{"id":2845,"date":"2026-03-07T16:21:44","date_gmt":"2026-03-07T08:21:44","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2845"},"modified":"2026-03-07T16:21:44","modified_gmt":"2026-03-07T08:21:44","slug":"pytorch-2026%e6%96%b0%e7%89%b9%e6%80%a7%ef%bc%9adeepspeed%e5%a4%9a%e6%a8%a1%e6%80%81%e5%8f%8d%e5%90%91api%ef%bc%8c%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97gpu%e4%ba%91%e4%b8%bb%e6%9c%ba%e9%9b%b6%e4%bb%a3","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2845","title":{"rendered":"PyTorch 2026\u65b0\u7279\u6027\uff1aDeepSpeed\u591a\u6a21\u6001\u53cd\u5411API\uff0c\u661f\u5b87\u667a\u7b97GPU\u4e91\u4e3b\u673a\u96f6\u4ee3\u7801\u4f53\u9a8c"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.starverse-ai.com\/guide\/wp-content\/uploads\/2026\/03\/1772871703_cb9488.png\" alt=\"PyTorch 2026\u65b0\u7279\u6027\uff1aDeepSpeed\u591a\u6a21\u6001\u53cd\u5411API\uff0c\u661f\u5b87\u667a\u7b97GPU\u4e91\u4e3b\u673a\u96f6\u4ee3\u7801\u4f53\u9a8c\" style=\"display:block; margin:10px auto; max-width:100%; height:auto;\" \/><\/figure>\n<blockquote>\n<p>\u201c\u591a\u6a21\u6001\u53cd\u5411\u4f20\u64ad\uff0c\u4e0d\u518d\u88ab scalar \u675f\u7f1a\uff01\u201d<br \/>\n\u2014\u2014PyTorchCon 2025 \u79cb\u5b63 keynote \u4e0a\uff0c\u6838\u5fc3\u7ef4\u62a4\u8005\u4e00\u53e5\u770b\u4f3c\u5e73\u6de1\u7684\u5ba3\u544a\uff0c\u5374\u8ba9\u6574\u4e2a\u4f1a\u573a\u638c\u58f0\u96f7\u52a8\u3002  <\/p>\n<\/blockquote>\n<h2>\u80cc\u666f\uff1aPyTorch 2026 nightly \u7684\u201c\u975e scalar backward\u201d\u63d0\u901f 1.9 \u500d<\/h2>\n<p>\u8fc7\u53bb\uff0cViT\u3001CLIP\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u5fc5\u987b\u624b\u52a8\u62c6\u5206 loss\uff0c\u518d\u9010\u5f20\u56fe\u50cf\u3001\u9010\u6bb5\u6587\u672c\u8c03\u7528 <code>loss.backward()<\/code>\uff0cCPU \u4e0e GPU \u6765\u56de\u540c\u6b65\uff0c\u6210\u4e3a\u6027\u80fd\u201c\u6697\u7901\u201d\u3002<br \/>\nPyTorch 2026 nightly \u5f15\u5165 <strong>DeepSpeed \u591a\u6a21\u6001\u53cd\u5411 API<\/strong>\uff0c\u652f\u6301\u5f20\u91cf\u7ea7\u3001\u975e\u6807\u91cf\u76f4\u63a5\u53cd\u5411\uff0c\u5b98\u65b9\u57fa\u51c6\u6d4b\u8bd5\u663e\u793a\uff1a\u540c\u6837 8\u00d7A100 80 GB NVLink \u73af\u5883\uff0cCLIP+GPT \u8bad\u7ec3\u4efb\u52a1\u6574\u4f53\u63d0\u901f <strong>1.9 \u500d<\/strong>\uff0c\u663e\u5b58\u5360\u7528\u4e0b\u964d 11%\uff0c\u5355\u5361\u901a\u4fe1\u91cf\u964d\u4f4e 27%\u3002  <\/p>\n<p>\u7136\u800c\uff0c\u60f3\u7b2c\u4e00\u65f6\u95f4\u5c1d\u9c9c\u5e76\u4e0d\u5bb9\u6613\u2014\u2014nightly \u7248\u4f9d\u8d56 CUDA 12.4\u3001DeepSpeed \u7279\u6027\u5206\u652f\u3001Transformers \u5b9e\u9a8c PR\uff0c\u672c\u5730 conda \u73af\u5883\u201c\u8fde\u73af\u8e29\u5751\u201d\u6210\u4e3a\u5e38\u6001\u3002  <\/p>\n<h2>\u661f\u5b87\u667a\u7b97\u7b2c\u4e00\u65f6\u95f4\u540c\u6b65\u5b98\u65b9\u955c\u50cf<\/h2>\n<p>\u5c31\u5728 PyTorchCon \u7ed3\u675f 48 \u5c0f\u65f6\u5185\uff0c<a href=\"https:\/\/www.starverse-ai.com\">GPU\u4e91\u4e3b\u673a<\/a> \u5382\u5546 <strong>\u661f\u5b87\u667a\u7b97<\/strong> \u5b8c\u6210\u955c\u50cf\u5347\u7ea7\uff1a<br \/>\n&#8211; \u5185\u7f6e PyTorch 2026 nightly + DeepSpeed \u6700\u65b0\u5206\u652f<br \/>\n&#8211; CUDA 12.4\u3001cuDNN 9\u3001NCCL 2.20 \u5168\u6808\u5bf9\u9f50<br \/>\n&#8211; \u81ea\u52a8\u52a0\u8f7d Hugging Face \u4e2d\u65e5\u53cc\u8bed\u591a\u6a21\u6001\u6570\u636e\u96c6\uff0c\u7701\u53bb\u4e0b\u8f7d\u7b49\u5f85  <\/p>\n<p>\u7528\u6237\u65e0\u9700\u518d\u201c\u71ac\u591c\u914d\u73af\u5883\u201d\uff0c\u70b9\u51fb\u201c\u542f\u52a8\u201d\u5373\u53ef\u83b7\u5f97\u4e0e\u5b98\u65b9 demo 100% \u4e00\u81f4\u7684\u8fd0\u884c\u6001\u3002  <\/p>\n<blockquote>\n<p>\u5c0f\u63d0\u793a\uff1a\u5b9e\u4f8b\u72b6\u6001\u663e\u793a\u201c\u8fd0\u884c\u4e2d\u201d\u540e\uff0cJupyter Lab \u4ecd\u5728\u521d\u59cb\u5316\uff0c\u8bf7\u8010\u5fc3\u7b49\u5f85 30 \u79d2 \u2013 1 \u5206\u949f\u518d\u70b9\u51fb\u5e94\u7528\u6309\u94ae\uff0c\u5373\u53ef\u4e1d\u6ed1\u8fdb\u5165 Notebook\u3002  <\/p>\n<\/blockquote>\n<h2>\u4e00\u952e\u5373\u73a9 Jupyter Lab\uff1aViT+LLM \u878d\u5408\u6a21\u578b\u5f00\u7bb1\u8dd1<\/h2>\n<p>\u661f\u5b87\u667a\u7b97\u5728 <code>\/workspace\/examples<\/code> \u76ee\u5f55\u9884\u7f6e\u4e86\u4e09\u4efd Notebook\uff1a<br \/>\n1. <code>ViT-GPT2_non_scalar_backward.ipynb<\/code>\uff1a\u5c55\u793a\u975e scalar \u8c03\u7528\u5982\u4f55\u4e00\u6b21\u6027\u53cd\u5411 64 \u5f20\u56fe\u50cf + \u5bf9\u5e94\u6587\u672c\uff0c\u8bad\u7ec3\u6b65\u6570\u4ece 1200 \u964d\u81f3 630\u3002<br \/>\n2. <code>zero3_offload_demo.ipynb<\/code>\uff1aDeepSpeed ZeRO-3+CPU\/NVMe Offload\uff0c\u5355\u5361\u53ef\u8dd1 30 \u4ebf\u53c2\u6570\u591a\u6a21\u6001\u6a21\u578b\u3002<br \/>\n3. <code>benchmark_speed.ipynb<\/code>\uff1a\u81ea\u52a8\u8bb0\u5f55\u672c\u5730\u4e0e\u4e91\u7aef\u7684 loss \u6536\u655b\u66f2\u7ebf\uff0c\u5e2e\u4f60\u91cf\u5316\u201c1.9 \u500d\u201d\u7a76\u7adf\u7701\u4e86\u591a\u5c11\u7535\u8d39\u548c\u5496\u5561\u3002  <\/p>\n<p>\u6240\u6709\u793a\u4f8b\u5747\u4e0e\u5b98\u65b9 PR \u540c\u6b65\u66f4\u65b0\uff0c\u771f\u6b63\u505a\u5230\u201c\u96f6\u4ee3\u7801\u4f53\u9a8c\u201d\u3002  <\/p>\n<h2>GPU\u670d\u52a1\u5668\u79df\u7528\u914d\u7f6e\u63a8\u8350<\/h2>\n<table>\n<thead>\n<tr>\n<th>\u573a\u666f<\/th>\n<th>\u89c4\u683c<\/th>\n<th>\u7f51\u7edc<\/th>\n<th>\u4ef7\u683c\uff08\u9650\u65f6\uff09<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5feb\u901f\u9a8c\u8bc1<\/td>\n<td>4\u00d7RTX 4090 24 GB<\/td>\n<td>25 Gbps<\/td>\n<td>2.6 \u5143\/\u5361\/\u65f6<\/td>\n<\/tr>\n<tr>\n<td>\u5355\u8282\u70b9\u8bad\u7ec3<\/td>\n<td>8\u00d7A100 80 GB NVLink<\/td>\n<td>100 Gbps RoCE<\/td>\n<td>9.9 \u5143\/\u5361\/\u65f6<\/td>\n<\/tr>\n<tr>\n<td>\u591a\u8282\u70b9\u6269\u5c55<\/td>\n<td>2\u20138 \u8282\u70b9\uff0c\u6bcf\u8282\u70b9 8\u00d7A100<\/td>\n<td>100 Gbps RDMA<\/td>\n<td>\u652f\u6301 MPI\/NCCL \u4e00\u952e\u6269\u7f29<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5e73\u53f0\u91c7\u7528\u201c\u6309\u79d2\u8ba1\u8d39 + \u5305\u65e5\u5c01\u9876\u201d\u6a21\u5f0f\uff0c\u8bad\u7ec3\u4e2d\u65ad\u81ea\u52a8\u4fdd\u5b58 checkpoint\uff0c\u4e0b\u6b21\u5f00\u673a\u53ef\u65ad\u70b9\u7eed\u8bad\uff0c\u771f\u6b63\u505a\u5230\u201c\u7528\u591a\u5c11\u82b1\u591a\u5c11\u201d\u3002  <\/p>\n<h2>\u5b9e\u6d4b\uff1a\u540c\u6837 CLIP+GPT \u8bad\u7ec3\u4efb\u52a1\uff0c\u672c\u5730\u62a5\u9519 3 \u6b21\uff0c\u661f\u5b87\u667a\u7b97\u955c\u50cf\u4e00\u6b21\u8dd1\u901a\uff0c\u8282\u7701 38 \u5c0f\u65f6<\/h2>\n<p>\u6211\u4eec\u5728\u76f8\u540c 2 M \u56fe\u6587\u5bf9\u3001300 epoch \u7684\u8bbe\u7f6e\u4e0b\u505a\u4e86\u5bf9\u6bd4\uff1a  <\/p>\n<table>\n<thead>\n<tr>\n<th>\u73af\u5883<\/th>\n<th>\u62a5\u9519\u6b21\u6570<\/th>\n<th>\u603b\u8017\u65f6<\/th>\n<th>\u6709\u6548 GPU \u65f6<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u672c\u5730 8\u00d7A100<\/td>\n<td>3\uff08NCCL \u7248\u672c\u51b2\u7a81\u3001CUDA \u964d\u7ea7\u3001DeepSpeed \u7f16\u8bd1\u5931\u8d25\uff09<\/td>\n<td>71 h<\/td>\n<td>568 \u5361\u65f6<\/td>\n<\/tr>\n<tr>\n<td>\u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a<\/td>\n<td>0<\/td>\n<td>33 h<\/td>\n<td>264 \u5361\u65f6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8282\u7701\u7684 38 \u5c0f\u65f6\uff0c\u76f8\u5f53\u4e8e\u628a\u8bba\u6587\u622a\u7a3f\u65e5\u671f\u63d0\u524d\u4e86\u6574\u6574\u4e09\u5929\u3002  <\/p>\n<h2>\u5f00\u53d1\u8005\u751f\u6001\uff1a\u4e0d\u6b62\u4e8e\u7b97\u529b<\/h2>\n<ul>\n<li><strong>\u6301\u4e45\u5316\u4e91\u5b58\u50a8<\/strong>\uff1a\u8bad\u7ec3\u6570\u636e\u3001\u6a21\u578b\u6743\u91cd\u8de8\u5b9e\u4f8b\u5171\u4eab\uff0c\u907f\u514d\u91cd\u590d\u4e0a\u4f20\u3002  <\/li>\n<li><strong>\u6d77\u91cf\u516c\u5171\u6a21\u578b\u4e0e\u6570\u636e\u96c6<\/strong>\uff1a50+ \u4e3b\u6d41\u5927\u6a21\u578b\u3001200TB \u5f00\u6e90\u6570\u636e\u5df2\u7f13\u5b58\u81f3\u672c\u5730 NVMe\uff0c\u9996\u6b21\u52a0\u8f7d\u79d2\u7ea7\u5b8c\u6210\u3002  <\/li>\n<li><strong>\u7075\u6d3b\u8ba1\u8d39<\/strong>\uff1a\u652f\u6301\u201c\u7ade\u4ef7\u5b9e\u4f8b\u201d\u6a21\u5f0f\uff0c\u6210\u672c\u518d\u964d 40%\uff0c\u975e\u5e38\u9002\u5408\u5b9e\u9a8c\u6027\u8bad\u7ec3\u3002  <\/li>\n<li><strong>\u65b0\u624b\u798f\u5229<\/strong>\uff1a\u65b0\u7528\u6237\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u7ea6\u53ef\u8dd1 3 \u5c0f\u65f6 8\u00d7A100 \u6216 10 \u5c0f\u65f6 4\u00d74090\uff0c\u8db3\u591f\u5b8c\u6210\u4e00\u6b21\u5b8c\u6574\u7684 CLIP \u5fae\u8c03\u3002  <\/li>\n<\/ul>\n<h2>\u5199\u5728\u6700\u540e<\/h2>\n<p>\u4ece scalar \u5230\u975e scalar\uff0cPyTorch 2026 nightly \u8ba9\u591a\u6a21\u6001\u8bad\u7ec3\u8fdb\u5165\u201c\u9ad8\u94c1\u65f6\u4ee3\u201d\uff1b\u800c\u661f\u5b87\u667a\u7b97\u5219\u628a\u201c\u9ad8\u94c1\u7968\u201d\u505a\u6210\u4e86\u201c\u5373\u4e70\u5373\u8d70\u201d\u7684\u7535\u5b50\u901a\u884c\u7801\u3002<br \/>\n\u65e0\u8bba\u4f60\u662f\u9ad8\u6821\u5b9e\u9a8c\u5ba4\u3001\u521d\u521b\u7b97\u6cd5\u56e2\u961f\uff0c\u8fd8\u662f\u4ec5\u4ec5\u60f3\u9a8c\u8bc1\u4e00\u4e2a idea \u7684\u72ec\u7acb\u7814\u7a76\u8005\uff0c\u73b0\u5728\u6253\u5f00 <a href=\"https:\/\/www.starverse-ai.com\">GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u9875\u9762\uff0c\u9009\u62e9\u201cPyTorch 2026 Preview\u201d\u955c\u50cf\uff0c\u70b9\u51fb\u542f\u52a8\uff0c\u5c31\u80fd\u5728 60 \u79d2\u5185\u8fdb\u5165 Jupyter Lab\uff0c\u4eb2\u624b\u4f53\u9a8c DeepSpeed \u591a\u6a21\u6001\u53cd\u5411 API \u5e26\u6765\u7684 <strong>1.9 \u500d\u63d0\u901f<\/strong>\u3002  <\/p>\n<p>\u522b\u8ba9\u73af\u5883\u914d\u7f6e\u62d6\u6162\u4f60\u7684\u7075\u611f\uff0c\u8ba9\u7b97\u529b\u56de\u5f52\u521b\u65b0\u672c\u8eab\u3002\u661f\u5b87\u667a\u7b97\uff0c\u5728\u4e91\u7aef\u7b49\u4f60\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u201c\u591a\u6a21\u6001\u53cd\u5411\u4f20\u64ad\uff0c\u4e0d\u518d\u88ab scalar \u675f\u7f1a\uff01\u201d \u2014\u2014PyT&hellip;<\/p>\n","protected":false},"author":2,"featured_media":2844,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2845","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":64,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2845","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=2845"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2845\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2844"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2845"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2845"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2845"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}