{"id":517,"date":"2026-07-08T03:05:42","date_gmt":"2026-07-07T19:05:42","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=517"},"modified":"2026-07-08T21:59:36","modified_gmt":"2026-07-08T13:59:36","slug":"dtensor%ef%bc%9a%e4%b8%80%e5%bc%a0%e5%9b%be%e7%ba%b8%ef%bc%8c%e5%a4%9a%e5%8d%a1%e5%85%b1%e4%ba%ab%e7%9a%84%e5%88%86%e5%b8%83%e5%bc%8f%e5%bc%a0%e9%87%8f%e6%8a%bd%e8%b1%a1","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/517\/","title":{"rendered":"(16) DTensor\uff1a\u4e00\u5f20\u56fe\u7eb8\uff0c\u591a\u5361\u5171\u4eab\u7684\u5206\u5e03\u5f0f\u5f20\u91cf\u62bd\u8c61"},"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\u5341\u516d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0a\u4e00\u7bc7\u6211\u4eec\u8bb2\u4e86 MemoryPlan \u5982\u4f55\u628a\u663e\u5b58\u6309\u8bed\u4e49\u5207\u6210\u4e00\u4e2a\u4e2a Region\uff0c\u5e76\u5728\u7f16\u8bd1\u671f\u628a\u6bcf\u5f20\u5f20\u91cf\u8be5\u653e\u5728\u54ea\u91cc\u3001\u5360\u591a\u5c11\u5b57\u8282\u3001\u751f\u547d\u5468\u671f\u591a\u957f\uff0c\u5168\u90e8\u5b89\u6392\u5f97\u660e\u660e\u767d\u767d\u3002\u4f46 MemoryPlan \u89e3\u51b3\u7684\u662f\u201c\u5355\u5361\u5185\u90e8\u201d\u7684\u5e03\u5c40\u95ee\u9898\uff1b\u4e00\u65e6\u8fdb\u5165\u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u516b\u5f20\u5361\u751a\u81f3\u66f4\u591a\u5361\u540c\u65f6\u5de5\u4f5c\uff0c\u95ee\u9898\u5c31\u4e0a\u5347\u4e86\u4e00\u4e2a\u7ef4\u5ea6\uff1a<strong>\u5982\u4f55\u8ba9\u6240\u6709 GPU \u5bf9\u540c\u4e00\u5f20\u201c\u5185\u5b58\u56fe\u7eb8\u201d\u8fbe\u6210\u4e00\u81f4\uff1f<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u56de\u7b54\u8fd9\u4e2a\u95ee\u9898\u4e4b\u524d\uff0c\u6211\u4eec\u4e0d\u59a8\u5148\u9000\u4e00\u6b65\uff0c\u770b\u770b\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u91cc\u7684\u201c\u5f20\u91cf\u201d\u5230\u5e95\u5728\u626e\u6f14\u4ec0\u4e48\u89d2\u8272\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u5f20\u91cf\u7684\u4e24\u79cd\u8eab\u4efd<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e2d\uff0c\u4e00\u4e2a\u5f20\u91cf\u5176\u5b9e\u627f\u8f7d\u7740\u4e24\u79cd\u622a\u7136\u4e0d\u540c\u7684\u8eab\u4efd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e00\u79cd\u8eab\u4efd\u662f\u6570\u636e\u5bb9\u5668\u3002<\/strong> \u5b83\u6301\u6709\u5b9e\u9645\u7684\u5185\u5b58\u7f13\u51b2\u533a\uff0c\u5b58\u50a8\u7740\u6d6e\u70b9\u6570\u6216\u6574\u6570\uff0c\u4f60\u53ef\u4ee5\u5f80\u91cc\u9762\u5199\u6570\u636e\uff0c\u4e5f\u53ef\u4ee5\u4ece\u4e2d\u8bfb\u6570\u636e\u3002\u5f53\u6211\u4eec\u8bf4\u201c\u628a\u4e00\u4e2a batch \u7684\u56fe\u50cf\u52a0\u8f7d\u5230 GPU \u663e\u5b58\u201d\uff0c\u6211\u4eec\u8c08\u8bba\u7684\u662f\u6570\u636e\u5bb9\u5668\u7684\u8eab\u4efd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e8c\u79cd\u8eab\u4efd\u662f\u5185\u5b58\u63cf\u8ff0\u7b26\u3002<\/strong> \u5b83\u63cf\u8ff0\u4e86\u4e00\u5757\u5185\u5b58\u7684\u903b\u8f91\u5c5e\u6027\uff1a\u5f62\u72b6\u662f\u4ec0\u4e48\u3001\u6570\u636e\u7c7b\u578b\u662f\u4ec0\u4e48\u3001\u5728\u663e\u5b58\u7684\u54ea\u4e2a\u4f4d\u7f6e\u3001stride \u662f\u591a\u5c11\u3002\u5f53\u6211\u4eec\u8bf4\u201c\u5377\u79ef\u5c42\u7684\u8f93\u5165\u5f20\u91cf\u5f62\u72b6\u662f [128, 224, 224, 3]\u201d\uff0c\u6211\u4eec\u8c08\u8bba\u7684\u662f\u63cf\u8ff0\u7b26\u7684\u8eab\u4efd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3b\u6d41\u6846\u67b6\u901a\u5e38\u628a\u8fd9\u4e24\u79cd\u8eab\u4efd\u5408\u4e8c\u4e3a\u4e00\u3002PyTorch \u7684 <code>torch.Tensor<\/code> \u540c\u65f6\u6301\u6709\u5143\u6570\u636e\uff08shape\u3001stride\u3001dtype\u3001device\uff09\u548c\u5b9e\u9645\u6570\u636e\u7f13\u51b2\u533a\uff08storage\uff09\u3002\u8fd9\u79cd\u8bbe\u8ba1\u7684\u597d\u5904\u662f\u76f4\u89c2\u2014\u2014\u4f60\u521b\u5efa\u4e00\u4e2a Tensor\uff0c\u5b83\u81ea\u52a8\u5206\u914d\u5185\u5b58\uff0c\u7136\u540e\u4f60\u5c31\u80fd\u76f4\u63a5\u64cd\u4f5c\u5b83\u3002\u8fd9\u79cd\u201c\u6240\u89c1\u5373\u6240\u5f97\u201d\u7684\u4f53\u9a8c\u975e\u5e38\u9002\u5408\u7814\u7a76\u548c\u539f\u578b\u5f00\u53d1\uff0c\u4e5f\u662f PyTorch \u80fd\u5728\u5b66\u672f\u754c\u8fc5\u901f\u6d41\u884c\u7684\u5173\u952e\u539f\u56e0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u52a3\u52bf\u540c\u6837\u660e\u663e\uff1a\u5143\u6570\u636e\u548c\u6570\u636e\u7ed1\u5b9a\u5728\u4e00\u8d77\uff0c\u610f\u5473\u7740\u4f60\u65e0\u6cd5\u5728\u4e0d\u89e6\u53ca\u5177\u4f53\u6570\u636e\u7684\u524d\u63d0\u4e0b\u53bb\u89c4\u5212\u6574\u4e2a\u663e\u5b58\u5e03\u5c40\u3002\u4f60\u5fc5\u987b\u5148\u521b\u5efa Tensor\u3001\u5206\u914d\u5185\u5b58\uff0c\u624d\u80fd\u786e\u5b9a\u5b83\u7684 shape \u548c stride\uff0c\u8fd9\u5c31\u628a\u201c\u5148\u89c4\u5212\u540e\u6267\u884c\u201d\u7684\u53ef\u80fd\u6027\u5835\u6b7b\u4e86\u3002\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e2d\uff0c8 \u5f20 GPU \u4e0a\u5404\u6709\u5404\u7684 Tensor \u5b9e\u4f8b\uff0c\u5b83\u4eec\u7684\u5143\u6570\u636e\u867d\u7136\u76f8\u540c\uff0c\u4f46\u90fd\u662f\u72ec\u7acb\u7ef4\u62a4\u7684\u526f\u672c\uff1b\u6bcf\u6b21\u8fdb\u7a0b\u542f\u52a8\u65f6 <code>cudaMalloc<\/code> \u8fd4\u56de\u7684\u5730\u5740\u4e5f\u53ef\u80fd\u4e0d\u540c\uff0c\u7ed9 CUDA Graph \u7684\u5730\u5740\u7a33\u5b9a\u6027\u5e26\u6765\u6311\u6218\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u9009\u62e9\u662f\u628a\u8fd9\u4e24\u79cd\u8eab\u4efd\u62c6\u5f00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Tensor<\/strong> \u662f\u6570\u636e\u5bb9\u5668\uff0c\u5b83\u53ea\u5728 CPU \u7aef\u5b58\u5728\uff0c\u4e13\u95e8\u8d1f\u8d23\u4e3b\u673a\u4e0e\u8bbe\u5907\u4e4b\u95f4\u7684\u6570\u636e\u642c\u8fd0\uff08H2D\/D2H\uff09\u3002\u5b83\u6301\u6709\u9875\u9501\u5b9a\u5185\u5b58\uff0c\u652f\u6301\u5404\u79cd\u521d\u59cb\u5316\u65b9\u5f0f\uff0c\u4f46\u5b83\u4e0d\u53c2\u4e0e\u4efb\u4f55 GPU \u7aef\u7684\u8fd0\u7b97\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>DTensor<\/strong>\uff08DistributedTensor\uff09\u662f\u7eaf\u5185\u5b58\u63cf\u8ff0\u7b26\u3002\u5b83\u53ea\u5b58\u50a8\u5f62\u72b6\u3001\u504f\u79fb\u91cf\u3001stride \u548c Region \u4fe1\u606f\uff0c\u4e0d\u6301\u6709\u4efb\u4f55\u5185\u5b58\u6307\u9488\uff0c\u4e0d\u5206\u914d\u4efb\u4f55\u7f13\u51b2\u533a\u3002\u5b83\u662f\u4e00\u4e2a\u201c\u865a\u62df\u5f20\u91cf\u201d\u2014\u2014\u5b83\u544a\u8bc9\u4f60\u201c\u6709\u8fd9\u6837\u4e00\u4e2a\u5f20\u91cf\uff0c\u5b83\u7684\u5f62\u72b6\u662f [128, 224, 224, 3]\uff0c\u6570\u636e\u7c7b\u578b\u662f FP16\uff0c\u4f4d\u4e8e\u663e\u5b58\u504f\u79fb\u91cf X \u5904\uff0cstride \u662f [HWC, WC, C, 1]\u201d\uff0c\u4f46\u5b83\u81ea\u5df1\u5e76\u4e0d\u62e5\u6709\u90a3\u5757\u5185\u5b58\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u5206\u79bb\u5e26\u6765\u4e86\u4e00\u4e2a\u6df1\u8fdc\u7684\u7ed3\u679c\uff1a<strong>\u540c\u4e00\u4efd DTensor \u63cf\u8ff0\uff0c\u5728\u6240\u6709 GPU \u4e0a\u5bf9\u5e94\u76f8\u540c\u7684\u5185\u5b58\u504f\u79fb\u4f4d\u7f6e\uff0c\u5e03\u5c40\u5b8c\u5168\u76f8\u540c\uff0c\u4f46\u5404\u5361\u663e\u5b58\u4e2d\u5b58\u50a8\u7684\u5b9e\u9645\u6570\u636e\u53ef\u4ee5\u6709\u522b\u3002<\/strong> \u8fd9\u5c31\u662f\u201c\u4e00\u5f20\u56fe\u7eb8\uff0c\u516b\u5361\u5171\u4eab\u201d\u7684\u771f\u6b63\u542b\u4e49\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6709\u4e86\u8fd9\u4e2a\u6982\u5ff5\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u8fdb\u5165 DTensor \u7684\u8bbe\u8ba1\u7ec6\u8282\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e3a\u4ec0\u4e48\u9700\u8981\u7edf\u4e00\u7684\u5f20\u91cf\u62bd\u8c61<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u5f88\u5c11\u53ea\u8dd1\u5728\u4e00\u5f20 GPU \u4e0a\u3002\u65e0\u8bba\u662f\u4e3a\u4e86\u7f29\u77ed\u8bad\u7ec3\u65f6\u95f4\uff0c\u8fd8\u662f\u4e3a\u4e86\u88c5\u4e0b\u66f4\u5927\u7684\u6a21\u578b\uff0c\u591a\u5361\u534f\u540c\u90fd\u662f\u5e38\u6001\u3002\u591a\u5361\u8bad\u7ec3\u6700\u4e3b\u6d41\u7684\u6a21\u5f0f\u662f<strong>\u6570\u636e\u5e76\u884c<\/strong>\uff1a\u6bcf\u5f20\u5361\u90fd\u6301\u6709\u5b8c\u6574\u6a21\u578b\u7684\u526f\u672c\uff0c\u5404\u81ea\u5904\u7406\u4e0d\u540c\u7684\u6570\u636e\u5b50\u96c6\uff0c\u7b97\u51fa\u68af\u5ea6\u540e\u518d\u901a\u8fc7 AllReduce \u628a\u6240\u6709\u5361\u4e0a\u7684\u68af\u5ea6\u52a0\u548c\u5e73\u5747\uff0c\u4fdd\u8bc1\u5404\u5361\u53c2\u6570\u59cb\u7ec8\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8fd9\u4e2a\u6a21\u5f0f\u4e0b\uff0c\u6bcf\u5f20\u5361\u6267\u884c\u7684\u8ba1\u7b97\u56fe\u672c\u8d28\u4e0a\u662f\u540c\u4e00\u5f20\u56fe\uff0c\u53ea\u662f\u8f93\u5165\u6570\u636e\u4e0d\u540c\u3002\u65e2\u7136\u5982\u6b64\uff0c\u7406\u60f3\u60c5\u51b5\u5c31\u662f\uff1a<strong>\u540c\u4e00\u6bb5\u4ee3\u7801\u3001\u540c\u4e00\u5f20\u8ba1\u7b97\u56fe\u3001\u540c\u4e00\u79cd\u663e\u5b58\u5e03\u5c40\uff0c\u5728\u6240\u6709\u5361\u4e0a\u5171\u4eab\u6267\u884c<\/strong>\uff0c\u800c\u4e0d\u662f\u4e3a\u6bcf\u5f20\u5361\u5355\u72ec\u6784\u56fe\u3001\u5355\u72ec\u5206\u914d\u3001\u5355\u72ec\u8c03\u4f18\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u8fd9\u9700\u8981\u89e3\u51b3\u51e0\u4e2a\u5173\u952e\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\uff0c<strong>\u5730\u5740\u5fc5\u987b\u7a33\u5b9a<\/strong>\u3002CUDA Graph \u6355\u83b7\u8981\u6c42\u91cd\u653e\u65f6\u7684\u5f20\u91cf\u5730\u5740\u4e0e\u6355\u83b7\u65f6\u5b8c\u5168\u4e00\u81f4\uff1b\u5982\u679c\u6bcf\u5f20\u5361\u5404\u81ea\u52a8\u6001\u5206\u914d\uff0c\u5730\u5740\u5c31\u96be\u4ee5\u4fdd\u8bc1\u4e00\u81f4\uff0c\u56fe\u4e5f\u65e0\u6cd5\u590d\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\uff0c<strong>\u5e03\u5c40\u5fc5\u987b\u4e00\u81f4<\/strong>\u3002\u6240\u6709\u5361\u4e0a\u540c\u4e00\u4e2a\u903b\u8f91\u5f20\u91cf\u2014\u2014\u6bd4\u5982\u7b2c 3 \u5c42\u7684\u5377\u79ef\u6743\u91cd\u2014\u2014\u5fc5\u987b\u4f4d\u4e8e\u5404\u81ea\u663e\u5b58\u6c60\u7684\u76f8\u540c\u504f\u79fb\u5904\uff0c\u5426\u5219\u540c\u4e00\u6bb5 kernel \u4ee3\u7801\u65e0\u6cd5\u5728\u6240\u6709\u5361\u4e0a\u540c\u6837\u89e3\u6790\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e09\uff0c<strong>\u53d8\u4f53\u5fc5\u987b\u517c\u5bb9<\/strong>\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u53ef\u80fd\u9047\u5230\u6700\u540e\u4e00\u4e2a\u4e0d\u5b8c\u6574\u7684 batch\u3001\u6e10\u8fdb\u5f0f\u5206\u8fa8\u7387\u5207\u6362\u3001\u9a8c\u8bc1\u4e0e\u8bad\u7ec3\u4e0d\u540c\u5c3a\u5bf8\u7b49\u60c5\u51b5\u3002\u5982\u679c\u6bcf\u79cd\u60c5\u51b5\u90fd\u5bfc\u81f4\u504f\u79fb\u6f02\u79fb\uff0c\u90a3\u56fe\u6355\u83b7\u548c\u8de8\u5361\u901a\u4fe1\u90fd\u4f1a\u4e71\u5957\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u56db\uff0c<strong>\u901a\u4fe1\u5fc5\u987b\u53ef\u8868\u8fbe<\/strong>\u3002\u6570\u636e\u5e76\u884c\u6700\u7ec8\u8981\u843d\u5b9e\u5728 NCCL AllReduce \u8fd9\u7c7b\u96c6\u5408\u901a\u4fe1\u4e0a\uff1b\u6846\u67b6\u9700\u8981\u77e5\u9053\u201c\u54ea\u4e9b\u5b57\u8282\u5c5e\u4e8e\u54ea\u4e2a\u5f20\u91cf\u3001\u54ea\u4e9b\u5f20\u91cf\u9700\u8981\u88ab\u4e00\u8d77\u5f52\u7ea6\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u4e3b\u6d41\u6846\u67b6\u91cc\uff0c\u8fd9\u4e9b\u95ee\u9898\u7684\u89e3\u6cd5\u5404\u4e0d\u76f8\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>PyTorch DTensor<\/strong> \u662f\u4e00\u4e2a\u57fa\u4e8e Placement \u7684\u5f20\u91cf\u62bd\u8c61\u3002\u5b83\u901a\u8fc7 <code>DeviceMesh<\/code>\u3001<code>Shard(dim)<\/code>\u3001<code>Replicate()<\/code>\u3001<code>Partial()<\/code> \u7b49\u6982\u5ff5\u63cf\u8ff0\u5f20\u91cf\u5728\u591a\u4e2a\u8bbe\u5907\u4e0a\u7684\u5206\u5e03\u65b9\u5f0f\uff0c\u7136\u540e\u7531\u6846\u67b6\u81ea\u52a8\u8fdb\u884c\u5206\u7247\u4f20\u64ad\u548c\u901a\u4fe1\u751f\u6210\u3002\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u901a\u7528\u7684\u8bbe\u8ba1\uff0c\u53ef\u4ee5\u652f\u6301\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\u7b49\u591a\u79cd\u7b56\u7565\uff0c\u4f46\u76f8\u5e94\u5730\uff0c\u5b83\u7684\u5b9e\u73b0\u590d\u6742\u5ea6\u9ad8\uff0c\u8fd0\u884c\u65f6\u4e5f\u9700\u8981\u505a\u5927\u91cf\u7684\u5206\u7247\u8ba1\u7b97\u548c\u8c03\u5ea6\u51b3\u7b56\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>JAX<\/strong> \u901a\u8fc7 <code>pjit<\/code> \u548c <code>NamedSharding<\/code> \u628a\u540c\u6837\u7684\u601d\u60f3\u505a\u8fdb\u4e86\u51fd\u6570\u5f0f\u7f16\u8bd1\u7ba1\u7ebf\uff1a\u7528\u6237\u7528 <code>jax.device_put<\/code> \u6216 <code>with_sharding_constraint<\/code> \u6807\u6ce8\u5f20\u91cf\u5982\u4f55\u653e\u7f6e\uff0cXLA \u5728\u7f16\u8bd1\u65f6\u51b3\u5b9a\u901a\u4fe1\u5e76\u751f\u6210 SPMD \u7a0b\u5e8f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>TensorFlow DTensor<\/strong> \u4e5f\u7c7b\u4f3c\uff0c\u56f4\u7ed5 <code>Layout<\/code> \u548c <code>Mesh<\/code> \u6784\u5efa\uff0c\u5f3a\u8c03\u8de8\u6846\u67b6\u7684\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u65b9\u6848\u7684\u5171\u540c\u70b9\u662f\uff1a<strong>\u628a\u201c\u5f20\u91cf\u5982\u4f55\u5206\u5e03\u5230\u8bbe\u5907\u201d\u4f5c\u4e3a\u4e00\u7b49\u516c\u6c11<\/strong>\u3002\u5b83\u4eec\u5f3a\u5927\u3001\u901a\u7528\uff0c\u4f46\u901a\u7528\u6027\u672c\u8eab\u4e5f\u5e26\u6765\u4e86\u5f00\u9500\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u9009\u62e9\u5219\u66f4\u4e3a\u76f4\u63a5\uff1a\u6211\u4eec\u5f53\u524d\u805a\u7126\u7684\u662f<strong>\u540c\u8d28 GPU \u96c6\u7fa4\u4e0a\u7684\u9ad8\u901f\u6570\u636e\u5e76\u884c\u8bad\u7ec3<\/strong>\u3002\u5728\u8fd9\u4e2a\u573a\u666f\u4e0b\uff0c\u6a21\u578b\u53ef\u4ee5\u653e\u8fdb\u5355\u5f20 GPU\uff0c\u6bcf\u5f20\u5361\u6301\u6709\u5b8c\u6574\u526f\u672c\uff0c\u901a\u4fe1\u4e3b\u4f53\u53ea\u6709\u68af\u5ea6 AllReduce \u548c BN \u7edf\u8ba1\u91cf\u540c\u6b65\u3002\u4e8e\u662f\u6211\u4eec\u4e0d\u9700\u8981\u4e00\u4e2a\u5b8c\u6574\u7684\u5206\u7247\/\u5e03\u5c40\u7cfb\u7edf\uff1b\u6211\u4eec\u9700\u8981\u7684\u662f\u4e00\u4e2a<strong>\u8de8\u5361\u4e00\u81f4\u7684\u5185\u5b58\u89c6\u56fe\u63cf\u8ff0\u7b26<\/strong>\u2014\u2014\u8fd9\u5c31\u662f DTensor\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001DTensor\uff1a\u7eaf\u63cf\u8ff0\u7b26\uff0c\u4e0d\u6301\u6709\u5185\u5b58<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>include\/renaissance\/tensor\/distributed_tensor.h<\/code> \u4e2d\uff0cDTensor \u88ab\u5b9a\u4e49\u4e3a\u4e00\u4e2a\u975e\u5e38\u8f7b\u91cf\u7684\u7ed3\u6784\u4f53\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=\"\">struct DistributedTensor {\n    int32_t  id      = -1;\n    Shape    shape;                   \/\/ [N, H, W, C] \u903b\u8f91\u7ef4\u5ea6\n    DType    dtype   = DType::FP32;\n    Region   region  = Region::DEFAULT;\n\n    int32_t  n_ = 1, h_ = 1, w_ = 1, c_ = 1;\n    InitConfig init_config;\n\n    int64_t numel() const noexcept;\n    uint64_t nbytes() const noexcept;\n    uint64_t padded_bytes() const noexcept;\n    uint64_t slot_bytes() const noexcept;\n\n    bool valid() const noexcept;\n    bool is_compact() const noexcept;\n\n    uint8_t cuda_alignment() const noexcept;\n    int64_t padded_c() const noexcept;\n\n    static uint64_t compute_slot_bytes(const Shape&amp; shape, DType dtype, Region region) noexcept;\n    uint64_t offset() const;\n\n    DistributedTensor(int32_t i, Shape s, DType d, Region r);\n    DistributedTensor(int32_t i, Shape s, DType d, Region r, uint64_t sb);\n};\n\nusing DTensor = DistributedTensor;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u5b83\u7684\u6838\u5fc3\u7279\u5f81\uff1a<strong>DTensor \u662f\u4e00\u4e2a\u7eaf\u865a\u62df\u6982\u5ff5\uff0c\u53ea\u5b58\u5f62\u72b6\u3001\u504f\u79fb\u91cf\u3001stride \u548c\u5185\u5b58\u533a\u57df\u4fe1\u606f\uff0c\u4e0d\u6301\u6709\u4efb\u4f55\u5b9e\u9645\u5185\u5b58\uff0c\u4e5f\u4e0d\u4fdd\u5b58\u8bbe\u5907\u6307\u9488\u3002<\/strong> \u540c\u4e00\u4e2a DTensor \u63cf\u8ff0\u7b26\u5728\u6240\u6709\u5361\u4e0a\u5bf9\u5e94\u76f8\u540c\u7684\u504f\u79fb\u4f4d\u7f6e\uff0c\u5e03\u5c40\u5b8c\u5168\u76f8\u540c\uff0c\u4f46\u5404\u5361\u663e\u5b58\u4e2d\u5b58\u50a8\u7684\u5b9e\u9645\u6570\u636e\u53ef\u4ee5\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8bbe\u8ba1\u628a\u201c\u5f20\u91cf\u662f\u4ec0\u4e48\u201d\u548c\u201c\u5f20\u91cf\u5b58\u5728\u54ea\u91cc\u201d\u89e3\u8026\u4e86\u3002DTensor \u63cf\u8ff0\u7684\u662f\u201c\u56fe\u7eb8\u4e0a\u7684\u4f4d\u7f6e\u201d\u2014\u2014\u5b83\u6709 id\u3001shape\u3001dtype\u3001region\u3001offset\u3001slot_bytes\uff1b\u800c\u771f\u6b63\u7684\u7269\u7406\u5185\u5b58\u7531\u540e\u7aef\u7684 <code>ArenaKeeper<\/code> \u7ba1\u7406\uff0c\u6bcf\u5f20\u5361\u6709\u4e00\u5757\u72ec\u7acb\u7684\u663e\u5b58\u6c60\u3002\u8fd0\u884c\u65f6\uff0c<code>DeviceContext::ptr_at(dtensor_id)<\/code> \u624d\u628a DTensor \u89e3\u6790\u4e3a\u5f53\u524d rank \u4e0a\u7684\u5b9e\u9645\u6307\u9488\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5bf9 DTensor \u6709\u4e00\u4e2a\u6838\u5fc3\u7ea6\u5b9a\uff1a<strong>\u6240\u6709 rank \u7684\u5f20\u91cf\u884c\u4e3a\u5fc5\u987b\u540c\u6b65\u2014\u2014\u53ea\u5141\u8bb8\u6570\u636e\u4e0d\u540c\uff0c\u6267\u884c\u7684\u6240\u6709\u64cd\u4f5c\u5fc5\u987b\u5b8c\u5168\u76f8\u540c\u3002<\/strong> \u8fd9\u4e0d\u662f\u4e00\u4e2a\u8f6f\u6027\u7684\u201c\u7ea6\u5b9a\u201d\uff0c\u800c\u662f\u7531 DTensor \u7684\u67b6\u6784\u4ece\u6839\u672c\u4e0a\u4fdd\u8bc1\u7684\u786c\u6027\u7ea6\u675f\u3002\u56e0\u4e3a\u6240\u6709 rank \u5171\u4eab\u540c\u4e00\u4efd MemoryPlan \u548c\u540c\u4e00\u4efd ComputationGraph\uff0c\u6240\u4ee5\u201c\u5bf9 DTensor #42 \u505a\u5377\u79ef\uff0c\u7ed3\u679c\u5199\u5165 DTensor #43\u201d\u8fd9\u6761\u6307\u4ee4\u5728\u6240\u6709\u5361\u4e0a\u5b8c\u5168\u4e00\u81f4\uff1b\u552f\u4e00\u7684\u533a\u522b\u53ea\u662f <code>ptr_at()<\/code> \u8fd4\u56de\u7684\u7269\u7406\u5730\u5740\u6307\u5411\u4e0d\u540c GPU \u7684\u663e\u5b58\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e0e PyTorch \u7684 <code>torch.Tensor<\/code> \u6216 JAX \u7684 <code>DeviceArray<\/code> \u5f88\u4e0d\u4e00\u6837\u3002\u540e\u4e24\u8005\u901a\u5e38\u76f4\u63a5\u6301\u6709\u6570\u636e\u6216\u6307\u5411\u6570\u636e\u7684\u6307\u9488\uff0c\u5e76\u9644\u5e26\u8bbe\u5907\u4fe1\u606f\uff1b\u800c DTensor \u66f4\u50cf\u662f\u84dd\u56fe\u91cc\u7684\u4e00\u4e2a\u7f16\u53f7\u548c\u5750\u6807\u3002\u4f60\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u4e3a\u5efa\u7b51\u56fe\u7eb8\u4e0a\u7684\u201c\u7b2c 15 \u53f7\u67f1\u5b50\u5728 3 \u8f74\u4ea4 B \u8f74\u3001\u622a\u9762 400\u00d7400\u201d\uff0c\u81f3\u4e8e\u5de5\u5730\u73b0\u573a\u7b2c 15 \u53f7\u67f1\u5b50\u5177\u4f53\u7531\u54ea\u6839\u94a2\u7b4b\u6c34\u6ce5\u6d47\u7b51\uff0c\u90a3\u662f\u65bd\u5de5\u961f\u6309\u56fe\u7eb8\u53bb\u627e\u7684\u4e8b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6846\u67b6\u91cc\u8fd8\u6709\u4e00\u4e2a <code>Tensor<\/code> \u7c7b\uff08<code>include\/renaissance\/tensor\/tensor.h<\/code>\uff09\uff0c\u5b83\u4e13\u95e8\u8d1f\u8d23 CPU \u7aef\u7684\u6570\u636e\u642c\u8fd0\uff1a\u52a0\u8f7d\u6570\u636e\u96c6\u3001\u505a\u9884\u5904\u7406\u3001H2D \u4f20\u8f93\u524d\u7684\u6682\u5b58\u3002Tensor \u662f\u7d27\u51d1\u7684\u3001 owning \u7684\u5185\u5b58\u5bb9\u5668\uff1bDTensor \u5219\u4ece\u4e0d owning \u5185\u5b58\u3002\u4e24\u8005\u804c\u8d23\u6e05\u6670\uff1aTensor \u7ba1\u201c\u6570\u636e\u4ece\u54ea\u91cc\u6765\u201d\uff0cDTensor \u7ba1\u201c\u6570\u636e\u5728 GPU \u4e0a\u6309\u4ec0\u4e48\u89c4\u5219\u5b58\u653e\u201d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001MemoryPlan + ArenaKeeper\uff1a\u56fe\u7eb8\u5982\u4f55\u843d\u5730\u5230\u6bcf\u5f20\u5361<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u80fd\u8de8\u5361\u5171\u4eab\u7684\u5173\u952e\uff0c\u662f <strong>MemoryPlan \u5728\u6240\u6709 rank \u4e0a\u5b8c\u5168\u4e00\u81f4<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u7f16\u8bd1\u9636\u6bb5\uff0c<code>Compiler<\/code> \u4f1a\u4e3a\u6bcf\u4e2a\u8bad\u7ec3\u53d8\u4f53\u751f\u6210\u4e00\u4e2a <code>MemoryPlan<\/code>\u3002MemoryPlan \u901a\u8fc7\u8bed\u4e49\u5316\u63a5\u53e3\u5206\u914d DTensor\uff0c\u4f8b\u5982\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=\"\">DTensor MemoryPlan::alloc_fc_weight(const Shape&amp; shape);\nDTensor MemoryPlan::alloc_deep_conv_weight(const Shape&amp; shape);\nDTensor MemoryPlan::alloc_feature(const Shape&amp; shape, DType dtype);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u63a5\u53e3\u5185\u90e8\u90fd\u8c03\u7528 <code>alloc_impl<\/code>\uff0c\u7ed9\u5f20\u91cf\u5206\u914d\u4e00\u4e2a\u5168\u5c40\u552f\u4e00\u7684 <code>id<\/code>\uff0c\u5e76\u8ba1\u7b97\u5b83\u7684 <code>slot_bytes<\/code>\u3002\u5f53\u6240\u6709 DTensor \u90fd\u5206\u914d\u5b8c\u6bd5\u540e\uff0c\u8c03\u7528 <code>MemoryPlan::finalize()<\/code>\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 MemoryPlan::finalize() {\n    uint64_t cursor = 0;\n    for (size_t ri = 0; ri &lt; static_cast&lt;size_t>(Region::NUM_REGIONS); ++ri) {\n        auto&amp; info = region_infos_[ri];\n        info.base_offset = cursor;\n\n        for (int32_t dt_id : region_dt_ids_[ri]) {\n            auto&amp; entry = entries_[id_to_idx_.at(dt_id)];\n            entry.dt.offset_ = cursor;\n            cursor += entry.dt.slot_bytes();\n        }\n\n        info.total_bytes = cursor - info.base_offset;\n    }\n    total_bytes_ = cursor;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6bb5\u4ee3\u7801\u505a\u7684\u4e8b\u60c5\u5f88\u7b80\u5355\uff1a\u6309 Region \u987a\u5e8f\u7ebf\u6027\u904d\u5386\uff0c\u628a\u6bcf\u4e2a DTensor \u7684 <code>offset_<\/code> \u8bbe\u4e3a\u5f53\u524d cursor\uff0c\u7136\u540e cursor \u524d\u8fdb <code>slot_bytes()<\/code>\u3002\u6700\u7ec8\u6240\u6709 DTensor \u7684\u504f\u79fb\u88ab\u4e00\u6b21\u6027\u786e\u5b9a\uff0c\u6ca1\u6709\u7a7a\u9699\u3001\u6ca1\u6709\u788e\u7247\u3001\u6ca1\u6709\u5bf9\u8fd0\u884c\u65f6\u5206\u914d\u5668\u7684\u4f9d\u8d56\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684 Region \u662f\u663e\u5b58\u7684\u8bed\u4e49\u5206\u533a\u3002\u5f53\u524d\u7248\u672c\u7684 <code>Region<\/code> \u679a\u4e3e\u5b9a\u4e49\u4e86 69 \u4e2a\u5206\u533a\uff08<code>NUM_REGIONS = 69<\/code>\uff09\uff0c\u6309 B\uff08BN \u7edf\u8ba1\u91cf\uff09\u3001W\uff08\u4e3b\u6743\u91cd\uff09\u3001E\uff08EMA \u6743\u91cd\uff09\u3001A\uff08AMP FP16 \u6743\u91cd\uff09\u3001G\uff08\u68af\u5ea6\uff09\u3001R\uff08\u7ed3\u679c\u533a\uff09\u3001M\uff08\u4e00\u9636\u52a8\u91cf\uff09\u3001V\uff08\u4e8c\u9636\u52a8\u91cf\uff09\u3001N\uff08LARS \u8303\u6570\uff09\u3001I\uff08\u8f93\u5165\u7f13\u51b2\uff09\u3001F\uff08\u7279\u5f81\u56fe\u4e0e\u68af\u5ea6\u69fd\uff09\u3001S\uff08\u6807\u91cf\u4e0e\u63a9\u7801\uff09\u3001T\uff08\u4e34\u65f6\u5f20\u91cf\uff09\u7b49\u5b57\u6bcd\u7cfb\u5217\u7ec4\u7ec7\u3002\u6bcf\u4e2a\u7cfb\u5217\u5185\u90e8\u518d\u6309\u5f20\u91cf\u7c7b\u578b\u7ec6\u5206\uff0c\u4f8b\u5982 W \u7cfb\u5217\u5305\u542b <code>W_BN_BIAS<\/code>\u3001<code>W_BN_WEIGHT<\/code>\u3001<code>W_FC_BIAS<\/code>\u3001<code>W_FC_WEIGHT<\/code>\u3001<code>W_FIRST_CONV<\/code>\u3001<code>W_DEEP_CONV<\/code> \u7b49\u3002\u8fd9\u79cd\u547d\u540d\u4e0d\u662f\u88c5\u9970\uff0c\u800c\u662f\u76f4\u63a5\u670d\u52a1\u4e8e RANGE OP\uff1a\u5f53\u6846\u67b6\u9700\u8981\u201c\u628a FP32 \u4e3b\u6743\u91cd\u5168\u90e8\u8f6c\u6210 FP16\u201d\u6216\u201c\u628a\u6df1\u5c42\u5377\u79ef\u68af\u5ea6\u505a\u4e00\u6b21 AllReduce\u201d\u65f6\uff0c\u7f16\u8bd1\u5668\u53ea\u9700\u7ed9\u51fa\u8d77\u59cb Region \u548c\u7ed3\u675f Region\uff0c\u5c31\u80fd\u8986\u76d6\u6574\u4e2a\u8fde\u7eed\u533a\u95f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cd\u8981\u7684\u662f\uff1a<strong>\u6bcf\u4e2a rank \u62ff\u5230\u7684\u662f\u540c\u4e00\u4efd MemoryPlan \u7684\u526f\u672c<\/strong>\u2014\u2014\u540c\u4e00\u5f20\u56fe\u7eb8\u3002\u4e8e\u662f\u7b2c 15 \u53f7 DTensor \u5728 rank 0 \u4e0a\u4f4d\u4e8e offset X\uff0c\u5728 rank 7 \u4e0a\u4e5f\u4f4d\u4e8e offset X\uff0c\u53ea\u662f\u5b83\u4eec\u5bf9\u5e94\u7684\u7269\u7406\u663e\u5b58\u6c60\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7269\u7406\u663e\u5b58\u6c60\u7531 <code>ArenaKeeper<\/code> \u7ba1\u7406\u3002\u521d\u59cb\u5316\u65f6\uff0c\u6846\u67b6\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=\"\">ArenaKeeper::initialize(true, gpu_ids, total_bytes_per_device);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u6bcf\u5f20 GPU \u72ec\u7acb <code>cudaMalloc<\/code> \u4e00\u5757\u5927\u5c0f\u5b8c\u5168\u76f8\u540c\u7684\u663e\u5b58\u3002\u8fd0\u884c\u65f6\uff0c<code>DeviceContext::ptr_at()<\/code> \u7684\u4ee3\u7801\u53ea\u6709\u4e24\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=\"\">void* DeviceContext::ptr_at(int dtensor_id) const noexcept {\n    const DTensor&amp; dt = current_mp_->get_dtensor(dtensor_id);\n    return ArenaKeeper::instance().ptr_at(rank_for_context_, static_cast&lt;size_t>(dt.offset()));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ArenaKeeper::ptr_at<\/code> \u66f4\u7eaf\u7cb9\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* ArenaKeeper::ptr_at(int rank, size_t offset) const noexcept {\n    return static_cast&lt;char*>(arenas_[rank]->base_ptr()) + offset;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u70ed\u8def\u5f84\u4e0a\u5c31\u662f\u4e00\u6b21\u57fa\u5740\u52a0\u504f\u79fb\u3002<strong>\u6ca1\u6709\u54c8\u5e0c\u67e5\u627e\uff0c\u6ca1\u6709\u8fd0\u884c\u65f6\u5206\u914d\uff0c\u6ca1\u6709\u8de8 rank \u7684\u5206\u7247\u8ba1\u7b97<\/strong>\u3002\u8fd9\u6b63\u662f\u6570\u636e\u5e76\u884c\u573a\u666f\u4e0b\u6700\u5e72\u51c0\u7684\u62bd\u8c61\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u8de8\u53d8\u4f53\u4e00\u81f4\u6027\uff1aslot_bytes \u4e0e stride \u7684\u5206\u79bb<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u4e2d\u5e38\u5e38\u9700\u8981\u5207\u6362\u5f62\u72b6\u53d8\u4f53\u3002\u4f8b\u5982\u6700\u540e\u4e00\u4e2a batch \u7684\u5927\u5c0f\u53ef\u80fd\u4e0d\u662f 128 \u800c\u662f 37\uff1b\u6e10\u8fdb\u5f0f\u5206\u8fa8\u7387\u8bad\u7ec3\u53ef\u80fd\u5148\u7528 224\u00d7224 \u518d\u7528 288\u00d7288\uff1b\u9a8c\u8bc1\u65f6\u901a\u5e38\u4e0d\u7528\u6570\u636e\u589e\u5f3a\u3001batch \u5c3a\u5bf8\u4e5f\u53ef\u80fd\u4e0d\u540c\u3002\u5982\u679c\u6bcf\u79cd\u53d8\u4f53\u90fd\u5bfc\u81f4 DTensor \u504f\u79fb\u53d8\u5316\uff0c\u90a3 CUDA Graph \u6355\u83b7\u5c31\u5f97\u6bcf\u79cd\u53d8\u4f53\u91cd\u65b0\u505a\u4e00\u904d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u89e3\u6cd5\u662f\uff1a<strong>\u8de8\u53d8\u4f53\u53d6\u6700\u5927 slot_bytes\uff0c\u4f46 stride \u4ecd\u6309\u5404\u81ea shape \u8ba1\u7b97<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u6709\u4e24\u4e2a\u6784\u9020\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=\"\">\/\/ \u6807\u51c6\u6784\u9020\uff1aslot_bytes \u4ece shape \u63a8\u5bfc\nDistributedTensor(int32_t i, Shape s, DType d, Region r);\n\n\/\/ \u53d8\u4f53\u6784\u9020\uff1aslot_bytes \u663e\u5f0f\u4f20\u5165\nDistributedTensor(int32_t i, Shape s, DType d, Region r, uint64_t sb);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>Compiler<\/code> \u7684\u7b2c\u4e8c\u9636\u6bb5\uff0c\u6846\u67b6\u4f1a\u5bf9\u6bcf\u4e2a (layer, tensor) \u4f4d\u7f6e\u8de8\u6240\u6709\u53d8\u4f53\u8ba1\u7b97\u6700\u5927 slot \u5b57\u8282\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=\"\">uint64_t max_slot = DTensor::compute_slot_bytes(shape_max, dtype, region);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u540e\u5728\u7b2c\u4e09\u9636\u6bb5\u4e3a\u6bcf\u4e2a\u53d8\u4f53\u521b\u5efa MemoryPlan \u65f6\uff0c\u7528\u540c\u4e00\u4e2a <code>max_slot<\/code> \u53bb\u6784\u9020 DTensor\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=\"\">DTensor dt = alloc_impl(shape_variant, dtype, region, max_slot);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u505a\u7684\u7ed3\u679c\u662f\uff1a<strong>\u4e0d\u540c\u53d8\u4f53\u4e0b\u540c\u4e00\u4e2a DTensor id \u7684 offset \u5b8c\u5168\u76f8\u540c<\/strong>\uff0c\u56e0\u4e3a <code>MemoryPlan::finalize()<\/code> \u53ea\u8ba4 <code>slot_bytes()<\/code>\u3002\u4e0e\u6b64\u540c\u65f6\uff0cDTensor \u7684 stride \u662f\u5728\u6784\u9020\u65f6\u6839\u636e\u81ea\u5df1\u7684 <code>shape<\/code> \u7b97\u51fa\u6765\u7684\uff0c\u6240\u4ee5\u6bcf\u4e2a\u53d8\u4f53\u8fd0\u884c\u65f6\u4f7f\u7528\u7684 stride \u4ecd\u7136\u662f\u6b63\u786e\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f\u4e00\u79cd\u975e\u5e38\u7cbe\u5de7\u7684\u89e3\u8026\uff1a<strong>slot_bytes \u51b3\u5b9a\u201c\u56fe\u7eb8\u4e0a\u7684\u5730\u76d8\u5927\u5c0f\u201d\uff0c\u4fdd\u8bc1\u504f\u79fb\u4e00\u81f4\uff1bstride \u51b3\u5b9a\u201c\u5b9e\u9645\u600e\u4e48\u8d70\u8def\u5f84\u201d\uff0c\u4fdd\u8bc1\u6570\u636e\u8bbf\u95ee\u6b63\u786e<\/strong>\u3002\u4e24\u8005\u4e92\u4e0d\u5e72\u6270\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001NHWC\u3001\u5bf9\u9f50\u4e0e padding \u7684\u5de5\u7a0b\u7ec6\u8282<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u7684 stride \u8ba1\u7b97\u4e0d\u662f\u7b80\u5355\u5730\u628a <code>shape.c()<\/code> \u5f53\u4f5c\u6700\u5185\u5c42\u7ef4\u5ea6\uff0c\u800c\u662f\u8981\u6839\u636e <code>cuda_alignment()<\/code> \u505a C \u901a\u9053 padding\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=\"\">uint8_t cuda_alignment() const noexcept {\n    if (dtype == DType::FP16) {\n        switch (region) {\n            case Region::I_A_DATA:\n            case Region::I_B_DATA:  return 4;\n            case Region::F_FEATURE_FP16:\n            case Region::F_GRAD_SLOT_FP16:\n                return 8;\n            default: return 1;\n        }\n    }\n    if (dtype == DType::INT8 &amp;&amp; region == Region::S_MASK) {\n        return 8;\n    }\n    return 1;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>padded_c()<\/code> \u628a\u539f\u59cb C \u901a\u9053\u5411\u4e0a\u5bf9\u9f50\u5230 <code>cuda_alignment<\/code> \u7684\u500d\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=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   C = shape.c()              \/\/ \u539f\u59cb\u901a\u9053\u6570\n\/\/   A = cuda_alignment()       \/\/ \u5bf9\u9f50\u56e0\u5b50\uff081\/4\/8\uff09\n\nint64_t padded_c = align_up(C, A);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e FP16 \u8f93\u5165\u7f13\u51b2\u533a\uff0cC \u901a\u9053\u6309 4 \u5bf9\u9f50\uff1b\u5bf9\u4e8e FP16 \u7279\u5f81\u56fe\u548c\u68af\u5ea6\u69fd\uff0c\u6309 8 \u5bf9\u9f50\u3002\u8fd9\u662f\u4e3a\u4e86\u5145\u5206\u5229\u7528 Tensor Core \u7684\u5411\u91cf\u5316\u8bbf\u95ee\u6a21\u5f0f\uff1a\u5f53\u901a\u9053\u6570\u662f 8 \u7684\u500d\u6570\u65f6\uff0c\u534a\u7cbe\u5ea6\u5377\u79ef\u548c\u77e9\u9635\u4e58\u53ef\u4ee5\u66f4\u9ad8\u6548\u5730\u8bfb\u53d6\u548c\u5199\u5165\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>slot_bytes()<\/code> \u7684\u8ba1\u7b97\u516c\u5f0f\u4e5f\u5f88\u6709\u8bb2\u7a76\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=\"\">static uint64_t compute_slot_bytes(const Shape&amp; shape, DType dtype, Region region) noexcept {\n    \/\/ ... alignment \u63a8\u65ad ...\n    int64_t pc = align_up(static_cast&lt;int64_t>(shape.c()), alignment);\n    uint64_t elems = static_cast&lt;uint64_t>(shape.n()) * shape.h() * shape.w() * pc;\n\n    if (dtype == DType::FP16) {\n        return utils::align_up_256(elems * 2 + 16);\n    } else if (dtype == DType::INT8) {\n        return utils::align_up_256(elems * 1 + 16);\n    } else {\n        return 2 * utils::align_up_256(elems * 2 + 16);  \/\/ FP32 = 2 \u00d7 FP16\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u51e0\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u8bbe\u8ba1\u70b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e00\uff0c\u672b\u5c3e\u9884\u7559 16 \u5b57\u8282\u3002<\/strong> \u8fd9\u4e0d\u662f\u6d6a\u8d39\uff0c\u800c\u662f\u4e3a\u4e86\u517c\u5bb9 XNNPACK \u7b49\u5e95\u5c42\u5e93\u7684\u8fb9\u754c\u8981\u6c42\u3002\u67d0\u4e9b kernel \u5728\u5904\u7406\u6700\u540e\u4e00\u884c\u65f6\u53ef\u80fd\u4f1a\u505a\u8d8a\u754c\u8bfb\u53d6\u6216\u5411\u91cf\u5316\u5b58\u53d6\uff0c\u9884\u7559 16 \u5b57\u8282\u53ef\u4ee5\u907f\u514d\u975e\u6cd5\u8bbf\u95ee\u98ce\u9669\uff0c\u540c\u65f6\u4e0d\u5f71\u54cd\u903b\u8f91\u6570\u636e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e8c\uff0cslot \u5927\u5c0f\u5411\u4e0a\u53d6\u6574\u5230 256 \u5b57\u8282\u3002<\/strong> \u8fd9\u4e0e\u6846\u67b6\u201c\u9996\u5730\u5740 256 \u5b57\u8282\u5bf9\u9f50\u201d\u7684\u539f\u5219\u4e00\u81f4\uff0c\u786e\u4fdd\u6240\u6709\u5f20\u91cf\u7684\u8d77\u59cb\u5730\u5740\u6ee1\u8db3 GPU \u7684\u5bf9\u9f50\u8981\u6c42\uff0c\u4e5f\u6709\u5229\u4e8e\u5185\u5b58\u5408\u5e76\u8bbf\u95ee\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e09\uff0cFP32 \u7684 slot \u662f FP16 \u7b49\u6548 slot \u7684\u4e24\u500d\u3002<\/strong> \u8fd9\u4e0d\u662f\u968f\u610f\u5b9a\u7684\u3002\u6846\u67b6\u5728 AMP \u8bad\u7ec3\u65f6\u9700\u8981\u5728 FP32 \u4e3b\u6743\u91cd\u548c FP16 \u8ba1\u7b97\u6743\u91cd\u4e4b\u95f4\u505a\u6574\u533a\u8f6c\u6362\uff1b\u5982\u679c FP32 \u533a\u57df\u7684\u603b\u5b57\u8282\u6570\u4e0e\u5bf9\u5e94 FP16 \u533a\u57df\u603b\u5b57\u8282\u6570\u59cb\u7ec8\u4fdd\u6301\u4e25\u683c\u7684 2:1 \u5173\u7cfb\uff0c\u90a3\u4e48\u4ece\u201cFP32 \u6743\u91cd\u533a\u201d\u5230\u201cFP16 \u6743\u91cd\u533a\u201d\u7684\u6279\u91cf\u7c7b\u578b\u8f6c\u6362\u5c31\u53ef\u4ee5\u7528\u4e00\u4e2a\u7b80\u5355\u7684 RANGE OP \u4e00\u6b21\u6027\u5b8c\u6210\u2014\u2014\u4e00\u4e2a kernel \u904d\u5386\u6240\u6709\u5c42\u7684\u6240\u6709\u53c2\u6570\uff0c\u6309\u56fa\u5b9a\u6bd4\u4f8b\u7f29\u653e\u5373\u53ef\u3002\u8fd9\u6bd4\u81ea\u5df1\u5199\u9010\u5c42\u8f6c\u6362\u5feb\u5f97\u591a\uff0c\u4e5f\u662f\u6846\u67b6\u901f\u5ea6\u5feb\u7684\u6839\u57fa\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u8fd8\u663e\u5f0f\u533a\u5206\u4e86 CUDA \u5bf9\u9f50 stride \u548c CPU \u7d27\u51d1 stride\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=\"\">int64_t n_stride_cuda() const noexcept;  \/\/ \u57fa\u4e8e padded_c\nint64_t n_stride_cpu() const noexcept;   \/\/ \u57fa\u4e8e shape.c()\uff0c\u6052\u7d27\u51d1<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CPU \u4e0a\u7684 DTensor \u6c38\u8fdc\u7d27\u51d1\uff0c\u4e0e <code>Tensor<\/code> \u7684\u5185\u5b58\u6392\u5e03\u4e00\u81f4\uff0c\u65b9\u4fbf H2D\/D2H \u4f20\u8f93\uff1bGPU \u4e0a\u5219\u5141\u8bb8 padding \u4ee5\u6362\u53d6\u5377\u79ef\u6548\u7387\u3002\u5f53 <code>is_compact()<\/code> \u8fd4\u56de false \u65f6\uff0c\u4f20\u8f93\u8def\u5f84\u4f1a\u505a\u5fc5\u8981\u7684 layout conversion\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u6784\u9020\u65f6\u540c\u65f6\u7b97\u51fa\u4e24\u5957 stride\u3002\u4ee5 CUDA \u5bf9\u9f50 stride \u4e3a\u4f8b\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=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   ac = padded_c()       \/\/ \u5bf9\u9f50\u540e\u7684 C \u901a\u9053\u6570\n\/\/   W  = shape.w()        \/\/ \u5bbd\u5ea6\n\/\/   H  = shape.h()        \/\/ \u9ad8\u5ea6\n\nint64_t c_stride_cuda_ = 1;\nint64_t w_stride_cuda_ = ac;\nint64_t h_stride_cuda_ = ac * W;\nint64_t n_stride_cuda_ = ac * W * H;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CPU \u7d27\u51d1 stride \u5219\u7528\u539f\u59cb <code>shape.c()<\/code> \u8ba1\u7b97\uff0c\u4e0d\u5f15\u5165 padding\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=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   C = shape.c()         \/\/ \u539f\u59cb\u901a\u9053\u6570\n\/\/   W = shape.w()\n\/\/   H = shape.h()\n\nint64_t c_stride_cpu_ = 1;\nint64_t w_stride_cpu_ = C;\nint64_t h_stride_cpu_ = C * W;\nint64_t n_stride_cpu_ = C * W * H;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u5bf9\u9f50\u89c4\u5219\u770b\u4f3c\u7410\u788e\uff0c\u5b9e\u5219\u5171\u540c\u670d\u52a1\u4e8e\u4e00\u4e2a\u76ee\u6807\uff1a<strong>\u8ba9\u6279\u91cf\u64cd\u4f5c\uff08RANGE OP\uff09\u6210\u4e3a\u53ef\u80fd<\/strong>\u3002DTensor \u7684 Region \u5206\u7c7b\u3001slot \u5bf9\u9f50\u3001FP32 \u4e0e FP16 \u4e4b\u95f4\u7684 2:1 \u5b57\u8282\u5173\u7cfb\uff0c\u90fd\u4e0d\u662f\u5b64\u7acb\u7684\u5de5\u7a0b\u7ec6\u8282\uff0c\u800c\u662f\u4e3a\u4e86\u8ba9\u7f16\u8bd1\u5668\u53ef\u4ee5\u7528\u4e00\u6b21 kernel launch \u8986\u76d6\u6574\u4e2a\u8bed\u4e49\u5206\u533a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5 AMP \u8bad\u7ec3\u7684\u6743\u91cd\u540c\u6b65\u4e3a\u4f8b\u3002\u6a21\u578b\u540c\u65f6\u7ef4\u62a4 FP32 \u4e3b\u6743\u91cd\uff08W \u7cfb\u5217 Region\uff09\u548c FP16 \u8ba1\u7b97\u6743\u91cd\uff08A \u7cfb\u5217 Region\uff09\u3002\u7531\u4e8e\u6bcf\u4e2a FP32 \u5f20\u91cf\u7684 <code>slot_bytes<\/code> \u6052\u7b49\u4e8e\u5176\u5bf9\u5e94 FP16 \u5f20\u91cf slot \u7684\u4e24\u500d\uff0c\u6240\u6709\u5c42\u7684\u4e3b\u6743\u91cd\u5728 W \u533a\u8fde\u7eed\u6392\u5217\uff0c\u6240\u6709\u5c42\u7684 AMP \u6743\u91cd\u5728 A \u533a\u8fde\u7eed\u6392\u5217\uff0cW \u533a\u603b\u5b57\u8282\u6570\u4e0e A \u533a\u603b\u5b57\u8282\u6570\u4e5f\u4fdd\u6301\u4e25\u683c\u7684 2:1 \u5173\u7cfb\u3002\u4e8e\u662f\u201cFP32\u2192FP16\u201d\u7684\u6574\u533a\u8f6c\u6362\u53ea\u9700\u8981\u4e00\u4e2a RANGE CAST \u7b97\u5b50\uff1a\u4e00\u4e2a kernel \u4ece W \u533a\u8d77\u70b9\u904d\u5386\u5230 W \u533a\u7ec8\u70b9\uff0c\u6309 2:1 \u6bd4\u4f8b\u628a\u6570\u636e\u5199\u5165 A \u533a\u3002\u540c\u7406\uff0c\u4f18\u5316\u5668\u66f4\u65b0\u6240\u6709 BN bias\u3001\u6240\u6709 Conv weight\u3001\u6240\u6709 FC weight\uff0c\u4e5f\u53ef\u4ee5\u6309 Region \u5206\u7ec4\uff0c\u7528 1~2 \u4e2a kernel \u5b8c\u6210\uff0c\u800c\u4e0d\u662f\u9010\u5c42\u904d\u5386\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6362\u53e5\u8bdd\u8bf4\uff0cDTensor \u7684\u5bf9\u9f50\u4e0e\u5206\u533a\u89c4\u5219\uff0c\u662f\u628a\u201c\u6a21\u578b\u6709\u591a\u5c11\u5c42\u201d\u8fd9\u4e2a\u590d\u6742\u5ea6\u4ece\u8fd0\u884c\u65f6\u62b9\u6389\u7684\u5173\u952e\u3002\u65e0\u8bba\u6a21\u578b\u662f 19 \u5c42\u7684 VGG16 \u8fd8\u662f 152 \u5c42\u7684 ResNet\uff0c\u53ea\u8981\u540c\u4e00\u7c7b\u5f20\u91cf\u843d\u5728\u540c\u4e00\u4e2a Region \u91cc\uff0c\u6279\u91cf\u64cd\u4f5c\u7684\u5f00\u9500\u5c31\u4e0d\u4f1a\u968f\u5c42\u6570\u7ebf\u6027\u589e\u957f\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001DTensor \u5982\u4f55\u652f\u6491\u5206\u5e03\u5f0f\u56fe\u6267\u884c<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u4e0d\u53ea\u662f\u5185\u5b58\u63cf\u8ff0\u7b26\uff0c\u5b83\u4e5f\u662f\u8ba1\u7b97\u56fe\u91cc\u7684\u201c\u5730\u5740\u7b26\u53f7\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>ComputationGraph<\/code> \u4e2d\uff0c\u6bcf\u4e2a\u7b97\u5b50\u8282\u70b9\u53ea\u4fdd\u5b58 DTensor \u7684 id\uff0c\u4e0d\u4fdd\u5b58\u6307\u9488\uff0c\u4e5f\u4e0d\u4fdd\u5b58\u5f62\u72b6\u3002\u4f8b\u5982\u4e00\u4e2a\u5377\u79ef\u8282\u70b9\u4f1a\u8bb0\u5f55 <code>input_ids = {3, 5}<\/code>\u3001<code>output_ids = {7}<\/code>\uff1b\u8fd0\u884c\u65f6\uff0c<code>DeviceContext::ptr_at(3)<\/code>\u3001<code>ptr_at(5)<\/code>\u3001<code>ptr_at(7)<\/code> \u628a\u8fd9\u4e9b id \u89e3\u6790\u4e3a\u5f53\u524d rank \u4e0a\u7684\u5b9e\u9645\u6307\u9488\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e\u6240\u6709 rank \u7684 MemoryPlan \u76f8\u540c\uff0c\u540c\u4e00\u4e2a DTensor id \u5728\u6240\u6709 rank \u4e0a\u89e3\u6790\u51fa\u7684\u504f\u79fb\u76f8\u540c\uff1b\u7531\u4e8e <code>ArenaKeeper<\/code> \u4e3a\u6bcf\u5f20\u5361\u5206\u914d\u4e86\u72ec\u7acb\u7684\u6c60\uff0c\u89e3\u6790\u51fa\u7684\u6307\u9488\u6307\u5411\u5404\u81ea\u5361\u4e0a\u7684\u6570\u636e\u3002\u4e8e\u662f<strong>\u540c\u4e00\u5f20\u8ba1\u7b97\u56fe\u53ef\u4ee5\u5728\u6240\u6709 rank \u4e0a\u65e0\u4fee\u6539\u5730\u6267\u884c<\/strong>\uff0c\u8fd9\u6b63\u662f\u6570\u636e\u5e76\u884c\u8bad\u7ec3\u7684\u57fa\u7840\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u96c6\u5408\u901a\u4fe1\uff0cDTensor \u7684 Region \u8bed\u4e49\u8fdb\u4e00\u6b65\u53d1\u6325\u4e86\u4f5c\u7528\u3002\u68af\u5ea6 AllReduce \u4e0d\u662f\u9010\u5f20\u91cf\u505a\u7684\uff0c\u800c\u662f\u6309 Region \u6279\u91cf\u505a\u7684\u3002\u5728 <code>src\/backend\/ops\/range\/allreduce_op.cpp<\/code> \u4e2d\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=\"\">static void launch_allreduce_cuda_impl(\n    const GraphNode&amp; node,\n    const MemoryPlan&amp; mp,\n    const DeviceContext&amp; ctx,\n    MultiStreamCaptureState&amp; state)\n{\n    cudaStream_t s = static_cast&lt;cudaStream_t>(ctx.stream(StreamKind::UPDATE));\n    \/\/ ...\n    for (size_t i = 0; i &lt; node.input_ranges.size(); ++i) {\n        auto [src_off, src_sz] = mp.resolve_region_bounds(\n            static_cast&lt;Region>(node.input_ranges[i].start_region_id),\n            static_cast&lt;Region>(node.input_ranges[i].end_region_id));\n        auto [dst_off, dst_sz] = mp.resolve_region_bounds(\n            static_cast&lt;Region>(node.output_ranges[i].start_region_id),\n            static_cast&lt;Region>(node.output_ranges[i].end_region_id));\n\n        void* src = ArenaKeeper::instance().ptr_at(ctx.rank_for_context(), src_off);\n        void* dst = ArenaKeeper::instance().ptr_at(ctx.rank_for_context(), dst_off);\n        size_t count = std::min(src_sz, dst_sz) \/ sizeof(float);\n\n        if (src != dst) cudaMemcpyAsync(dst, src, count * sizeof(float),\n                                        cudaMemcpyDeviceToDevice, s);\n\n        ncclAllReduce(dst, dst, count, ncclFloat32, ncclSum,\n                      static_cast&lt;ncclComm_t>(ctx.nccl_comm()), s);\n\n        if (do_mean &amp;&amp; world_size > 1) {\n            float inv = 1.0f \/ static_cast&lt;float>(world_size);\n            launch_tr_scale_fp32_kernel(static_cast&lt;float*>(dst), inv, count, s);\n        }\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc <code>GraphNode<\/code> \u7684 <code>input_ranges<\/code> \u548c <code>output_ranges<\/code> \u662f Region \u8303\u56f4\u3002<code>MemoryPlan::resolve_region_bounds<\/code> \u628a Region \u533a\u95f4\u89e3\u6790\u6210 <code>(offset, size)<\/code>\uff0c\u7136\u540e <code>ArenaKeeper::ptr_at<\/code> \u628a offset \u53d8\u6210\u5f53\u524d rank \u7684\u6307\u9488\u3002\u4e00\u6b21 <code>ncclAllReduce<\/code> \u53ef\u4ee5\u8986\u76d6\u6574\u4e2a\u68af\u5ea6\u6876\u2014\u2014\u4f8b\u5982 <code>G_DEEP_CONV<\/code> \u533a\u6240\u6709\u6df1\u5c42\u5377\u79ef\u68af\u5ea6\u3001\u6216 <code>G_FC_WEIGHT<\/code> \u533a\u6240\u6709\u5168\u8fde\u63a5\u6743\u91cd\u68af\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a DTensor \u7684 Region \u662f\u7f16\u8bd1\u671f\u786e\u5b9a\u7684\uff0c\u6240\u4ee5\u8fd9\u4e9b\u8303\u56f4\u5728\u56fe\u6355\u83b7\u9636\u6bb5\u5c31\u89e3\u6790\u597d\u4e86\uff1bCUDA Graph \u6355\u83b7\u65f6\uff0cNCCL \u8c03\u7528\u7684\u6307\u9488\u5730\u5740\u4e5f\u662f\u56fa\u5b9a\u7684\u3002\u8fd0\u884c\u65f6\u53ea\u9700\u8981 <code>cudaGraphLaunch<\/code>\uff0c\u4e0d\u9700\u8981\u518d\u505a\u4efb\u4f55 Region \u5230\u6307\u9488\u7684\u8f6c\u6362\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u4e0e\u4e3b\u6d41\u5f20\u91cf\u62bd\u8c61\u7684\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6709\u4eba\u53ef\u80fd\u4f1a\u95ee\uff1aTech-Renaissance \u7684 DTensor \u4e3a\u4ec0\u4e48\u4e0d\u652f\u6301 PyTorch \u90a3\u79cd <code>Shard(dim)<\/code> \/ <code>Replicate()<\/code> \/ <code>Partial()<\/code> \u7684\u901a\u7528\u5206\u7247\u8bed\u4e49\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b54\u6848\u4e0e\u5b9a\u4f4d\u6709\u5173\u3002PyTorch DTensor\u3001JAX \u7684 <code>NamedSharding<\/code>\u3001TensorFlow DTensor \u90fd\u628a\u201c\u5f20\u91cf\u5982\u4f55\u5206\u5e03\u5230\u8bbe\u5907\u201d\u4f5c\u4e3a\u4e00\u7b49\u516c\u6c11\uff0c\u76ee\u6807\u662f\u652f\u6301\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\u7b49\u591a\u79cd\u7b56\u7565\u3002\u8fd9\u79cd\u901a\u7528\u6027\u5f88\u5f3a\u5927\uff0c\u4f46\u8fd0\u884c\u65f6\u4e5f\u9700\u8981\u505a\u5206\u7247\u4f20\u64ad\u3001\u5e03\u5c40\u8ba1\u7b97\u548c\u901a\u4fe1\u751f\u6210\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5f53\u524d\u805a\u7126\u7684\u662f<strong>\u540c\u8d28 GPU \u96c6\u7fa4\u4e0a\u7684\u9ad8\u901f\u6570\u636e\u5e76\u884c\u8bad\u7ec3<\/strong>\uff1a\u6a21\u578b\u80fd\u653e\u8fdb\u5355\u5f20 GPU\uff0c\u6bcf\u5f20\u5361\u6301\u6709\u5b8c\u6574\u526f\u672c\uff0c\u901a\u4fe1\u6a21\u5f0f\u56fa\u5b9a\u4e3a\u201c\u672c\u5730\u7b97\u68af\u5ea6 \u2192 \u5168\u5c40 AllReduce \u2192 \u672c\u5730\u66f4\u65b0\u201d\u3002\u5728\u8fd9\u4e2a\u573a\u666f\u4e0b\uff0c\u6211\u4eec\u4e0d\u9700\u8981\u4e00\u4e2a\u5b8c\u6574\u7684 Placement \u7cfb\u7edf\uff1b\u6211\u4eec\u9700\u8981\u7684\u662f\u4e00\u4e2a<strong>\u8de8\u5361\u4e00\u81f4\u7684\u5185\u5b58\u89c6\u56fe\u63cf\u8ff0\u7b26<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5 Tech-Renaissance \u7684 DTensor \u662f\u4e00\u4e2a\u201c\u9762\u5411\u6570\u636e\u5e76\u884c\u7684\u5185\u5b58\u5e03\u5c40\u63cf\u8ff0\u7b26\u201d\uff0c\u800c\u4e0d\u662f\u201c\u9762\u5411\u901a\u7528\u5e76\u884c\u7684\u5206\u7247\u5f20\u91cf\u201d\u3002\u5b83\u7684\u4f18\u52bf\u5728\u4e8e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6307\u9488\u89e3\u6790\u6781\u7b80<\/strong>\uff1a\u57fa\u5740\u52a0\u504f\u79fb\uff0c\u4e00\u6b21\u52a0\u6cd5\u3002<\/li>\n\n\n\n<li><strong>\u56fe\u5171\u4eab\u65e0\u6b67\u4e49<\/strong>\uff1a\u540c\u4e00\u5f20 ComputationGraph \u53ef\u4ee5\u5728\u6240\u6709 rank \u4e0a\u590d\u7528\u3002<\/li>\n\n\n\n<li><strong>CUDA Graph \u53cb\u597d<\/strong>\uff1a\u5730\u5740\u7a33\u5b9a\u3001Region \u8303\u56f4\u9759\u6001\u786e\u5b9a\u3001\u901a\u4fe1\u53ef\u6355\u83b7\u3002<\/li>\n\n\n\n<li><strong>\u53d8\u4f53\u5207\u6362\u96f6\u504f\u79fb\u6f02\u79fb<\/strong>\uff1a\u8de8\u53d8\u4f53\u6700\u5927 slot \u4fdd\u8bc1\u540c\u4e00 id \u5728\u4e0d\u540c shape \u4e0b offset \u4e00\u81f4\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u4ee3\u4ef7\u662f<strong>\u4e0d\u652f\u6301\u6a21\u578b\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c<\/strong>\u3002\u5982\u679c\u672a\u6765\u8981\u628a\u5355\u4e2a\u5927\u6a21\u578b\u62c6\u5230\u591a\u5f20\u5361\u4e0a\uff0c\u8fd9\u4e2a\u62bd\u8c61\u5c31\u9700\u8981\u6269\u5c55\u3002\u4f46\u5c31\u5f53\u524d\u4efb\u52a1\u800c\u8a00\u2014\u2014\u8ba9\u80fd\u653e\u8fdb\u5355\u5361\u7684\u6a21\u578b\u5728\u516b\u5361\u4e0a\u8dd1\u5f97\u66f4\u5feb\u2014\u2014\u5b83\u5df2\u7ecf\u8db3\u591f\u597d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u53e6\u4e00\u4e2a\u89d2\u5ea6\u770b\uff0cTech-Renaissance \u7684 DTensor \u4e0e PyTorch \u7684 <code>torch.Tensor<\/code> \u4ee3\u8868\u4e86\u4e24\u79cd\u8bbe\u8ba1\u54f2\u5b66\u3002PyTorch \u7684 Tensor \u662f<strong>\u72b6\u6001\u578b\u5bf9\u8c61<\/strong>\uff1a\u5143\u6570\u636e\u548c\u6570\u636e\u7ed1\u5b9a\uff0c\u521b\u5efa\u5373\u5206\u914d\uff0c\u9500\u6bc1\u5373\u91ca\u653e\uff0c\u76f4\u89c2\u4f46\u628a\u201c\u5e03\u5c40\u89c4\u5212\u201d\u7559\u5728\u4e86\u8fd0\u884c\u65f6\u3002DTensor \u662f<strong>\u7eaf\u63cf\u8ff0\u7b26<\/strong>\uff1a\u4e0d\u6301\u6709\u5185\u5b58\uff0c\u7f16\u8bd1\u671f\u5373\u53ef\u786e\u5b9a offset\u3001stride\u3001slot_bytes\uff0c\u727a\u7272\u4e86\u4e00\u90e8\u5206\u76f4\u63a5\u53ef\u64cd\u4f5c\u6027\uff0c\u4f46\u6362\u6765\u4e86\u5e03\u5c40\u786e\u5b9a\u6027\u3001\u8de8\u5361\u4e00\u81f4\u6027\u548c\u6279\u91cf\u64cd\u4f5c\u6548\u7387\u3002\u4e24\u8005\u6ca1\u6709\u7edd\u5bf9\u7684\u4f18\u52a3\uff0c\u53ea\u662f\u4e3a\u4e0d\u540c\u7684\u6267\u884c\u6a21\u578b\u505a\u4e86\u4e0d\u540c\u7684\u53d6\u820d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">DTensor \u662f Tech-Renaissance \u91cc\u8fde\u63a5\u201c\u5355\u5361\u5185\u5b58\u89c4\u5212\u201d\u4e0e\u201c\u591a\u5361\u5206\u5e03\u5f0f\u6267\u884c\u201d\u7684\u6865\u6881\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u4e0d\u6301\u6709\u5185\u5b58\uff0c\u53ea\u63cf\u8ff0\u5185\u5b58\uff1aid\u3001shape\u3001dtype\u3001region\u3001offset\u3001slot_bytes\u3001stride\u3002MemoryPlan \u8d1f\u8d23\u5728\u6240\u6709 rank \u4e0a\u751f\u6210\u540c\u4e00\u4efd\u56fe\u7eb8\uff1bArenaKeeper \u4e3a\u6bcf\u5f20\u5361\u5206\u914d\u72ec\u7acb\u4f46\u5927\u5c0f\u76f8\u540c\u7684\u663e\u5b58\u6c60\uff1bDeviceContext \u5728\u8fd0\u884c\u65f6\u628a DTensor id \u89e3\u6790\u4e3a\u5f53\u524d rank \u7684\u5b9e\u9645\u6307\u9488\u3002\u56e0\u4e3a\u56fe\u7eb8\u4e00\u81f4\uff0c\u6240\u4ee5\u540c\u4e00\u5f20\u8ba1\u7b97\u56fe\u53ef\u4ee5\u516b\u5361\u5171\u4eab\uff1b\u56e0\u4e3a slot_bytes \u8de8\u53d8\u4f53\u53d6\u6700\u5927\uff0c\u6240\u4ee5\u4e0d\u540c batch \u5927\u5c0f\u3001\u4e0d\u540c\u5206\u8fa8\u7387\u5207\u6362\u65f6\u504f\u79fb\u4e0d\u4f1a\u6f02\u79fb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5957\u8bbe\u8ba1\u7684\u6838\u5fc3\u601d\u60f3\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e00\u53e5\u8bdd\uff1a<strong>\u4e00\u5f20\u56fe\u7eb8\uff0c\u591a\u5361\u5171\u4eab\uff1b\u6570\u636e\u4e0d\u540c\uff0c\u6267\u884c\u76f8\u540c\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u5c06\u8fdb\u5165\u8fd0\u884c\u65f6\u9636\u6bb5\uff0c\u5148\u8bb2 Tech-Renaissance \u7684\u591a\u6d41\u5e76\u53d1\u67b6\u6784\u2014\u2014\u4e09\u6761\u8ba1\u7b97\u6d41\u3001\u4e00\u6761\u4f20\u8f93\u6d41\u3001\u4e00\u6761\u66f4\u65b0\u6d41\u5982\u4f55\u534f\u540c\u5de5\u4f5c\uff0c\u628a\u6570\u636e\u642c\u8fd0\u3001\u68af\u5ea6\u901a\u4fe1\u548c\u8ba1\u7b97\u6700\u5927\u9650\u5ea6\u5730\u91cd\u53e0\u8d77\u6765\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\u5341\u516d \u4e0a\u4e00\u7bc7\u6211\u4eec\u8bb2\u4e86 Memory [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":703,"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-517","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\/517","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=517"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/517\/revisions"}],"predecessor-version":[{"id":718,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/517\/revisions\/718"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/703"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=517"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=517"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=517"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}