\u8c03\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u590d\u73b0\u6027\u6709\u4ee3\u4ef7\uff1a\u5b83\u8981\u6c42\u4f60\u653e\u5f03\u4e00\u4e9b\u56e0\u4e71\u5e8f\u6267\u884c\u800c\u83b7\u5f97\u7684\u6027\u80fd\uff0c\u8981\u6c42\u6240\u6709\u968f\u673a\u6d88\u8d39\u70b9\u90fd\u88ab\u6846\u67b6\u7cbe\u786e\u8ba1\u6570\uff0c\u4e5f\u8981\u6c42\u5e95\u5c42\u7b97\u5b50\u4e0d\u80fd\u5f15\u5165\u975e\u786e\u5b9a\u6027\u3002\u4f46\u5728\u8c03\u8bd5\u548c\u79d1\u7814\u573a\u666f\u91cc\uff0c\u8fd9\u79cd\u4ee3\u4ef7\u901a\u5e38\u662f\u503c\u5f97\u7684\u3002\u6bd5\u7adf\uff0c\u4e00\u4e2a\u8dd1\u5f97\u5feb\u5374\u4e0d\u53ef\u89e3\u91ca\u7684\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u5f88\u96be\u88ab\u79f0\u4e3a\u771f\u6b63\u53ef\u9760\u7684\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u4e00\u4e0b\u672c\u6587\u7684\u8981\u70b9\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u4f20\u7edf\u72b6\u6001\u673a\u5f0f RNG \u5728\u5e76\u53d1\u573a\u666f\u4e0b\u5929\u751f\u96be\u4ee5\u4fdd\u8bc1\u53ef\u590d\u73b0\uff0c\u7ebf\u7a0b\u8c03\u5ea6\u987a\u5e8f\u4f1a\u5f71\u54cd\u72b6\u6001\u66f4\u65b0\u987a\u5e8f\uff1b<\/li>\n\n\n\n<li>Philox \u662f\u57fa\u4e8e\u8ba1\u6570\u5668\u7684\u65e0\u72b6\u6001 RNG\uff0c\u76f8\u540c <code>(seed, offset)<\/code> \u6c38\u8fdc\u5f97\u5230\u76f8\u540c\u968f\u673a\u6570\uff0c\u5929\u7136\u652f\u6301\u65e0\u9501\u5e76\u884c\uff1b<\/li>\n\n\n\n<li>Tech-Renaissance \u7684 <code>Generator<\/code> \u7528\u539f\u5b50 <code>offset_<\/code> \u9884\u7559\u533a\u95f4\uff0c\u914d\u5408 Pimpl \u6a21\u5f0f\u89e3\u51b3 MUSA \u517c\u5bb9\u6027\uff1b<\/li>\n\n\n\n<li><code>manual_seed()<\/code> \u4e00\u884c\u8c03\u7528\u5373\u53ef\u5f00\u542f\u7aef\u5230\u7aef\u53ef\u590d\u73b0\u6a21\u5f0f\uff1b<\/li>\n\n\n\n<li>Dropout \u7684 per-rank seed \u7528 SplitMix64 \u6d3e\u751f\u3001Xorshift64* \u81ea\u65cb\u8f6c\uff0c\u5b8c\u5168\u5728 CUDA Graph \u5185\u5b8c\u6210\uff1b<\/li>\n\n\n\n<li>\u53ef\u590d\u73b0\u6027\u662f\u5168\u94fe\u8def\u95ee\u9898\uff0c\u8fd8\u9700\u8981\u9759\u6001\u56fe\u3001\u786e\u5b9a\u6027\u5f52\u7ea6\u3001TF32 \u5173\u95ed\u7b49\u914d\u5408\uff1b<\/li>\n\n\n\n<li>\u5728\u76f8\u540c\u4ee3\u7801\u3001\u76f8\u540c\u786c\u4ef6\u3001\u76f8\u540c\u79cd\u5b50\u4e0b\uff0cTech-Renaissance \u53ef\u4ee5\u4fdd\u8bc1\u591a\u6b21\u8fd0\u884c\u7ed3\u679c\u4e00\u81f4\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u5c06\u628a\u89c6\u89d2\u4ece\u5355\u5361\u62c9\u5411\u591a\u5361\uff0c\u804a\u804a NCCL \u5206\u5e03\u5f0f\u901a\u4fe1\u4e0e\u6570\u636e\u5e76\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\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\u4e09 \u8bad\u7ec3\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u65f6\uff0c\u968f [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":602,"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-559","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\/559","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=559"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/559\/revisions"}],"predecessor-version":[{"id":695,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/559\/revisions\/695"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/602"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=559"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=559"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=559"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}