{"id":499,"date":"2026-07-08T03:02:29","date_gmt":"2026-07-07T19:02:29","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=499"},"modified":"2026-07-09T00:32:24","modified_gmt":"2026-07-08T16:32:24","slug":"archplan%e4%b8%8e%e7%bc%96%e8%af%91%e7%ae%a1%e7%ba%bf%ef%bc%9a%e4%bb%8e%e6%a8%a1%e5%9e%8b%e5%ae%9a%e4%b9%89%e5%88%b0%e5%8f%af%e6%89%a7%e8%a1%8c%e5%9b%be","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/499\/","title":{"rendered":"(13) ArchPlan\u4e0e\u7f16\u8bd1\u7ba1\u7ebf\uff1a\u4ece\u6a21\u578b\u5b9a\u4e49\u5230\u53ef\u6267\u884c\u56fe"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u5341\u4e09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0a\u4e00\u7bc7\u6211\u4eec\u8bb2\u4e86 BluePrint DSL\uff1a\u7528\u6237\u53ef\u4ee5\u7528 <code>seq<\/code>\u3001<code>cbr<\/code>\u3001<code>block<\/code>\u3001<code>gap_fc<\/code> \u8fd9\u4e9b\u5de5\u5382\u51fd\u6570\uff0c\u7528\u5341\u51e0\u884c\u4ee3\u7801\u642d\u51fa ResNet-50 \u6216\u8005 MobileNet \u8fd9\u6837\u7684\u590d\u6742\u7f51\u7edc\u3002\u4f46 BluePrint \u672c\u8d28\u4e0a\u53ea\u662f\u4e00\u68f5\u63cf\u8ff0&#8221;\u6a21\u578b\u957f\u4ec0\u4e48\u6837&#8221;\u7684\u6811\uff0c\u5b83\u8ddd\u79bb\u771f\u6b63\u80fd\u5728 GPU \u4e0a\u9ad8\u6548\u6267\u884c\u7684\u4ee3\u7801\u8fd8\u5f88\u8fdc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a\u9ad8\u5c42\u6a21\u578b\u5b9a\u4e49\u8f6c\u6362\u6210\u53ef\u6267\u884c\u56fe\uff0c\u8fd9\u6b63\u662f<strong>\u7f16\u8bd1\u7ba1\u7ebf<\/strong>\u8981\u89e3\u51b3\u7684\u95ee\u9898\u3002\u5728 Tech-Renaissance \u91cc\uff0c\u8fd9\u6761\u7ba1\u7ebf\u7531\u4e24\u4e2a\u6838\u5fc3\u7ec4\u4ef6\u627f\u8f7d\uff1a<code>ArchPlan<\/code> \u8d1f\u8d23\u628a BluePrint \u5c55\u5f00\u3001\u89c4\u8303\u5316\u3001\u878d\u5408\u6210\u6807\u51c6\u5316\u7684\u67b6\u6784\u63cf\u8ff0\uff1b<code>Compiler<\/code> \u8d1f\u8d23\u5728\u8fd9\u4e2a\u67b6\u6784\u4e0a\u505a\u5f62\u72b6\u63a8\u5bfc\u3001\u663e\u5b58\u89c4\u5212\u3001\u8ba1\u7b97\u56fe\u6784\u5efa\u548c\u53d8\u4f53\u53bb\u91cd\u3002\u672c\u6587\u5c31\u6cbf\u7740\u8fd9\u6761\u94fe\u8def\uff0c\u770b\u770b\u4e00\u5f20\u56fe\u5230\u5e95\u662f\u600e\u4e48\u4ece&#8221;\u58f0\u660e\u5f0f\u8bed\u6cd5&#8221;\u4e00\u6b65\u6b65\u53d8\u6210&#8221;\u53ef\u6267\u884c\u8ba1\u5212&#8221;\u7684\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e3a\u4ec0\u4e48\u9700\u8981\u7f16\u8bd1<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f88\u591a\u4eba\u5728\u7b2c\u4e00\u6b21\u63a5\u89e6\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u65f6\u4f1a\u6709\u4e2a\u9519\u89c9\uff1a\u6846\u67b6\u4e0d\u5c31\u662f\u8c03\u7528 cuDNN\u3001cuBLAS \u8fd9\u4e9b\u5e93\u5417\uff1f\u6211\u5199 <code>torch.nn.Conv2d<\/code>\uff0c\u6846\u67b6\u5e2e\u6211\u8c03 <code>cudnnConvolutionForward<\/code>\uff0c\u8fd9\u7b97\u4ec0\u4e48\u7f16\u8bd1\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u53ea\u662f\u8dd1\u4e00\u4e2a\u5b64\u7acb\u7b97\u5b50\uff0c\u786e\u5b9e\u8c08\u4e0d\u4e0a\u7f16\u8bd1\u3002\u4f46\u73b0\u4ee3\u795e\u7ecf\u7f51\u7edc\u5f80\u5f80\u5305\u542b\u51e0\u5341\u5230\u51e0\u767e\u4e2a\u7b97\u5b50\uff0c\u8fd9\u4e9b\u7b97\u5b50\u4e4b\u95f4\u6709\u6570\u636e\u4f9d\u8d56\u3001\u6709\u4e2d\u95f4\u5f20\u91cf\u3001\u6709\u53cd\u5411\u4f20\u64ad\u3001\u6709\u6743\u91cd\u66f4\u65b0\u3002\u6846\u67b6\u5982\u679c\u53ea\u505a\u4e00\u4e2a&#8221;\u7b97\u5b50\u4e00\u4e2a\u7b97\u5b50\u5730\u8c03\u7528&#8221;\u7684\u8c03\u5ea6\u5668\uff0c\u5c31\u4f1a\u9762\u4e34\u51e0\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e00\uff0c\u4e2d\u95f4\u5f20\u91cf\u53cd\u590d\u8bfb\u5199\u663e\u5b58\u3002<\/strong> \u6bd4\u5982 Conv \u2192 BN \u2192 ReLU \u4e09\u4e2a\u7b97\u5b50\u5982\u679c\u72ec\u7acb\u6267\u884c\uff0cConv \u7684\u8f93\u51fa\u8981\u5148\u5199\u5230\u663e\u5b58\uff0cBN \u8bfb\u51fa\u6765\u518d\u5199\u56de\u53bb\uff0cReLU \u518d\u8bfb\u4e00\u6b21\u3002\u5982\u679c\u80fd\u628a\u5b83\u4eec\u5408\u5e76\u6210\u4e00\u4e2a kernel\uff0c\u4e2d\u95f4\u7ed3\u679c\u5c31\u53ef\u4ee5\u7559\u5728\u5bc4\u5b58\u5668\u6216\u5171\u4eab\u5185\u5b58\u91cc\uff0c\u663e\u5b58\u5e26\u5bbd\u76f4\u63a5\u7701\u6389\u4e00\u5927\u534a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e8c\uff0c\u5168\u5c40\u5185\u5b58\u5e03\u5c40\u65e0\u6cd5\u9884\u5148\u89c4\u5212\u3002<\/strong> \u52a8\u6001\u56fe\u6846\u67b6\u901a\u5e38\u7531\u8fd0\u884c\u65f6\u5206\u914d\u5668\u6309\u9700\u7533\u8bf7\u663e\u5b58\uff0c\u8fd9\u4f1a\u5e26\u6765\u5206\u914d\u5ef6\u8fdf\u3001\u5185\u5b58\u788e\u7247\uff0c\u4ee5\u53ca\u5730\u5740\u4e0d\u7a33\u5b9a\u3002\u800c\u9759\u6001\u7f16\u8bd1\u53ef\u4ee5\u5728\u8fd0\u884c\u524d\u5c31\u7b97\u51fa\u6bcf\u4e2a\u5f20\u91cf\u5e94\u8be5\u653e\u5728\u54ea\u91cc\u3001\u4ec0\u4e48\u65f6\u5019\u53ef\u4ee5\u590d\u7528\u522b\u4eba\u7684\u4f4d\u7f6e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e09\uff0ckernel launch \u548c Python \u6d3e\u53d1\u5f00\u9500\u3002<\/strong> \u6bcf\u4e2a\u7b97\u5b50\u90fd\u8d70\u4e00\u904d dispatcher\u3001\u751f\u6210\u4e00\u6b21 CUDA kernel\u3001\u505a\u4e00\u6b21 host-device \u540c\u6b65\uff0c\u5c0f\u7b97\u5b50\u5bc6\u96c6\u65f6 CPU \u8c03\u5ea6\u672c\u8eab\u5c31\u4f1a\u6210\u4e3a\u74f6\u9888\u3002\u628a\u4e00\u6574\u6bb5\u8ba1\u7b97\u6355\u83b7\u6210 CUDA Graph \u91cd\u653e\uff0c\u662f\u6d88\u9664\u8fd9\u7c7b\u5f00\u9500\u7684\u6709\u6548\u624b\u6bb5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\uff0c<strong>\u7f16\u8bd1\u662f\u6846\u67b6\u4ece&#8221;\u80fd\u8dd1&#8221;\u8d70\u5411&#8221;\u8dd1\u5f97\u5feb&#8221;\u7684\u5173\u952e\u4e00\u6b65<\/strong>\u3002\u5b83\u4e0d\u4e00\u5b9a\u9700\u8981\u751f\u6210\u6c47\u7f16\u6216 LLVM IR\uff0c\u800c\u662f\u8981\u5728\u8fd0\u884c\u524d\u628a\u7b97\u5b50\u3001\u5f20\u91cf\u3001\u6267\u884c\u987a\u5e8f\u3001\u5185\u5b58\u5e03\u5c40\u3001\u8c03\u5ea6\u7b56\u7565\u90fd\u786e\u5b9a\u4e0b\u6765\u3002\u6253\u4e2a\u6bd4\u65b9\uff1a\u7528\u6237\u7ed9\u7684\u662f\u5efa\u7b51\u8349\u56fe\uff0cArchPlan \u505a\u7ed3\u6784\u5206\u6790\u4e0e\u89c4\u8303\u5316\uff0cCompiler \u753b\u65bd\u5de5\u56fe\u5e76\u6807\u51fa\u6bcf\u6839\u94a2\u7b4b\u3001\u6bcf\u5757\u6df7\u51dd\u571f\u7684\u4f4d\u7f6e\uff0cComputationGraph \u52a0 CUDA Graph \u6355\u83b7\u624d\u662f\u6700\u7ec8\u4ea4\u4ed8\u7684\u53ef\u6267\u884c\u5efa\u7b51\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u4e3b\u6d41\u6846\u67b6\u7684\u7f16\u8bd1\u65b9\u6848<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u540c\u6846\u67b6\u5904\u7406\u8fd9\u4e2a\u95ee\u9898\u7684\u601d\u8def\u5dee\u5f02\u5f88\u5927\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>PyTorch Eager<\/strong> \u662f\u5178\u578b\u7684&#8221;\u8fb9\u6267\u884c\u8fb9\u5efa\u56fe&#8221;\u6a21\u5f0f\uff0c\u6bcf\u4e2a\u7b97\u5b50\u72ec\u7acb\u8c03\u5ea6\u3002\u5b83\u7684\u4f18\u52bf\u662f\u7075\u6d3b\u3001\u8c03\u8bd5\u65b9\u4fbf\uff0c\u4f46\u4ee3\u4ef7\u662f\u4e0a\u9762\u63d0\u5230\u7684\u4e09\u70b9 overhead\u3002\u4e3a\u4e86\u5f25\u8865\u6027\u80fd\uff0cPyTorch 2.0 \u5f15\u5165\u4e86 <code>torch.compile<\/code>\uff1aTorchDynamo \u6355\u83b7 FX \u56fe\uff0cAOTAutograd \u5904\u7406\u524d\u5411\u548c\u53cd\u5411\uff0cInductor \u505a\u7b97\u5b50\u878d\u5408\u4e0e kernel \u751f\u6210\uff0c\u5e76\u5c1d\u8bd5 CUDA Graph \u6355\u83b7\u3002\u8fd9\u662f\u4e00\u4e2a\u5728\u52a8\u6001\u56fe\u4e4b\u4e0a\u53e0\u52a0 JIT \u7f16\u8bd1\u5668\u7684\u65b9\u6848\uff0c\u4f46 graph break\u3001\u52a8\u6001\u5f62\u72b6\u3001\u5730\u5740\u7a33\u5b9a\u6027\u7b49\u95ee\u9898\u4e5f\u8ba9\u5b83\u65f6\u4e0d\u65f6\u56de\u9000\u5230 eager\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>TensorFlow<\/strong> \u4ece 1.x \u65f6\u4ee3\u5c31\u8d70\u9759\u6001\u56fe\u8def\u7ebf\uff0c<code>tf.Graph<\/code> + <code>tf.Session<\/code> \u7684\u6a21\u5f0f\u628a\u6574\u4e2a\u8ba1\u7b97\u56fe\u4ea4\u7ed9\u6846\u67b6\u7f16\u8bd1\u3002TensorFlow 2.x \u9ed8\u8ba4 eager\uff0c\u4f46 <code>@tf.function<\/code> \u53ef\u4ee5\u628a Python \u51fd\u6570 trace \u6210\u9759\u6001\u56fe\u4ea4\u7ed9 XLA \u4f18\u5316\u3002XLA \u662f\u4e00\u4e2a\u4e13\u95e8\u9762\u5411\u7ebf\u6027\u4ee3\u6570\u7684\u7f16\u8bd1\u5668\uff0c\u80fd\u505a\u7b97\u5b50\u878d\u5408\u3001\u5e03\u5c40\u4f18\u5316\u3001\u5185\u5b58\u590d\u7528\u7b49\u5168\u5c40\u4f18\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>JAX<\/strong> \u7684\u8bbe\u8ba1\u54f2\u5b66\u662f&#8221;Python \u5199\u52a8\u6001\u903b\u8f91\uff0cjit \u7f16\u8bd1\u6210\u9759\u6001 XLA \u56fe&#8221;\u3002<code>jax.jit<\/code> \u4f1a\u7528 XLA \u628a\u51fd\u6570\u7f16\u8bd1\u6210\u9488\u5bf9\u7279\u5b9a\u8bbe\u5907\u7684\u53ef\u6267\u884c\u7a0b\u5e8f\uff0c\u9002\u5408\u9700\u8981\u9ad8\u9636\u5bfc\u6570\u3001\u51fd\u6570\u53d8\u6362\u7684\u524d\u6cbf\u7814\u7a76\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>ONNX Runtime<\/strong> \u5219\u662f\u4e00\u79cd\u90e8\u7f72\u4fa7\u7f16\u8bd1\u601d\u8def\uff1a\u628a\u6a21\u578b\u5bfc\u51fa\u6210 ONNX \u4e2d\u95f4\u8868\u793a\uff0c\u518d\u7531\u8fd0\u884c\u65f6\u505a\u56fe\u4f18\u5316\u3001\u7b97\u5b50\u9009\u62e9\u3001\u5185\u5b58\u89c4\u5212\u3002\u5b83\u7684\u4f18\u5316\u7a7a\u95f4\u53d6\u51b3\u4e8e ONNX \u56fe\u672c\u8eab\u7684\u4fe1\u606f\u5b8c\u6574\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u4ee5\u770b\u5230\uff0c\u4e0d\u7ba1\u662f\u9759\u6001\u6846\u67b6\u8fd8\u662f\u52a8\u6001\u6846\u67b6\uff0c<strong>\u6700\u7ec8\u90fd\u5728\u5f80&#8221;\u56fe\u7ea7\u7f16\u8bd1&#8221;\u8fd9\u4e2a\u65b9\u5411\u9760\u62e2<\/strong>\u3002\u5dee\u522b\u53ea\u5728\u4e8e\uff1a\u9759\u6001\u6846\u67b6\u662f&#8221;\u5148\u5b9a\u4e49\u518d\u7f16\u8bd1&#8221;\uff0c\u52a8\u6001\u6846\u67b6\u662f&#8221;\u5148\u8fd0\u884c\u518d\u6355\u83b7&#8221;\u3002Tech-Renaissance \u5c5e\u4e8e\u524d\u8005\uff0c\u5e76\u4e14\u628a\u7f16\u8bd1\u8fd9\u4ef6\u4e8b\u505a\u5f97\u5f88\u91cd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u4e00\u4e2a\u503c\u5f97\u4e13\u95e8\u63d0\u4e00\u4e0b\u7684\u6982\u5ff5\uff1a<strong>\u4e2d\u95f4\u8868\u793a\uff08Intermediate Representation\uff0cIR\uff09<\/strong>\u3002\u7f16\u8bd1\u5668\u901a\u5e38\u4e0d\u4f1a\u76f4\u63a5\u628a\u6e90\u4ee3\u7801\u7ffb\u8bd1\u6210\u673a\u5668\u7801\uff0c\u800c\u662f\u5148\u7ffb\u8bd1\u6210\u4e00\u79cd\u4ecb\u4e8e\u6e90\u8bed\u8a00\u548c\u76ee\u6807\u8bed\u8a00\u4e4b\u95f4\u7684\u8868\u793a\u3002IR \u7684\u597d\u5904\u662f\u65e2\u80fd\u4fdd\u7559\u9ad8\u5c42\u8bed\u4e49\uff0c\u53c8\u4fbf\u4e8e\u505a\u5404\u79cd\u5206\u6790\u548c\u53d8\u6362\u3002\u5728 Tech-Renaissance \u91cc\uff0c<code>ArchPlan<\/code> \u5c31\u662f\u4e00\u79cd\u9ad8\u5c42 IR\uff1a\u5b83\u4fdd\u7559\u4e86\u5c42\u7684\u6982\u5ff5\u3001\u53c2\u6570\u3001\u8fde\u63a5\u5173\u7cfb\uff0c\u4f46\u5df2\u7ecf\u53bb\u6389\u4e86 BluePrint \u4e2d\u90a3\u4e9b\u4e3a\u4e86\u7528\u6237\u53cb\u597d\u800c\u5f15\u5165\u7684\u8bed\u6cd5\u7cd6\u3002Compiler \u5728\u8fd9\u4e2a IR \u4e0a\u505a\u5f62\u72b6\u63a8\u5bfc\u3001\u7b97\u5b50\u5c55\u5f00\u3001\u5185\u5b58\u5206\u914d\uff0c\u6700\u7ec8\u751f\u6210\u4f4e\u5c42\u7684 <code>ComputationGraph<\/code>\u3002\u6ca1\u6709 IR \u8fd9\u4e2a\u4e2d\u95f4\u5c42\uff0c\u524d\u7aef DSL \u7684\u6bcf\u4e00\u6b21\u8bed\u6cd5\u53d8\u5316\u90fd\u53ef\u80fd\u8981\u6c42\u540e\u7aef\u91cd\u65b0\u9002\u914d\uff1b\u6709\u4e86 IR\uff0c\u524d\u7aef\u548c\u540e\u7aef\u5c31\u53ef\u4ee5\u5404\u81ea\u72ec\u7acb\u6f14\u5316\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001Tech-Renaissance \u7684\u7f16\u8bd1\u5168\u666f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u4e2d\uff0c\u4e00\u4e2a\u6a21\u578b\u4ece\u7528\u6237\u4ee3\u7801\u5230\u53ef\u6267\u884c\u56fe\u7684\u5b8c\u6574\u65c5\u7a0b\u662f\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"bash\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">BluePrint (Layer \u6811)\n    \u2193 ArchPlan::from_blueprint\nArchPlan (ArchLayer \u5e8f\u5217)\n    \u2193 ArchPlan::build\n\u6807\u51c6\u5316 + \u878d\u5408\u7684 ArchPlan\n    \u2193 Compiler::compile\nMemoryPlan[6] + ComputationGraph \u00d72\n    \u2193 GraphAtlas::build\nGraphAtlas (\u53d8\u4f53 \u00d7 \u5b50\u56fe\u6620\u5c04)\n    \u2193 CapturedGraph::capture\nCUDA Graph \/ CPU fallback<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u6d41\u7a0b\u6709\u51e0\u4e2a\u5173\u952e\u8bbe\u8ba1\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>BluePrint \u53ea\u8d1f\u8d23\u63cf\u8ff0\u6a21\u578b\u62d3\u6251<\/strong>\uff0c\u4e0d\u5173\u5fc3\u8bbe\u5907\u3001\u5f62\u72b6\u3001\u7cbe\u5ea6\u3002<\/li>\n\n\n\n<li><strong>ArchPlan \u662f\u7f16\u8bd1\u4e2d\u95f4\u8868\u793a\uff08IR\uff09<\/strong>\uff0c\u628a\u6811\u5f62 DSL \u5c55\u5f00\u6210\u7ebf\u6027\u5c42\u5e8f\u5217\uff0c\u5e76\u505a\u89c4\u8303\u5316\u4e0e\u878d\u5408\u3002<\/li>\n\n\n\n<li><strong>Compiler \u4e94\u9636\u6bb5<\/strong>\u628a ArchPlan \u8f6c\u6362\u6210\u663e\u5b58\u5e03\u5c40\uff08MemoryPlan\uff09\u548c\u96f6\u5f62\u72b6\u8ba1\u7b97\u56fe\uff08ComputationGraph\uff09\u3002<\/li>\n\n\n\n<li><strong>ComputationGraph \u4e0e\u5f62\u72b6\u89e3\u8026<\/strong>\uff0c\u540c\u4e00\u4efd\u56fe\u53ef\u4ee5\u670d\u52a1\u591a\u4e2a shape-only \u53d8\u4f53\u3002<\/li>\n\n\n\n<li><strong>\u6700\u7ec8\u7531 GraphAtlas \u7ba1\u7406\u53d8\u4f53\u4e0e\u5b50\u56fe\u7684\u6620\u5c04\uff0cCapturedGraph \u5b8c\u6210 CUDA Graph \u6355\u83b7\u3002<\/strong><\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u9762\u4e24\u8282\u5206\u522b\u5c55\u5f00 ArchPlan \u548c Compiler\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001ArchPlan\uff1a\u5341\u6b65\u7ba1\u7ebf\u7684\u4e2d\u95f4\u8868\u793a<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ArchPlan<\/code> \u662f BluePrint \u4e4b\u540e\u7684\u7b2c\u4e00\u7ad9\u3002\u5b83\u7684\u4efb\u52a1\u662f\u628a\u7528\u6237\u53cb\u597d\u7684 DSL \u6811\uff0c\u8f6c\u6362\u6210\u540e\u7aef\u80fd\u591f\u7406\u89e3\u7684\u3001\u89c4\u8303\u5316\u8fc7\u7684\u5c42\u5e8f\u5217\u3002\u7c7b\u7684\u6ce8\u91ca\u628a\u5b83\u8bf4\u5f97\u5f88\u6e05\u695a\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=\"\">\/\/ include\/renaissance\/graph\/arch_plan.h:4-5\n\/**\n * @brief \u67b6\u6784\u89c4\u5212\uff1a\u5c42\u63cf\u8ff0\u3001\u7f16\u8bd1\u7ba1\u7ebf\u548c\u5e8f\u5217\u5316\n * @details \u4ece BluePrint \u63a5\u6536\u6a21\u578b\u5b9a\u4e49\uff0c\u7ecf 10 \u6b65\u7ba1\u7ebf\u751f\u6210\u6807\u51c6\u5316\u3001\u5168\u878d\u5408\u7684\u67b6\u6784\u63cf\u8ff0\n *\/<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u8fd9\u91cc\u8bf4\u662f 10 \u6b65\uff0c\u4f46\u516c\u5f00 API \u53ea\u66b4\u9732\u4e86 step2 \u5230 step10\u3002step1 \u5bf9\u5e94\u7684\u662f <code>from_blueprint<\/code> \u91cc\u8c03\u7528\u7684 <code>expand_tree<\/code>\uff0c\u4e0d\u7b97\u5728 <code>build()<\/code> \u5185\u90e8\u3002\u6211\u4eec\u6309\u5b9e\u9645\u6267\u884c\u987a\u5e8f\u6765\u770b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 Step 1\uff1a\u6811\u5c55\u5f00<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ArchPlan::from_blueprint()<\/code> \u63a5\u6536 <code>BluePrint<\/code> \u548c <code>InputSpec<\/code>\uff0c\u9012\u5f52\u904d\u5386 Layer \u6811\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=\"\">\/\/ src\/graph\/arch_plan_expand.cpp:470-489\nArchPlan ArchPlan::from_blueprint(const BluePrint&amp; bp, const InputSpec&amp; input, bool fuse) {\n    if (fuse &amp;&amp; !GlobalRegistry::instance().using_amp()) {\n        TR_VALUE_ERROR(\n            \"ArchPlan::from_blueprint: operator fusion (fuse=true) requires AMP mode.\"\n        );\n    }\n\n    ArchPlan arch;\n    arch.input_ = Shape(1, input.h, input.w, input.c);\n    arch.fuse_ = fuse;\n\n    int current_c = input.c;\n    expand_tree(bp.root_, arch.layers_, current_c, fuse);\n    return arch;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>expand_tree<\/code> \u4f1a\u628a <code>Sequential<\/code>\u3001<code>Repeat<\/code>\u3001<code>Add2<\/code>\u3001<code>Block<\/code> \u8fd9\u4e9b\u9ad8\u5c42\u8282\u70b9\u5168\u90e8\u5c55\u5f00\u6210 <code>ArchLayer<\/code>\u3002\u4f8b\u5982\u4e00\u4e2a ResNet Bottleneck \u5757\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=\"\">block(64, 256, RESNET_1_3_1)<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c <code>fuse=true<\/code>\uff0c\u5b83\u4f1a\u76f4\u63a5\u751f\u6210\u4e00\u4e2a <code>LayerKind::BottleneckIdentity<\/code> \u6216 <code>BottleneckProjection<\/code> \u7684 ArchLayer\uff1b\u5982\u679c <code>fuse=false<\/code>\uff0c\u5b83\u4f1a\u5c55\u5f00\u6210 <code>Add2Start<\/code>\u3001shortcut\u3001<code>Add2ShortcutEnd<\/code>\u3001\u4e3b\u5e72 <code>Conv-BN-ReLU<\/code>\u3001\u4e3b\u5e72 <code>Conv-BN-ReLU<\/code>\u3001\u4e3b\u5e72 <code>Conv-BN<\/code>\u3001<code>Add2End<\/code>\u3001<code>ReLU<\/code> \u7b49\u539f\u59cb\u5e8f\u5217\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">CBR \u5de5\u5382\u51fd\u6570\u5728 <code>fuse=true<\/code> \u65f6\u4e5f\u4f1a\u4fdd\u6301\u4e3a\u5355\u4e2a <code>LayerKind::CBR<\/code>\uff0c\u800c\u4e0d\u662f\u62c6\u6210 Conv+BN+ReLU\u3002\u8fd9\u4fdd\u8bc1\u4e86\u540e\u7aef\u878d\u5408\u7b97\u5b50\u4e0d\u4f1a\u88ab\u7f16\u8bd1\u5668\u524d\u7aef\u7834\u574f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 Step 2\uff1aBN \u91cd\u547d\u540d<\/h3>\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=\"\">\/\/ src\/graph\/arch_plan_normalize.cpp:53-73\nvoid ArchPlan::step2_rename_bn() {\n    for (size_t i = 0; i &lt; layers_.size(); ++i) {\n        if (layers_[i].kind != LayerKind::Bn2d) continue;\n        LayerKind prev = LayerKind::Conv;\n        for (int j = static_cast&lt;int>(i) - 1; j >= 0; --j) {\n            auto k = layers_[j].kind;\n            if (k == LayerKind::ReLU || k == LayerKind::Tanh || k == LayerKind::Identity\n                || k == LayerKind::ChannelPadding) continue;\n            if (k == LayerKind::Add2ShortcutEnd || k == LayerKind::Add2Start) {\n                prev = LayerKind::Conv; break;\n            }\n            prev = k; break;\n        }\n        bool is_1d = (prev == LayerKind::FC || prev == LayerKind::Flatten || prev == LayerKind::GAP);\n        layers_[i].kind = is_1d ? LayerKind::Bn1d : LayerKind::Bn2d;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u6b65\u5411\u540e\u770b BN \u7684\u524d\u9a71\u7b97\u5b50\uff0c\u8df3\u8fc7 ReLU\u3001Tanh\u3001Identity \u4ee5\u53ca <code>ChannelPadding<\/code> \u7b49\u4e0d\u6539\u53d8\u5f20\u91cf\u8bed\u4e49\u7684\u5c42\uff1b\u5982\u679c\u9047\u5230 <code>Add2Start<\/code> \u6216 <code>Add2ShortcutEnd<\/code>\uff0c\u5219\u6309\u5377\u79ef\u573a\u666f\u5904\u7406\u3002\u5982\u679c BN \u524d\u9762\u662f FC\u3001Flatten \u6216 GAP\uff0c\u5c31\u6807\u8bb0\u4e3a <code>Bn1d<\/code>\uff1b\u5426\u5219\u4fdd\u6301 <code>Bn2d<\/code>\u3002\u8fd9\u6837\u540e\u7aef\u53ef\u4ee5\u4e3a 1D \u548c 2D \u573a\u666f\u9009\u62e9\u4e0d\u540c\u7684 kernel\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 Step 3\uff1aSoftmaxCE \u89c4\u8303\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>SoftmaxCE<\/code> \u5728 BluePrint \u91cc\u53ef\u4ee5\u51fa\u73b0\u5728\u4efb\u610f\u4f4d\u7f6e\uff0c\u4f46 Tech-Renaissance \u5f3a\u5236\u628a\u5b83\u653e\u5230\u6700\u540e\uff0c\u5e76\u7edf\u4e00\u53c2\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/graph\/arch_plan_normalize.cpp:75-80\nvoid ArchPlan::step3_normalize_softmax_ce(int num_classes) {\n    layers_.erase(std::remove_if(layers_.begin(), layers_.end(),\n        [](const ArchLayer&amp; l) { return l.kind == LayerKind::SoftmaxCE; }),\n        layers_.end());\n    layers_.push_back({LayerKind::SoftmaxCE, SoftmaxCELayerParams{num_classes}, \"softmax_ce\"});\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u7f16\u8bd1\u5668\u5c31\u77e5\u9053\u635f\u5931\u51fd\u6570\u4e00\u5b9a\u5728\u94fe\u5c3e\uff0c\u4e0d\u9700\u8981\u5904\u7406\u5206\u652f\u6216\u591a\u8f93\u51fa loss\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 Step 4\uff1aIdentity \u89c4\u8303\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>add2<\/code> \u5728 BluePrint \u4e2d\u8868\u793a\u6b8b\u5dee\u8fde\u63a5\u3002\u5c55\u5f00\u540e\u4f1a\u4ea7\u751f <code>Add2Start<\/code>\u3001<code>Add2ShortcutEnd<\/code>\u3001<code>Add2End<\/code> \u4e09\u4e2a\u6807\u8bb0\u5c42\u3002\u8fd9\u4e00\u6b65\u786e\u4fdd shortcut \u5206\u652f\u81f3\u5c11\u6709\u4e00\u4e2a <code>Identity<\/code> \u5360\u4f4d\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=\"\">\/\/ src\/graph\/arch_plan_normalize.cpp:82-105\nvoid ArchPlan::step4_normalize_identity() {\n    layers_.erase(std::remove_if(..., Identity), layers_.end());\n    for (size_t i = 0; i &lt; layers_.size(); ++i) {\n        if (layers_[i].kind != LayerKind::Add2Start) continue;\n        size_t sc_end = find_add2_boundary(i + 1, LayerKind::Add2ShortcutEnd);\n        if (sc_end == i + 1) {\n            layers_.insert(layers_.begin() + sc_end,\n                {LayerKind::Identity, EmptyParams{}, \"identity\"});\n        }\n        \/\/ ...\n    }\n}<\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.5 Step 5\uff1aFlatten \u89c4\u8303\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Flatten<\/code> \u5728 BluePrint \u91cc\u662f\u53ef\u9009\u7684\u3002\u7f16\u8bd1\u5668\u4f1a\u6839\u636e\u5b9e\u9645\u9700\u8981\u51b3\u5b9a\u662f\u5426\u63d2\u5165\uff1a\u5982\u679c FC \u524d\u9762\u5df2\u7ecf\u662f <code>(1,1,C)<\/code> \u4e14 <code>C % 8 == 0<\/code>\uff0c\u5c31\u4e0d\u9700\u8981 Flatten\uff1b\u5426\u5219\u81ea\u52a8\u5728 FC \u6216 SoftmaxCE \u524d\u63d2\u5165 Flatten\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=\"\">\/\/ src\/graph\/arch_plan_normalize.cpp:149-199\nvoid ArchPlan::step5_normalize_flatten() {\n    layers_.erase(std::remove_if(..., Flatten), layers_.end());\n    \/\/ ... \u6839\u636e in_shape \u5224\u65ad\u662f\u5426\u9700\u8981\u63d2\u5165 Flatten\n    if (needs_flatten) {\n        layers_.insert(layers_.begin() + insert_at,\n            {LayerKind::Flatten, EmptyParams{}, \"flatten\"});\n        recompute_shapes_from(insert_at);\n    }\n}<\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.6 Step 6\uff1a\u5f62\u72b6\u63a8\u5bfc<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>build()<\/code> \u91cc\uff0c\u5f62\u72b6\u63a8\u5bfc\u5148\u4e8e Flatten \u89c4\u8303\u5316\u6267\u884c\uff0c\u56e0\u4e3a Flatten \u662f\u5426\u63d2\u5165\u672c\u8eab\u9700\u8981\u77e5\u9053\u5f62\u72b6\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=\"\">\/\/ src\/graph\/arch_plan_normalize.cpp:35-51\nvoid ArchPlan::build(int num_classes) {\n    step2_rename_bn();\n    step3_normalize_softmax_ce(num_classes);\n    step4_normalize_identity();\n    align_amp_input_channels_for_first_conv();  \/\/ AMP \u9996\u5c42 C \u5bf9\u9f50\u5230 4\n    step6_deduce_shapes();      \/\/ \u5148\u63a8\u5bfc\u5b9e\u9645 H\/W\/C\n    step5_normalize_flatten();  \/\/ \u518d\u51b3\u5b9a Flatten \u662f\u5426\u771f\u6b63\u9700\u8981\n    if (fuse_) {\n        step7_merge_blocks();\n        step8_merge_quadruple();\n        step9_merge_triple();\n        step10_merge_binary_and_mark();\n    } else {\n        mark_first_layer();\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>step6_deduce_shapes()<\/code> \u5c31\u662f\u6309 NHWC \u683c\u5f0f\u94fe\u5f0f\u8ba1\u7b97\u6bcf\u5c42\u8f93\u51fa\u5f62\u72b6\u3002\u5377\u79ef\u3001\u6c60\u5316\u3001GAP\u3001Flatten\u3001Bottleneck\u3001InvResidual \u90fd\u6709\u5404\u81ea\u7684\u63a8\u5bfc\u516c\u5f0f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.7 Step 7\uff1aBlock \u5408\u5e76<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f ArchPlan \u4e2d\u6700\u91cd\u8981\u7684\u878d\u5408\u6b65\u9aa4\u4e4b\u4e00\u3002\u7f16\u8bd1\u5668\u626b\u63cf <code>Add2Start ... Add2End<\/code> \u533a\u95f4\uff0c\u5c1d\u8bd5\u5339\u914d\u4e09\u79cd\u7ecf\u5178\u6b8b\u5dee\u6a21\u5f0f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Bottleneck<\/strong>\uff1aConv-BN-ReLU-Conv-BN-ReLU-Conv\uff087 \u5c42\u4e3b\u5e72\uff1b\u7b2c 3 \u4e2a Conv \u540e\u901a\u5e38\u4ecd\u8ddf\u6709 BN\uff0c\u4f46\u6a21\u5f0f\u8bc6\u522b\u53ea\u6821\u9a8c\u8fd9 7 \u5c42\uff09<\/li>\n\n\n\n<li><strong>BasicBlock<\/strong>\uff1aConv-BN-ReLU-Conv-BN<\/li>\n\n\n\n<li><strong>InvResidual<\/strong>\uff08MobileNetV2\uff09\uff1aConv-BN-ReLU-Conv-BN-ReLU-Conv-BN<\/li>\n<\/ul>\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=\"\">\/\/ src\/graph\/arch_plan_merge.cpp:162-204\nvoid ArchPlan::step7_merge_blocks() {\n    \/\/ ... \u626b\u63cf Add2 \u533a\u95f4 ...\n    if (try_merge_bottleneck(i, sc_end, add2_end, result)) { \/* ... *\/ }\n    if (try_merge_basic_block(i, sc_end, add2_end, result)) { \/* ... *\/ }\n    if (try_merge_inv_residual(i, sc_end, add2_end, result)) { \/* ... *\/ }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5339\u914d\u6210\u529f\u540e\uff0c\u539f\u672c\u5341\u51e0\u4e2a ArchLayer \u4f1a\u88ab\u538b\u7f29\u6210\u4e00\u4e2a <code>BottleneckProjection\/Identity<\/code>\u3001<code>BasicBlockProjection\/Identity<\/code> \u6216 <code>InvResidualNoShortcut\/Identity<\/code>\u3002\u8fd9\u4e9b\u878d\u5408\u5c42\u5728\u540e\u7aef\u6709\u4e13\u95e8\u7684 AMP kernel\uff0c\u53ef\u4ee5\u4e00\u6b21\u6027\u5b8c\u6210\u6574\u4e2a\u6b8b\u5dee\u5757\u7684\u524d\u5411\u548c\u53cd\u5411\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.8 Step 8-10\uff1a\u4e8c\u5143\/\u4e09\u5143\u878d\u5408<\/h3>\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=\"\">\/\/ src\/graph\/arch_plan_merge.cpp:268-300\nvoid ArchPlan::step8_merge_quadruple() {\n    \/\/ CBRP fusion removed\n}\n\nvoid ArchPlan::step9_merge_triple() {\n    if (!GlobalRegistry::instance().using_amp()) return;\n    merge_pattern_triple(LayerKind::Conv, LayerKind::Bn2d, LayerKind::ReLU,\n                         LayerKind::CBR, build_cbr);\n}\n\nvoid ArchPlan::step10_merge_binary_and_mark() {\n    merge_pattern_binary(LayerKind::GAP, LayerKind::FC,\n        LayerKind::GapFC, build_gapfc);\n    mark_first_layer();\n}<\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>step8<\/strong>\uff1a\u56db\u5143\u878d\u5408\uff08Conv+BN+ReLU+MaxPool\uff09\u5f53\u524d\u5df2\u79fb\u9664\uff0c\u4fdd\u6301\u72ec\u7acb\u5c42\u3002<\/li>\n\n\n\n<li><strong>step9<\/strong>\uff1a\u5728 AMP \u6a21\u5f0f\u4e0b\u628a\u672a\u8fdb\u5165 Block \u7684 <code>Conv+BN+ReLU<\/code> \u4e09\u5143\u7ec4\u878d\u5408\u6210 <code>CBR<\/code>\u3002<\/li>\n\n\n\n<li><strong>step10<\/strong>\uff1a\u628a <code>GAP+FC<\/code> \u878d\u5408\u6210 <code>GapFC<\/code>\uff0c\u5e76\u6807\u8bb0\u9996\u5c42\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ec8\u8f93\u51fa\u7684 ArchPlan \u662f\u4e00\u4e2a\u89c4\u8303\u5316\u7684\u3001\u5c3d\u53ef\u80fd\u878d\u5408\u7684\u5c42\u5e8f\u5217\uff0c\u6bcf\u4e00\u5c42\u90fd\u5e26\u6709\u8f93\u5165\/\u8f93\u51fa\u5f62\u72b6\u3001\u662f\u5426\u9996\u5c42\u6807\u8bb0\u3001\u53c2\u6570\u7b49\u5b8c\u6574\u4fe1\u606f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.9 \u4ece BluePrint \u5230 ArchLayer\uff1a\u4e00\u4e2a Bottleneck \u5757\u7684\u65c5\u7a0b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u66f4\u76f4\u89c2\u5730\u7406\u89e3 ArchPlan \u5728\u505a\u4ec0\u4e48\uff0c\u6211\u4eec\u6765\u770b\u4e00\u4e2a ResNet-50 \u4e2d\u5e38\u89c1\u7684 Bottleneck \u5757\u3002\u7528\u6237\u5728 BluePrint \u91cc\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=\"\">seq(\n    conv(64, 1, 1, 0), bn(), relu(),\n    conv(64, 3, 1, 1), bn(), relu(),\n    conv(256, 1, 1, 0), bn()\n)<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6216\u8005\u66f4\u7b80\u6d01\u5730\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=\"\">block(64, 256, RESNET_1_3_1)<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>fuse=true<\/code> \u7684 AMP \u6a21\u5f0f\u4e0b\uff0c<code>expand_tree<\/code> \u4f1a\u76f4\u63a5\u628a\u5b83\u6620\u5c04\u6210\u4e00\u4e2a <code>ArchLayer<\/code>\uff0c\u5176 <code>kind<\/code> \u4e3a <code>BottleneckIdentity<\/code> \u6216 <code>BottleneckProjection<\/code>\uff0c\u53c2\u6570\u4e2d\u8bb0\u5f55 bottleneck \u901a\u9053\u6570\u3001\u8f93\u51fa\u901a\u9053\u6570\u3001stride\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u5982\u679c\u7528\u6237\u5199\u6210\u672a\u878d\u5408\u7684\u539f\u59cb\u5e8f\u5217\uff0c\u6216\u8005 <code>fuse=false<\/code>\uff0c<code>expand_tree<\/code> \u4f1a\u628a\u5b83\u5c55\u5f00\u6210\u66f4\u957f\u7684\u5e8f\u5217\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"bash\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">Add2Start\nIdentity (\u6216 conv+bn \u4e0b\u91c7\u6837)\nAdd2ShortcutEnd\nConv + Bn2d + ReLU\nConv + Bn2d + ReLU\nConv + Bn2d\nAdd2End\nReLU<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7136\u540e <code>step7_merge_blocks<\/code> \u4f1a\u53cd\u5411\u626b\u63cf\u8fd9\u4e2a\u533a\u95f4\uff0c\u8bc6\u522b\u51fa&#8221;7 \u5c42\u4e3b\u5e72 Conv-BN-ReLU-Conv-BN-ReLU-Conv + shortcut&#8221;\u7684 Bottleneck \u6a21\u5f0f\uff0c\u518d\u628a\u5b83\u538b\u7f29\u56de\u4e00\u4e2a <code>BottleneckIdentity<\/code> \u6216 <code>BottleneckProjection<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a&#8221;\u5c55\u5f00\u2014\u518d\u5408\u5e76&#8221;\u770b\u4f3c\u7ed5\u8def\uff0c\u5176\u5b9e\u5f88\u6709\u5fc5\u8981\u3002BluePrint \u5141\u8bb8\u7528\u6237\u7528\u5404\u79cd\u65b9\u5f0f\u8868\u8fbe\u540c\u4e00\u4e2a\u7ed3\u6784\uff1a\u65e2\u53ef\u4ee5\u7528 <code>block()<\/code>\uff0c\u4e5f\u53ef\u4ee5\u624b\u5199 <code>cbr<\/code> \u5e8f\u5217\uff0c\u8fd8\u53ef\u4ee5\u6df7\u5408 <code>add2<\/code>\u3002ArchPlan \u7684\u4efb\u52a1\u5c31\u662f\u628a\u6240\u6709\u8fd9\u4e9b\u7b49\u4ef7\u5199\u6cd5\u5f52\u4e00\u5316\u5230\u540c\u4e00\u79cd\u540e\u7aef\u53cb\u597d\u7684\u8868\u793a\u3002\u540e\u7aef\u53ea\u9700\u8981\u4e3a <code>BottleneckIdentity<\/code>\u3001<code>BottleneckProjection<\/code>\u3001<code>CBR<\/code>\u3001<code>GapFC<\/code> \u7b49\u5c11\u6570\u51e0\u79cd\u878d\u5408\u7b97\u5b50\u5b9e\u73b0\u9ad8\u6027\u80fd kernel\uff0c\u800c\u4e0d\u9700\u8981\u5904\u7406\u7528\u6237\u53ef\u80fd\u5199\u51fa\u7684\u65e0\u7a77\u591a\u79cd\u6392\u5217\u7ec4\u5408\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.10 YAML \u5e8f\u5217\u5316\uff1a\u8ba9\u7f16\u8bd1\u7ed3\u679c\u53ef\u68c0\u67e5<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ArchPlan<\/code> \u8fd8\u63d0\u4f9b\u4e86 <code>to_yaml()<\/code> \u548c <code>from_yaml()<\/code>\uff0c\u53ef\u4ee5\u628a\u89c4\u8303\u5316\u540e\u7684\u67b6\u6784\u5bfc\u51fa\u6210\u4eba\u7c7b\u53ef\u8bfb\u7684 YAML \u6587\u4ef6\u3002\u8fd9\u5bf9\u4e8e\u8c03\u8bd5\u975e\u5e38\u6709\u7528\uff1a\u4f60\u53ef\u4ee5\u770b\u5230\u6a21\u578b\u7ecf\u8fc7\u878d\u5408\u540e\u5230\u5e95\u957f\u4ec0\u4e48\u6837\uff0c\u6bcf\u4e00\u5c42\u7684\u8f93\u5165\u8f93\u51fa\u5f62\u72b6\u662f\u4ec0\u4e48\uff0c\u9996\u5c42\u5728\u54ea\u91cc\uff0c\u54ea\u4e9b\u5757\u88ab\u5408\u5e76\u4e86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u529f\u80fd\u4e5f\u4f53\u73b0\u4e86 ArchPlan \u4f5c\u4e3a&#8221;\u7f16\u8bd1\u4e2d\u95f4\u8868\u793a&#8221;\u7684\u5b9a\u4f4d\u2014\u2014\u5b83\u4e0d\u4ec5\u662f\u7ed9 Compiler \u6d88\u8d39\u7684\u5185\u90e8\u7ed3\u6784\uff0c\u4e5f\u662f\u4e00\u4e2a\u53ef\u4ee5\u88ab\u68c0\u67e5\u3001\u88ab\u6301\u4e45\u5316\u3001\u88ab\u7248\u672c\u63a7\u5236\u7684\u6a21\u578b\u63cf\u8ff0\u5c42\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001Compiler\uff1a\u4e94\u9636\u6bb5\u7f16\u8bd1\u7f16\u6392\u5668<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">ArchPlan \u5b8c\u6210\u540e\uff0c\u771f\u6b63\u7684&#8221;\u56fe\u7f16\u8bd1&#8221;\u624d\u521a\u521a\u5f00\u59cb\u3002<code>Compiler::compile()<\/code> \u7684\u6ce8\u91ca\u628a\u4e94\u4e2a\u9636\u6bb5\u5199\u5f97\u5f88\u6e05\u695a\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=\"\">\/\/ include\/renaissance\/graph\/compiler.h:38-42\n\/**\n * Phase 1: derive_all_shapes      \u2014 \u5bf9\u6240\u6709 CompileSpec \u8c03\u7528 LayerDescriptor::infer_tensors\n * Phase 2: compute_max_slot_bytes \u2014 \u9010 (layer, tensor) \u8de8\u53d8\u4f53\u53d6 max slot_bytes\n * Phase 3: create_memory_plans    \u2014 \u7528 max_slot_bytes \u6784\u9020 6 \u4e2a MemoryPlan\n * Phase 4: build_computation_graph \u2014 \u904d\u5386 LayerDescriptor \u5e8f\u5217\u6784\u5efa GraphNode \u62d3\u6251\n * Phase 5: share_or_clone         \u2014 shape-only \u53d8\u4f53\u5171\u4eab\u6307\u9488\uff0cgraph-change \u53d8\u4f53\u72ec\u7acb new\n *\/<\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 CompileSpec \u4e0e\u516d\u53d8\u4f53<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u8fdb\u5165 Compiler \u4e4b\u524d\uff0c\u5148\u7406\u89e3 <code>CompileSpec<\/code>\u3002\u5b83\u63cf\u8ff0\u5355\u4e2a\u7f16\u8bd1\u53d8\u4f53\u7684\u5f62\u72b6\u76f8\u5173\u53c2\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/graph\/compile_spec.h:35-41\nstruct CompileSpec {\n    bool amp_enabled = false;        \/\/ \u662f\u5426\u542f\u7528\u6df7\u5408\u7cbe\u5ea6\n    int  max_sample_resolution = 0;  \/\/ MemoryPlan \u6700\u5927\u69fd\u4f4d\u9884\u7559\n    int  actual_resolution     = 0;  \/\/ \u8be5\u53d8\u4f53\u7684\u5b9e\u9645\u5206\u8fa8\u7387\n    int  batch_size            = 0;  \/\/ batch \u5927\u5c0f\n    int  num_color_channels    = 0;  \/\/ \u989c\u8272\u901a\u9053\u6570\n    bool freeze_first_layer = false; \/\/ \u8fd0\u884c\u65f6\u6807\u5fd7\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u516d\u53d8\u4f53\u5206\u522b\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>variants[0]\uff1atrain_base \u2014 \u6807\u51c6\u8bad\u7ec3 batch \u00d7 \u8d77\u59cb\u5206\u8fa8\u7387<\/li>\n\n\n\n<li>variants[1]\uff1atrain_last \u2014 \u8bad\u7ec3\u6700\u540e\u4e00\u4e2a\u4e0d\u5b8c\u6574 batch<\/li>\n\n\n\n<li>variants[2]\uff1atrain_lowres \u2014 \u6807\u51c6 batch \u00d7 \u7ed3\u675f\u5206\u8fa8\u7387\uff08\u5982\u6e10\u8fdb resize\uff09<\/li>\n\n\n\n<li>variants[3]\uff1atrain_lowres_last<\/li>\n\n\n\n<li>variants[4]\uff1aval_base \u2014 \u9a8c\u8bc1\u6807\u51c6 batch<\/li>\n\n\n\n<li>variants[5]\uff1aval_last \u2014 \u9a8c\u8bc1\u6700\u540e\u4e00\u4e2a batch<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u53d8\u4f53\u5171\u4eab\u540c\u4e00\u4efd\u8ba1\u7b97\u56fe\u62d3\u6251\uff0c\u4f46\u5404\u81ea\u6709\u72ec\u7acb\u7684 MemoryPlan\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 Phase 1\uff1a\u5f62\u72b6\u63a8\u5bfc\u4e0e\u8de8\u53d8\u4f53\u4e00\u81f4\u6027\u7ea6\u675f<\/h3>\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=\"\">\/\/ src\/graph\/compiler.cpp:568-612\nvoid Compiler::derive_all_shapes(...) {\n    for (size_t s = 0; s &lt; specs.size(); ++s) {\n        Shape cur_shape;\n        for (size_t l = 0; l &lt; arch.layers().size(); ++l) {\n            const auto&amp; layer = arch.layers()[l];\n            const LayerDescriptor&amp; descriptor = get_layer_descriptor(layer.kind);\n            \/\/ ...\n            all_shapes[s][l] = descriptor.infer_tensors(input, op_params, ctx);\n            if (!all_shapes[s][l].empty()) {\n                cur_shape = get_output_shape(layer.kind, all_shapes[s][l]);\n            }\n        }\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a <code>LayerKind<\/code> \u5bf9\u5e94\u4e00\u4e2a\u4e0d\u53ef\u53d8\u7684 <code>LayerDescriptor<\/code>\uff0c\u5b83\u63d0\u4f9b <code>infer_tensors<\/code>\u3001<code>build_forward<\/code>\u3001<code>build_backward<\/code>\u3001<code>build_inference<\/code> \u56db\u4e2a\u51fd\u6570\u6307\u9488\u3002<code>infer_tensors<\/code> \u8fd4\u56de\u8be5\u5c42\u5728\u4e09\u79cd\u6a21\u5f0f\u4e0b\u6240\u9700\u5f20\u91cf\u7684\u5e76\u96c6\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=\"\">\/\/ include\/renaissance\/graph\/layer_descriptor.h:107-120\nstruct LayerDescriptor {\n    using InferFn = std::vector&lt;TensorDesc> (*)(const Shape&amp; input,\n                                                  const OpParams&amp; params,\n                                                  const InferContext&amp; ctx);\n    using BuildFn = SubgraphPattern (*)(const OpParams&amp; params,\n                                         const std::vector&lt;TensorDesc>&amp; descs);\n\n    InferFn infer_tensors;    \/\/ \u8fd4\u56de\u4e09\u6a21\u5f0f\uff08train fwd\/bwd + inf\uff09\u5f20\u91cf\u5e76\u96c6\n    BuildFn build_forward;    \/\/ \u6784\u5efa\u524d\u5411\u5b50\u56fe\u6a21\u5f0f\n    BuildFn build_backward;   \/\/ \u6784\u5efa\u53cd\u5411\u5b50\u56fe\u6a21\u5f0f\n    BuildFn build_inference;  \/\/ \u6784\u5efa\u63a8\u7406\u5b50\u56fe\u6a21\u5f0f\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>get_layer_descriptor(LayerKind)<\/code> \u662f\u4e00\u4e2a\u96c6\u4e2d\u5f0f\u7684 switch \u6ce8\u518c\u8868\uff0c\u5b9a\u4e49\u5728 <code>src\/graph\/layer_descriptor_registry.cpp<\/code> \u4e2d\uff0c\u6bcf\u4e2a <code>LayerKind<\/code> \u5bf9\u5e94\u4e00\u4e2a\u9759\u6001 <code>LayerDescriptor<\/code>\u3002\u8fd9\u79cd\u8bbe\u8ba1\u8ba9 Compiler \u4e0d\u5fc5\u77e5\u9053\u6bcf\u4e2a\u7b97\u5b50\u7684\u5185\u90e8\u5b9e\u73b0\u7ec6\u8282\uff0c\u53ea\u9700\u8981\u8c03\u7528\u7edf\u4e00\u7684\u51fd\u6570\u6307\u9488\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f8b\u5982\uff0c\u4e00\u4e2a <code>LayerKind::Conv<\/code> \u7684 <code>infer_tensors<\/code> \u4f1a\u8fd4\u56de\uff1a\u8f93\u5165\u7279\u5f81\u56fe\u3001\u6743\u91cd\u3001\u8f93\u51fa\u7279\u5f81\u56fe\u3001\u6743\u91cd\u68af\u5ea6\u69fd\u3001\u8f93\u51fa\u68af\u5ea6\u69fd\u7b49\uff1b<code>build_forward<\/code> \u4f1a\u8fd4\u56de\u4e00\u4e2a\u5305\u542b <code>CONV_AMP_FWD<\/code> \u6216 <code>CONV_FP32_FWD<\/code> \u7b97\u5b50\u7684 <code>SubgraphPattern<\/code>\u3002Compiler \u628a\u8fd9\u4e9b\u6a21\u5f0f\u4e2d\u7684\u5f20\u91cf\u7d22\u5f15\u66ff\u6362\u6210\u771f\u5b9e\u7684 DTensor ID\uff0c\u5c31\u5f97\u5230\u4e86\u53ef\u6267\u884c\u7684 <code>GraphNode<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd&#8221;\u63cf\u8ff0\u7b26 + \u6a21\u5f0f&#8221;\u7684\u8bbe\u8ba1\u662f Compile \u9636\u6bb5\u4e0e Backend \u89e3\u8026\u7684\u5173\u952e\u3002\u5982\u679c\u672a\u6765\u8981\u652f\u6301\u65b0\u7684\u7b97\u5b50\u6216\u65b0\u7684\u786c\u4ef6\u540e\u7aef\uff0c\u53ea\u9700\u8981\u65b0\u589e <code>LayerKind<\/code> \u548c\u5bf9\u5e94\u7684 <code>LayerDescriptor<\/code>\uff0c\u4e0d\u9700\u8981\u6539\u52a8 Compiler \u7684\u4e3b\u6d41\u7a0b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u6b65\u4f1a\u9a8c\u8bc1\u8de8\u53d8\u4f53\u4e00\u81f4\u6027\u3002\u4ee3\u7801\u4e2d\u663e\u5f0f\u68c0\u67e5\u4e09\u6761\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5f20\u91cf\u6570\u91cf\u4e00\u81f4<\/strong>\uff1a\u540c\u4e00\u5c42\u5728\u6bcf\u4e2a\u53d8\u4f53\u4e0b\u8fd4\u56de\u7684\u5f20\u91cf\u4e2a\u6570\u5fc5\u987b\u76f8\u540c\uff1b<\/li>\n\n\n\n<li><strong>\u5f20\u91cf\u540d\u79f0\u4e00\u81f4<\/strong>\uff1a\u540c\u4e00\u4f4d\u7f6e\u7684\u5f20\u91cf\u540d\u79f0\u5fc5\u987b\u76f8\u540c\uff1b<\/li>\n\n\n\n<li><strong>\u5f20\u91cf\u6240\u5c5e Region \u4e00\u81f4<\/strong>\uff1a\u540c\u4e00\u4f4d\u7f6e\u7684\u5f20\u91cf\u5fc5\u987b\u5c5e\u4e8e\u540c\u4e00\u4e2a Region\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f20\u91cf\u7684\u987a\u5e8f\u7531 <code>std::vector<\/code> \u4e0b\u6807\u5929\u7136\u4fdd\u8bc1\uff1b\u800c\u5f62\u72b6\u548c dtype \u6b63\u662f\u5141\u8bb8\u4e0d\u540c\uff08\u54cd\u5e94 batch\u3001resolution\u3001AMP \u7684\u53d8\u5316\uff09\uff0c\u6240\u4ee5\u4e0d\u4f1a\u88ab\u5f3a\u5236\u68c0\u67e5\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=\"\">\/\/ src\/graph\/compiler.cpp:657-699\nvoid Compiler::validate_tensor_consistency(...) {\n    \/\/ \u663e\u5f0f\u68c0\u67e5\uff1a\u6bcf\u5c42\u5f20\u91cf\u6570\u91cf\u3001region\u3001name \u8de8 spec \u4e00\u81f4\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u4e00\u81f4\u6027\u7ea6\u675f\u662f\u540e\u7eed\u8de8\u53d8\u4f53\u5171\u4eab\u8ba1\u7b97\u56fe\u548c MemoryPlan offset \u5bf9\u9f50\u7684\u57fa\u7840\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 Phase 2\uff1a\u8de8\u53d8\u4f53\u53d6\u6700\u5927 slot<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u540c\u53d8\u4f53\u7684 batch \u6216\u5206\u8fa8\u7387\u4e0d\u540c\uff0c\u540c\u4e00\u5c42\u540c\u4e00\u4f4d\u7f6e\u7684\u5f20\u91cf\u5927\u5c0f\u4e5f\u4e0d\u540c\u3002\u4e3a\u4e86\u8ba9\u5927\u5bb6\u5171\u7528\u540c\u4e00\u4efd\u8ba1\u7b97\u56fe\uff0c\u53c8\u5404\u81ea\u6709\u72ec\u7acb\u7684\u663e\u5b58\u5e03\u5c40\uff0cCompiler \u4f1a\u5bf9\u6bcf\u4e2a <code>(layer, tensor)<\/code> \u4f4d\u7f6e\u8de8\u53d8\u4f53\u53d6\u6700\u5927\u5b57\u8282\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/graph\/compiler.cpp:618-651\nvoid Compiler::compute_max_slot_bytes(...) {\n    for (size_t l = 0; l &lt; num_layers; ++l) {\n        for (size_t t = 0; t &lt; num_tensors; ++t) {\n            uint64_t max_bytes = 0;\n            for (size_t s = 0; s &lt; all_shapes.size(); ++s) {\n                uint64_t bytes = DTensor::compute_slot_bytes(desc.shape, desc.dtype, desc.region);\n                max_bytes = std::max(max_bytes, bytes);\n            }\n            max_slots[l][t] = max_bytes;\n        }\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u6bcf\u4e2a\u53d8\u4f53\u7684 MemoryPlan \u5728\u8be5\u4f4d\u7f6e\u5206\u914d\u540c\u6837\u5927\u5c0f\u7684\u69fd\u4f4d\uff0c\u4ece\u800c\u4fdd\u8bc1 DTensor ID \u548c offset \u8de8\u53d8\u4f53\u4e00\u81f4\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.4 Phase 3\uff1a\u521b\u5efa MemoryPlan<\/h3>\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=\"\">\/\/ src\/graph\/compiler.cpp:705-977\nvoid Compiler::create_memory_plans(...) {\n    for (size_t s = 0; s &lt; all_shapes.size(); ++s) {\n        memory_plans[s] = std::make_unique&lt;MemoryPlan>(plan_config);\n        \/\/ ... \u5206\u914d baseline dtensors ...\n        for (size_t l = 0; l &lt; num_layers; ++l) {\n            for (size_t t = 0; t &lt; tensors.size(); ++t) {\n                DTensor dt = memory_plans[s]->alloc(alloc_shape, desc.dtype, desc.region, slot_bytes);\n                \/\/ ...\n            }\n            \/\/ \u6761\u4ef6\u5206\u914d M\/V\/N \u7cfb\u5217\n        }\n        memory_plans[s]->finalize();\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u53d8\u4f53\u5f97\u5230\u72ec\u7acb\u7684 <code>MemoryPlan<\/code>\u3002\u7b2c 14 \u7bc7\u4f1a\u4e13\u95e8\u8bb2 MemoryPlan \u7684 Region \u5206\u533a\uff0868 \u4e2a\u547d\u540d\u8bed\u4e49\u533a + 1 \u4e2a\u8fb9\u754c\u54e8\u5175\uff0c\u5171 69 \u4e2a\u69fd\u4f4d\uff09\uff0c\u8fd9\u91cc\u53ea\u9700\u8981\u77e5\u9053\uff1aCompiler \u5728\u8fd9\u4e00\u6b65\u628a\u6240\u6709\u5f20\u91cf\u3001\u68af\u5ea6\u3001\u52a8\u91cf\u3001EMA\u3001\u4e34\u65f6\u7f13\u51b2\u533a\u3001\u6807\u91cf\u7684 offset \u90fd\u786e\u5b9a\u4e0b\u6765\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cCompiler \u8fd8\u4f1a\u6839\u636e <code>GlobalRegistry<\/code> \u4e2d\u5f53\u524d\u7684\u4f18\u5316\u5668\u7c7b\u578b\u8986\u76d6 <code>PlanConfig<\/code> \u7684\u76f8\u5173\u6807\u5fd7\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=\"\">\/\/ src\/graph\/compiler.cpp:2454-2462\nauto opt = GlobalRegistry::instance().optimizer_kind();\npc.use_momentum = (opt != OptimizerKind::SGD);\npc.use_adam     = (opt == OptimizerKind::ADAM || opt == OptimizerKind::ADAMW);\npc.use_lars     = (opt == OptimizerKind::LARS || opt == OptimizerKind::LARS_NESTEROV);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\uff0c\u52a8\u91cf\u7f13\u51b2\u533a\u3001Adam \u4e8c\u9636\u77e9\u3001LARS \u8303\u6570\u533a\u7684\u5206\u914d\u4e0d\u662f\u786c\u7f16\u7801\u7684\uff0c\u800c\u662f\u7531\u5f53\u524d\u4f18\u5316\u5668\u81ea\u52a8\u51b3\u5b9a\uff1b\u8c03\u7528\u8005\u4f20\u5165\u7684 <code>PlanConfig<\/code> \u4e2d\u5bf9\u5e94\u5b57\u6bb5\u4f1a\u88ab\u8986\u76d6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.5 Phase 4\uff1a\u6784\u5efa ComputationGraph<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f\u628a ArchPlan \u53d8\u6210\u53ef\u6267\u884c\u56fe\u7684\u6838\u5fc3\u6b65\u9aa4\u3002<code>build_computation_graph<\/code> \u505a\u4e09\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6b63\u5411\u904d\u5386<\/strong>\uff1a\u5bf9\u6bcf\u5c42\u8c03\u7528 <code>build_forward<\/code>\uff0c\u751f\u6210\u524d\u5411\u8282\u70b9\uff0c\u6ce8\u5165\u8de8\u5c42\u6570\u636e\u94fe\uff1b<\/li>\n\n\n\n<li><strong>\u53cd\u5411\u904d\u5386<\/strong>\uff1a\u5bf9\u6bcf\u5c42\u8c03\u7528 <code>build_backward<\/code>\uff0c\u751f\u6210\u53cd\u5411\u8282\u70b9\uff0c\u6ce8\u5165\u68af\u5ea6\u94fe\uff1b<\/li>\n\n\n\n<li><strong>\u8f85\u52a9\u56fe<\/strong>\uff1a\u6784\u5efa <code>TRANSFER_A\/B<\/code>\u3001<code>ZERO_GRAD<\/code>\u3001<code>FIRST_COMM<\/code>\/<code>DEEP_COMM<\/code>\u3001<code>STATS_COMM<\/code>\u3001<code>OPTIMIZER<\/code>\uff08\u542b LARS \u65f6\u7684 <code>LARS_*_OPT<\/code>\uff09\u3001<code>EMA_UPDATE<\/code>\u3001AMP \u7c7b\u578b\u8f6c\u6362\uff08<code>CAST_*<\/code>\uff09\u3001<code>NAN_CHECK_AND_GRAD_SCALING<\/code>\u3001<code>UPDATE_STATS<\/code>\u3001<code>UPDATE_BN_INF_PARAMS<\/code>\u3001<code>CLEAR_METRICS<\/code>\u3001<code>ACCUM_METRICS_*<\/code>\u3001<code>VAL_RESULT_COMM<\/code> \u7b49\u56fe\uff1b<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u56fe<\/strong>\uff1a\u518d\u6b63\u5411\u904d\u5386\u4e00\u6b21\uff0c\u6784\u5efa <code>INF_MAIN_A\/B<\/code>\uff1b<code>INF_EMA_A\/B<\/code> \u7b49 ID \u867d\u5df2\u9884\u7559\uff0c\u4f46\u5f53\u524d Compiler \u9636\u6bb5\u5c1a\u672a\u586b\u5145\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u53cd\u5411\u904d\u5386\u6bd4\u524d\u5411\u66f4\u590d\u6742\uff0c\u56e0\u4e3a\u9700\u8981\u5904\u7406\u68af\u5ea6\u94fe\u548c in-place \u5199\u56de\u3002Compiler \u7528 <code>layer_input_ids<\/code> \u8bb0\u5f55\u4e86\u6bcf\u5c42\u524d\u5411\u8f93\u5165 X \u7684 DTensor ID\uff0c\u53cd\u5411\u65f6\u628a\u8fd9\u4e9b ID \u4f5c\u4e3a <code>dX<\/code> \u7684\u8f93\u51fa\u76ee\u6807\u3002\u4f8b\u5982\u5bf9\u4e8e Conv \u7684\u53cd\u5411\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=\"\">\/\/ src\/graph\/compiler.cpp:1375-1397\nif (is_conv_bwd_op(gn.compute_op) || is_cbr_bwd_op(gn.compute_op)) {\n    if (layer.is_first_layer) {\n        \/\/ \u9996\u5c42\uff1adX \u5199\u56de I_A_DATA \/ I_B_DATA\n        gn_a.output_ids.insert(gn_a.output_ids.begin(), b.data_a);\n        train_cg.append(GraphId::FIRST_LAYER_BWD_A, gn_a);\n        gn_b.output_ids.insert(gn_b.output_ids.begin(), b.data_b);\n        train_cg.append(GraphId::FIRST_LAYER_BWD_B, gn_b);\n    } else {\n        \/\/ \u975e\u9996\u5c42\uff1adX \u5199\u56de\u524d\u4e00\u5c42\u8f93\u51fa\uff08\u5373\u672c\u5c42\u8f93\u5165 X\uff09\n        gn.output_ids.insert(gn.output_ids.begin(), it->second);\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">SoftmaxCE \u7684\u53cd\u5411\u4f1a\u6ce8\u5165 <code>scaling<\/code>\u3001<code>label_smce<\/code>\u3001<code>label_smoothing<\/code> \u7b49\u57fa\u7ebf\u6807\u91cf\uff1bBN \u7684\u53cd\u5411\u4f1a\u8ffd\u52a0\u539f\u59cb\u8f93\u5165 X \u5e76\u628a dX \u5199\u56de X\u3002\u8fd9\u4e9b\u7ec6\u8282\u4fdd\u8bc1\u4e86\u524d\u5411\u548c\u53cd\u5411\u5728\u6570\u636e\u4f9d\u8d56\u4e0a\u7684\u5b8c\u5168\u5bf9\u5e94\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8f85\u52a9\u56fe\u4e2d\u5305\u542b\u5927\u91cf <code>RANGE<\/code> \u8282\u70b9\uff0c\u5b83\u4eec\u4e0d\u64cd\u4f5c\u5355\u4e2a DTensor\uff0c\u800c\u662f\u4e00\u6b21\u6027\u64cd\u4f5c\u6574\u4e2a Region\u3002\u6bd4\u5982 <code>ZERO_GRAD<\/code> \u56fe\u4f1a\u6e05\u7a7a\u6240\u6709\u68af\u5ea6\u533a\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=\"\">\/\/ src\/graph\/compiler.cpp:1304-1311\nGraphNode zg_node;\nzg_node.kind = GraphNode::Kind::RANGE;\nzg_node.range_op = RangeOp::RANGE_CLEAR;\nzg_node.output_ranges.push_back(\n    memory_plan.region_range(Region::G_BN_BIAS, Region::G_DEEP_CONV_FP16));\ntrain_cg.append(GraphId::ZERO_GRAD, zg_node);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd&#8221;Region \u7ea7\u6279\u91cf\u64cd\u4f5c&#8221;\u662f Tech-Renaissance \u663e\u5b58\u5206\u533a\u8bbe\u8ba1\u5e26\u6765\u7684\u7ea2\u5229\uff0c\u6211\u4eec\u5728\u7b2c 14 \u7bc7\u4f1a\u6df1\u5165\u5c55\u5f00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>ComputationGraph<\/code> \u662f<strong>\u96f6\u5f62\u72b6\u4fe1\u606f<\/strong>\u7684\u7eaf\u62d3\u6251\u5bb9\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=\"\">\/\/ include\/renaissance\/graph\/computation_graph.h:239-246\n\/**\n * @brief \u7eaf\u7b97\u5b50\u62d3\u6251\u5bb9\u5668\n * \u6838\u5fc3\u7279\u5f81\uff1a\u96f6\u5f62\u72b6\u4fe1\u606f\n *   - \u8282\u70b9\u4e2d\u53ea\u5b58 (GraphNode, OpParams, tensor_id)\n *   - Shape \/ DType \/ Region \u5168\u90e8\u4ece MemoryPlan \u7684 DTensor \u83b7\u53d6\n *   - \u4e00\u4efd\u56fe\u4f9b\u591a\u4e2a shape-only \u53d8\u4f53\u5171\u4eab\n *\/<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8282\u70b9\u5206\u4e3a\u4e24\u7c7b\uff1a<code>COMPUTE<\/code> \u8282\u70b9\u4f7f\u7528 DTensor ID\uff1b<code>RANGE<\/code> \u8282\u70b9\u4f7f\u7528\u9884\u8ba1\u7b97\u7684\u5185\u5b58\u8303\u56f4\u3002\u8fd9\u6837\u540c\u4e00\u4efd\u56fe\u65e2\u53ef\u4ee5\u7528\u5728 base \u53d8\u4f53\uff0c\u4e5f\u53ef\u4ee5\u7528\u5728 last\/lowres \u53d8\u4f53\uff0c\u53ea\u8981\u5b83\u4eec\u7684\u62d3\u6251\u7ed3\u6784\u76f8\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Compiler \u8fd8\u4e3a\u6bcf\u4e2a\u5b50\u56fe\u505a\u4e86\u804c\u8d23\u5212\u5206\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=\"\">\/\/ include\/renaissance\/graph\/computation_graph.h:73-108\nenum class GraphId : uint8_t {\n    TRANSFER_A, TRANSFER_B,\n    FIRST_LAYER_FWD_A, FIRST_LAYER_FWD_B,\n    DEEP_FWD_BWD,\n    ZERO_GRAD,\n    FIRST_LAYER_BWD_A, FIRST_LAYER_BWD_B,\n    FIRST_COMM, DEEP_COMM,\n    CAST_DEEP_GRAD_FP16_TO_FP32,\n    CAST_FIRST_GRAD_FP16_TO_FP32,\n    NAN_CHECK_AND_GRAD_SCALING,\n    STATS_COMM, UPDATE_STATS,\n    OPTIMIZER, EMA_UPDATE,\n    INF_MAIN_A, INF_MAIN_B,\n    INF_EMA_A, INF_EMA_B,\n    CAST_MAIN_FP32_TO_FP16, CAST_EMA_FP32_TO_FP16,\n    ACCUM_METRICS, ACCUM_METRICS_TRAIN_LAST, ACCUM_METRICS_VAL_LAST,\n    VAL_RESULT_COMM, CLEAR_METRICS,\n    SIMPLE_TASK_GRAPH,\n    LARS_FC_OPT, LARS_FIRST_CONV_OPT, LARS_DEEP_CONV_OPT,\n    UPDATE_BN_INF_PARAMS,\n    COUNT              \/\/ = 33\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>GraphId::COUNT = 33<\/code>\uff0c\u610f\u5473\u7740\u6574\u4e2a <code>ComputationGraph<\/code> \u5185\u90e8\u6309 33 \u4e2a\u6876\u6765\u7ec4\u7ec7\u5b50\u56fe\u3002\u5f53\u524d <code>Compiler<\/code> \u5b9e\u9645\u586b\u5145\u4e86\u8bad\u7ec3\u524d\u5411\/\u53cd\u5411\u3001\u6570\u636e\u642c\u8fd0\u3001\u901a\u4fe1\u3001\u4f18\u5316\u5668\u66f4\u65b0\u3001AMP \u7c7b\u578b\u8f6c\u6362\u3001\u68af\u5ea6 NaN \u68c0\u67e5\u3001\u7edf\u8ba1\u91cf\u66f4\u65b0\u3001\u6307\u6807\u7d2f\u79ef\u3001\u63a8\u7406\u4e3b\u4f53\u7b49\u5b50\u56fe\uff1b<code>INF_EMA_A\/B<\/code> \u7b49 ID \u867d\u5df2\u9884\u7559\uff0c\u4f46\u5c1a\u672a\u5728\u6b64\u9636\u6bb5\u586b\u5145\u3002\u5176\u4e2d shape \u65e0\u5173\u7684\u5b50\u56fe\uff08\u5982 <code>TRANSFER<\/code>\u3001<code>COMM<\/code>\u3001<code>OPTIMIZER<\/code>\u3001<code>EMA_UPDATE<\/code>\uff09\u5728\u6240\u6709\u53d8\u4f53\u95f4\u5171\u4eab\uff1bshape \u76f8\u5173\u7684\u5b50\u56fe\uff08\u5982 <code>FIRST_LAYER_FWD<\/code>\u3001<code>DEEP_FWD_BWD<\/code>\uff09\u6309 <code>ShapeId<\/code> \u53bb\u91cd\u6355\u83b7\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.6 Phase 5\uff1a\u53d8\u4f53\u5171\u4eab<\/h3>\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=\"\">\/\/ src\/graph\/compiler.cpp:2348-2367\nvoid Compiler::share_or_clone(Result&amp; result, ...) {\n    for (size_t i = 0; i &lt; result.variants.size(); ++i) {\n        auto&amp; v = result.variants[i];\n        if (v.name.find(\"val_\") == 0) {\n            v.train = nullptr;        \/\/ \u9a8c\u8bc1\u53d8\u4f53\u4e0d\u9700\u8981\u8bad\u7ec3\u56fe\n        } else {\n            v.train = &amp;train_cg;\n        }\n        v.inference = &amp;infer_cg;      \/\/ \u6240\u6709\u53d8\u4f53\u5171\u4eab\u63a8\u7406\u56fe\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Result<\/code> \u53ea\u6301\u6709\u4e24\u4efd <code>ComputationGraph<\/code>\uff08train \u548c infer\uff09\uff0c\u516d\u4e2a\u53d8\u4f53\u901a\u8fc7\u6307\u9488\u5171\u4eab\u5b83\u4eec\uff0c\u5404\u81ea\u6301\u6709\u72ec\u7acb\u7684 <code>MemoryPlan<\/code>\u3002\u8fd9\u907f\u514d\u4e86\u4e3a\u6bcf\u4e2a\u53d8\u4f53\u91cd\u590d\u5efa\u56fe\u3002\u51fd\u6570\u540d <code>share_or_clone<\/code> \u867d\u7136\u6697\u793a\u4e86&#8221;\u5171\u4eab\u6216\u514b\u9686&#8221;\u4e24\u79cd\u53ef\u80fd\uff0c\u4f46\u5f53\u524d\u5b9e\u73b0\u53ea\u8d70\u5171\u4eab\u5206\u652f\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u5173\u952e\u8bbe\u8ba1\u601d\u60f3\uff1a\u56fe\u4e0e\u5f62\u72b6\u89e3\u8026<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7f16\u8bd1\u7ba1\u7ebf\u7684\u6838\u5fc3\u8bbe\u8ba1\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e00\u53e5\u8bdd\uff1a<strong>\u628a&#8221;\u56fe\u62d3\u6251&#8221;\u548c&#8221;\u5f62\u72b6\/\u5185\u5b58\u5e03\u5c40&#8221;\u89e3\u8026<\/strong>\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>ArchPlan<\/code> \u6709\u5f62\u72b6\u4fe1\u606f\uff0c\u4f46\u5b83\u662f\u89c4\u8303\u5316\u540e\u7684\u5c42\u5e8f\u5217\uff0c\u4e0d\u662f\u6267\u884c\u56fe\u3002<\/li>\n\n\n\n<li><code>ComputationGraph<\/code> \u6ca1\u6709\u5f62\u72b6\u4fe1\u606f\uff0c\u53ea\u4fdd\u5b58\u7b97\u5b50\u62d3\u6251\u548c DTensor ID\u3002<\/li>\n\n\n\n<li><code>MemoryPlan<\/code> \u4fdd\u5b58\u6bcf\u4e2a DTensor \u7684\u5f62\u72b6\u3001dtype\u3001offset\u3001stride\u3002<\/li>\n\n\n\n<li><code>CompileSpec<\/code> \u53ea\u63cf\u8ff0\u5f62\u72b6\u76f8\u5173\u7684\u53c2\u6570\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u89e3\u8026\u7684\u597d\u5904\u662f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u540c\u4e00\u4efd\u56fe\u670d\u52a1\u591a\u4e2a\u53d8\u4f53<\/strong>\u3002train_base \u548c train_last \u7684 batch \u4e0d\u540c\uff0c\u4f46\u56fe\u62d3\u6251\u4e00\u6837\uff0c\u76f4\u63a5\u5171\u4eab\u3002<\/li>\n\n\n\n<li><strong>\u5f62\u72b6\u53d8\u5316\u4e0d\u9700\u8981\u91cd\u65b0\u5efa\u56fe<\/strong>\u3002\u6e10\u8fdb\u5f0f resize \u8bad\u7ec3\u4e2d\uff0c\u5206\u8fa8\u7387\u53d8\u5316\u53ea\u5f71\u54cd MemoryPlan \u548c\u6355\u83b7\u53c2\u6570\uff0c\u4e0d\u5f71\u54cd ComputationGraph\u3002<\/li>\n\n\n\n<li><strong>\u4fbf\u4e8e CUDA Graph \u6355\u83b7<\/strong>\u3002\u56fe\u62d3\u6251\u56fa\u5b9a\uff0c\u5730\u5740\u7531 MemoryPlan \u56fa\u5b9a\uff0c\u6355\u83b7\u548c\u91cd\u653e\u90fd\u7a33\u5b9a\u3002<\/li>\n\n\n\n<li><strong>\u4fbf\u4e8e\u5206\u5e03\u5f0f\u6269\u5c55<\/strong>\u3002\u540c\u4e00 MemoryPlan \u56fe\u7eb8\u53ef\u4ee5\u5728\u591a\u5f20 GPU \u4e0a\u5171\u4eab\u6267\u884c\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u6574\u4e2a\u7ba1\u7ebf\uff0c\u4f60\u8fd8\u4f1a\u53d1\u73b0\u53e6\u4e00\u4e2a\u7279\u70b9\uff1a<strong>\u591a\u8f6e\u89c4\u5212\u3001\u5c42\u5c42\u9012\u8fdb<\/strong>\u3002BluePrint \u7684 Block \u5b8f\u4e3a ArchPlan \u7684\u878d\u5408\u4fdd\u7559\u4e86&#8221;\u610f\u56fe&#8221;\uff1bArchPlan \u7684\u5f62\u72b6\u63a8\u5bfc\u4e3a Compiler \u7684\u663e\u5b58\u5206\u914d\u63d0\u4f9b\u4e86&#8221;\u5c3a\u5bf8&#8221;\uff1bCompiler \u7684 MemoryPlan \u4e3a ComputationGraph \u63d0\u4f9b\u4e86&#8221;\u5730\u5740&#8221;\uff1bComputationGraph \u53c8\u4e3a CUDA Graph \u6355\u83b7\u63d0\u4f9b\u4e86&#8221;\u6307\u4ee4\u5e8f\u5217&#8221;\u3002\u6bcf\u4e00\u6b65\u53ea\u505a\u81ea\u5df1\u8be5\u505a\u7684\u4e8b\uff0c\u4ea7\u51fa\u6e05\u6670\u5730\u4f20\u9012\u7ed9\u4e0b\u4e00\u6b65\u3002\u8fd9\u79cd\u5206\u5c42\u5e26\u6765\u7684\u53ef\u6d4b\u8bd5\u6027\u3001\u53ef\u6269\u5c55\u6027\u548c\u786e\u5b9a\u6027\uff0c\u662f\u9759\u6001\u56fe\u6846\u67b6\u7684\u5178\u578b\u4f18\u52bf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u4e0e PyTorch torch.compile \u7684\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5199\u5230\u8fd9\u91cc\uff0c\u53ef\u80fd\u6709\u4eba\u4f1a\u95ee\uff1aTech-Renaissance \u7684\u7f16\u8bd1\u548c PyTorch <code>torch.compile<\/code> \u6709\u4ec0\u4e48\u4e0d\u540c\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u5927\u7684\u533a\u522b\u5728\u4e8e\uff1a<strong>Tech-Renaissance \u4ece\u8bbe\u8ba1\u4e4b\u521d\u5c31\u662f\u9759\u6001\u7684\uff0c\u800c torch.compile \u662f\u5728\u52a8\u6001\u56fe\u4e4b\u4e0a\u4e8b\u540e\u8865\u7f16\u8bd1\u3002<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u56fe\u6355\u83b7\u65b9\u5f0f<\/strong>\uff1aTech-Renaissance \u7684\u56fe\u7531 <code>ArchPlan<\/code> + <code>Compiler<\/code> \u663e\u5f0f\u6784\u5efa\uff1btorch.compile \u9700\u8981 Dynamo \u5728\u8fd0\u884c\u65f6\u6355\u83b7 Python \u6267\u884c\u75d5\u8ff9\u3002<\/li>\n\n\n\n<li><strong>\u5f62\u72b6\u5904\u7406<\/strong>\uff1aTech-Renaissance \u5728\u7f16\u8bd1\u671f\u5c31\u5904\u7406\u591a\u53d8\u4f53\uff1btorch.compile \u9047\u5230\u52a8\u6001\u5f62\u72b6\u4f1a\u91cd\u65b0\u7f16\u8bd1\u6216 specialization\u3002<\/li>\n\n\n\n<li><strong>\u5185\u5b58\u5730\u5740<\/strong>\uff1aTech-Renaissance \u901a\u8fc7 MemoryPlan \u9759\u6001\u5206\u533a\u4fdd\u8bc1\u5730\u5740\u7a33\u5b9a\uff1btorch.compile \u9700\u8981\u989d\u5916\u7684 CUDA Graph Trees \u7b49\u673a\u5236\u6765\u7ef4\u62a4\u5730\u5740\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li><strong>\u7075\u6d3b\u6027<\/strong>\uff1atorch.compile \u53ef\u4ee5\u5904\u7406\u4efb\u610f Python \u63a7\u5236\u6d41\uff08\u867d\u7136\u4f1a graph break\uff09\uff1bTech-Renaissance \u76ee\u524d\u4e0d\u652f\u6301\u6570\u636e\u4f9d\u8d56\u63a7\u5236\u6d41\u548c\u8fd0\u884c\u671f\u52a8\u6001\u4fee\u6539\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u518d\u6b21\u8bf4\u660e\u4e86\u4e00\u4e2a\u9053\u7406\uff1a<strong>\u6ca1\u6709\u5b8c\u7f8e\u7684\u6846\u67b6\uff0c\u53ea\u6709\u9762\u5411\u7279\u5b9a\u573a\u666f\u7684\u53d6\u820d<\/strong>\u3002Tech-Renaissance \u727a\u7272\u4e86\u90e8\u5206\u7075\u6d3b\u6027\uff0c\u6362\u53d6\u4e86\u5168\u5c40\u4f18\u5316\u7a7a\u95f4\u548c\u8fd0\u884c\u671f\u786e\u5b9a\u6027\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">ArchPlan \u548c Compiler \u662f Tech-Renaissance \u4ece&#8221;\u6a21\u578b\u5b9a\u4e49&#8221;\u5230&#8221;\u53ef\u6267\u884c\u56fe&#8221;\u7684\u6865\u6881\u3002ArchPlan \u7528\u5341\u6b65\u7ba1\u7ebf\u628a BluePrint \u6811\u5c55\u5f00\u3001\u89c4\u8303\u5316\u3001\u878d\u5408\u6210\u6807\u51c6\u5316\u7684\u5c42\u5e8f\u5217\uff1bCompiler \u7528\u4e94\u9636\u6bb5\u628a\u5c42\u5e8f\u5217\u8f6c\u6362\u6210 MemoryPlan \u548c\u96f6\u5f62\u72b6\u7684 ComputationGraph\uff0c\u5e76\u5b9e\u73b0\u591a\u53d8\u4f53\u5171\u4eab\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\u8981\u70b9\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>ArchPlan \u662f\u7f16\u8bd1 IR<\/strong>\uff1a\u8d1f\u8d23 Block \u5408\u5e76\u3001CBR \u878d\u5408\u3001GapFC \u878d\u5408\u3001BN \u91cd\u547d\u540d\u3001\u5f62\u72b6\u63a8\u5bfc\u7b49\u67b6\u6784\u7ea7\u53d8\u6362\u3002<\/li>\n\n\n\n<li><strong>Compiler \u662f\u4e94\u9636\u6bb5\u7f16\u6392\u5668<\/strong>\uff1a\u5f62\u72b6\u63a8\u5bfc \u2192 \u6700\u5927\u69fd\u4f4d\u8ba1\u7b97 \u2192 MemoryPlan \u521b\u5efa \u2192 \u8ba1\u7b97\u56fe\u6784\u5efa \u2192 \u53d8\u4f53\u5171\u4eab\u3002<\/li>\n\n\n\n<li><strong>\u4e09\u6761\u4e00\u81f4\u6027\u7ea6\u675f<\/strong>\uff08\u6570\u91cf\u3001\u540d\u79f0\u3001Region\uff09\u4fdd\u8bc1\u8de8\u53d8\u4f53\u5f20\u91cf\u63cf\u8ff0\u4e00\u81f4\uff0c\u662f\u56fe\u5171\u4eab\u548c offset \u5bf9\u9f50\u7684\u57fa\u7840\u3002<\/li>\n\n\n\n<li><strong>\u56fe\u4e0e\u5f62\u72b6\u89e3\u8026<\/strong>\uff1aComputationGraph \u96f6\u5f62\u72b6\uff0cMemoryPlan \u7ba1\u5f62\u72b6\u548c\u5730\u5740\uff0c\u4e00\u4efd\u56fe\u670d\u52a1\u591a\u4efd\u53d8\u4f53\u3002<\/li>\n\n\n\n<li><strong>\u9759\u6001\u7f16\u8bd1\u6362\u53d6\u5168\u5c40\u4f18\u5316\u7a7a\u95f4<\/strong>\uff1a\u8fd9\u662f\u540e\u7eed CUDA Graph \u5168\u6355\u83b7\u3001\u9759\u6001\u663e\u5b58\u89c4\u5212\u3001\u786e\u5b9a\u6027\u8bad\u7ec3\u7684\u5171\u540c\u5e95\u5ea7\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u5c06\u4e13\u95e8\u8bb2 <code>ComputationGraph<\/code> \u548c <code>GraphAtlas<\/code> \u7684\u8bbe\u8ba1\u2014\u2014\u4e00\u4efd\u56fe\u7eb8\uff0c\u5982\u4f55\u88ab\u591a\u5904\u590d\u7528\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\u5341\u4e09 \u4e0a\u4e00\u7bc7\u6211\u4eec\u8bb2\u4e86 BluePr [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":729,"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-499","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\/499","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=499"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/499\/revisions"}],"predecessor-version":[{"id":724,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/499\/revisions\/724"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/729"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=499"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=499"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=499"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}