{"id":27082,"date":"2025-02-27T22:22:25","date_gmt":"2025-02-27T14:22:25","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=27082"},"modified":"2025-02-27T22:23:55","modified_gmt":"2025-02-27T14:23:55","slug":"dualpipe","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/dualpipe\/","title":{"rendered":"DualPipe\uff1a\u53cc\u5411\u6d41\u6c34\u7ebf\u5e76\u884c\u7b97\u6cd5\uff0c\u63d0\u5347\u5927\u89c4\u6a21AI\u6a21\u578b\u8bad\u7ec3\u6548\u7387\uff08DeepSeek \u5f00\u6e90\u5468\u7b2c\u56db\u5929\uff09"},"content":{"rendered":"<p>DualPipe \u662f\u7531 DeepSeek-AI \u56e2\u961f\u5f00\u53d1\u7684\u4e00\u9879\u5f00\u6e90\u6280\u672f\uff0c\u4e13\u6ce8\u4e8e\u63d0\u5347\u5927\u89c4\u6a21 AI \u6a21\u578b\u8bad\u7ec3\u7684\u6548\u7387\u3002\u5b83\u662f\u4e00\u4e2a\u521b\u65b0\u7684\u53cc\u5411\u6d41\u6c34\u7ebf\u5e76\u884c\u7b97\u6cd5\uff0c\u4e3b\u8981\u7528\u4e8e\u5728 DeepSeek-V3 \u548c R1 \u6a21\u578b\u8bad\u7ec3\u4e2d\u5b9e\u73b0\u8ba1\u7b97\u4e0e\u901a\u4fe1\u7684\u5b8c\u5168\u91cd\u53e0\uff0c\u6709\u6548\u51cf\u5c11\u6d41\u6c34\u7ebf\u4e2d\u7684\u201c\u6c14\u6ce1\u201d\uff08\u5373\u7b49\u5f85\u65f6\u95f4\uff09\uff0c\u4ece\u800c\u52a0\u901f\u8bad\u7ec3\u8fc7\u7a0b\u3002\u8be5\u9879\u76ee\u7531 Jiashi Li\u3001Chengqi Deng \u548c Wenfeng Liang \u4e09\u4eba\u5f00\u53d1\uff0c\u5df2\u5728 GitHub \u4e0a\u5f00\u6e90\uff0c\u53d7\u5230 AI \u6280\u672f\u793e\u533a\u7684\u5173\u6ce8\u3002DualPipe \u7684\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e\u901a\u8fc7\u4f18\u5316\u8c03\u5ea6\uff0c\u4f7f\u6a21\u578b\u8bad\u7ec3\u80fd\u5728\u591a\u8282\u70b9 GPU \u96c6\u7fa4\u4e2d\u9ad8\u6548\u8fd0\u884c\uff0c\u9002\u7528\u4e8e\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\u7684\u6a21\u578b\u8bad\u7ec3\u573a\u666f\uff0c\u4e3aAI\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u63d0\u4f9b\u4e86\u65b0\u7684\u5e76\u884c\u8303\u5f0f\u3002<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u53cc\u5411\u6d41\u6c34\u7ebf\u8c03\u5ea6<\/strong>\uff1a\u652f\u6301\u540c\u65f6\u4ece\u6d41\u6c34\u7ebf\u4e24\u7aef\u8f93\u5165\u5fae\u6279\u6b21\uff0c\u5b9e\u73b0\u8ba1\u7b97\u4e0e\u901a\u4fe1\u7684\u9ad8\u5ea6\u91cd\u53e0\u3002<\/li>\n<li><strong>\u51cf\u5c11\u6d41\u6c34\u7ebf\u6c14\u6ce1<\/strong>\uff1a\u901a\u8fc7\u7b97\u6cd5\u4f18\u5316\uff0c\u964d\u4f4e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u7a7a\u95f2\u7b49\u5f85\u65f6\u95f4\u3002<\/li>\n<li><strong>\u652f\u6301\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3<\/strong>\uff1a\u9002\u914d DeepSeek-V3 \u7b49\u8d85\u5927\u89c4\u6a21\u6a21\u578b\uff0c\u5e94\u5bf9\u4e07\u4ebf\u53c2\u6570\u8bad\u7ec3\u9700\u6c42\u3002<\/li>\n<li><strong>\u8ba1\u7b97\u4e0e\u901a\u4fe1\u91cd\u53e0<\/strong>\uff1a\u5728\u6b63\u5411\u548c\u53cd\u5411\u4f20\u64ad\u4e2d\u5e76\u884c\u5904\u7406\u8ba1\u7b97\u4e0e\u901a\u4fe1\u4efb\u52a1\uff0c\u63d0\u5347 GPU \u5229\u7528\u7387\u3002<\/li>\n<li><strong>\u5f00\u6e90\u4ee3\u7801\u652f\u6301<\/strong>\uff1a\u63d0\u4f9b\u5b8c\u6574 Python \u5b9e\u73b0\uff0c\u5f00\u53d1\u8005\u53ef\u81ea\u7531\u4e0b\u8f7d\u3001\u4fee\u6539\u548c\u96c6\u6210\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>DualPipe \u662f\u4e00\u4e2a\u9762\u5411\u5f00\u53d1\u8005\u7684\u9ad8\u7ea7\u5de5\u5177\uff0c\u4f5c\u4e3a GitHub \u5f00\u6e90\u9879\u76ee\uff0c\u5b83\u6ca1\u6709\u72ec\u7acb\u7684\u56fe\u5f62\u754c\u9762\uff0c\u800c\u662f\u4ee5\u4ee3\u7801\u5e93\u7684\u5f62\u5f0f\u63d0\u4f9b\u7ed9\u7528\u6237\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u7684\u4f7f\u7528\u6307\u5357\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u5feb\u901f\u4e0a\u624b\u5e76\u5c06\u5176\u96c6\u6210\u5230\u81ea\u5df1\u7684 AI \u8bad\u7ec3\u9879\u76ee\u4e2d\u3002<\/p>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>DualPipe \u7684\u5b89\u88c5\u9700\u8981\u4e00\u5b9a\u7684 Python \u548c\u6df1\u5ea6\u5b66\u4e60\u73af\u5883\u57fa\u7840\u3002\u4ee5\u4e0b\u662f\u5177\u4f53\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li><strong>\u73af\u5883\u51c6\u5907<\/strong>\n<ul>\n<li>\u786e\u4fdd\u7cfb\u7edf\u5df2\u5b89\u88c5 Python 3.8 \u6216\u4ee5\u4e0a\u7248\u672c\u3002<\/li>\n<li>\u5b89\u88c5 Git\uff0c\u7528\u4e8e\u4ece GitHub \u4e0b\u8f7d\u4ee3\u7801\u3002<\/li>\n<li>\u5efa\u8bae\u4f7f\u7528\u865a\u62df\u73af\u5883\u4ee5\u907f\u514d\u4f9d\u8d56\u51b2\u7a81\uff0c\u547d\u4ee4\u5982\u4e0b\uff1a\n<pre><code>python -m venv dualpipe_env\r\nsource dualpipe_env\/bin\/activate  # Linux\/Mac\r\ndualpipe_env\\Scripts\\activate  # Windows\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u514b\u9686\u4ee3\u7801\u4ed3\u5e93<\/strong><br \/>\n\u5728\u7ec8\u7aef\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\uff0c\u5c06 DualPipe \u4ed3\u5e93\u4e0b\u8f7d\u5230\u672c\u5730\uff1a<\/p>\n<pre><code>git clone https:\/\/github.com\/deepseek-ai\/DualPipe.git\r\ncd DualPipe<\/code><\/pre>\n<\/li>\n<\/ol>\n<ol start=\"3\">\n<li><strong>\u5b89\u88c5\u4f9d\u8d56<\/strong><br \/>\nDualPipe \u4f9d\u8d56\u5e38\u89c1\u7684\u6df1\u5ea6\u5b66\u4e60\u5e93\uff0c\u5177\u4f53\u4f9d\u8d56\u672a\u5728\u4ed3\u5e93\u4e2d\u660e\u786e\u5217\u51fa\uff0c\u4f46\u6839\u636e\u5176\u529f\u80fd\u63a8\u6d4b\u9700\u8981 PyTorch \u7b49\u73af\u5883\u3002\u53ef\u4ee5\u5c1d\u8bd5\u4ee5\u4e0b\u547d\u4ee4\u5b89\u88c5\u57fa\u7840\u4f9d\u8d56\uff1a<\/p>\n<pre><code>pip install torch torchvision\r\n<\/code><\/pre>\n<p>\u5982\u679c\u9047\u5230\u7f3a\u5c11\u7279\u5b9a\u5e93\u7684\u62a5\u9519\uff0c\u53ef\u6839\u636e\u63d0\u793a\u8fdb\u4e00\u6b65\u5b89\u88c5\u3002<\/li>\n<li><strong>\u9a8c\u8bc1\u5b89\u88c5<\/strong><br \/>\n\u7531\u4e8e DualPipe \u662f\u7b97\u6cd5\u4ee3\u7801\u800c\u975e\u72ec\u7acb\u5e94\u7528\uff0c\u65e0\u6cd5\u76f4\u63a5\u8fd0\u884c\u9a8c\u8bc1\u3002\u4f46\u53ef\u4ee5\u901a\u8fc7\u67e5\u770b\u4ee3\u7801\u6587\u4ef6\uff08\u5982\u00a0<code>dualpipe.py<\/code>\uff09\u786e\u8ba4\u662f\u5426\u5b8c\u6574\u4e0b\u8f7d\u3002<\/li>\n<\/ol>\n<h3>\u4f7f\u7528\u65b9\u6cd5<\/h3>\n<p>DualPipe \u7684\u6838\u5fc3\u662f\u4e00\u4e2a\u8c03\u5ea6\u7b97\u6cd5\uff0c\u5f00\u53d1\u8005\u9700\u8981\u5c06\u5176\u96c6\u6210\u5230\u73b0\u6709\u7684\u6a21\u578b\u8bad\u7ec3\u6846\u67b6\u4e2d\uff08\u5982 PyTorch \u6216 DeepSpeed\uff09\u3002\u4ee5\u4e0b\u662f\u64cd\u4f5c\u6d41\u7a0b\uff1a<\/p>\n<h4>1. \u7406\u89e3\u4ee3\u7801\u7ed3\u6784<\/h4>\n<ul>\n<li>\u6253\u5f00\u00a0<code>DualPipe<\/code>\u00a0\u6587\u4ef6\u5939\uff0c\u4e3b\u8981\u4ee3\u7801\u53ef\u80fd\u4f4d\u4e8e\u00a0<code>dualpipe.py<\/code>\u00a0\u6216\u7c7b\u4f3c\u6587\u4ef6\u4e2d\u3002<\/li>\n<li>\u9605\u8bfb\u4ee3\u7801\u6ce8\u91ca\u548c DeepSeek-V3 \u6280\u672f\u62a5\u544a\uff08\u94fe\u63a5\u89c1 GitHub \u4ed3\u5e93\u63cf\u8ff0\uff09\uff0c\u4e86\u89e3\u7b97\u6cd5\u903b\u8f91\u3002\u62a5\u544a\u4e2d\u63d0\u5230 DualPipe \u7684\u8c03\u5ea6\u793a\u4f8b\uff08\u5982 8 \u4e2a\u6d41\u6c34\u7ebf\u7b49\u7ea7\u548c 20 \u4e2a\u5fae\u6279\u6b21\uff09\u3002<\/li>\n<\/ul>\n<h4>2. \u96c6\u6210\u5230\u8bad\u7ec3\u6846\u67b6<\/h4>\n<ul>\n<li><strong>\u51c6\u5907\u6a21\u578b\u548c\u6570\u636e<\/strong>\uff1a\u5047\u8bbe\u4f60\u5df2\u6709\u57fa\u4e8e PyTorch \u7684\u6a21\u578b\u548c\u6570\u636e\u96c6\u3002<\/li>\n<li><strong>\u4fee\u6539\u8bad\u7ec3\u5faa\u73af<\/strong>\uff1a\u5c06 DualPipe \u7684\u8c03\u5ea6\u903b\u8f91\u5d4c\u5165\u5230\u8bad\u7ec3\u4ee3\u7801\u4e2d\u3002\u4ee5\u4e0b\u662f\u4e00\u4e2a\u7b80\u5316\u7684\u793a\u4f8b\uff1a\n<pre><code># \u4f2a\u4ee3\u7801\u793a\u4f8b\r\nfrom dualpipe import DualPipeScheduler  # \u5047\u8bbe\u6a21\u5757\u540d\r\nimport torch\r\n# \u521d\u59cb\u5316\u6a21\u578b\u548c\u6570\u636e\r\nmodel = MyModel().cuda()\r\noptimizer = torch.optim.Adam(model.parameters())\r\ndata_loader = MyDataLoader()\r\n# \u521d\u59cb\u5316 DualPipe \u8c03\u5ea6\u5668\r\nscheduler = DualPipeScheduler(num_ranks=8, num_micro_batches=20)\r\n# \u8bad\u7ec3\u5faa\u73af\r\nfor epoch in range(num_epochs):\r\nscheduler.schedule(model, data_loader, optimizer)  # \u8c03\u7528 DualPipe \u8c03\u5ea6\r\n<\/code><\/pre>\n<\/li>\n<li>\u5177\u4f53\u5b9e\u73b0\u9700\u8981\u6839\u636e\u5b9e\u9645\u4ee3\u7801\u8c03\u6574\uff0c\u5efa\u8bae\u53c2\u8003 GitHub \u4ed3\u5e93\u4e2d\u7684\u793a\u4f8b\uff08\u82e5\u6709\uff09\u3002<\/li>\n<\/ul>\n<h4>3. \u914d\u7f6e\u786c\u4ef6\u73af\u5883<\/h4>\n<ul>\n<li>DualPipe \u8bbe\u8ba1\u7528\u4e8e\u591a\u8282\u70b9 GPU \u96c6\u7fa4\uff0c\u63a8\u8350\u4f7f\u7528\u81f3\u5c11 8 \u4e2a GPU\uff08\u5982 NVIDIA H800\uff09\u3002<\/li>\n<li>\u786e\u4fdd\u96c6\u7fa4\u652f\u6301 InfiniBand \u6216 NVLink\uff0c\u4ee5\u4fbf\u5145\u5206\u5229\u7528\u901a\u4fe1\u4f18\u5316\u3002<\/li>\n<\/ul>\n<h4>4. \u8fd0\u884c\u4e0e\u8c03\u8bd5<\/h4>\n<ul>\n<li>\u5728\u7ec8\u7aef\u8fd0\u884c\u8bad\u7ec3\u811a\u672c\uff1a\n<pre><code>python train_with_dualpipe.py\r\n<\/code><\/pre>\n<\/li>\n<li>\u89c2\u5bdf\u65e5\u5fd7\u8f93\u51fa\uff0c\u68c0\u67e5\u8ba1\u7b97\u4e0e\u901a\u4fe1\u662f\u5426\u6210\u529f\u91cd\u53e0\u3002\u82e5\u6709\u6027\u80fd\u74f6\u9888\uff0c\u53ef\u8c03\u6574\u5fae\u6279\u6b21\u6570\u91cf\u6216\u6d41\u6c34\u7ebf\u7b49\u7ea7\u3002<\/li>\n<\/ul>\n<h3>\u7279\u8272\u529f\u80fd\u64cd\u4f5c<\/h3>\n<h4>\u53cc\u5411\u6d41\u6c34\u7ebf\u8c03\u5ea6<\/h4>\n<ul>\n<li>\u5728\u914d\u7f6e\u6587\u4ef6\u6216\u4ee3\u7801\u4e2d\u8bbe\u7f6e\u00a0<code>num_ranks<\/code>\uff08\u6d41\u6c34\u7ebf\u7b49\u7ea7\u6570\uff09\u548c\u00a0<code>num_micro_batches<\/code>\uff08\u5fae\u6279\u6b21\u6570\uff09\u3002<\/li>\n<li>\u793a\u4f8b\u914d\u7f6e\uff1a8 \u4e2a\u7b49\u7ea7\u300120 \u4e2a\u5fae\u6279\u6b21\uff0c\u53ef\u53c2\u8003\u6280\u672f\u62a5\u544a\u4e2d\u7684\u8c03\u5ea6\u56fe\u3002<\/li>\n<\/ul>\n<h4>\u8ba1\u7b97\u901a\u4fe1\u91cd\u53e0<\/h4>\n<ul>\n<li>\u65e0\u9700\u624b\u52a8\u5e72\u9884\uff0cDualPipe \u81ea\u52a8\u5c06\u6b63\u5411\u8ba1\u7b97\uff08\u5982\u00a0<code>F<\/code>\uff09\u4e0e\u53cd\u5411\u8ba1\u7b97\uff08\u5982\u00a0<code>B<\/code>\uff09\u7684\u901a\u4fe1\u4efb\u52a1\u91cd\u53e0\u3002<\/li>\n<li>\u68c0\u67e5\u65e5\u5fd7\u4e2d\u7684\u65f6\u95f4\u6233\uff0c\u786e\u8ba4\u901a\u4fe1\u65f6\u95f4\u88ab\u9690\u85cf\u5728\u8ba1\u7b97\u4e2d\u3002<\/li>\n<\/ul>\n<h4>\u51cf\u5c11\u6d41\u6c34\u7ebf\u6c14\u6ce1<\/h4>\n<ul>\n<li>\u901a\u8fc7\u8c03\u6574\u5fae\u6279\u6b21\u5927\u5c0f\uff08\u5982\u4ece 20 \u8c03\u6574\u5230 16\uff09\uff0c\u89c2\u5bdf\u8bad\u7ec3\u65f6\u95f4\u53d8\u5316\uff0c\u627e\u5230\u6700\u4f73\u914d\u7f6e\u3002<\/li>\n<\/ul>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li><strong>\u786c\u4ef6\u8981\u6c42<\/strong>\uff1a\u5355\u673a\u5355\u5361\u65e0\u6cd5\u5145\u5206\u53d1\u6325 DualPipe \u4f18\u52bf\uff0c\u5efa\u8bae\u591a GPU \u73af\u5883\u3002<\/li>\n<li><strong>\u6587\u6863\u652f\u6301<\/strong>\uff1a\u76ee\u524d GitHub \u9875\u9762\u4fe1\u606f\u8f83\u5c11\uff0c\u5efa\u8bae\u7ed3\u5408 DeepSeek-V3 \u6280\u672f\u62a5\u544a\uff08arXiv: 2412.19437\uff09\u6df1\u5165\u5b66\u4e60\u3002<\/li>\n<li><strong>\u793e\u533a\u652f\u6301<\/strong>\uff1a\u53ef\u5728 GitHub Issues \u9875\u9762\u63d0\u95ee\uff0c\u6216\u53c2\u8003 X \u5e73\u53f0\u4e0a\u76f8\u5173\u8ba8\u8bba\uff08\u5982 @deepseek_ai \u7684\u5e16\u5b50\uff09\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u4e0a\u8ff0\u6b65\u9aa4\uff0c\u5f00\u53d1\u8005\u53ef\u4ee5\u5c06 DualPipe \u96c6\u6210\u5230\u81ea\u5df1\u7684\u9879\u76ee\u4e2d\uff0c\u663e\u8457\u63d0\u5347\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>DualPipe \u662f\u7531 DeepSeek-AI \u56e2\u961f\u5f00\u53d1\u7684\u4e00\u9879\u5f00\u6e90\u6280\u672f\uff0c\u4e13\u6ce8\u4e8e\u63d0\u5347\u5927\u89c4\u6a21 AI \u6a21\u578b\u8bad\u7ec3\u7684\u6548\u7387\u3002\u5b83\u662f\u4e00\u4e2a\u521b\u65b0\u7684\u53cc\u5411\u6d41\u6c34\u7ebf\u5e76\u884c\u7b97\u6cd5\uff0c\u4e3b\u8981\u7528\u4e8e\u5728 DeepSeek-V3 \u548c R1 \u6a21\u578b\u8bad\u7ec3\u4e2d\u5b9e\u73b0\u8ba1\u7b97\u4e0e\u901a\u4fe1\u7684\u5b8c\u5168\u91cd\u53e0\uff0c\u6709\u6548\u51cf\u5c11\u6d41\u6c34&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61930,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230],"class_list":["post-27082","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/27082","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=27082"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/27082\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media\/61930"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=27082"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=27082"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=27082"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}