{"id":565,"date":"2026-07-08T03:15:02","date_gmt":"2026-07-07T19:15:02","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=565"},"modified":"2026-07-08T14:26:29","modified_gmt":"2026-07-08T06:26:29","slug":"nccl%e9%80%9a%e4%bf%a1%e4%b8%8e%e5%88%86%e5%b8%83%e5%bc%8f%e6%95%b0%e6%8d%ae%e5%b9%b6%e8%a1%8c","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/565\/","title":{"rendered":"(24) NCCL\u901a\u4fe1\u4e0e\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c"},"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\u56db<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u4f60\u628a ResNet-50 \u5728\u5355\u5f20 A100 \u4e0a\u63a8\u5230\u6781\u81f4\uff0c\u4ecd\u7136\u4f1a\u89c9\u5f97\u8bad\u7ec3\u4e00\u4e2a useful \u7684\u6a21\u578b\u592a\u6162\uff1b\u5f53\u4f60\u60f3\u5c1d\u8bd5\u66f4\u5927\u7684 batch size\u3001\u66f4\u9ad8\u7684\u5206\u8fa8\u7387\uff0c\u6216\u8005\u5e72\u8106\u5c31\u662f\u66f4\u5927\u7684\u6a21\u578b\uff0c\u5355\u5f20\u5361\u7684\u663e\u5b58\u548c\u7b97\u529b\u5f88\u5feb\u5c31\u4f1a\u6210\u4e3a\u5929\u82b1\u677f\u3002\u8fd9\u65f6\u5019\uff0c\u552f\u4e00\u80fd\u7ee7\u7eed\u653e\u5927\u8bad\u7ec3\u541e\u5410\u91cf\u7684\u529e\u6cd5\uff0c\u5c31\u662f\u8ba9\u591a\u5f20 GPU \u4e00\u8d77\u5e72\u6d3b\u3002\u8fd9\u5c31\u662f\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c\uff08Distributed Data Parallel\uff0cDDP\uff09\u8981\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5355\u5f20 GPU \u7684\u663e\u5b58\u548c\u7b97\u529b\u7ec8\u5f52\u6709\u9650\u3002\u5f53\u6a21\u578b\u8d8a\u6765\u8d8a\u5927\u3001\u6570\u636e\u8d8a\u6765\u8d8a\u591a\uff0c\u5c31\u5fc5\u987b\u8ba9\u591a\u5f20 GPU \u534f\u540c\u5de5\u4f5c\u3002\u5728\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4e2d\uff0cDDP \u662f\u6700\u5e38\u7528\u3001\u4e5f\u6700\u5bb9\u6613\u843d\u5730\u7684\u7b56\u7565\uff1a\u6bcf\u4e2a GPU \u6301\u6709\u5b8c\u6574\u7684\u6a21\u578b\u526f\u672c\uff0c\u5404\u81ea\u5904\u7406\u4e0d\u540c\u7684\u6570\u636e\u5b50\u96c6\uff0c\u72ec\u7acb\u5b8c\u6210\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\uff0c\u7136\u540e\u901a\u8fc7\u96c6\u5408\u901a\u4fe1\u628a\u6240\u6709 GPU \u4e0a\u7684\u68af\u5ea6\u52a0\u548c\u5e73\u5747\uff0c\u4fdd\u8bc1\u53c2\u6570\u59cb\u7ec8\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">DDP \u542c\u8d77\u6765\u7b80\u5355\u2014\u2014&#8221;\u7b97\u5b8c\u68af\u5ea6\uff0c\u505a\u4e2a AllReduce\uff0c\u518d\u66f4\u65b0\u53c2\u6570&#8221;\u3002\u4f46\u4e00\u4e2a\u597d\u7684\u5206\u5e03\u5f0f\u5b9e\u73b0\u8fdc\u4e0d\u6b62\u4e8e\u6b64\u3002\u5982\u679c\u628a\u901a\u4fe1\u548c\u8ba1\u7b97\u4e32\u884c\u6267\u884c\uff0c\u901a\u4fe1\u65f6\u95f4\u4f1a\u5b8c\u6574\u5730\u66b4\u9732\u5728\u8bad\u7ec3\u8017\u65f6\u4e2d\u3002\u5bf9\u4e8e CNN \u8bad\u7ec3\uff0c\u68af\u5ea6\u7684\u6570\u636e\u91cf\u4e0e\u53c2\u6570\u91cf\u76f8\u5f53\uff0c\u800c\u4e00\u6b21 ring AllReduce \u6bcf\u5f20\u5361\u8981\u6536\u53d1\u7ea6\u4e24\u500d\u4e8e\u68af\u5ea6\u5927\u5c0f\u7684\u6570\u636e\uff0c\u901a\u4fe1\u5f00\u9500\u5728\u591a\u5361\u573a\u666f\u4e0b\u4f1a\u8fc5\u901f\u51f8\u663e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c<strong>&#8220;\u597d\u7684\u5206\u5e03\u5f0f\u5b9e\u73b0\u4e0d\u662f\u7b97\u5b8c\u518d\u4f20\uff0c\u800c\u662f\u8fb9\u7b97\u8fb9\u4f20\uff0c\u5c3d\u91cf\u628a\u901a\u4fe1\u9690\u85cf\u5728\u8ba1\u7b97\u80cc\u540e&#8221;<\/strong>\u3002\u8fd9\u662f\u672c\u6587\u7684\u6838\u5fc3\u8bba\u70b9\uff0c\u4e5f\u662f Tech-Renaissance \u5206\u5e03\u5f0f\u901a\u4fe1\u8bbe\u8ba1\u7684\u6307\u5bfc\u601d\u60f3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6570\u636e\u5e76\u884c\u7684\u672c\u8d28\uff1a\u628a\u5927 batch \u62c6\u6210\u82e5\u5e72\u5c0f batch\uff0c\u518d\u8ba9\u68af\u5ea6\u4f1a\u5e08<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u5e76\u884c\u7684\u601d\u8def\u975e\u5e38\u6734\u7d20\uff1a\u628a\u6574\u4e2a\u5168\u5c40 batch \u5207\u6210\u82e5\u5e72\u4efd\uff0c\u6bcf\u4efd\u4ea4\u7ed9\u4e00\u5f20 GPU\uff1b\u6bcf\u5f20\u5361\u90fd\u4fdd\u5b58\u4e00\u4efd\u5b8c\u6574\u7684\u6a21\u578b\u526f\u672c\uff0c\u72ec\u7acb\u5b8c\u6210\u524d\u5411\u548c\u53cd\u5411\u8ba1\u7b97\uff1b\u7136\u540e\uff0c\u628a\u6240\u6709\u5361\u7b97\u51fa\u6765\u7684\u68af\u5ea6\u6c47\u603b\u3001\u53d6\u5e73\u5747\uff0c\u518d\u5404\u81ea\u66f4\u65b0\u81ea\u5df1\u7684\u90a3\u4efd\u53c2\u6570\u3002\u56e0\u4e3a\u5404\u5361\u7684\u8d77\u59cb\u53c2\u6570\u76f8\u540c\u3001\u66f4\u65b0\u516c\u5f0f\u76f8\u540c\u3001\u5b66\u4e60\u7387\u76f8\u540c\uff0c\u66f4\u65b0\u540e\u7684\u53c2\u6570\u4ecd\u7136\u4fdd\u6301\u4e00\u81f4\u3002\u4e8e\u662f\uff0cN \u5f20\u5361\u76f8\u5f53\u4e8e\u5e76\u884c\u5904\u7406\u4e86\u4e00\u4e2a N \u500d\u5927\u7684 batch\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684\u5173\u952e\u5728\u4e8e\u68af\u5ea6\u540c\u6b65\u3002\u5047\u8bbe\u7b2c <code>r<\/code> \u5f20\u5361\u4e0a\u7684\u672c\u5730\u68af\u5ea6\u662f <code>grad[r]<\/code>\uff0c\u90a3\u4e48\u5168\u5c40\u5e73\u5747\u68af\u5ea6\u53ef\u4ee5\u7528\u4e00\u6bb5\u7b80\u5355\u7684\u4f2a\u4ee3\u7801\u8868\u793a\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=\"\">\/\/ N        : world_size\uff0c\u53c2\u4e0e\u8bad\u7ec3\u7684 GPU \u6570\u91cf\n\/\/ grad[r]  : \u7b2c r \u5f20\u5361\u4e0a\u7684\u672c\u5730\u68af\u5ea6\uff08FP32\uff09\n\/\/ grad_sync: \u540c\u6b65\u540e\u7684\u5168\u5c40\u5e73\u5747\u68af\u5ea6\nTensor grad_sync = zeros_like(grad[0]);\nfor (int r = 0; r &lt; N; ++r) {\n    grad_sync += grad[r];            \/\/ \u5148\u5bf9\u6240\u6709\u5361\u7684\u68af\u5ea6\u6c42\u548c\n}\ngrad_sync \/= static_cast&lt;float>(N);  \/\/ \u518d\u9664\u4ee5\u5361\u6570\u53d6\u5e73\u5747<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a&#8221;\u6c42\u548c\u518d\u5e7f\u64ad\u5230\u6240\u6709\u5361&#8221;\u7684\u64cd\u4f5c\uff0c\u5728\u5206\u5e03\u5f0f\u8ba1\u7b97\u91cc\u53eb <strong>AllReduce<\/strong>\u3002\u53ea\u8981\u5e26\u5bbd\u8db3\u591f\uff0c\u5b83\u80fd\u8ba9 N \u5f20\u5361\u7684\u8bad\u7ec3\u5728\u6570\u5b66\u4e0a\u7b49\u4ef7\u4e8e\u5355\u5361\u8bad\u7ec3\u4e00\u4e2a N \u500d\u5927\u7684 batch\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001NCCL\uff1aGPU \u4e4b\u95f4\u7684\u9ad8\u901f\u516c\u8def<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 GPU \u4e0a\u505a AllReduce\uff0c\u6700\u6734\u7d20\u7684\u505a\u6cd5\u662f\u628a\u68af\u5ea6\u4ece\u663e\u5b58\u62f7\u5230 CPU \u5185\u5b58\uff0c\u5728 CPU \u4e0a\u6c42\u548c\uff0c\u518d\u62f7\u56de\u53bb\u3002\u8fd9\u79cd&#8221;\u7ed5\u7ecf CPU&#8221;\u7684\u65b9\u6848\u5728 PCIe \u5e26\u5bbd\u548c\u7f51\u7edc\u5ef6\u8fdf\u9762\u524d\u4f1a\u8fc5\u901f\u6210\u4e3a\u74f6\u9888\u3002NVIDIA \u7684 <strong>NCCL\uff08NVIDIA Collective Communications Library\uff09<\/strong> \u6b63\u662f\u4e3a\u4e86\u7ed5\u8fc7 CPU\u3001\u76f4\u63a5\u5728 GPU \u4e4b\u95f4\u505a\u96c6\u5408\u901a\u4fe1\u800c\u8bbe\u8ba1\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">NCCL \u521d\u59cb\u5316\u65f6\u4f1a\u505a\u51e0\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u62d3\u6251\u63a2\u6d4b<\/strong>\uff1a\u626b\u63cf PCIe\u3001NVLink\u3001\u7f51\u7edc\u7f51\u5361\uff0c\u753b\u51fa GPU \u4e4b\u95f4\u7684\u8fde\u63a5\u56fe\uff1b<\/li>\n\n\n\n<li><strong>\u901a\u9053\u6784\u5efa<\/strong>\uff1a\u6839\u636e\u62d3\u6251\u5efa\u7acb\u591a\u6761\u5e76\u884c ring \u6216 tree \u901a\u9053\uff0c\u628a SMs \u548c\u62f7\u8d1d\u5f15\u64ce\u6620\u5c04\u5230\u4e0d\u540c\u94fe\u8def\u4e0a\uff1b<\/li>\n\n\n\n<li><strong>\u4f20\u8f93\u8def\u5f84\u9009\u62e9<\/strong>\uff1a\u540c\u8282\u70b9\u5185\u4f18\u5148\u7528 NVLink \u6216 P2P\uff0c\u8de8\u8282\u70b9\u7528 InfiniBand \/ RoCE \u914d\u5408 GPUDirect RDMA\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e\u6bcf\u4e00\u6b21 AllReduce\uff0cNCCL \u4f1a\u6839\u636e\u6d88\u606f\u5927\u5c0f\u548c\u7f51\u7edc\u7c7b\u578b\u81ea\u52a8\u9009\u62e9\u7b97\u6cd5\u548c\u534f\u8bae\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Ring \u7b97\u6cd5<\/strong>\uff1a\u628a\u6570\u636e\u5207\u6210 N \u6bb5\uff0c\u6cbf\u73af\u5f62\u4f9d\u6b21\u505a reduce-scatter \u548c all-gather\uff0c\u9002\u5408\u5927\u6d88\u606f\u3001\u9ad8\u5e26\u5bbd\u573a\u666f\uff1b<\/li>\n\n\n\n<li><strong>Tree \u7b97\u6cd5<\/strong>\uff1a\u6309\u6811\u5f62\u7ed3\u6784\u5411\u4e0a\u5f52\u7ea6\u3001\u5411\u4e0b\u5e7f\u64ad\uff0clog(N) \u8df3\u5b8c\u6210\uff0c\u9002\u5408\u5c0f\u6d88\u606f\u6216\u8bbe\u5907\u6570\u5f88\u591a\u65f6\uff1b<\/li>\n\n\n\n<li><strong>NVLS \/ CollNet<\/strong>\uff1a\u5229\u7528 NVSwitch \u6216 InfiniBand SHARP \u505a\u7f51\u5185\u5f52\u7ea6\uff0c\u628a\u90e8\u5206\u8ba1\u7b97 offload \u5230\u4ea4\u6362\u673a\u4e0a\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u534f\u8bae\u65b9\u9762\uff0cNCCL \u63d0\u4f9b LL\uff08Latency-sensitive\uff0c8 \u5b57\u8282\u5e26 flag\uff09\u3001LL128\uff08128 \u5b57\u8282\u7ebf\uff0c\u517c\u987e\u5ef6\u8fdf\u4e0e\u5e26\u5bbd\uff09\u548c Simple\uff08\u5927\u5757\u4f20\u8f93\uff0c\u5e26\u5bbd\u6700\u4f18\uff09\u3002\u8fd9\u4e9b\u9009\u62e9\u901a\u5e38\u4e0d\u9700\u8981\u7528\u6237\u5e72\u9884\uff0c\u4f46\u7406\u89e3\u5b83\u4eec\u6709\u52a9\u4e8e\u89e3\u91ca\u4e3a\u4ec0\u4e48\u540c\u6837\u5927\u5c0f\u7684\u68af\u5ea6\u5728\u4e0d\u540c\u786c\u4ef6\u4e0a\u901a\u4fe1\u65f6\u95f4\u4f1a\u6709\u663e\u8457\u5dee\u5f02\u3002\u4f8b\u5982\uff0c\u5728 8 \u5361 NVLink \u8282\u70b9\u5185\uff0cNCCL \u5e38\u5e38\u4e3a\u5927\u68af\u5ea6\u9009\u62e9 Ring + Simple\uff0c\u800c\u4e3a\u5c0f\u68af\u5ea6\u9009\u62e9 Tree + LL128\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u4e3b\u6d41\u6846\u67b6\u7684\u505a\u6cd5\uff1a\u68af\u5ea6\u5206\u6876\u4e0e\u8ba1\u7b97\u901a\u4fe1\u91cd\u53e0<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u7684 <code>DistributedDataParallel<\/code> \u662f\u76ee\u524d\u5de5\u4e1a\u754c\u548c\u5b66\u672f\u754c\u6700\u4e3b\u6d41\u7684\u6570\u636e\u5e76\u884c\u5b9e\u73b0\u3002\u5b83\u7684\u6838\u5fc3\u6280\u5de7\u662f <strong>\u68af\u5ea6\u5206\u6876\uff08gradient bucketing\uff09<\/strong>\uff1a\u628a\u6240\u6709\u53c2\u6570\u6309\u53cd\u5411\u4f20\u64ad\u4e2d\u68af\u5ea6\u5c31\u7eea\u7684\u5148\u540e\u987a\u5e8f\u5206\u7ec4\uff0c\u6bcf\u7ec4\u9ed8\u8ba4\u7ea6 25 MB\u3002\u5f53\u4e00\u4e2a\u6876\u91cc\u7684\u6240\u6709\u68af\u5ea6\u90fd ready\uff0c\u5c31\u7acb\u5373\u53d1\u8d77\u4e00\u6b21 NCCL AllReduce\u3002\u8fd9\u6837\uff0c\u8f93\u51fa\u5c42\u9644\u8fd1\u7684\u68af\u5ea6\u901a\u4fe1\u53ef\u4ee5\u548c\u9760\u8fd1\u8f93\u5165\u5c42\u7684\u53cd\u5411\u8ba1\u7b97\u91cd\u53e0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u673a\u5236\u4f9d\u8d56 autograd hook\uff1a\u6bcf\u4e2a\u68af\u5ea6\u7d2f\u52a0\u5668\u5b8c\u6210\u8ba1\u7b97\u540e\u89e6\u53d1 hook\uff0chook \u68c0\u67e5\u5176\u6240\u5c5e\u6876\u662f\u5426\u5df2\u6ee1\uff0c\u5982\u679c\u6876\u5185\u6240\u6709\u68af\u5ea6\u90fd\u5df2\u5c31\u7eea\uff0c\u5219\u7acb\u5373\u542f\u52a8\u8be5\u6876\u7684\u5f02\u6b65 AllReduce\u3002\u6876\u7684\u5927\u5c0f\u662f\u4e00\u4e2a\u6743\u8861\uff1a\u6876\u592a\u5c0f\u4f1a\u589e\u5927 NCCL kernel \u542f\u52a8\u6b21\u6570\uff0c\u6876\u592a\u5927\u4f1a\u5ef6\u8fdf\u9996\u6b21\u901a\u4fe1\u7684\u5f00\u59cb\u65f6\u95f4\u5e76\u589e\u52a0\u5cf0\u503c\u663e\u5b58\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BatchNorm \u5219\u662f\u53e6\u4e00\u4e2a\u5206\u5e03\u5f0f\u75db\u70b9\u3002\u6bcf\u5f20\u5361\u53ea\u770b\u5230\u672c\u5730 mini-batch\uff0c\u5982\u679c\u5404\u81ea\u7b97\u5404\u81ea\u7684\u5747\u503c\u548c\u65b9\u5dee\uff0c\u7edf\u8ba1\u91cf\u4f1a\u5728\u5361\u95f4\u53d1\u6563\u3002PyTorch \u63d0\u4f9b\u4e86 <code>SyncBatchNorm<\/code>\uff0c\u901a\u8fc7 AllGather\/AllReduce \u628a\u5404\u5361\u7684 batch \u7edf\u8ba1\u91cf\u6c47\u603b\u6210\u5168\u5c40\u7edf\u8ba1\u91cf\u3002\u8fd9\u4e2a\u673a\u5236\u5bf9 batch size \u8f83\u5c0f\u7684\u4efb\u52a1\u5c24\u5176\u91cd\u8981\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">TensorFlow \u7684 <code>MultiWorkerMirroredStrategy<\/code>\u3001JAX \u7684 <code>pmap<\/code>\/<code>pjit<\/code> \u4e5f\u4f9d\u8d56 NCCL \u6216\u7b49\u4ef7\u7684\u96c6\u5408\u901a\u4fe1\u540e\u7aef\u3002\u5b83\u4eec\u7684\u5171\u540c\u70b9\u662f\uff1a\u5728\u8fd0\u884c\u65f6\u52a8\u6001\u7ec4\u7ec7\u901a\u4fe1\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001Tech-Renaissance \u7684\u5206\u5e03\u5f0f\u8bbe\u8ba1\u603b\u89c8<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u9009\u62e9\u9759\u6001\u56fe\u7f16\u8bd1\u8def\u7ebf\uff0c\u8fd9\u4ef6\u4e8b\u7ed9\u5206\u5e03\u5f0f\u901a\u4fe1\u5e26\u6765\u4e86\u5929\u7136\u4f18\u52bf\uff1a\u6240\u6709\u5f20\u91cf\u7684\u5f62\u72b6\u3001\u5185\u5b58\u504f\u79fb\u3001\u751f\u547d\u5468\u671f\u5728\u7f16\u8bd1\u671f\u5c31\u5df2\u7ecf\u786e\u5b9a\uff0c\u56e0\u6b64\u540c\u4e00\u5f20 <code>MemoryPlan<\/code> \u53ef\u4ee5\u76f4\u63a5\u590d\u5236\u5230\u6240\u6709 GPU \u4e0a\uff0c\u6240\u6709 rank \u5bf9\u540c\u4e00\u4e2a DTensor \u7684\u504f\u79fb\u8fbe\u6210\u4e00\u81f4\u3002NCCL \u901a\u4fe1\u4e0d\u518d\u9700\u8981\u8fd0\u884c\u65f6\u53bb\u62fc\u88c5\u5f20\u91cf\u5730\u5740\uff0c\u800c\u662f\u76f4\u63a5\u5bf9\u56fa\u5b9a\u7684 Region \u8303\u56f4\u505a AllReduce\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6765\u8bf4\uff0c\u6846\u67b6\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u7531\u4ee5\u4e0b\u51e0\u4e2a\u73af\u8282\u5171\u540c\u6784\u6210\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><code>GlobalRegistry<\/code> \u901a\u8fc7 <code>use_gpu()<\/code> \u8bbe\u7f6e world size\uff0c\u5e76\u901a\u8fc7 <code>global_batch_size()<\/code> \u628a\u5168\u5c40 batch \u5747\u5206\u5230\u6bcf\u5f20\u5361\uff1b<\/li>\n\n\n\n<li><code>TaskBase::compile_alloc_hardware()<\/code> \u8c03\u7528 <code>ncclCommInitAll()<\/code> \u4e3a\u6bcf\u4e2a rank \u521b\u5efa NCCL \u901a\u4fe1\u5668\uff1b<\/li>\n\n\n\n<li><code>Compiler<\/code> \u5728\u6784\u5efa\u8bad\u7ec3\u56fe\u65f6\uff0c\u628a\u68af\u5ea6\u533a\u5206\u6210\u4e24\u4e2a\u6876\uff0c\u5206\u522b\u751f\u6210 <code>DEEP_COMM<\/code> \u548c <code>FIRST_COMM<\/code> \u4e24\u5f20 AllReduce \u5b50\u56fe\uff1b<\/li>\n\n\n\n<li><code>allreduce_op.cpp<\/code> \u5728 <code>UPDATE<\/code> \u6d41\u4e0a\u6267\u884c <code>ncclAllReduce<\/code>\uff0c\u968f\u540e\u7528\u4e00\u4e2a scale kernel \u5b8c\u6210&#8221;\u6c42\u548c\u53d6\u5e73\u5747&#8221;\uff1b<\/li>\n\n\n\n<li><code>CapturedGraph<\/code> \u628a\u6240\u6709 rank \u7684 NCCL \u8c03\u7528\u540c\u65f6\u6355\u83b7\u8fdb CUDA Graph\uff0c\u8fd0\u884c\u65f6\u53ea\u505a <code>cudaGraphLaunch<\/code>\uff1b<\/li>\n\n\n\n<li><code>DeepLearningTask::run_train_epoch_gpu()<\/code> \u6309\u56fa\u5b9a\u987a\u5e8f\u8c03\u5ea6\u8fd9\u4e9b captured graph\uff0c\u5b9e\u73b0\u8ba1\u7b97\u4e0e\u901a\u4fe1\u7684\u7cbe\u786e\u91cd\u53e0\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u9762\u6309\u8fd9\u4e2a\u94fe\u8def\u9010\u5c42\u5c55\u5f00\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001world size \u4e0e local batch size\uff1a\u4ece\u7528\u6237\u914d\u7f6e\u5230\u8fd0\u884c\u65f6<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u91cc\uff0c\u5f00\u542f\u591a\u5361\u8bad\u7ec3\u53ea\u9700\u8981\u5728\u7a0b\u5e8f\u5f00\u5934\u5199\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(\"0-3\")          \/\/ world_size = 4\n    .global_batch_size(128); \/\/ local_batch_size = 128 \/ 4 = 32<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>use_gpu()<\/code> \u4f1a\u89e3\u6790 GPU ID \u5b57\u7b26\u4e32\uff0c\u6821\u9a8c\u6570\u91cf\u4e3a 2 \u7684\u5e42\u4e14\u4e0d\u8d85\u8fc7\u5f53\u524d\u7248\u672c\u652f\u6301\u7684\u4e0a\u9650\uff0c\u7136\u540e\u5728 <code>GlobalRegistry<\/code> \u4e2d\u5199\u5165 world size\u3002<code>global_batch_size()<\/code> \u5219\u8981\u6c42\u5168\u5c40 batch \u5fc5\u987b\u80fd\u88ab world size \u6574\u9664\uff0c\u5426\u5219\u76f4\u63a5\u629b\u9519\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::global_batch_size(int value) {\n    int ws = world_size();\n    if (value % ws != 0) {\n        TR_VALUE_ERROR(\"global_batch_size \" &lt;&lt; value\n                      &lt;&lt; \" is not divisible by world_size \" &lt;&lt; ws);\n    }\n    int local_bs = value \/ ws;\n    local_batch_size(local_bs);\n    return *this;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u6b65\u770b\u8d77\u6765\u53ea\u662f\u7b80\u5355\u7684\u9664\u6cd5\uff0c\u4f46\u5b83\u5960\u5b9a\u4e86\u6570\u636e\u5e76\u884c\u7684\u6570\u5b66\u7b49\u4ef7\u6027\uff1a\u53ea\u8981\u5404\u5361\u5904\u7406 <code>global_batch_size \/ world_size<\/code> \u4e2a\u6837\u672c\uff0c\u5e76\u4e14\u68af\u5ea6\u505a\u7b97\u672f\u5e73\u5747\uff0c\u90a3\u4e48 N \u5361\u8bad\u7ec3\u5c31\u7b49\u4ef7\u4e8e\u5355\u5361\u8bad\u7ec3\u4e00\u4e2a N \u500d\u5927\u7684 batch\u3002\u6846\u67b6\u91cc\u6ca1\u6709&#8221;\u6bcf\u4e2a\u8fdb\u7a0b\u81ea\u5df1\u51b3\u5b9a batch size&#8221;\u7684\u7075\u6d3b\u6027\uff0c\u56e0\u4e3a\u4efb\u4f55\u4e0d\u4e00\u81f4\u90fd\u4f1a\u7834\u574f AllReduce \u7684\u524d\u63d0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001NCCL \u901a\u4fe1\u5668\u7684\u521d\u59cb\u5316<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">NCCL \u7684\u4f7f\u7528\u4ece\u521b\u5efa communicator \u5f00\u59cb\u3002\u5728 <code>TaskBase::compile_alloc_hardware()<\/code> \u4e2d\uff0c\u5f53 GPU \u6570\u91cf\u5927\u4e8e 1 \u65f6\uff0c\u6846\u67b6\u8c03\u7528 <code>ncclCommInitAll()<\/code> \u4e00\u6b21\u6027\u4e3a\u6240\u6709\u9009\u4e2d\u7684 GPU \u5efa\u7acb\u901a\u4fe1\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=\"\">#ifdef TR_USE_NCCL\nif (gpu_ids.size() > 1) {\n    std::vector&lt;ncclComm_t> comms(gpu_ids.size());\n    ncclResult_t nccl_result = ncclCommInitAll(\n        comms.data(),\n        static_cast&lt;int>(gpu_ids.size()),\n        gpu_ids.data());\n    \/\/ \u9519\u8bef\u5904\u7406 ...\n    for (size_t i = 0; i &lt; gpu_ids.size(); ++i) {\n        backend_->contexts[i]->set_nccl_comm(comms[i]);\n    }\n}\n#endif<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ncclCommInitAll<\/code> \u4f1a\u81ea\u5df1\u5b8c\u6210\u62d3\u6251\u63a2\u6d4b\u3001\u901a\u9053\u6784\u5efa\u548c\u4f20\u8f93\u8def\u5f84\u9009\u62e9\u3002\u521b\u5efa\u597d\u7684 <code>ncclComm_t<\/code> \u88ab\u5b58\u8fdb\u6bcf\u4e2a <code>DeviceContext<\/code>\uff0c\u540e\u7eed\u6240\u6709 NCCL \u96c6\u5408\u901a\u4fe1\u90fd\u901a\u8fc7\u5b83\u53d1\u8d77\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u8fd9\u91cc\u7528\u7684\u662f\u5355\u8fdb\u7a0b\u591a\u7ebf\u7a0b\u6a21\u578b\uff1a\u4e00\u4e2a\u8fdb\u7a0b\u7ba1\u7406\u6240\u6709 GPU\uff0c<code>DeepLearningTask::run_train_epoch_gpu()<\/code> \u4e3a\u6bcf\u4e2a rank \u521b\u5efa\u4e00\u4e2a\u7ebf\u7a0b\uff0c\u5404\u81ea\u5728\u5bf9\u5e94\u7684 GPU \u4e0a launch CUDA Graph\u3002\u8fd9\u4e0e PyTorch DDP \u5e38\u89c1\u7684&#8221;\u6bcf\u4e2a GPU \u4e00\u4e2a\u8fdb\u7a0b&#8221;\u6a21\u578b\u4e0d\u540c\uff0c\u4f46 NCCL \u7684 <code>ncclCommInitAll<\/code> \u5bf9\u8fd9\u79cd\u540c\u8fdb\u7a0b\u591a\u5361\u573a\u666f\u540c\u6837\u652f\u6301\u5f97\u5f88\u597d\u3002\u5355\u8fdb\u7a0b\u6a21\u578b\u7684\u4f18\u70b9\u662f\u6240\u6709 rank \u5171\u4eab\u540c\u4e00\u4efd <code>MemoryPlan<\/code> \u548c\u4e3b\u673a\u7aef\u6570\u636e\u7ed3\u6784\uff0c\u8c03\u8bd5\u548c\u90e8\u7f72\u90fd\u66f4\u7b80\u5355\uff1b\u5728 C++ \u539f\u751f\u5b9e\u73b0\u91cc\u4e5f\u53ef\u4ee5\u907f\u5f00 Python \u751f\u6001\u4e2d\u67d0\u4e9b\u9700\u8981 per-process \u521d\u59cb\u5316\u7684\u7b2c\u4e09\u65b9\u5e93\u9650\u5236\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001DTensor + MemoryPlan\uff1a\u8de8\u5361\u540c\u4e00\u4efd\u56fe\u7eb8<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8981\u8ba9 AllReduce \u5de5\u4f5c\uff0c\u4e00\u4e2a\u5fc5\u8981\u6761\u4ef6\u662f\uff1a\u5404\u5361\u8981\u901a\u4fe1\u7684\u6570\u636e\u5728\u5404\u81ea\u663e\u5b58\u91cc\u7684\u5b57\u8282\u504f\u79fb\u5fc5\u987b\u4e00\u81f4\u3002Tech-Renaissance \u80fd\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u662f\u56e0\u4e3a <code>DTensor<\/code> \u672c\u8eab\u4e0d\u6301\u6709\u5185\u5b58\uff0c\u5b83\u53ea\u662f <code>(id, shape, dtype, region, offset, stride)<\/code> \u7684\u63cf\u8ff0\u7b26\uff1b\u771f\u6b63\u5185\u5b58\u7531 <code>ArenaKeeper<\/code> \u6309 <code>MemoryPlan<\/code> \u7684\u5e03\u5c40\u5206\u914d\u3002\u540c\u4e00\u5f20 <code>MemoryPlan<\/code> \u88ab\u6240\u6709 rank \u5171\u4eab\uff0c\u56e0\u6b64\u540c\u4e00 DTensor \u5728\u6bcf\u4e2a rank \u4e0a\u7684 <code>offset<\/code> \u5b8c\u5168\u4e00\u6837\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>DeviceContext::ptr_at()<\/code> \u7684\u8fd0\u884c\u65f6\u89e3\u6790\u8bc1\u5b9e\u4e86\u8fd9\u4e00\u70b9\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_,\n                                          static_cast&lt;size_t>(dt.offset()));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fdb\u4e00\u6b65\u5730\uff0c\u6846\u67b6\u628a\u68af\u5ea6\u6309\u8bed\u4e49\u96c6\u4e2d\u5230 G-Series Region\uff1a<code>G_BN_BIAS<\/code>\u3001<code>G_BN_WEIGHT<\/code>\u3001<code>G_FC_BIAS<\/code>\u3001<code>G_FC_WEIGHT<\/code>\u3001<code>G_FIRST_CONV<\/code>\u3001<code>G_DEEP_CONV<\/code>\u3002BN \u7edf\u8ba1\u91cf\u5219\u653e\u5728 <code>B_NEXT_MEAN<\/code> \u548c <code>B_NEXT_VAR<\/code>\u3002\u7f16\u8bd1\u5668\u4e0d\u9700\u8981\u77e5\u9053\u6bcf\u4e2a\u5177\u4f53\u5f20\u91cf\u7684\u540d\u5b57\uff0c\u53ea\u9700\u8981\u77e5\u9053&#8221;\u4ece Region X \u5230 Region Y \u7684\u8fde\u7eed\u5b57\u8282\u8303\u56f4&#8221;\uff0c\u5c31\u80fd\u76f4\u63a5\u751f\u6210 NCCL AllReduce\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd&#8221;\u6309 Region \u901a\u4fe1&#8221;\u7684\u8bbe\u8ba1\uff0c\u662f\u5f20\u91cf\u62bd\u8c61\u548c\u5185\u5b58\u5206\u533a\u5171\u540c\u51b3\u5b9a\u7684\uff1aRegion \u91cc\u7684\u5f20\u91cf\u5df2\u7ecf\u8fde\u7eed\u6392\u653e\uff0cAllReduce \u53ef\u4ee5\u4e00\u6b21\u6027\u626b\u8fc7\u6574\u4e2a\u8303\u56f4\uff0c\u65e0\u9700\u50cf PyTorch DDP \u90a3\u6837\u5728\u8fd0\u884c\u65f6\u628a\u4e00\u4e2a\u4e2a\u53c2\u6570\u7684\u68af\u5ea6\u62f7\u8d1d\u5230\u901a\u4fe1\u6876\u91cc\u3002\u5bf9\u4e8e ResNet-50 \u8fd9\u6837\u51e0\u5341\u5c42\u3001\u51e0\u767e\u4e2a\u53c2\u6570\u5f20\u91cf\u7684\u7f51\u7edc\uff0c\u7701\u6389\u7684\u62f7\u8d1d\u548c\u5730\u5740\u89e3\u6790\u5f00\u9500\u76f8\u5f53\u53ef\u89c2\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cTech-Renaissance \u7684 <code>MemoryPlan<\/code> \u8fd8\u4f1a\u901a\u8fc7\u8de8\u53d8\u4f53\u53d6\u6700\u5927 slot bytes \u7684\u65b9\u5f0f\uff0c\u4fdd\u8bc1\u540c\u4e00\u4e2a DTensor \u5728\u8bad\u7ec3\/\u9a8c\u8bc1\u3001\u6b63\u5e38 batch\/\u672b batch\u3001\u4e0d\u540c\u5206\u8fa8\u7387\u53d8\u4f53\u4e4b\u95f4\u7684 offset \u59cb\u7ec8\u4e00\u81f4\u3002\u8fd9\u610f\u5473\u7740\u901a\u4fe1\u5b50\u56fe\u53ef\u4ee5\u5728\u4e0d\u540c\u53d8\u4f53\u4e4b\u95f4\u5171\u4eab\uff0c\u8fdb\u4e00\u6b65\u51cf\u5c11\u4e86 CUDA Graph \u7684\u6355\u83b7\u6570\u91cf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u4e24\u6876\u68af\u5ea6 AllReduce\uff1a\u6734\u7d20\u7684\u5212\u5206\uff0c\u9ad8\u6548\u7684\u9690\u85cf<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u6ca1\u6709\u91c7\u7528 PyTorch \u90a3\u79cd\u6309 25 MB \u52a8\u6001\u5206\u6876\u7684\u7b56\u7565\uff0c\u800c\u662f\u505a\u4e86\u4e00\u4e2a\u66f4\u6734\u7d20\u7684\u4e24\u6876\u5212\u5206\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6876 1\uff08DEEP_COMM\uff09<\/strong>\uff1a<code>G_DEEP_CONV<\/code> \u5230 <code>R_RESULT<\/code> \u7684\u6df1\u5c42\u5377\u79ef\u68af\u5ea6\uff1b<\/li>\n\n\n\n<li><strong>\u6876 2\uff08FIRST_COMM\uff09<\/strong>\uff1a<code>G_BN_BIAS<\/code> \u5230 <code>G_FIRST_CONV<\/code> \u7684\u9996\u5c42 + BN + FC \u68af\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><code>MemoryPlan<\/code> \u66b4\u9732\u4e86\u4e24\u4e2a\u67e5\u8be2\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=\"\">CommRange MemoryPlan::get_comm_range_bucket1() const {\n    auto&amp; r = region_infos_[static_cast&lt;size_t>(Region::G_DEEP_CONV)];\n    auto&amp; e = region_infos_[static_cast&lt;size_t>(Region::R_RESULT)];\n    uint64_t start = r.base_offset;\n    uint64_t end   = e.base_offset + e.total_bytes;\n    return {start, end - start};\n}\n\nCommRange MemoryPlan::get_comm_range_bucket2() const {\n    auto&amp; s = region_infos_[static_cast&lt;size_t>(Region::G_BN_BIAS)];\n    auto&amp; e = region_infos_[static_cast&lt;size_t>(Region::G_FIRST_CONV)];\n    return {s.base_offset,\n            e.base_offset + e.total_bytes - s.base_offset};\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Compiler<\/code> \u76f4\u63a5\u628a\u8fd9\u4e24\u4e2a\u8303\u56f4\u6302\u5230\u4e24\u5f20\u901a\u4fe1\u5b50\u56fe\u4e0a\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=\"\">MemRange r_first = memory_plan.region_range(\n    Region::G_BN_BIAS, Region::G_FIRST_CONV);\ntrain_cg.append_range(GraphId::FIRST_COMM, RangeOp::RANGE_MEAN_ALLREDUCE,\n                      {r_first}, {r_first});\n\nMemRange r_deep = memory_plan.region_range(\n    Region::G_DEEP_CONV, Region::R_RESULT);\ntrain_cg.append_range(GraphId::DEEP_COMM, RangeOp::RANGE_MEAN_ALLREDUCE,\n                      {r_deep}, {r_deep});<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\u8981\u8fd9\u6837\u5206\uff1f\u56e0\u4e3a CNN \u7684\u53cd\u5411\u4f20\u64ad\u4ece loss \u5f80\u56de\u8d70\uff0c\u6df1\u5c42\u5377\u79ef\u7684\u68af\u5ea6\u5148\u7b97\u51fa\u6765\uff0c\u800c\u9996\u5c42\u5377\u79ef\u7684\u68af\u5ea6\u8981\u5230\u6700\u540e\u624d\u7b97\u51fa\u6765\u3002\u5982\u679c\u6211\u4eec\u628a\u6df1\u5c42\u68af\u5ea6\u4f5c\u4e3a\u4e00\u6876\uff0c\u5728\u9996\u5c42\u53cd\u5411\u4f20\u64ad\u8fd8\u5728\u8ba1\u7b97\u7684\u65f6\u5019\uff0c\u5c31\u5f02\u6b65\u53d1\u8d77\u6df1\u5c42\u68af\u5ea6\u7684 AllReduce\uff0c\u90a3\u4e48\u901a\u4fe1\u65f6\u95f4\u51e0\u4e4e\u53ef\u4ee5\u5b8c\u5168\u88ab\u9996\u5c42\u53cd\u5411\u8ba1\u7b97\u76d6\u4f4f\u3002\u6876 2 \u5219\u5fc5\u987b\u7b49\u9996\u5c42\u53cd\u5411\u5b8c\u6210\u540e\u624d\u80fd\u5f00\u59cb\u3002\u6876\u6570\u8d8a\u591a\uff0c\u542f\u52a8 NCCL kernel \u7684\u6b21\u6570\u5c31\u8d8a\u591a\uff0cCPU \u6d3e\u53d1\u5f00\u9500\u4e5f\u4f1a\u53d8\u5927\uff1b\u4e24\u6876\u662f\u4e00\u4e2a\u5728 CNN \u573a\u666f\u4e0b\u7b80\u5355\u4e14\u9ad8\u6548\u7684\u9009\u62e9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ec6\u5fc3\u7684\u8bfb\u8005\u4f1a\u6ce8\u610f\u5230\uff0c\u6876 1 \u7684\u8303\u56f4\u4ece <code>G_DEEP_CONV<\/code> \u4e00\u76f4\u5ef6\u4f38\u5230 <code>R_RESULT<\/code>\u3002<code>R_RESULT<\/code> \u662f\u6bcf batch \u7684 loss \u548c top-1\/top-5 \u6807\u91cf\u533a\uff0c\u628a\u5b83\u4e5f\u7eb3\u5165 AllReduce \u8303\u56f4\u662f\u56e0\u4e3a\u5b83\u5728\u5185\u5b58\u5e03\u5c40\u4e0a\u4e0e\u6df1\u5c42\u68af\u5ea6\u8fde\u7eed\uff0c\u6570\u636e\u91cf\u6781\u5c0f\uff0c\u4e0d\u4f1a\u589e\u52a0\u5b9e\u9645\u901a\u4fe1\u8d1f\u62c5\uff0c\u53cd\u800c\u8ba9\u7f16\u8bd1\u5668\u53ef\u4ee5\u7528\u4e00\u6b21 <code>region_range<\/code> \u8c03\u7528\u8986\u76d6\u6574\u4e2a\u533a\u95f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5e76\u4e0d\u610f\u5473\u7740\u52a8\u6001\u5206\u6876\u6ca1\u6709\u4ef7\u503c\u3002\u5bf9\u4e8e Transformer \u8fd9\u7c7b\u53c2\u6570\u91cf\u5206\u5e03\u66f4\u5747\u5300\u3001\u53cd\u5411\u8def\u5f84\u66f4\u957f\u7684\u6a21\u578b\uff0c\u66f4\u7ec6\u7c92\u5ea6\u7684\u5206\u6876\u53ef\u80fd\u4f1a\u5e26\u6765\u66f4\u597d\u7684\u91cd\u53e0\u3002Tech-Renaissance \u7684\u4e24\u6876\u7b56\u7565\u662f\u5728\u5f53\u524d CNN \u5de5\u4f5c\u8d1f\u8f7d\u548c\u9759\u6001\u56fe\u7ea6\u675f\u4e0b\u7684\u52a1\u5b9e\u9009\u62e9\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001AllReduce \u7b97\u5b50\uff1a\u5148\u6c42\u548c\uff0c\u518d\u5e73\u5747<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9645\u6267\u884c AllReduce \u7684\u662f <code>src\/backend\/ops\/range\/allreduce_op.cpp<\/code>\u3002\u5b83\u9996\u5148\u901a\u8fc7 <code>GlobalRegistry::instance().world_size()<\/code> \u62ff\u5230 world size\uff0c\u5982\u679c\u53ea\u6709\u4e00\u5361\u5c31\u76f4\u63a5\u8df3\u8fc7\uff1b\u5426\u5219\u5bf9\u6bcf\u4e2a\u8f93\u5165\u8303\u56f4\u6267\u884c <code>ncclAllReduce(..., ncclSum, ...)<\/code>\uff0c\u7136\u540e\u7528\u4e00\u4e2a scale kernel \u628a\u7ed3\u679c\u9664\u4ee5 world size\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\uff1a\u5728\u540e\u7eed\u6781\u9650\u4f18\u5316\u7248\u672c\u4e2d\uff0c\u8fd9\u6b65\u989d\u5916\u7684 scale kernel \u53ef\u4ee5\u79fb\u9664\uff0c\u76f4\u63a5\u8c03\u7528\u652f\u6301\u5e73\u5747\u5f52\u7ea6\u7684 <code>ncclAvg<\/code>\uff08NCCL 2.10+\uff09\uff0c\u6216\u8005\u5c06\u9664\u4ee5 <code>world_size<\/code> \u7684\u64cd\u4f5c\u878d\u8fdb\u4f18\u5316\u5668\u66f4\u65b0 kernel\uff0c\u4ee5\u7701\u6389\u4e00\u6b21\u5168\u5c40\u663e\u5b58\u8bfb\u5199\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=\"\">int world_size = GlobalRegistry::instance().world_size();\nif (world_size &lt;= 1) return;\n\nbool do_mean = (node.range_op == RangeOp::RANGE_MEAN_ALLREDUCE ||\n                node.range_op == RangeOp::RANGE_BN_STATS_ALLREDUCE);\n\n\/\/ ... \u89e3\u6790 src\/dst \u504f\u79fb ...\n\nncclResult_t res = ncclAllReduce(\n    dst, dst, count, ncclFloat32, ncclSum,\n    static_cast&lt;ncclComm_t>(ctx.nccl_comm()), s);\n\nif (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,\n                                static_cast&lt;int64_t>(count), s);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u51e0\u4e2a\u7ec6\u8282\u503c\u5f97\u6ce8\u610f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>ncclAllReduce<\/code> \u7684 <code>sendbuf<\/code> \u548c <code>recvbuf<\/code> \u6307\u5411\u540c\u4e00\u5757\u663e\u5b58\uff0c\u662f in-place \u64cd\u4f5c\uff1b<\/li>\n\n\n\n<li>\u901a\u4fe1\u6d41\u56fa\u5b9a\u5728 <code>StreamKind::UPDATE<\/code>\uff0c\u4e0e\u8ba1\u7b97\u6d41 <code>COMP_1\/2\/3<\/code> \u5206\u79bb\uff0c\u4fbf\u4e8e\u91cd\u53e0\uff1b<\/li>\n\n\n\n<li>\u7f29\u653e\u7531\u81ea\u5b9a\u4e49 CUDA kernel \u5b8c\u6210\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56 NCCL \u7684 reduce \u5e73\u5747\u8bed\u4e49\uff0c\u8fd9\u6837\u903b\u8f91\u66f4\u6e05\u6670\uff0c\u4e5f\u4fbf\u4e8e\u540e\u7eed\u6269\u5c55\uff0c\u6bd4\u5982\u652f\u6301\u4e0d\u540c\u7684\u5f52\u7ea6\u56e0\u5b50\u6216\u6df7\u5408\u7cbe\u5ea6\u7f29\u653e\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u3001\u628a NCCL \u4e5f\u6355\u83b7\u8fdb CUDA Graph<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u4e00\u5927\u7279\u8272\u662f\u8bad\u7ec3\u5faa\u73af\u5168\u9636\u6bb5 CUDA Graph \u6355\u83b7\u3002NCCL \u8c03\u7528\u80fd\u4e0d\u80fd\u8fdb Graph\uff1f\u53ef\u4ee5\uff0c\u4f46\u5fc5\u987b\u6240\u6709 rank \u540c\u65f6\u5f00\u59cb capture\u3001\u540c\u65f6\u7ed3\u675f\uff0c\u5e76\u4e14 NCCL \u8c03\u7528\u8981\u5305\u5728 <code>ncclGroupStart()<\/code> \/ <code>ncclGroupEnd()<\/code> \u4e4b\u95f4\u3002<code>src\/graph\/captured_graph.cpp<\/code> \u4e2d\u7684 <code>capture_nccl_graph_coordinated()<\/code> \u4e13\u95e8\u5904\u7406\u4e86\u8fd9\u79cd\u60c5\u51b5\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=\"\">\/\/ 1. \u6240\u6709 rank \u540c\u65f6\u5f00\u59cb\u6355\u83b7\nfor (int r = 0; r &lt; num_ranks; ++r) {\n    cudaSetDevice(contexts[r]->device_id());\n    cudaStreamBeginCapture(cap_streams[r], cudaStreamCaptureModeThreadLocal);\n}\n\n\/\/ 2. \u91cd\u653e\u6bcf\u4e2a rank \u7684\u5b50\u56fe\u8282\u70b9\uff0cNCCL \u8c03\u7528\u5305\u5728 group \u4e2d\nncclGroupStart();\nfor (int r = 0; r &lt; num_ranks; ++r) {\n    DeviceContext&amp; dc = *contexts[r];\n    cudaSetDevice(dc.device_id());\n    \/\/ \u4f9d\u6b21\u91cd\u653e GraphId::FIRST_COMM \/ DEEP_COMM \/ ... \u4e2d\u7684\u8282\u70b9\n    \/\/ \u4e3b\u8981\u5305\u62ec ncclAllReduce\n}\nncclGroupEnd();\n\n\/\/ 3. \u6240\u6709 rank \u540c\u65f6\u7ed3\u675f\u6355\u83b7\u5e76\u5b9e\u4f8b\u5316\nfor (int r = 0; r &lt; num_ranks; ++r) {\n    cudaSetDevice(contexts[r]->device_id());\n    cudaStreamEndCapture(cap_streams[r], &amp;captured_graphs[r]);\n}\nfor (int r = 0; r &lt; num_ranks; ++r) {\n    cudaGraphInstantiate(&amp;exec, captured_graphs[r], ...);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ncclGroupStart\/End<\/code> \u7684\u542b\u4e49\u662f\uff1a\u628a\u63a5\u4e0b\u6765\u6240\u6709 rank \u7684 NCCL \u8c03\u7528\u6807\u8bb0\u4e3a\u540c\u4e00\u4e2a\u534f\u8c03\u7ec4\uff0c\u8fd9\u6837 NCCL \u624d\u80fd\u5728\u591a rank \u4e4b\u95f4\u4fdd\u8bc1\u6b63\u786e\u7684\u540c\u6b65\u987a\u5e8f\u3002\u5982\u679c\u67d0\u4e2a rank \u6f0f\u6389\u4e86\u4e00\u4e2a NCCL \u8c03\u7528\uff0c\u6216\u8005 capture \u987a\u5e8f\u4e0d\u4e00\u81f4\uff0c\u6574\u4e2a communicator \u5c31\u4f1a\u6b7b\u9501\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u901a\u4fe1\u5b50\u56fe\uff08<code>FIRST_COMM<\/code>\u3001<code>DEEP_COMM<\/code>\u3001<code>STATS_COMM<\/code>\u3001<code>VAL_RESULT_COMM<\/code>\uff09\u662f shape-invariant \u7684\uff0c\u6240\u6709\u53d8\u4f53\u5171\u4eab\u540c\u4e00\u4efd captured graph\uff0c\u4e0d\u4f1a\u88ab\u91cd\u590d\u6355\u83b7\u3002\u8fd0\u884c\u65f6\u53ea\u9700 <code>cudaGraphLaunch<\/code>\uff0c\u4e0d\u9700\u8981\u6bcf\u6b21\u91cd\u65b0\u7ec4\u7ec7 NCCL \u8c03\u7528\uff0cCPU \u6d3e\u53d1\u5f00\u9500\u88ab\u964d\u5230\u6700\u4f4e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>is_shape_invariant_graph()<\/code> \u7684\u679a\u4e3e\u4e5f explicitly \u5305\u542b\u4e86\u8fd9\u51e0\u5f20\u901a\u4fe1\u5b50\u56fe\u3002\u4e4b\u6240\u4ee5 shape-invariant\uff0c\u662f\u56e0\u4e3a\u5b83\u4eec\u7684\u8f93\u5165\u8f93\u51fa\u53ea\u4f9d\u8d56 Region \u504f\u79fb\u548c world size\uff0c\u4e0d\u4f9d\u8d56 batch size \u6216\u8f93\u5165\u5206\u8fa8\u7387\u3002\u56e0\u6b64\u5373\u4f7f\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9047\u5230\u6700\u540e\u4e00\u4e2a\u4e0d\u5b8c\u6574\u7684 batch\uff0c\u6216\u8005\u5728\u4e0d\u540c\u5206\u8fa8\u7387\u53d8\u4f53\u4e4b\u95f4\u5207\u6362\uff0cAllReduce \u5b50\u56fe\u4e5f\u4e0d\u9700\u8981\u91cd\u65b0 capture\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u70b9\u4e0e PyTorch DDP \u5f62\u6210\u9c9c\u660e\u5bf9\u6bd4\uff1aPyTorch DDP \u7684 AllReduce \u7531 C++ Reducer \u5728\u8fd0\u884c\u65f6\u52a8\u6001\u53d1\u8d77\uff0c\u6bcf\u6b21 step \u90fd\u8981\u8d70\u4e00\u904d\u6876\u7ba1\u7406\u3001hook \u89e6\u53d1\u3001kernel launch \u7684\u6d41\u7a0b\u3002Tech-Renaissance \u5728\u7f16\u8bd1\u671f\u5c31\u628a\u8fd9\u4e9b\u5168\u90e8\u56fa\u5b9a\u4e0b\u6765\uff0c\u8fd0\u884c\u671f\u53ea\u5269\u56fe\u7684\u56de\u653e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e00\u3001\u8bad\u7ec3\u5faa\u73af\u4e2d\u7684\u8c03\u5ea6\uff1a\u8ba1\u7b97\u4e0e\u901a\u4fe1\u7684\u91cd\u53e0<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>DeepLearningTask::run_train_epoch_gpu()<\/code> \u7528\u591a\u7ebf\u7a0b\u9a71\u52a8\u6bcf\u5f20\u5361\uff0c\u6bcf\u4e2a\u7ebf\u7a0b\u6309\u56fa\u5b9a\u987a\u5e8f launch \u5404\u4e2a captured graph\u3002\u4e00\u4e2a\u5e38\u89c4 batch \u7684\u5173\u952e\u7247\u6bb5\u5982\u4e0b\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=\"\">\/\/ 1. \u6df1\u5c42\u524d\u5411 + \u53cd\u5411\uff08\u8ba1\u7b97\u6d41 COMP_1\uff09\nif (g_deep) cudaGraphLaunch(g_deep, s_c1);\nsync_comp();\n\n\/\/ 2. \u9996\u5c42\u53cd\u5411 + AMP \u6df1\u5c42\u68af\u5ea6 cast + \u6df1\u5c42\u68af\u5ea6 AllReduce\n\/\/    \u9996\u5c42\u53cd\u5411\u5728 COMP_1\uff0ccast \/ AllReduce \u5728 UPDATE\uff0c\u4e24\u8005\u53ef\u91cd\u53e0\nif (!frozen &amp;&amp; g_first) cudaGraphLaunch(g_first, s_c1);\nif (using_amp &amp;&amp; n_cdg) cudaGraphLaunch(n_cdg, s_up);\nif (n_dar) cudaGraphLaunch(n_dar, s_up);\n\nsync_up(); sync_comp();\n\n\/\/ 3. \u9996\u5c42\u68af\u5ea6 cast + \u9996\u5c42 AllReduce + \u6307\u6807\u7d2f\u52a0 + NaN \u68c0\u6d4b + BN \u7edf\u8ba1\u540c\u6b65\nif (using_amp &amp;&amp; n_cfg) cudaGraphLaunch(n_cfg, s_up);\nif (n_far) cudaGraphLaunch(n_far, s_up);\nif (n_accum) cudaGraphLaunch(n_accum, s_up);\nif (n_ncg) cudaGraphLaunch(n_ncg, s_up);\nif (n_sc) cudaGraphLaunch(n_sc, s_up);\nif (n_us) cudaGraphLaunch(n_us, s_up);\nsync_up();\n\n\/\/ 4. \u4f18\u5316\u5668\u66f4\u65b0 + LARS \u4e09\u6d41\u5e76\u884c\nif (n_wu) cudaGraphLaunch(n_wu, s_up);\nif (n_lars_fc)  cudaGraphLaunch(n_lars_fc,  s_c1);\nif (n_lars_fc2) cudaGraphLaunch(n_lars_fc2, s_c2);\nif (n_lars_dc)  cudaGraphLaunch(n_lars_dc,  s_c3);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u6b65\u9aa4 2 \u4e2d <code>g_first<\/code>\uff08\u9996\u5c42\u53cd\u5411\uff09\u548c <code>n_dar<\/code>\uff08\u6df1\u5c42\u68af\u5ea6 AllReduce\uff09\u662f\u80cc\u9760\u80cc launch \u7684\uff1a\u9996\u5c42\u53cd\u5411\u8dd1\u5728 <code>COMP_1<\/code>\uff0c\u6df1\u5c42 AllReduce \u8dd1\u5728 <code>UPDATE<\/code>\uff0c\u4e24\u6761\u6d41\u5e76\u884c\u3002\u7b49 <code>sync_up(); sync_comp();<\/code> \u540c\u6b65\u65f6\uff0c\u6df1\u5c42\u901a\u4fe1\u5f88\u53ef\u80fd\u5df2\u7ecf\u5b8c\u6210\u3002\u8fd9\u5c31\u662f&#8221;\u628a\u901a\u4fe1\u85cf\u5728\u8ba1\u7b97\u80cc\u540e&#8221;\u7684\u5177\u4f53\u5b9e\u73b0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e8c\u3001BN \u7edf\u8ba1\u91cf\u7684\u8de8\u5361\u540c\u6b65<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">BatchNorm \u5728\u8bad\u7ec3\u65f6\u8981\u7ef4\u62a4 running mean \u548c running variance\u3002\u591a\u5361\u8bad\u7ec3\u4e0b\uff0c\u5982\u679c\u6bcf\u5f20\u5361\u53ea\u66f4\u65b0\u81ea\u5df1\u7684 running stats\uff0c\u90a3\u4e48\u5404\u5361\u7684 BN \u53c2\u6570\u4f1a\u53d1\u6563\u3002Tech-Renaissance \u7684\u505a\u6cd5\u662f\uff1a\u6bcf\u4e2a batch \u7684\u524d\u5411\u4f20\u64ad\u628a\u5f53\u524d batch \u7684\u7edf\u8ba1\u91cf\u5199\u5230 <code>B_NEXT_MEAN<\/code> \u548c <code>B_NEXT_VAR<\/code>\uff1b\u5728\u5b8c\u6210\u8be5 batch \u7684\u53cd\u5411\u4f20\u64ad\u540e\uff0c\u7528 <code>STATS_COMM<\/code> \u505a\u4e00\u6b21 AllReduce\uff1b\u7136\u540e <code>UPDATE_STATS<\/code> \u628a\u540c\u6b65\u540e\u7684 <code>B_NEXT_*<\/code> \u590d\u5236\u5230 <code>B_PREV_*<\/code>\uff0c\u4f5c\u4e3a\u4e0b\u4e00 batch \u7684 running stats\u3002\u63a8\u7406\u524d\u518d\u7528 <code>UPDATE_BN_INF_PARAMS<\/code> \u8ba1\u7b97\u6700\u7ec8\u7684 <code>eq_scale<\/code> \/ <code>eq_bias<\/code>\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=\"\">if (has_bn) {\n    MemRange r_next = memory_plan.region_range(\n        Region::B_NEXT_MEAN, Region::B_NEXT_VAR);\n    train_cg.append_range(GraphId::STATS_COMM,\n        RangeOp::RANGE_BN_STATS_ALLREDUCE, {r_next}, {r_next});\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\uff0c<code>STATS_COMM<\/code> \u548c <code>UPDATE_STATS<\/code> \u662f<strong>\u6bcf\u4e2a\u8bad\u7ec3 batch<\/strong>\u90fd\u4f1a\u6267\u884c\u4e00\u6b21\u7684\uff0c\u800c\u4e0d\u662f\u6bcf\u4e2a epoch \u624d\u6267\u884c\u4e00\u6b21\u3002\u8fd9\u79cd per-batch \u540c\u6b65\u7684\u662f BN \u7684 running statistics\uff0c\u4f7f\u5404 rank \u5728\u540e\u7eed\u63a8\u7406\u548c\u7edf\u8ba1\u91cf\u63a8\u8fdb\u4e0a\u4fdd\u6301\u4e00\u81f4\uff1b\u5b83\u5e76\u4e0d\u5b8c\u5168\u7b49\u4ef7\u4e8e PyTorch <code>SyncBatchNorm<\/code> \u90a3\u79cd\u5728\u524d\u5411\u5f52\u4e00\u5316\u524d\u540c\u6b65\u5f53\u524d batch \u7edf\u8ba1\u91cf\u7684\u5b9e\u73b0\u3002\u82e5\u8981\u505a\u5230\u4e25\u683c\u7684 SyncBatchNorm \u8bed\u4e49\uff0c\u9700\u8981\u5728 BN finalize \/ apply \u4e4b\u524d\u5bf9\u5404\u5361\u7684 <code>sum<\/code> \u548c <code>sq_sum<\/code> \u505a AllReduce\u3002Tech-Renaissance \u76ee\u524d\u7531\u4e8e running stats \u7684\u66f4\u65b0\u8282\u594f\u548c\u9759\u6001\u56fe\u7f16\u8bd1\u7279\u70b9\uff0c\u628a running stats \u7684\u540c\u6b65\u56fa\u5316\u6210\u4e86\u4e00\u5f20\u5728\u6bcf\u4e2a batch \u672b\u5c3e\u6267\u884c\u7684\u901a\u4fe1\u5b50\u56fe\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e09\u3001\u53c2\u6570\u521d\u59cb\u5316\u4e0e\u9a8c\u8bc1\u6307\u6807\u7684\u5168\u5c40\u5e7f\u64ad<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u9664\u4e86\u68af\u5ea6\u540c\u6b65\uff0c\u5206\u5e03\u5f0f\u8bad\u7ec3\u8fd8\u9700\u8981\u4fdd\u8bc1\u6240\u6709 rank \u4ece\u76f8\u540c\u7684\u521d\u59cb\u53c2\u6570\u51fa\u53d1\u3002<code>TaskBase::init()<\/code> \u4f1a\u5148\u5728 CPU \u7aef\u7528 Philox \u751f\u6210\u968f\u673a\u6570\uff0cH2D \u5230 rank 0\uff0c\u518d\u8c03\u7528 <code>broadcast_from_rank0()<\/code> \u628a\u6743\u91cd\u5e7f\u64ad\u7ed9\u6240\u6709 rank\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 TaskBase::broadcast_from_rank0(const DTensor&amp; dt) {\n    \/\/ ... dtype \/ count \u8f6c\u6362 ...\n    ncclGroupStart();\n    for (int rank = 0; rank &lt; num_gpus_; ++rank) {\n        cudaSetDevice(reg.gpu_ids()[rank]);\n        void* ptr = backend_->contexts[rank]->ptr_at(dt.id);\n        cudaStream_t update_stream = static_cast&lt;cudaStream_t>(\n            backend_->contexts[rank]->stream(StreamKind::UPDATE));\n\n        ncclBroadcast(\n            ptr, ptr, nccl_count, nccl_type, 0,\n            backend_->contexts[rank]->nccl_comm(),\n            update_stream);\n    }\n    ncclGroupEnd();\n    for (int rank = 0; rank &lt; num_gpus_; ++rank) {\n        backend_->contexts[rank]->synchronize_stream(StreamKind::UPDATE);\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u9a8c\u8bc1\u9636\u6bb5\uff0c\u6bcf\u5f20\u5361\u5404\u81ea\u7d2f\u79ef\u81ea\u5df1\u7684 <code>sum_loss<\/code>\u3001<code>sum_top1<\/code>\u3001<code>sum_top5<\/code>\uff0c\u6700\u540e\u901a\u8fc7 <code>VAL_RESULT_COMM<\/code> \u5bf9 <code>R_RESULT_ACCUMULATED<\/code> \u533a\u57df\u505a\u4e00\u6b21 <code>RANGE_MEAN_ALLREDUCE<\/code>\uff0c\u5f97\u5230\u5168\u5c40\u5e73\u5747\u6307\u6807\u3002\u8fd9\u6837\u6bcf\u5f20\u5361\u770b\u5230\u7684\u9a8c\u8bc1 loss \u548c top-1\/top-5 \u90fd\u662f\u5168\u5c40\u4e00\u81f4\u7684\u7ed3\u679c\uff0c\u800c\u4e0d\u662f\u5404\u81ea\u5c40\u90e8\u7684\u8fd1\u4f3c\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>VAL_RESULT_COMM<\/code> \u4e0e\u8bad\u7ec3\u65f6\u7684\u68af\u5ea6 AllReduce \u5171\u7528\u540c\u4e00\u5957 RangeOp \u57fa\u7840\u8bbe\u65bd\uff0c\u533a\u522b\u53ea\u5728\u4e8e\u5b83\u64cd\u4f5c\u7684\u662f <code>R_RESULT_ACCUMULATED<\/code> Region\uff0c\u5e76\u4e14\u53ea\u5728\u6bcf\u4e2a epoch \u7684\u9a8c\u8bc1\u7ed3\u675f\u65f6\u89e6\u53d1\u4e00\u6b21\u3002\u7531\u4e8e\u5b83\u5df2\u7ecf\u88ab\u9884\u5148\u6355\u83b7\u8fdb CUDA Graph\uff0c\u9a8c\u8bc1\u6307\u6807\u7684\u540c\u6b65\u4e0d\u4f1a\u5f15\u5165\u989d\u5916\u7684 host-device \u540c\u6b65\u70b9\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u56db\u3001\u901a\u4fe1\u4f1a\u4e0d\u4f1a\u6210\u4e3a\u74f6\u9888<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u5e76\u884c\u7684\u6269\u5c55\u6548\u7387\u5e76\u4e0d\u603b\u662f\u7ebf\u6027\u7684\u3002\u51b3\u5b9a\u74f6\u9888\u7684\u662f&#8221;\u8ba1\u7b97\u91cf \/ \u901a\u4fe1\u91cf&#8221;\u8fd9\u4e2a\u6bd4\u503c\u3002\u5bf9 AllReduce \u800c\u8a00\uff0c\u5047\u8bbe\u5355\u5361\u6a21\u578b\u68af\u5ea6\u7684\u603b\u5b57\u8282\u6570\u4e3a <code>|g|<\/code>\uff0cworld size \u4e3a <code>N<\/code>\uff0c\u90a3\u4e48\u5355\u6b21 AllReduce \u4e2d\u6bcf\u5f20\u5361\u9700\u8981\u53d1\u9001\u548c\u63a5\u6536\u7684\u6570\u636e\u91cf\u5927\u81f4\u4e3a\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=\"\">\/\/ |g| : \u5355\u5361\u6a21\u578b\u68af\u5ea6\u7684\u603b\u5b57\u8282\u6570\n\/\/ N   : world_size\n\/\/ \u5355\u6b21 AllReduce \u4e2d\uff0c\u6bcf\u5f20\u5361\u9700\u8981\u53d1\u9001\/\u63a5\u6536\u7684\u6570\u636e\u91cf\u7ea6\u4e3a\uff1a\ndouble traffic_per_rank = 2.0 * (N - 1) \/ N * |g|;\n\/\/ \u5f53 N \u8f83\u5927\u65f6\uff0c\u8fd1\u4f3c\u4e3a 2 * |g|<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u6a21\u578b\u5f88\u5927\u3001batch \u5f88\u5c0f\uff0c\u901a\u4fe1\u5360\u6bd4\u5c31\u4f1a\u5347\u9ad8\uff1b\u5982\u679c batch \u8db3\u591f\u5927\uff0c\u6bcf\u5f20\u5361\u82b1\u5728 forward\/backward \u4e0a\u7684\u65f6\u95f4\u8fdc\u8d85\u901a\u4fe1\uff0c\u6269\u5c55\u6548\u7387\u5c31\u9ad8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u5916\uff0c\u8282\u70b9\u5185 NVLink \u7684\u5e26\u5bbd\u901a\u5e38\u5728 600\u2013900 GB\/s\uff0c\u800c\u8de8\u8282\u70b9 InfiniBand \u53ef\u80fd\u53ea\u6709 25\u2013100 GB\/s\uff0c\u540e\u8005\u66f4\u5bb9\u6613\u6210\u4e3a\u74f6\u9888\u3002Tech-Renaissance \u5f53\u524d\u7248\u672c\u4e3b\u8981\u9762\u5411\u5355\u8282\u70b9\u591a\u5361\u573a\u666f\uff0c<code>ncclCommInitAll()<\/code> \u4f1a\u81ea\u52a8\u9009\u62e9 NVLink\/P2P \u8def\u5f84\uff0c\u901a\u4fe1\u6548\u7387\u901a\u5e38\u5f88\u9ad8\u3002\u591a\u8282\u70b9\u8bad\u7ec3\u9700\u8981\u628a\u521d\u59cb\u5316\u6539\u6210\u57fa\u4e8e TCP\/RDMA \u7684\u8fdb\u7a0b\u53d1\u73b0\uff0c\u8fd9\u662f\u540e\u7eed\u53ef\u80fd\u7684\u6269\u5c55\u65b9\u5411\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u4e94\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c\u4e0d\u662f\u7b80\u5355\u5730\u5728\u53cd\u5411\u4f20\u64ad\u540e\u52a0\u4e00\u53e5 <code>ncclAllReduce<\/code>\uff0c\u800c\u662f\u628a\u901a\u4fe1\u548c\u9759\u6001\u56fe\u7f16\u8bd1\u3001MemoryPlan \u5206\u533a\u3001CUDA Graph \u6355\u83b7\u3001\u591a\u6d41\u8c03\u5ea6\u6574\u5408\u5728\u4e00\u8d77\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>GlobalRegistry<\/code> \u7edf\u4e00\u914d\u7f6e world size \u548c local batch size\uff1b<\/li>\n\n\n\n<li><code>ncclCommInitAll()<\/code> \u5728\u7f16\u8bd1\u671f\u5efa\u7acb NCCL \u901a\u4fe1\u5668\uff1b<\/li>\n\n\n\n<li>\u540c\u4e00\u5f20 <code>MemoryPlan<\/code> \u4fdd\u8bc1\u6240\u6709 rank \u7684 DTensor \u504f\u79fb\u4e00\u81f4\uff1b<\/li>\n\n\n\n<li>\u4e24\u6876\u68af\u5ea6 AllReduce \u628a\u6df1\u5c42\u901a\u4fe1\u9690\u85cf\u5728\u9996\u5c42\u53cd\u5411\u8ba1\u7b97\u4e2d\uff1b<\/li>\n\n\n\n<li>\u6240\u6709\u901a\u4fe1\u5b50\u56fe\u90fd\u88ab\u6355\u83b7\u8fdb CUDA Graph\uff0c\u8fd0\u884c\u65f6\u628a\u6bcf\u6b21 step \u7684 CPU \u6d3e\u53d1\u538b\u7f29\u6210\u4e00\u6b21 <code>cudaGraphLaunch<\/code>\uff1b<\/li>\n\n\n\n<li>BN \u7edf\u8ba1\u91cf\u3001\u521d\u59cb\u53c2\u6570\u3001\u9a8c\u8bc1\u6307\u6807\u90fd\u901a\u8fc7 NCCL \u505a\u5168\u5c40\u540c\u6b65\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u8fd9\u4e9b\u8bbe\u8ba1\uff0c\u6709\u51e0\u4e2a\u8d2f\u7a7f\u59cb\u7ec8\u7684\u539f\u5219\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e00\uff0c\u9759\u6001\u4f18\u4e8e\u52a8\u6001\u3002<\/strong> \u6240\u6709\u901a\u4fe1\u56fe\u5728\u7f16\u8bd1\u671f\u786e\u5b9a\uff0c\u68af\u5ea6\u5206\u6876\u7b56\u7565\u57fa\u4e8e Region \u800c\u975e\u8fd0\u884c\u65f6\u5f20\u91cf\u5927\u5c0f\uff0c\u907f\u514d\u4e86\u52a8\u6001\u8c03\u6574\u5e26\u6765\u7684\u4e0d\u786e\u5b9a\u6027\u548c\u8c03\u5ea6\u5f00\u9500\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e8c\uff0c\u8fde\u7eed\u4f18\u4e8e\u79bb\u6563\u3002<\/strong> \u5229\u7528 MemoryPlan \u7684\u8bed\u4e49\u5206\u533a\u8bbe\u8ba1\uff0c\u5c06\u540c\u7c7b\u578b\u68af\u5ea6\u8fde\u7eed\u5b58\u653e\uff0c\u4f7f\u5f97\u4e00\u6b21 <code>ncclAllReduce<\/code> \u5c31\u80fd\u5b8c\u6210\u6574\u4e2a\u6876\u7684\u901a\u4fe1\u3002\u8fd9\u4e0d\u4ec5\u51cf\u5c11\u4e86 kernel \u542f\u52a8\u6b21\u6570\uff0c\u4e5f\u907f\u514d\u4e86\u5c0f\u6570\u636e\u5757\u7684\u901a\u4fe1\u6548\u7387\u635f\u5931\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e09\uff0c\u91cd\u53e0\u4f18\u4e8e\u4e32\u884c\u3002<\/strong> \u4e24\u6876\u8bbe\u8ba1\u914d\u5408\u591a\u6d41\u67b6\u6784\uff0c\u8ba9\u6df1\u5c42\u68af\u5ea6\u901a\u4fe1\u4e0e\u9996\u5c42\u53cd\u5411\u8ba1\u7b97\u5728\u7269\u7406\u4e0a\u5e76\u884c\u6267\u884c\u3002\u901a\u4fe1\u65f6\u95f4\u88ab\u8ba1\u7b97\u65f6\u95f4\u8986\u76d6\uff0c\u51e0\u4e4e\u4e0d\u589e\u52a0\u7aef\u5230\u7aef\u8017\u65f6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u56db\uff0c\u591f\u7528\u5c31\u884c\u3002<\/strong> \u4e24\u6876\u4e0d\u662f\u5bf9\u6240\u6709\u573a\u666f\u90fd\u6700\u4f18\u7684\u65b9\u6848\u2014\u2014\u5bf9\u4e8e\u67d0\u4e9b\u7279\u6b8a\u7684\u6a21\u578b\u7ed3\u6784\uff0c\u66f4\u591a\u6876\u6216\u8bb8\u80fd\u5b9e\u73b0\u66f4\u597d\u7684\u91cd\u53e0\u3002\u4f46\u5bf9\u4e8e CNN \u8bad\u7ec3\u8fd9\u4e00\u6838\u5fc3\u573a\u666f\uff0c\u4e24\u6876\u662f\u7b80\u6d01\u4e14\u9ad8\u6548\u7684\u3002\u5b83\u4e0d\u9700\u8981\u590d\u6742\u7684\u6876\u586b\u5145\u903b\u8f91\uff0c\u4e0d\u9700\u8981\u8fd0\u884c\u65f6 hook\uff0c\u4e0d\u9700\u8981\u52a8\u6001\u9608\u503c\u8c03\u6574\u3002\u7b80\u5355\u610f\u5473\u7740\u6b63\u786e\uff0c\u6b63\u786e\u610f\u5473\u7740\u53ef\u590d\u73b0\uff0c\u53ef\u590d\u73b0\u610f\u5473\u7740\u79d1\u5b66\u53ef\u4fe1\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5206\u5e03\u5f0f\u8bad\u7ec3\u662f\u4e00\u4e2a\u590d\u6742\u7684\u7cfb\u7edf\u5de5\u7a0b\u95ee\u9898\uff0c\u4f46\u597d\u7684\u8bbe\u8ba1\u5f80\u5f80\u4e0d\u662f\u590d\u6742\u7684\u3002Tech-Renaissance \u7528\u4e24\u6876\u68af\u5ea6\u901a\u4fe1\u3001\u4e94\u4e2a\u6d41\u3001\u4e00\u5957 Region \u7b97\u5b50\uff0c\u5c31\u628a DDP \u8bad\u7ec3\u505a\u5f97\u9ad8\u6548\u800c\u53ef\u9760\u2014\u2014\u8fd9\u6b63\u662f\u6846\u67b6\u8bbe\u8ba1\u54f2\u5b66\u7684\u4e00\u4e2a\u7f29\u5f71\uff1a<strong>\u7528\u6700\u5c11\u7684\u62bd\u8c61\uff0c\u505a\u6700\u9ad8\u6548\u7684\u4e8b\u60c5<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u4f1a\u628a\u6240\u6709\u8fd9\u4e9b\u8bbe\u8ba1\u4e32\u8d77\u6765\uff0c\u7528\u5b9e\u6d4b\u6570\u636e\u56de\u7b54\u90a3\u4e2a\u6700\u5173\u952e\u7684\u95ee\u9898\uff1a\u4e00\u4e2a\u4eba\u7528 AI \u5199\u51fa\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u5230\u5e95\u80fd\u6bd4 PyTorch \u5feb\u591a\u5c11\uff1f<\/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\u56db \u5f53\u4f60\u628a ResNet-50 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":599,"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-565","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\/565","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=565"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/565\/revisions"}],"predecessor-version":[{"id":660,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/565\/revisions\/660"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/599"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=565"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=565"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=565"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}