{"id":449,"date":"2026-07-08T02:46:48","date_gmt":"2026-07-07T18:46:48","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=449"},"modified":"2026-07-08T16:01:27","modified_gmt":"2026-07-08T08:01:27","slug":"%e9%9d%99%e6%80%81%e5%9b%be-vs-%e5%8a%a8%e6%80%81%e5%9b%be","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/449\/","title":{"rendered":"(5) \u9759\u6001\u56fe Vs \u52a8\u6001\u56fe\uff1a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u4e24\u79cd\u4e16\u754c\u89c2"},"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\u4e94<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0a\u4e00\u7bc7\u6211\u4eec\u628a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u62c6\u89e3\u6210\u4e86\u5f20\u91cf\u3001\u7b97\u5b50\u3001\u81ea\u52a8\u5fae\u5206\u3001\u8ba1\u7b97\u56fe\u3001\u6570\u636e\u7ba1\u7ebf\u3001\u5206\u5e03\u5f0f\u8bad\u7ec3\u8fd9\u51e0\u4e2a\u6838\u5fc3\u90e8\u4ef6\u3002\u5176\u4e2d\u6709\u4e00\u4e2a\u6982\u5ff5\u88ab\u6211\u5f53\u4f5c&#8221;\u80cc\u666f&#8221;\u4e00\u7b14\u5e26\u8fc7\uff0c\u4f46\u5b83\u5b9e\u9645\u4e0a\u51b3\u5b9a\u4e86\u4e00\u4e2a\u6846\u67b6\u7684<strong>\u57fa\u672c\u6c14\u8d28<\/strong>\u2014\u2014\u90a3\u5c31\u662f\u8ba1\u7b97\u56fe\u5230\u5e95\u662f<strong>\u9759\u6001<\/strong>\u7684\uff0c\u8fd8\u662f<strong>\u52a8\u6001<\/strong>\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u9700\u8981\u660e\u767d\u7684\u662f\uff0c\u8fd9\u662f\u4e24\u79cd\u4e0d\u540c\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e16\u754c\u89c2\uff0c\u5404\u6709\u957f\u5904\uff0c\u4f46\u4f9d\u7136\u6709\u529e\u6cd5\u517c\u5177\u4e24\u8005\u7684\u4f18\u70b9\uff0c\u4ee5\u81f3\u4e8e\u5728\u5f53\u4e0b\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u9886\u57df\u91cc\uff0c\u4e24\u8005\u7684\u5dee\u8ddd\u6b63\u5728\u7f29\u5c0f\u3001\u8fb9\u754c\u6b63\u5728\u53d8\u5f97\u6a21\u7cca\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u8ba1\u7b97\u56fe\uff1a\u6846\u67b6\u91cc\u7684&#8221;\u65bd\u5de5\u56fe\u7eb8&#8221;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8ba8\u8bba\u9759\u6001\u56fe\u548c\u52a8\u6001\u56fe\u4e4b\u524d\uff0c\u5148\u7edf\u4e00\u4e00\u4e0b\u4ec0\u4e48\u53eb<strong>\u8ba1\u7b97\u56fe\uff08Computational Graph\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u65e0\u8bba\u591a\u590d\u6742\uff0c\u672c\u8d28\u4e0a\u90fd\u53ef\u4ee5\u62c6\u6210\u4e00\u7cfb\u5217\u57fa\u672c\u8fd0\u7b97\uff1a\u77e9\u9635\u4e58\u6cd5\u3001\u5377\u79ef\u3001\u52a0\u6cd5\u3001\u6fc0\u6d3b\u51fd\u6570\u3001\u635f\u5931\u51fd\u6570\u2026\u2026\u5982\u679c\u628a\u8fd9\u4e9b\u8fd0\u7b97\u753b\u6210\u8282\u70b9\uff0c\u628a\u5f20\u91cf\u6570\u636e\u6d41\u52a8\u753b\u6210\u6709\u5411\u8fb9\uff0c\u5c31\u5f97\u5230\u4e00\u5f20<strong>\u6709\u5411\u65e0\u73af\u56fe<\/strong>\u2014\u2014\u8fd9\u5c31\u662f\u8ba1\u7b97\u56fe\u3002\u8282\u70b9\u662f\u7b97\u5b50\uff0c\u8fb9\u662f\u6570\u636e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba1\u7b97\u56fe\u7684\u4ef7\u503c\u5728\u4e8e\u5b83\u628a&#8221;\u6a21\u578b\u957f\u4ec0\u4e48\u6837&#8221;\u548c&#8221;\u600e\u4e48\u6267\u884c&#8221;\u5206\u79bb\u5f00\u4e86\u3002\u6709\u4e86\u8fd9\u5f20\u56fe\uff0c\u6846\u67b6\u5c31\u53ef\u4ee5\u505a\u5f88\u591a\u4e8b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u81ea\u52a8\u5fae\u5206<\/strong>\uff1a\u4ece\u635f\u5931\u8282\u70b9\u53cd\u5411\u904d\u5386\u56fe\uff0c\u6cbf\u94fe\u5f0f\u6cd5\u5219\u4f20\u64ad\u68af\u5ea6\uff1b<\/li>\n\n\n\n<li><strong>\u56fe\u4f18\u5316<\/strong>\uff1a\u5e38\u91cf\u6298\u53e0\u3001\u6b7b\u4ee3\u7801\u6d88\u9664\u3001\u516c\u5171\u5b50\u8868\u8fbe\u5f0f\u6d88\u9664\uff1b<\/li>\n\n\n\n<li><strong>\u7b97\u5b50\u878d\u5408<\/strong>\uff1a\u628a\u76f8\u90bb\u7684 conv\u3001bias\u3001batchnorm\u3001relu \u5408\u5e76\u6210\u4e00\u4e2a kernel\uff1b<\/li>\n\n\n\n<li><strong>\u5185\u5b58\u89c4\u5212<\/strong>\uff1a\u5206\u6790\u5f20\u91cf\u751f\u547d\u5468\u671f\uff0c\u590d\u7528\u663e\u5b58\uff1b<\/li>\n\n\n\n<li><strong>\u5e76\u884c\u8c03\u5ea6<\/strong>\uff1a\u628a\u4e0d\u76f8\u5173\u7684\u5b50\u56fe\u6254\u5230\u4e0d\u540c\u6d41\u4e0a\u5e76\u53d1\u6267\u884c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6362\u53e5\u8bdd\u8bf4\uff0c\u8ba1\u7b97\u56fe\u662f\u6846\u67b6\u505a\u5168\u5c40\u4f18\u5316\u7684\u6293\u624b\u3002\u6ca1\u6709\u56fe\uff0c\u6846\u67b6\u5c31\u53ea\u80fd\u5728\u5c40\u90e8&#8221;\u89c1\u4e00\u4e2a\u7b97\u5b50\u4f18\u5316\u4e00\u4e2a\u7b97\u5b50&#8221;\uff1b\u6709\u4e86\u56fe\uff0c\u6846\u67b6\u624d\u80fd&#8221;\u653e\u773c\u5168\u5c40&#8221;\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u9759\u6001\u56fe\u4e0e\u52a8\u6001\u56fe\uff1a\u4e24\u79cd\u4e16\u754c\u89c2<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5386\u53f2\u4e0a\uff0c\u6846\u67b6\u56f4\u7ed5&#8221;\u8fd9\u5f20\u56fe\u4ec0\u4e48\u65f6\u5019\u5efa\u3001\u600e\u4e48\u7528&#8221;\u5206\u6210\u4e86\u4e24\u6d3e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 \u9759\u6001\u56fe\uff1a\u5148\u753b\u56fe\uff0c\u518d\u65bd\u5de5<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9759\u6001\u56fe\uff08Static Graph\uff0cdefine-and-run\uff09<\/strong>\u7684\u601d\u8def\u662f\uff1a\u7528\u6237\u5148\u7528\u67d0\u79cd\u65b9\u5f0f\u628a\u6574\u5f20\u8ba1\u7b97\u56fe\u63cf\u8ff0\u51fa\u6765\uff0c\u6846\u67b6\u5bf9\u8fd9\u5f20\u56fe\u505a\u7f16\u8bd1\u3001\u4f18\u5316\u3001\u5185\u5b58\u89c4\u5212\uff0c\u7136\u540e\u518d\u5582\u6570\u636e\u6267\u884c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">TensorFlow 1.x \u7684 <code>tf.Graph<\/code> + <code>tf.Session<\/code>\u3001Theano \u7684\u7b26\u53f7\u56fe\uff0c\u90fd\u662f\u8fd9\u4e2a\u8def\u7ebf\u7684\u4ee3\u8868\u3002\u7528\u6237\u5199\u7684\u4ee3\u7801\u66f4\u50cf\u662f\u5728&#8221;\u58f0\u660e&#8221;\u4e00\u5f20\u56fe\uff0c\u800c\u4e0d\u662f\u5728&#8221;\u6267\u884c&#8221;\u6307\u4ee4\u3002\u4f18\u70b9\u548c\u7f3a\u70b9\u90fd\u5f88\u9c9c\u660e\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4f18\u70b9\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5168\u5c40\u4f18\u5316\u7a7a\u95f4\u5927<\/strong>\u3002\u56e0\u4e3a\u56fe\u5728\u8fd0\u884c\u524d\u5df2\u7ecf\u5b8c\u6574\u53ef\u89c1\uff0c\u7f16\u8bd1\u5668\u53ef\u4ee5\u505a\u8de8\u7b97\u5b50\u7684\u5e38\u91cf\u6298\u53e0\u3001\u7b97\u5b50\u878d\u5408\u3001\u5e03\u5c40\u8f6c\u6362\u3001\u5185\u5b58\u590d\u7528\u89c4\u5212\u3002<\/li>\n\n\n\n<li><strong>\u90e8\u7f72\u53cb\u597d<\/strong>\u3002\u56fe\u53ef\u4ee5\u5bfc\u51fa\u4e3a SavedModel\u3001ONNX\u3001TensorRT Plan \u7b49\u683c\u5f0f\uff0c\u8131\u79bb Python \u8fd0\u884c\u3002<\/li>\n\n\n\n<li><strong>\u66f4\u5bb9\u6613\u505a\u786e\u5b9a\u6027\u8bad\u7ec3<\/strong>\u3002\u56e0\u4e3a\u56fe\u7ed3\u6784\u548c kernel \u5e8f\u5217\u5728\u8fd0\u884c\u524d\u5df2\u7ecf\u786e\u5b9a\uff0c\u5982\u679c\u518d\u914d\u5408\u9759\u6001\u5185\u5b58\u89c4\u5212\u3001\u56fa\u5b9a\u8c03\u5ea6\u548c\u786e\u5b9a\u6027\u7b97\u5b50\u9009\u62e9\uff0c\u5c31\u66f4\u5bb9\u6613\u8ba9\u540c\u4e00\u7a0b\u5e8f\u591a\u6b21\u8fd0\u884c\u5f97\u5230\u4e00\u81f4\u7ed3\u679c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7f3a\u70b9\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5bf9\u521d\u5b66\u8005\u4e0d\u53cb\u597d<\/strong>\u3002\u9759\u6001\u56fe\u7684\u601d\u8def\u5bf9\u4e8e\u5c1a\u4e0d\u7406\u89e3\u8ba1\u7b97\u56fe\u4e3a\u4f55\u7269\u7684\u521d\u5b66\u8005\u800c\u8a00\u53ef\u8c13\u76f8\u5f53\u8be1\u5f02\uff0c\u5f88\u591a\u4eba\u65e0\u6cd5\u7406\u89e3\u201c\u4e3a\u4ec0\u4e48\u6211\u8981\u5148\u5b9a\u4e49session\u518drun\u201d\u3001\u201c\u4e3a\u4ec0\u4e48\u6211\u5199\u4e86\u4e00\u4e2a\u8fd0\u7b97\u5374\u6ca1\u6709\u6267\u884c\u201d\u3002<\/li>\n\n\n\n<li><strong>\u8c03\u8bd5\u4f53\u9a8c\u76f8\u5bf9\u8f83\u5dee<\/strong>\u3002\u4f60\u5199\u7684\u4e0d\u662f&#8221;\u666e\u901a Python \u7a0b\u5e8f&#8221;\uff0c\u800c\u662f&#8221;\u63cf\u8ff0\u56fe\u7684\u7a0b\u5e8f&#8221;\u3002\u51fa\u9519\u65f6\u5f88\u96be\u76f4\u63a5 <code>print<\/code> \u4e2d\u95f4\u5f20\u91cf\uff0c\u4e5f\u5f88\u96be\u7528 <code>pdb<\/code> \u5355\u6b65\u8ddf\u8fdb\u3002<\/li>\n\n\n\n<li><strong>\u52a8\u6001\u903b\u8f91\u8868\u8fbe\u56f0\u96be<\/strong>\u3002Python \u91cc\u7684 <code>if<\/code>\u3001<code>for<\/code>\u3001\u9012\u5f52\u3001\u6570\u636e\u4f9d\u8d56\u63a7\u5236\u6d41\uff0c\u5728\u9759\u6001\u56fe\u91cc\u9700\u8981 special \u5904\u7406\uff08\u6bd4\u5982 <code>tf.cond<\/code>\u3001<code>tf.while_loop<\/code>\uff09\uff0c\u5b66\u4e60\u66f2\u7ebf\u9661\u5ced\u3002<\/li>\n\n\n\n<li><strong>\u56fe\u6784\u5efa\u6709\u5f00\u9500<\/strong>\u3002\u590d\u6742\u6a21\u578b\u6784\u56fe\u548c\u7f16\u8bd1\u53ef\u80fd\u8981\u51e0\u5341\u79d2\u751a\u81f3\u66f4\u4e45\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u52a8\u6001\u56fe\uff1a\u8fb9\u65bd\u5de5\uff0c\u8fb9\u753b\u56fe<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u52a8\u6001\u56fe\uff08Dynamic Graph\uff0cdefine-by-run\uff09<\/strong>\u5219\u53cd\u5176\u9053\u800c\u884c\u4e4b\uff1a\u4f60\u5199\u4e00\u884c\u4ee3\u7801\uff0c\u6846\u67b6\u5c31\u6267\u884c\u4e00\u884c\u3001\u5efa\u4e00\u4e2a\u8282\u70b9\uff1b\u4e2d\u95f4\u5f20\u91cf\u7acb\u523b\u53ef\u89c1\uff0c\u8c03\u8bd5\u70b9\u54ea\u91cc\u6253\u54ea\u91cc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch Eager\u3001TensorFlow Eager\u3001Chainer\u3001DyNet \u90fd\u5c5e\u4e8e\u8fd9\u4e00\u6d3e\u3002PyTorch \u7684\u6210\u529f\u5f88\u5927\u7a0b\u5ea6\u4e0a\u5f52\u529f\u4e8e\u52a8\u6001\u56fe\u5e26\u6765\u7684&#8221;\u6240\u5199\u5373\u6240\u5f97&#8221;\u4f53\u9a8c\u3002\u7814\u7a76\u8005\u53ef\u4ee5\u50cf\u5199 NumPy \u4e00\u6837\u5199\u6a21\u578b\uff0c\u53ef\u4ee5\u7528\u539f\u751f Python \u63a7\u5236\u6d41\uff0c\u53ef\u4ee5\u968f\u65f6\u968f\u5730 <code>print(x.shape)<\/code> \u6216\u8005 <code>import pdb; pdb.set_trace()<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4f18\u70b9\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7b26\u5408\u76f4\u89c9<\/strong>\u3002\u548c\u5199\u666e\u901a Python \u7a0b\u5e8f\u51e0\u4e4e\u6ca1\u6709\u533a\u522b\u3002<\/li>\n\n\n\n<li><strong>\u8c03\u8bd5\u65b9\u4fbf<\/strong>\u3002\u4efb\u4f55\u4e2d\u95f4\u7ed3\u679c\u90fd\u80fd\u7acb\u5373\u67e5\u770b\u3002<\/li>\n\n\n\n<li><strong>\u52a8\u6001\u6a21\u578b\u53cb\u597d<\/strong>\u3002\u53d8\u957f\u5e8f\u5217\u3001\u52a8\u6001\u7ed3\u6784\u3001\u6761\u4ef6\u5206\u652f\u3001\u7a00\u758f\u8ba1\u7b97\u90fd\u53ef\u4ee5\u81ea\u7136\u8868\u8fbe\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7f3a\u70b9\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7f3a\u4e4f\u5168\u5c40\u89c6\u91ce<\/strong>\u3002\u6846\u67b6\u770b\u5230\u7684\u662f\u4e00\u6b21\u4e00\u6b21\u72ec\u7acb\u7684\u7b97\u5b50\u8c03\u7528\uff0c\u5f88\u96be\u505a\u8de8\u7b97\u5b50\u7684\u5168\u5c40\u4f18\u5316\u3002<\/li>\n\n\n\n<li><strong>\u8c03\u5ea6\u5f00\u9500\u5927<\/strong>\u3002\u6bcf\u4e2a\u7b97\u5b50\u90fd\u8981\u8d70\u4e00\u904d Python \u6d3e\u53d1\u3001C++ dispatcher\u3001CUDA kernel launch\uff0c\u5c0f\u7b97\u5b50\u5bc6\u96c6\u65f6 CPU \u8c03\u5ea6\u4f1a\u6210\u4e3a\u74f6\u9888\u3002<\/li>\n\n\n\n<li><strong>\u4e2d\u95f4\u7ed3\u679c\u5e38\u9a7b\u663e\u5b58<\/strong>\u3002\u52a8\u6001\u56fe\u4e3a\u4e86\u53cd\u5411\u4f20\u64ad\uff0c\u901a\u5e38\u8981\u628a\u524d\u5411\u4e2d\u95f4\u7ed3\u679c\u90fd\u4fdd\u7559\u4e0b\u6765\uff0c\u5185\u5b58\u5360\u7528\u66f4\u5927\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 \u8fb9\u754c\u6b63\u5728\u6a21\u7cca<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6709\u610f\u601d\u7684\u662f\uff0c\u8fd9\u573a&#8221;\u9759\u6001 vs \u52a8\u6001&#8221;\u4e4b\u4e89\u5e76\u6ca1\u6709\u4ee5\u4e00\u65b9\u7684\u5f7b\u5e95\u80dc\u5229\u544a\u7ec8\uff0c\u800c\u662f\u8d70\u5411\u4e86<strong>\u878d\u5408<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">TensorFlow 2.x \u9ed8\u8ba4\u4f7f\u7528 Eager Execution\uff0c\u4f46\u63d0\u4f9b\u4e86 <code>@tf.function<\/code>\uff0c\u5728\u4fdd\u7559\u52a8\u6001\u7f16\u7a0b\u4f53\u9a8c\u7684\u540c\u65f6\u628a\u51fd\u6570\u7f16\u8bd1\u6210\u9759\u6001\u56fe\u3002PyTorch 2.0 \u5219\u5728 2023 \u5e74\u63a8\u51fa\u4e86 <code>torch.compile<\/code>\uff1a\u524d\u7aef TorchDynamo \u6355\u83b7 FX \u56fe\uff0cAOTAutograd \u5904\u7406\u524d\u5411\u4e0e\u53cd\u5411\uff0c\u540e\u7aef Inductor \u505a\u7b97\u5b50\u878d\u5408\u3001kernel \u751f\u6210\uff0c\u751a\u81f3\u5c1d\u8bd5 CUDA Graph \u6355\u83b7\u3002JAX \u4ece\u8bbe\u8ba1\u4e4b\u521d\u5c31\u9f13\u52b1\u7528\u6237\u7528 <code>jax.jit<\/code> \u628a\u52a8\u6001 Python \u51fd\u6570\u7f16\u8bd1\u6210 XLA \u9759\u6001\u56fe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u73b0\u5728\u7684\u5c40\u9762\u4e0d\u662f&#8221;\u9759\u6001\u56fe\u6846\u67b6&#8221;\u548c&#8221;\u52a8\u6001\u56fe\u6846\u67b6&#8221;\u6cfe\u6e2d\u5206\u660e\uff0c\u800c\u662f<strong>\u52a8\u6001\u6846\u67b6\u7eb7\u7eb7\u957f\u51fa\u7f16\u8bd1\u6a21\u5f0f\uff0c\u9759\u6001\u6846\u67b6\u4e5f\u7eb7\u7eb7\u8865 eager \u4f53\u9a8c<\/strong>\u3002\u8fd9\u662f\u4e00\u4e2a\u5927\u5bb6\u90fd\u627f\u8ba4\u7684\u4e8b\u5b9e\uff1a<strong>\u52a8\u6001\u56fe\u7684\u7075\u6d3b\u6027\u662f\u4ee5\u727a\u7272\u5168\u5c40\u4f18\u5316\u7a7a\u95f4\u4e3a\u4ee3\u4ef7\u7684\uff1b\u800c\u9759\u6001\u56fe\u7684\u6027\u80fd\u4f18\u52bf\uff0c\u5219\u5efa\u7acb\u5728&#8221;\u7f16\u8bd1\u671f\u786e\u5b9a\u4e00\u5207&#8221;\u7684\u524d\u63d0\u4e4b\u4e0a\u3002<\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u4e3a\u4ec0\u4e48 Tech-Renaissance \u9009\u62e9\u9759\u6001\u56fe\u7f16\u8bd1<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bb2\u6e05\u695a\u4e86\u901a\u7528\u80cc\u666f\uff0c\u73b0\u5728\u56de\u5230\u6211\u4eec\u7684\u6846\u67b6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u5b9a\u4f4d\u975e\u5e38\u660e\u786e\uff1a\u5b83\u4e0d\u662f\u7ed9\u7814\u7a76\u8005\u63d0\u4f9b\u7075\u6d3b\u8c03\u8bd5\u73a9\u5177\u7684\u6846\u67b6\uff0c\u800c\u662f\u4e00\u4e2a<strong>\u8ffd\u6c42\u6781\u81f4\u8bad\u7ec3\u541e\u5410\u91cf\u7684\u751f\u4ea7\u7ea7\u8bad\u7ec3\u6846\u67b6<\/strong>\u3002\u8fd9\u4e2a\u5b9a\u4f4d\u51b3\u5b9a\u4e86\u6211\u4eec\u5fc5\u987b\u4ece\u67b6\u6784\u4e0a\u4f18\u5148\u4fdd\u7559\u5168\u5c40\u4f18\u5316\u7a7a\u95f4\u3002\u9759\u6001\u56fe\u4e0d\u662f&#8221;\u4e3a\u4e86\u4e0d\u540c\u800c\u4e0d\u540c&#8221;\uff0c\u800c\u662f\u6211\u4eec\u540e\u7eed\u4e00\u7cfb\u5217\u6838\u5fc3\u80fd\u529b\u7684<strong>\u5fc5\u8981\u524d\u63d0<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u8bf4\uff0c\u4ee5\u4e0b\u4e09\u4ef6\u4e8b\uff0c\u79bb\u5f00\u4e86\u9759\u6001\u56fe\u51e0\u4e4e\u4e0d\u53ef\u80fd\u505a\u597d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 CUDA Graph \u5168\u6355\u83b7<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">CUDA Graph \u662f NVIDIA GPU \u63d0\u4f9b\u7684\u4e00\u79cd\u673a\u5236\uff1a\u628a\u4e00\u6bb5 GPU \u5de5\u4f5c\uff08\u591a\u4e2a kernel\u3001memcpy\u3001\u4e8b\u4ef6\u540c\u6b65\uff09<strong>\u4e00\u6b21\u6027\u6355\u83b7\u4e0b\u6765<\/strong>\uff0c\u4e4b\u540e\u901a\u8fc7 <code>cudaGraphLaunch<\/code> \u53cd\u590d\u91cd\u653e\uff0c\u4ece\u800c\u6d88\u9664\u6bcf\u6b21 kernel \u7684 CPU \u6d3e\u53d1\u5f00\u9500\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46 CUDA Graph \u6709\u51e0\u4e2a\u786c\u6027\u7ea6\u675f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8f93\u5165\u8f93\u51fa\u5f20\u91cf\u7684\u865a\u62df\u5730\u5740\u5fc5\u987b\u56fa\u5b9a<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>\u6355\u83b7\u671f\u95f4\u4e0d\u80fd\u52a8\u6001\u5206\u914d\u663e\u5b58<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>\u63a7\u5236\u6d41\u5fc5\u987b\u662f\u9759\u6001\u7684<\/strong>\uff0c\u4e0d\u80fd\u6709\u6761\u4ef6\u5206\u652f\u6216\u52a8\u6001\u5faa\u73af\uff1b<\/li>\n\n\n\n<li><strong>\u4e0d\u80fd\u51fa\u73b0 host-device \u540c\u6b65<\/strong>\uff08\u6bd4\u5982 <code>.item()<\/code> \u53d6\u6807\u91cf\u5230 CPU\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u7ea6\u675f\u5929\u7136\u5730\u6307\u5411\u9759\u6001\u56fe\uff1a\u53ea\u6709\u5728\u7f16\u8bd1\u671f\u5c31\u786e\u5b9a\u6240\u6709\u5f20\u91cf\u7684\u5f62\u72b6\u3001\u5185\u5b58\u504f\u79fb\u3001kernel \u5e8f\u5217\uff0c\u624d\u80fd\u5728\u8fd0\u884c\u671f\u653e\u5fc3\u5730\u628a\u6574\u4e2a\u8bad\u7ec3\u5faa\u73af\u6355\u83b7\u6210 CUDA Graph\u3002\u52a8\u6001\u56fe\u6846\u67b6\u5982\u679c\u60f3\u505a\u8fd9\u4ef6\u4e8b\uff0c\u5c31\u5fc5\u987b\u5148\u628a\u52a8\u6001\u6267\u884c\u75d5\u8ff9\u51bb\u7ed3\u6210\u9759\u6001\u5b50\u56fe\u2014\u2014\u8fd9\u6b63\u662f <code>torch.compile<\/code> \u5728\u505a\u7684\u4e8b\uff0c\u4f46\u5b83\u7ecf\u5e38\u56e0\u4e3a graph break\u3001\u52a8\u6001\u5f62\u72b6\u3001\u5730\u5740\u4e0d\u7a33\u5b9a\u800c\u88ab\u8feb\u56de\u9000\u5230 eager\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u91cc\uff0c\u6211\u4eec\u4e0d\u662f&#8221;\u4e8b\u540e\u6355\u83b7&#8221;\uff0c\u800c\u662f&#8221;\u4ece\u8bbe\u8ba1\u4e4b\u521d\u5c31\u4e3a\u6355\u83b7\u800c\u5efa&#8221;\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 MemoryPlan \u9759\u6001\u663e\u5b58\u89c4\u5212<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u6709\u4e00\u4e2a\u975e\u5e38\u6838\u5fc3\u7684\u8bbe\u8ba1\uff1a<strong>\u8fd0\u884c\u671f\u4e0d\u505a\u52a8\u6001\u663e\u5b58\u5206\u914d<\/strong>\u3002\u6240\u6709\u5f20\u91cf\u7684\u504f\u79fb\u3001\u5927\u5c0f\u3001\u751f\u547d\u5468\u671f\u90fd\u5728\u7f16\u8bd1\u671f\u7531 <code>MemoryPlan<\/code> \u786e\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>MemoryPlan<\/code> \u628a\u663e\u5b58\u6309\u8bed\u4e49\u5212\u5206\u4e3a <strong>68 \u4e2a\u547d\u540d Region<\/strong>\uff08\u679a\u4e3e\u69fd\u4f4d\u5171 69 \u4e2a\uff0c\u6700\u540e\u4e00\u4e2a\u4e3a\u6570\u7ec4\u8fb9\u754c\u54e8\u5175\uff09\uff1a\u6743\u91cd\u533a\u3001\u68af\u5ea6\u533a\u3001\u4e00\u9636\/\u4e8c\u9636\u52a8\u91cf\u533a\u3001EMA \u6743\u91cd\u533a\u3001AMP FP16 \u6743\u91cd\u533a\u3001\u8f93\u5165\u7f13\u51b2\u533a\u3001\u7279\u5f81\u56fe\u533a\u3001\u4e34\u65f6\u5f20\u91cf\u533a\u3001\u7ed3\u679c\u533a\u7b49\u7b49\u3002\u6bcf\u4e2a Region \u7684\u504f\u79fb\u5728 <code>finalize()<\/code> \u4e4b\u540e\u56fa\u5b9a\u4e0b\u6765\uff0c\u6240\u6709 <code>DTensor<\/code> \u53ea\u662f\u8fd9\u4e2a\u5e03\u5c40\u91cc\u7684\u4e00\u4e2a <code>(offset, stride, shape)<\/code> \u63cf\u8ff0\u7b26\uff0c\u4e0d\u6301\u6709\u5b9e\u9645\u5185\u5b58\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u8bbe\u8ba1\u5e26\u6765\u4e86\u51e0\u4e2a\u76f4\u63a5\u597d\u5904\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CUDA Graph \u6240\u9700\u7684\u5730\u5740\u7a33\u5b9a\u6027\u5929\u7136\u6ee1\u8db3<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>\u6d88\u9664\u52a8\u6001\u5206\u914d\u5ef6\u8fdf\u548c\u5185\u5b58\u788e\u7247<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>\u53ef\u4ee5\u6309 Region \u505a\u6279\u91cf\u64cd\u4f5c<\/strong>\uff0c\u6bd4\u5982\u4e00\u6b21\u6027\u6e05\u96f6\u5168\u90e8\u68af\u5ea6\u3001\u4e00\u6b21\u6027\u5207\u6362 EMA \u6743\u91cd\u3001\u4e00\u6b21\u6027\u5b8c\u6210\u6574\u533a\u66f4\u65b0\uff1b<\/li>\n\n\n\n<li><strong>\u591a\u5361\u5206\u5e03\u5f0f\u5e03\u5c40\u4e00\u81f4<\/strong>\uff0c\u540c\u4e00\u5f20 MemoryPlan \u56fe\u7eb8\u53ef\u4ee5\u5728\u591a\u5f20 GPU \u4e0a\u5171\u4eab\u6267\u884c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u6846\u67b6\u662f\u52a8\u6001\u56fe\uff0c\u663e\u5b58\u901a\u5e38\u7531\u8fd0\u884c\u65f6\u5206\u914d\u5668\u6309\u9700\u7533\u8bf7\uff0c\u5730\u5740\u6bcf\u6b21\u53ef\u80fd\u4e0d\u540c\uff0c\u5c31\u5f88\u96be\u505a\u5230\u8fd9\u79cd\u9759\u6001\u5206\u533a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u786e\u5b9a\u6027\u8bad\u7ec3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5f3a\u8c03<strong>\u53ef\u590d\u73b0\u6027<\/strong>\uff1a\u5728\u76f8\u540c\u786c\u4ef6\u3001\u76f8\u540c\u79cd\u5b50\u3001\u4e0d\u8c03\u7528\u975e\u786e\u5b9a\u6027\u7b97\u5b50\u7684\u60c5\u51b5\u4e0b\uff0c\u540c\u4e00\u7a0b\u5e8f\u591a\u6b21\u8fd0\u884c\u53ef\u4ee5\u5f97\u5230\u5b57\u8282\u7ea7\u4e00\u81f4\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4ef6\u4e8b\u4e5f\u4f9d\u8d56\u4e8e\u9759\u6001\u56fe\u3002\u56e0\u4e3a\u53ea\u6709\u56fe\u7ed3\u6784\u3001\u5185\u5b58\u5e03\u5c40\u3001kernel \u5e8f\u5217\u3001\u8c03\u5ea6\u987a\u5e8f\u5728\u7f16\u8bd1\u671f\u5c31\u56fa\u5b9a\u4e0b\u6765\uff0c\u6211\u4eec\u624d\u80fd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7528 Philox \u8ba1\u6570\u5668\u968f\u673a\u6570\u751f\u6210\u5668\u7cbe\u786e\u63a7\u5236\u6bcf\u4e2a\u968f\u673a\u70b9\u7684 seed \u548c offset\uff1b<\/li>\n\n\n\n<li>\u5728\u6570\u636e\u589e\u5f3a\u3001\u53c2\u6570\u521d\u59cb\u5316\u3001Dropout \u7b49\u6240\u6709\u968f\u673a\u70b9\u4fdd\u6301\u4e00\u81f4\uff1b<\/li>\n\n\n\n<li>\u907f\u514d\u8fd0\u884c\u65f6\u52a8\u6001\u8c03\u5ea6\u5e26\u6765\u7684\u4e0d\u786e\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u52a8\u6001\u56fe\u6846\u67b6\u7406\u8bba\u4e0a\u4e5f\u53ef\u4ee5\u505a\u786e\u5b9a\u6027\u8bad\u7ec3\uff0c\u4f46\u7531\u4e8e Python \u6267\u884c\u8def\u5f84\u3001\u5185\u5b58\u5206\u914d\u65f6\u673a\u3001\u5f02\u6b65\u8c03\u5ea6\u987a\u5e8f\u66f4\u5bb9\u6613\u6ce2\u52a8\uff0c\u5de5\u7a0b\u4e0a\u8981\u4fdd\u8bc1\u7aef\u5230\u7aef\u4e00\u81f4\u9700\u8981\u4ed8\u51fa\u66f4\u591a\u4ee3\u4ef7\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u4ee3\u7801\u5c42\u9762\u770b Tech-Renaissance \u7684\u9759\u6001\u56fe\u5b9e\u73b0<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7406\u8bba\u8bb2\u5b8c\uff0c\u770b\u4e00\u4e0b\u4ee3\u7801\u662f\u600e\u4e48\u843d\u5730\u7684\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u7528\u6237\u4fa7\uff1a\u5148\u5b9a\u4e49\uff0c\u540e\u7f16\u8bd1\uff0c\u518d\u8fd0\u884c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5 <code>tests\/example\/mlp_mnist.cpp<\/code> \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=\"\">BluePrint mlp = seq(\n    fc(1024, true), relu(),\n    fc(512, true), relu(),\n    fc(256, true), relu(),\n    fc(10, true)\n);\n\nDeepLearningTask task;\ntask.model(mlp)\n    .loss(CrossEntropyLoss().label_smoothing(0.1f))\n    .total_epochs(kTotalEpochs)\n    .optimizer(AdamW().weight_decay(1e-4f))\n    .scheduler(CosineAnnealingLR().base_lr(0.001f).warmup(5));\n\ntask.compile();\nauto result = task.run();<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u8fd9\u91cc\u7684 <code>task.compile()<\/code>\uff1a\u5b83\u4e0d\u662f\u4e00\u4e2a\u7b80\u5355\u7684&#8221;\u6253\u5370\u914d\u7f6e&#8221;\uff0c\u800c\u662f\u771f\u6b63\u628a BluePrint \u63cf\u8ff0\u7684\u6a21\u578b\u7f16\u8bd1\u6210\u5185\u90e8 IR\u3001MemoryPlan\u3001ComputationGraph\uff0c\u5e76\u6355\u83b7\u6210\u53ef\u6267\u884c\u56fe\u3002\u8c03\u7528 <code>task.run()<\/code> \u65f6\uff0c\u6846\u67b6\u53ea\u662f\u5728\u91cd\u590d\u6267\u884c\u5df2\u7ecf\u7f16\u8bd1\u597d\u7684\u56fe\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u4e09\u9636\u6bb5\u72b6\u6001\u673a<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>TaskBase<\/code> \u5f3a\u5236\u4e86\u4e00\u4e2a\u4e09\u9636\u6bb5\u72b6\u6001\u673a\uff0c\u5b9a\u4e49\u5728 <code>include\/renaissance\/core\/types.h<\/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=\"\">enum class Phase : uint8_t { PLANNING, MEMORY_LOCKED, COMPILED };<\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>PLANNING \u9636\u6bb5<\/strong>\uff1a\u7528\u6237\u6216\u7f16\u8bd1\u5668\u5206\u914d DTensor\u3001\u914d\u7f6e\u521d\u59cb\u5316\u7b56\u7565\uff1b<\/li>\n\n\n\n<li><strong>MEMORY_LOCKED \u9636\u6bb5<\/strong>\uff1a\u5185\u5b58\u5e03\u5c40\u9501\u5b9a\uff0c\u6ce8\u518c\u8ba1\u7b97\u56fe\uff1b<\/li>\n\n\n\n<li><strong>COMPILED \u9636\u6bb5<\/strong>\uff1a\u786c\u4ef6\u521d\u59cb\u5316\u3001CUDA Graph \u6355\u83b7\u5b8c\u6210\uff0c\u8fdb\u5165\u8fd0\u884c\u671f\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u72b6\u6001\u673a\u5728 <code>include\/renaissance\/task\/task_base.h<\/code> \u4e2d\u6709\u660e\u786e\u7ea6\u675f\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=\"\">\/\/ PLANNING \u9636\u6bb5\n[[nodiscard]] DTensor alloc(const Shape&amp; shape, DType dtype, Region region);\nvoid finalize_memory();\n\n\/\/ MEMORY_LOCKED \u9636\u6bb5\nvoid add_graph(const std::string&amp; name, ComputationGraph graph,\n               StreamKind stream = StreamKind::COMP_1);\n\n\/\/ COMPILED \u9636\u6bb5\u56fe\u6267\u884c\nvoid run(const std::string&amp; name);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u9636\u6bb5\u80fd\u8c03\u7528\u7684 API \u662f\u53d7\u9650\u7684\u3002\u8fd9\u79cd\u5f3a\u7ea6\u675f\u4fdd\u8bc1\u4e86\u6211\u4eec\u4e0d\u4f1a\u5728\u8fd0\u884c\u671f\u7a81\u7136&#8221;\u63d2\u5165&#8221;\u4e00\u5f20\u65b0\u56fe\u6216&#8221;\u52a8\u6001&#8221;\u7533\u8bf7\u4e00\u5757\u65b0\u5185\u5b58\u2014\u2014\u90a3\u4e0d\u662f\u8fd9\u4e2a\u6846\u67b6\u7684\u8bbe\u8ba1\u76ee\u6807\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 \u4e94\u9636\u6bb5\u7f16\u8bd1\u5668<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Compiler::compile()<\/code> \u662f\u9759\u6001\u56fe\u7f16\u8bd1\u7684\u6838\u5fc3\u5165\u53e3\uff0c\u5b9a\u4e49\u5728 <code>include\/renaissance\/graph\/compiler.h<\/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=\"\">static Result compile(const ArchPlan&amp; arch,\n                      const CompileSpec&amp; base_spec,\n                      const PlanConfig&amp; plan_config,\n                      Initializer&amp; initializer,\n                      const std::vector&lt;CompileSpec>&amp; variant_specs = {});<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u5185\u90e8\u7ba1\u7ebf\u5206\u4e3a\u4e94\u4e2a\u9636\u6bb5\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>derive_all_shapes<\/strong>\uff1a\u5bf9 6 \u4e2a\u53d8\u4f53\uff08base\u3001train_last\u3001train_lowres\u3001train_lowres_last\u3001val_base\u3001val_last\uff09\u5206\u522b\u505a\u5f62\u72b6\u63a8\u5bfc\uff1b<\/li>\n\n\n\n<li><strong>compute_max_slot_bytes<\/strong>\uff1a\u5bf9\u6bcf\u4e2a (layer, tensor) \u4f4d\u7f6e\u8de8\u53d8\u4f53\u53d6\u6700\u5927 slot \u5b57\u8282\u6570\uff1b<\/li>\n\n\n\n<li><strong>create_memory_plans<\/strong>\uff1a\u7528\u6700\u5927 slot \u6784\u9020 6 \u4e2a\u72ec\u7acb\u7684 MemoryPlan\uff0c\u4fdd\u8bc1\u540c\u4e00 DTensor \u5728\u4e0d\u540c\u53d8\u4f53\u4e0b offset \u4e00\u81f4\uff1b<\/li>\n\n\n\n<li><strong>build_computation_graph<\/strong>\uff1a\u904d\u5386 LayerDescriptor \u5e8f\u5217\uff0c\u6784\u5efa\u8bad\u7ec3\u56fe\u548c\u63a8\u7406\u56fe\u7684\u62d3\u6251\uff1b<\/li>\n\n\n\n<li><strong>share_or_clone<\/strong>\uff1ashape-only \u53d8\u4f53\u5171\u4eab\u540c\u4e00\u4efd ComputationGraph\uff0cval \u53d8\u4f53\u4e0d\u586b\u8bad\u7ec3\u56fe\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a<strong>\u628a&#8221;\u56fe\u62d3\u6251&#8221;\u548c&#8221;\u5f62\u72b6\/\u5185\u5b58\u5e03\u5c40&#8221;\u89e3\u8026<\/strong>\u3002<code>ComputationGraph<\/code> \u662f\u96f6\u5f62\u72b6\u4fe1\u606f\u7684\u7eaf\u62d3\u6251\u5bb9\u5668\uff0c\u4e00\u4efd\u56fe\u53ef\u4ee5\u670d\u52a1\u591a\u4e2a shape-only \u53d8\u4f53\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 \u96f6\u5f62\u72b6\u4fe1\u606f\u7684\u8ba1\u7b97\u56fe\u4e0e\u56fe\u96c6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>include\/renaissance\/graph\/computation_graph.h<\/code> \u4e2d\u7684\u6ce8\u91ca\u8bf4\u5f97\u5f88\u6e05\u695a\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\u7279\u5f81\uff1a\u96f6\u5f62\u72b6\u4fe1\u606f\u3002\u8282\u70b9\u4e2d\u53ea\u5b58 (GraphNode, OpParams, tensor_id)\uff1bShape \/ DType \/ Region \u5168\u90e8\u4ece MemoryPlan \u7684 DTensor \u83b7\u53d6\uff1b\u4e00\u4efd\u56fe\u4f9b\u591a\u4e2a shape-only \u53d8\u4f53\u5171\u4eab\u3002<\/p>\n<\/blockquote>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">class ComputationGraph {\n    \/\/ \u6bcf\u4e2a GraphId \u6876\u5404\u5b58\u4e00\u4e2a\u8282\u70b9\u5e8f\u5217\n    std::array&lt;std::vector&lt;GraphNode>, static_cast&lt;size_t>(GraphId::COUNT)> graphs_;\n    std::vector&lt;GraphNode> linear_nodes_;\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd0\u884c\u671f\uff0c\u54ea\u4e2a\u53d8\u4f53\u3001\u54ea\u4e2a\u5b50\u56fe\u8be5\u6267\u884c\uff0c\u7531 <code>GraphAtlas<\/code> \u51b3\u5b9a\u3002<code>GraphAtlas<\/code> \u662f\u4e00\u5f20 6 \u53d8\u4f53 \u00d7 \u591a\u5b50\u56fe\u7684\u6620\u5c04\u8868\uff0c\u5728 <code>include\/renaissance\/graph\/graph_atlas.h<\/code> \u4e2d\u5b9e\u73b0\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">class GraphAtlas {\n    static constexpr size_t kMaxVariants = 6;\n    static constexpr size_t kMaxGraphIds = static_cast&lt;size_t>(GraphId::COUNT);\n    \/\/ Phase A \u586b\u5165 cg \/ mp \/ shape_id\n    \/\/ Phase B \u9884\u6355\u83b7\u540e\u586b\u5165 captured_idx\n    \/\/ Phase C \u8fd0\u884c\u65f6 O(1) \u6570\u7ec4\u8bbf\u95ee\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">shape \u65e0\u5173\u7684\u5b50\u56fe\uff08\u6bd4\u5982 TRANSFER\u3001COMM\u3001OPTIMIZER\u3001EMA_UPDATE\uff09\u5728\u6240\u6709\u53d8\u4f53\u95f4\u5171\u4eab\uff1bshape \u76f8\u5173\u7684\u8bad\u7ec3\/\u63a8\u7406\u56fe\u6309 ShapeId \u53bb\u91cd\u3002\u8fd0\u884c\u671f\u53ea\u9700\u8981\u4e00\u6b21\u6570\u7ec4\u7d22\u5f15\uff0c\u96f6 hash\u3001\u96f6\u5206\u652f\u3001\u96f6\u52a8\u6001\u67e5\u627e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.5 CUDA Graph \u6355\u83b7<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>src\/graph\/capture_cuda.cpp<\/code> \u8d1f\u8d23\u628a ComputationGraph \u4e2d\u7684\u8282\u70b9\u5e8f\u5217\u6355\u83b7\u4e3a CUDA Graph\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 CapturedGraph::capture_cuda(const ComputationGraph&amp; cg,\n                                  const MemoryPlan&amp; mp,\n                                  GraphId gid,\n                                  StreamKind stream_kind,\n                                  const DeviceContext&amp; ctx,\n                                  CapturedGraph&amp; result) {\n    cudaStream_t primary_stream = select_primary_capture_stream(stream_kind, ctx);\n    const auto&amp; nodes = cg.linear_nodes().empty() ? cg.nodes(gid) : cg.linear_nodes();\n\n    cudaStreamBeginCapture(primary_stream, cudaStreamCaptureModeThreadLocal);\n    \/\/ ... \u904d\u5386\u8282\u70b9\uff0c\u5904\u7406\u591a\u6d41\u540c\u6b65\uff0ccudaStreamEndCapture, cudaGraphInstantiate\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6355\u83b7\u524d\uff0c\u6211\u4eec\u4f1a\u505a warmup\uff1b\u6355\u83b7\u65f6\uff0c<code>MultiStreamCaptureState<\/code> \u4f1a\u7ba1\u7406\u4e09\u6761\u8ba1\u7b97\u6d41\u3001\u4e00\u6761\u4f20\u8f93\u6d41\u3001\u4e00\u6761\u66f4\u65b0\u6d41\u4e4b\u95f4\u7684\u4f9d\u8d56\u548c\u4e8b\u4ef6\u540c\u6b65\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 Tech-Renaissance \u80fd\u505a\u5230<strong>\u8bad\u7ec3\u5faa\u73af\u5168\u9636\u6bb5 CUDA Graph \u6355\u83b7<\/strong>\u2014\u2014\u4ece H2D \u4f20\u8f93\u3001\u9996\u5c42\u524d\u5411\u3001\u6df1\u5c42\u524d\u5411+\u53cd\u5411\u3001\u68af\u5ea6\u901a\u4fe1\u3001\u4f18\u5316\u5668\u66f4\u65b0\u3001BN \u7edf\u8ba1\u91cf\u540c\u6b65\uff0c\u5168\u90e8\u53ef\u4ee5\u5728\u6355\u83b7\u8303\u56f4\u5185\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u4e0e PyTorch Eager \u548c torch.compile \u7684\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bf4\u5230\u8fd9\u91cc\uff0c\u53ef\u80fd\u6709\u4eba\u4f1a\u95ee\uff1aPyTorch \u4e0d\u662f\u4e5f\u6709 <code>torch.compile<\/code> \u5417\uff1f\u5b83\u4e0d\u4e5f\u662f\u5728\u8bd5\u56fe\u505a\u9759\u6001\u4f18\u5316\u548c CUDA Graph \u5417\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u662f\u7684\uff0c\u4f46\u8def\u5f84\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch Eager \u662f<strong>\u52a8\u6001\u56fe\u7684\u539f\u6559\u65e8\u4e3b\u4e49\u8005<\/strong>\uff1a\u6bcf\u4e2a\u7b97\u5b50\u90fd\u7ecf\u8fc7 Python dispatcher\u3001Autograd\u3001C++ kernel launch\uff0c\u9010\u4e2a\u63d0\u4ea4\u7ed9 GPU\u3002\u5b83\u7684\u4f18\u52bf\u662f\u7075\u6d3b\u548c\u6613\u7528\uff0c\u4ee3\u4ef7\u662f\u8c03\u5ea6\u5f00\u9500\u548c\u5168\u5c40\u4f18\u5316\u7a7a\u95f4\u53d7\u9650\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>torch.compile<\/code> \u662f\u5728\u52a8\u6001\u56fe\u4e4b\u4e0a\u53e0\u52a0\u7684\u4e00\u5c42<strong>JIT \u7f16\u8bd1\u5668<\/strong>\u3002\u5b83\u7528 TorchDynamo \u6355\u83b7 FX \u56fe\uff0c\u7528 Inductor \u505a\u878d\u5408\u548c kernel \u751f\u6210\uff0c\u5e76\u5c1d\u8bd5 CUDA Graph \u6355\u83b7\u3002\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u6709\u5de5\u7a0b\u4ef7\u503c\u7684\u65b9\u5411\uff0c\u4f46\u5b83\u4e5f\u9762\u4e34\u51e0\u4e2a\u7ed3\u6784\u6027\u6311\u6218\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Graph break<\/strong>\u3002\u9047\u5230 <code>torch.save<\/code>\u3001<code>.item()<\/code>\u3001\u6570\u636e\u4f9d\u8d56\u63a7\u5236\u6d41\u7b49\u4e0d\u652f\u6301\u7684\u64cd\u4f5c\u65f6\uff0cDynamo \u4f1a\u4e2d\u65ad\u5f53\u524d\u56fe\uff0c\u4e2d\u95f4\u7247\u6bb5\u56de\u9000\u5230 eager \u6267\u884c\u3002\u6bcf\u4e2a graph break \u90fd\u610f\u5473\u7740\u4e00\u6b21 host-device \u540c\u6b65\u548c\u4f18\u5316\u673a\u4f1a\u7684\u635f\u5931\u3002<\/li>\n\n\n\n<li><strong>\u52a8\u6001\u5f62\u72b6<\/strong>\u3002\u8f93\u5165\u5f62\u72b6\u53d8\u5316\u4f1a\u89e6\u53d1\u91cd\u65b0\u7f16\u8bd1\u3002\u867d\u7136 PyTorch \u6709\u7b26\u53f7\u5316\u5f62\u72b6\u7684\u652f\u6301\uff0c\u4f46\u590d\u6742\u573a\u666f\u4e0b\u4ecd\u4f1a\u5bfc\u81f4\u53cd\u590d\u7f16\u8bd1\u6216 shape specialization\u3002<\/li>\n\n\n\n<li><strong>\u5730\u5740\u7a33\u5b9a\u6027<\/strong>\u3002CUDA Graph \u8981\u6c42\u6355\u83b7\u65f6\u7684\u5730\u5740\u5728\u91cd\u653e\u65f6\u4e0d\u53d8\u3002PyTorch \u7684\u7f13\u5b58\u5206\u914d\u5668\u901a\u5e38\u662f&#8221;\u4f2a\u786e\u5b9a\u6027&#8221;\u7684\uff0c\u9700\u8981\u989d\u5916\u673a\u5236\uff08\u5982 private memory pool\u3001CUDA Graph Trees\uff09\u6765\u4fdd\u8bc1\uff0c\u7406\u8bba\u4e0a\u5e76\u975e 100% \u7a33\u5b9a\u3002<\/li>\n\n\n\n<li><strong>\u6b63\u786e\u6027\u98ce\u9669<\/strong>\u3002<code>torch.compile<\/code> \u7684\u7f16\u8bd1\u7ed3\u679c\u4e0e eager \u5728\u6570\u5b66\u4e0a\u5e94\u5f53\u4e00\u81f4\uff0c\u4f46\u793e\u533a\u957f\u671f\u62a5\u544a\u8fc7\u9759\u9ed8\u9519\u8bef\u3001\u68af\u5ea6\u5f02\u5e38\u3001\u4e0e CUDA Graph \u642d\u914d\u62a5\u9519\u7b49\u95ee\u9898\u3002\u8fd9\u4e0d\u662f\u8d2c\u4f4e PyTorch\uff0c\u800c\u662f\u8bf4\u660e<strong>\u5728\u52a8\u6001\u56fe\u57fa\u7840\u4e0a\u4e8b\u540e\u8865\u9759\u6001\u4f18\u5316\uff0c\u5929\u7136\u6bd4\u4ece\u9759\u6001\u56fe\u51fa\u53d1\u66f4\u56f0\u96be<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u9009\u62e9\u662f\uff1a<strong>\u4e0d\u8981\u4e8b\u540e\u6355\u83b7\uff0c\u800c\u662f\u4ece\u5934\u9759\u6001<\/strong>\u3002\u6211\u4eec\u628a\u6240\u6709\u5f62\u72b6\u3001\u5185\u5b58\u3001\u56fe\u62d3\u6251\u3001\u8c03\u5ea6\u5173\u7cfb\u90fd\u5728\u7f16\u8bd1\u671f\u6839\u636e\u4e00\u5b9a\u7684\u89c4\u5219\u786e\u5b9a\u4e0b\u6765\uff0c\u56e0\u6b64 CUDA Graph \u5168\u6355\u83b7\u3001MemoryPlan \u9759\u6001\u5206\u533a\u3001\u786e\u5b9a\u6027\u8bad\u7ec3\u8fd9\u4e9b\u80fd\u529b\u90fd\u80fd\u81ea\u7136\u843d\u5730\uff0c\u4e0d\u9700\u8981\u5728\u7075\u6d3b\u6027\u548c\u6027\u80fd\u4e4b\u95f4\u53cd\u590d\u62c9\u626f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u663e\u7136\u8fd9\u5b58\u5728\u4ee3\u4ef7\uff1aTech-Renaissance \u4e0d\u652f\u6301\u8fd0\u884c\u671f\u52a8\u6001\u4fee\u6539\u6a21\u578b\u7ed3\u6784\uff0c\u4e0d\u652f\u6301\u6570\u636e\u4f9d\u8d56\u63a7\u5236\u6d41\uff0c\u8c03\u8bd5\u4f53\u9a8c\u66f4\u63a5\u8fd1&#8221;\u5148\u7f16\u8bd1\u3001\u518d\u8fd0\u884c&#8221;\u7684 C++ \u7a0b\u5e8f\uff0c\u800c\u4e0d\u662f PyTorch Eager \u7684&#8221;\u6240\u89c1\u5373\u6240\u5f97&#8221;\u3002\u8fd9\u662f\u6211\u4eec\u6709\u610f\u8bc6\u5730\u63a5\u53d7\u7684 trade-off\u3002\u4f46\u662f<strong>\u8fd9\u5e26\u6765\u7684\u6027\u80fd\u4e0a\u7684\u6536\u76ca\uff0c\u5f88\u5feb\u5927\u5bb6\u5c31\u53ef\u4ee5\u89c1\u8bc6\u5230<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u9759\u6001\u56fe\u548c\u52a8\u6001\u56fe\u6ca1\u6709\u7edd\u5bf9\u7684\u4f18\u52a3\uff0c\u53ea\u6709<strong>\u573a\u666f\u53d6\u820d<\/strong>\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5982\u679c\u4f60\u7684\u9996\u8981\u8bc9\u6c42\u662f\u7814\u7a76\u5feb\u901f\u8fed\u4ee3\u3001\u8c03\u8bd5\u65b9\u4fbf\u3001\u52a8\u6001\u6a21\u578b\uff0c\u52a8\u6001\u56fe\u662f\u66f4\u81ea\u7136\u7684\u9009\u62e9\uff1b<\/li>\n\n\n\n<li>\u5982\u679c\u4f60\u7684\u9996\u8981\u8bc9\u6c42\u662f\u8bad\u7ec3\u541e\u5410\u3001\u90e8\u7f72\u6548\u7387\u3001\u786e\u5b9a\u6027\u3001\u7aef\u5230\u7aef\u590d\u73b0\uff0c\u9759\u6001\u56fe\u80fd\u7ed9\u4f60\u66f4\u5927\u7684\u4f18\u5316\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5c5e\u4e8e\u540e\u8005\u3002\u5b83\u4e0d\u662f&#8221;\u9759\u6001\u56fe\u6846\u67b6&#8221;\u56db\u4e2a\u5b57\u53ef\u4ee5\u6982\u62ec\u7684\uff0c\u800c\u662f\u56f4\u7ed5\u9759\u6001\u56fe\u7f16\u8bd1\u8fd9\u4e00\u9009\u62e9\uff0c\u628a <strong>Compiler \u4e94\u9636\u6bb5\u7f16\u8bd1\u300168-Region MemoryPlan\uff08\u542b\u54e8\u5175\u5171 69 \u4e2a\u69fd\u4f4d\uff09\u3001\u96f6\u5f62\u72b6 ComputationGraph\u3001GraphAtlas \u53d8\u4f53\u590d\u7528\u3001\u591a\u6d41 CUDA Graph \u5168\u6355\u83b7<\/strong> \u4e32\u6210\u4e86\u4e00\u6761\u5b8c\u6574\u94fe\u8def\u3002\u540e\u7eed\u51e0\u7bc7\u6587\u7ae0\u4f1a\u5206\u522b\u5c55\u5f00\u8fd9\u4e9b\u90e8\u4ef6\uff0c\u4f46\u5b83\u4eec\u7684\u5171\u540c\u5e95\u5ea7\uff0c\u5c31\u662f\u672c\u7bc7\u8bb2\u7684<strong>\u9759\u6001\u56fe\u7f16\u8bd1<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u628a\u89c6\u89d2\u62c9\u9ad8\uff0c\u7528\u4e00\u6761\u6837\u672c\u4ece\u78c1\u76d8\u5230\u6307\u6807\u8f93\u51fa\u7684\u5b8c\u6574\u65c5\u7a0b\uff0c\u4e32\u8d77 Tech-Renaissance \u7684\u4e03\u5927\u6a21\u5757\uff0c\u7ed9\u4f60\u4e00\u5f20\u5168\u5c40\u5730\u56fe\u3002<\/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\u4e94 \u4e0a\u4e00\u7bc7\u6211\u4eec\u628a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u62c6\u89e3\u6210 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":667,"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-449","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\/449","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=449"}],"version-history":[{"count":20,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/449\/revisions"}],"predecessor-version":[{"id":649,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/449\/revisions\/649"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/667"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=449"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=449"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=449"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}