{"id":547,"date":"2026-07-08T03:11:47","date_gmt":"2026-07-07T19:11:47","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=547"},"modified":"2026-07-08T21:28:22","modified_gmt":"2026-07-08T13:28:22","slug":"%e8%9e%8d%e5%90%88%e4%bc%98%e5%8c%96%e5%99%a8%ef%bc%9asgd%e3%80%81adamw%e4%b8%8elars%e7%9a%84%e6%95%b4%e5%8c%ba%e6%89%b9%e9%87%8f%e6%9b%b4%e6%96%b0","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/547\/","title":{"rendered":"(21) \u878d\u5408\u4f18\u5316\u5668\uff1aSGD\u3001AdamW\u4e0eLARS\u7684\u6574\u533a\u6279\u91cf\u66f4\u65b0"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341\u4e00<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\uff0c\u672c\u8d28\u4e0a\u5c31\u662f\u65e0\u6570\u6b21\u5730\u201c\u7b97\u68af\u5ea6\u3001\u6539\u6743\u91cd\u201d\u3002\u524d\u5411\u3001\u53cd\u5411\u3001\u901a\u4fe1\u8fd9\u4e9b\u9636\u6bb5\u56fa\u7136\u70ed\u95f9\uff0c\u4f46\u771f\u6b63\u51b3\u5b9a\u53c2\u6570\u600e\u4e48\u8d70\u7684\uff0c\u662f\u4f18\u5316\u5668\u8fd9\u4e00\u6b65\u3002\u5f88\u591a\u4eba\u5bf9\u4f18\u5316\u5668\u7684\u5370\u8c61\u8fd8\u505c\u7559\u5728 <code>optimizer.step()<\/code> \u8fd9\u4e00\u884c\u4ee3\u7801\u4e0a\uff0c\u89c9\u5f97\u5b83\u4e0d\u8fc7\u662f\u5728\u68af\u5ea6\u540e\u9762\u4e58\u4e2a\u5b66\u4e60\u7387\u3001\u518d\u51cf\u5230\u6743\u91cd\u4e0a\u800c\u5df2\u3002\u4f46\u5982\u679c\u4f60\u771f\u7684\u53bb profile \u4e00\u6b21\u8bad\u7ec3\uff0c\u5c31\u4f1a\u53d1\u73b0\u4f18\u5316\u5668 step \u5e38\u5e38\u662f\u5bb9\u6613\u88ab\u4f4e\u4f30\u7684\u6027\u80fd\u70ed\u70b9\uff1a\u539f\u56e0\u4e0d\u5728\u8ba1\u7b97\u91cf\uff0c\u800c\u5728<strong>\u8c03\u5ea6\u5f00\u9500<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u4ee3\u7f51\u7edc\u52a8\u8f84\u51e0\u5341\u3001\u4e0a\u767e\u4e2a\u53c2\u6570\u5f20\u91cf\u3002\u5728\u4e3b\u6d41\u6846\u67b6\u91cc\uff0c\u6bcf\u4e2a\u5f20\u91cf\u90fd\u8981\u5355\u72ec\u542f\u52a8\u4e00\u4e2a CUDA kernel\uff0c\u4ece GPU \u4e0a\u8bfb\u6743\u91cd\u3001\u8bfb\u68af\u5ea6\u3001\u5199\u6743\u91cd\u3001\u5199\u52a8\u91cf\uff0c\u51e0\u5341\u6b21 kernel launch \u6392\u961f\u7b49 CPU \u6d3e\u53d1\u3002PyTorch \u7684 <code>torch.optim.AdamW<\/code> \u867d\u7136\u63d0\u4f9b\u4e86 <code>fused=True<\/code>\uff0c\u80fd\u628a\u5355\u4e2a\u53c2\u6570\u5f20\u91cf\u5185\u90e8\u7684\u8bfb\u5199\u5408\u5e76\u5230\u4e00\u4e2a kernel \u91cc\uff0c\u4e14\u501f\u52a9 multi-tensor apply \u628a\u8bb8\u591a\u53c2\u6570\u5f20\u91cf\u6253\u5305\u8fdb\u5c11\u91cf kernel\uff0c\u4f46\u8fd9\u4e9b\u5f20\u91cf\u5728\u663e\u5b58\u4e2d\u5e76\u4e0d\u8fde\u7eed\uff0c\u6bcf\u6b21 launch \u90fd\u8981\u4f20\u5165\u9010\u5f20\u91cf\u7684\u6307\u9488\u5143\u6570\u636e\uff0c\u4e5f\u65e0\u6cd5\u50cf\u6574\u533a\u8fde\u7eed\u5185\u5b58\u90a3\u6837\u4e00\u6b21\u7ebf\u6027\u626b\u8fc7\uff1b\u53c2\u6570\u7ec4\u95f4\u7684\u5faa\u73af\u4e0e Python \u4fa7\u7684\u8c03\u5ea6\u5f00\u9500\u4f9d\u7136\u5b58\u5728\u3002\u5bf9\u4e8e\u4e00\u4e2a\u62e5\u6709\u5927\u91cf\u5c0f\u53c2\u6570\u5f20\u91cf\u7684\u7f51\u7edc\uff08\u6bd4\u5982 ResNet-50 \u91cc\u5927\u91cf\u7684 BN gamma\/beta \u548c\u5377\u79ef\u6743\u91cd\uff09\uff0c\u8fd9\u4e9b\u5f00\u9500\u52a0\u8d77\u6765\u76f8\u5f53\u53ef\u89c2\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u6ca1\u6709\u8d70\u8fd9\u6761\u8def\u3002\u5f97\u76ca\u4e8e MemoryPlan \u7684\u663e\u5b58\u5206\u533a\u8bbe\u8ba1\uff0c\u4f18\u5316\u5668\u66f4\u65b0\u88ab\u8868\u8fbe\u6210\u5bf9<strong>\u8fde\u7eed\u663e\u5b58\u533a\u57df<\/strong>\u7684\u6279\u91cf\u64cd\u4f5c\uff1a\u4e0d\u7ba1\u6a21\u578b\u6709\u591a\u5c11\u5c42\uff0c\u6743\u91cd\u66f4\u65b0\u5f80\u5f80\u53ea\u9700\u4e00\u4e24\u4e2a kernel\uff0c\u5c31\u80fd\u628a\u5168\u90e8\u53ef\u8bad\u7ec3\u53c2\u6570\u904d\u5386\u4e00\u904d\u3002\u8fd9\u7bc7\u6587\u7ae0\u5c31\u6765\u804a\u804a Tech-Renaissance \u7684\u878d\u5408\u4f18\u5316\u5668\u8bbe\u8ba1\u2014\u2014SGD\u3001AdamW\u3001LARS \u4e09\u79cd\u7b97\u6cd5\uff0c\u5982\u4f55\u88ab\u6620\u5c04\u5230\u540c\u4e00\u5f20\u201c\u6574\u533a\u66f4\u65b0\u201d\u7684\u56fe\u91cc\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u4f18\u5316\u5668\u5230\u5e95\u5728\u7b97\u4ec0\u4e48\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6df1\u5165\u5b9e\u73b0\u4e4b\u524d\uff0c\u5148\u5feb\u901f\u56de\u987e\u4e00\u4e0b\u6211\u4eec\u8981\u652f\u6301\u7684\u4e09\u79cd\u4f18\u5316\u5668\u3002\u4e0b\u9762\u7684\u4f2a\u4ee3\u7801\u90fd\u7528 <code>w<\/code> \u8868\u793a\u6743\u91cd\u3001<code>g<\/code> \u8868\u793a\u68af\u5ea6\u3001<code>lr<\/code> \u8868\u793a\u5b66\u4e60\u7387\u3001<code>wd<\/code> \u8868\u793a weight decay\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>SGD\uff08\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff09<\/strong> \u662f\u6700\u57fa\u7840\u7684\u53c2\u6570\u66f4\u65b0\u7b56\u7565\uff0c\u6bcf\u4e00\u6b65\u6cbf\u68af\u5ea6\u53cd\u65b9\u5411\u8d70\u4e00\u4e2a\u6b65\u957f\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=\"\">\/\/ SGD: w = w - lr * g\nfor (size_t i = 0; i &lt; n; ++i) {\n    w[i] = w[i] - lr * g[i];\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f15\u5165<strong>\u52a8\u91cf<\/strong>\u4e4b\u540e\uff0cSGD \u4e0d\u518d\u53ea\u770b\u5f53\u524d\u68af\u5ea6\uff0c\u800c\u662f\u7ef4\u62a4\u4e00\u4e2a\u4e00\u9636\u52a8\u91cf\u7f13\u51b2\u533a <code>m<\/code>\uff0c\u628a\u5386\u53f2\u68af\u5ea6\u65b9\u5411\u4ee5\u7cfb\u6570 <code>beta<\/code> \u7d2f\u79ef\u8fdb\u6765\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=\"\">\/\/ Momentum:\n\/\/   m = beta * m + g\n\/\/   w = w - lr * m\nfor (size_t i = 0; i &lt; n; ++i) {\n    m[i] = beta * m[i] + g[i];\n    w[i] = w[i] - lr * m[i];\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u52a8\u91cf\u8ba9\u66f4\u65b0\u66f4\u5e73\u6ed1\uff0c\u6709\u52a9\u4e8e\u7a7f\u8d8a\u5e73\u7f13\u533a\u57df\u3001\u6291\u5236\u9707\u8361\u3002<strong>Nesterov \u52a8\u91cf<\/strong>\u5219\u662f\u628a\u201c\u5148\u770b\u672a\u6765\u4e00\u6b65\u7684\u52a8\u91cf\u65b9\u5411\u201d\u518d\u6c42\u68af\u5ea6\u3002\u5728\u6211\u4eec\u7684 kernel \u91cc\uff0c\u5b83\u7b49\u4ef7\u4e8e\u4e0b\u9762\u8fd9\u79cd\u66f4\u5229\u4e8e\u878d\u5408\u7684\u5f62\u5f0f\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=\"\">\/\/ Nesterov:\n\/\/   m_new = beta * m + g\n\/\/   w     = w - lr * (m_new * beta + g)\n\/\/   m     = m_new\nfor (size_t i = 0; i &lt; n; ++i) {\n    float g_i  = g[i];\n    float m_new = beta * m[i] + g_i;\n    w[i] = w[i] - lr * (m_new * beta + g_i);\n    m[i] = m_new;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6bb5\u903b\u8f91\u5728\u4ee3\u6570\u4e0a\u5b8c\u5168\u7b49\u4ef7\u4e8e\u6807\u51c6\u7684 Nesterov \u52a8\u91cf\uff0c\u53ea\u662f\u6211\u4eec\u628a\u5b83\u5c55\u5f00\u5e76\u91cd\u65b0\u7ed3\u5408\u3002\u8fd9\u6837\u505a\u7684\u5999\u5904\u5728\u4e8e\uff1a\u5b83\u5141\u8bb8\u5728\u4e00\u4e2a kernel \u7684\u5355\u6b21\u5faa\u73af\u5185\uff0c\u5229\u7528\u5bc4\u5b58\u5668\uff08<code>m_new<\/code>\uff09\u540c\u65f6\u5b8c\u6210\u524d\u540e\u6b65\u52a8\u91cf\u7684\u63a8\u5bfc\u4e0e\u6743\u91cd\u7684\u5199\u5165\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u663e\u5b58\u8bfb\u5199\u7684\u4f9d\u8d56\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Adam<\/strong> \u5728\u52a8\u91cf\u57fa\u7840\u4e0a\u53c8\u52a0\u4e86\u4e8c\u9636\u77e9\u4f30\u8ba1 <code>v<\/code>\uff0c\u5b83\u8bb0\u5f55\u68af\u5ea6\u5e73\u65b9\u7684\u6307\u6570\u79fb\u52a8\u5e73\u5747\uff0c\u4ece\u800c\u5bf9\u6bcf\u4e2a\u53c2\u6570\u5355\u72ec\u7f29\u653e\u5b66\u4e60\u7387\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=\"\">\/\/ Adam:\n\/\/   m = beta1 * m + (1 - beta1) * g\n\/\/   v = beta2 * v + (1 - beta2) * g * g\n\/\/   w = w - lr * m_hat \/ (sqrt(v_hat) + eps)\n\/\/ \u5176\u4e2d m_hat = m \/ (1 - beta1^t), v_hat = v \/ (1 - beta2^t)\nfor (size_t i = 0; i &lt; n; ++i) {\n    float g_i = g[i];\n    m[i] = beta1 * m[i] + (1.0f - beta1) * g_i;\n    v[i] = beta2 * v[i] + (1.0f - beta2) * g_i * g_i;\n    float m_hat = m[i] * bias_corr1;   \/\/ bias_corr1 = 1 \/ (1 - beta1^t)\n    float v_hat = v[i] * bias_corr2;   \/\/ bias_corr2 = 1 \/ (1 - beta2^t)\n    w[i] = w[i] - lr * m_hat \/ (sqrtf(v_hat) + eps);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e <code>m<\/code> \u548c <code>v<\/code> \u5728\u8bad\u7ec3\u521d\u671f\u90fd\u504f\u5411 0\uff0cAdam \u8fd8\u9700\u8981\u505a<strong>\u504f\u5dee\u4fee\u6b63<\/strong>\uff08bias correction\uff09\uff0c\u628a <code>m<\/code> \u548c <code>v<\/code> \u5206\u522b\u9664\u4ee5 <code>1 - beta1^t<\/code> \u548c <code>1 - beta2^t<\/code>\uff0c\u5176\u4e2d <code>t<\/code> \u662f\u8bad\u7ec3\u6b65\u6570\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>AdamW<\/strong> \u4e0e Adam \u7684\u6570\u5b66\u533a\u522b\u5f88\u5c0f\uff0c\u5374\u5f71\u54cd\u6df1\u8fdc\uff1a\u5b83\u628a\u6743\u91cd\u8870\u51cf\u4ece\u68af\u5ea6\u91cc\u89e3\u8026\u51fa\u6765\uff0c\u76f4\u63a5\u4f5c\u7528\u5728\u53c2\u6570\u4e0a\uff0c\u800c\u4e0d\u662f\u52a0\u5728\u68af\u5ea6\u91cc\u518d\u88ab\u4e8c\u9636\u77e9\u653e\u5927\u3002AdamW \u7684\u66f4\u65b0\u53ef\u4ee5\u5199\u6210\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=\"\">\/\/ AdamW:\n\/\/   w = w * (1 - lr * wd)\n\/\/   m = beta1 * m + (1 - beta1) * g\n\/\/   v = beta2 * v + (1 - beta2) * g * g\n\/\/   w = w - lr * m_hat \/ (sqrt(v_hat) + eps)\nfor (size_t i = 0; i &lt; n; ++i) {\n    float g_i = g[i];\n    w[i] = w[i] * (1.0f - lr * wd);\n    m[i] = beta1 * m[i] + (1.0f - beta1) * g_i;\n    v[i] = beta2 * v[i] + (1.0f - beta2) * g_i * g_i;\n    float m_hat = m[i] * bias_corr1;\n    float v_hat = v[i] * bias_corr2;\n    w[i] = w[i] - lr * m_hat \/ (sqrtf(v_hat) + eps);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48 AdamW \u5728\u5927\u6a21\u578b\u8bad\u7ec3\u91cc\u6210\u4e86\u4e8b\u5b9e\u6807\u51c6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>LARS\uff08Layer-wise Adaptive Rate Scaling\uff09<\/strong> \u7684\u601d\u8def\u5219\u5b8c\u5168\u4e0d\u540c\u3002\u5b83\u4e0d\u662f\u7ed9\u6bcf\u4e2a\u53c2\u6570\u5355\u72ec\u7684\u5b66\u4e60\u7387\uff0c\u800c\u662f\u7ed9<strong>\u6bcf\u4e00\u5c42<\/strong>\u4e00\u4e2a trust ratio\u3002\u5bf9\u7b2c <code>l<\/code> \u5c42\u7684\u6743\u91cd <code>w_l<\/code> \u548c\u68af\u5ea6 <code>g_l<\/code>\uff0c\u5148\u8ba1\u7b97\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=\"\">\/\/ LARS trust ratio:\n\/\/   eta_l = trust_coefficient * ||w_l|| \/ (||g_l|| + wd * ||w_l|| + eps)\n\/\/ \u7136\u540e\u7528 eta_l * lr \u4f5c\u4e3a\u8be5\u5c42\u7684\u6709\u6548\u5b66\u4e60\u7387<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">LARS \u7684\u521d\u8877\u662f\u89e3\u51b3\u8d85\u5927 batch \u8bad\u7ec3\u65f6\u7684\u4f18\u5316\u7a33\u5b9a\u6027\u95ee\u9898\uff1a\u4e0d\u540c\u5c42\u7684\u6743\u91cd\u548c\u68af\u5ea6\u91cf\u7ea7\u5dee\u5f02\u5de8\u5927\uff0c\u7edf\u4e00\u7684\u5168\u5c40\u5b66\u4e60\u7387\u5f80\u5f80\u987e\u6b64\u5931\u5f7c\u3002LARS \u7684\u9010\u5c42\u7f29\u653e\u8ba9\u5927 batch \u8bad\u7ec3\u53d8\u5f97\u53ef\u884c\uff0c\u4e5f\u662f ResNet \u5728 batch size \u8fbe\u5230\u6570\u5343\u4e43\u81f3\u4e0a\u4e07\u65f6\u4ecd\u80fd\u6536\u655b\u7684\u5173\u952e\u6280\u672f\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u7684 foreach\/fused optimizer \u5df2\u7ecf\u80fd\u628a\u591a\u4e2a\u5f20\u91cf\u7684\u66f4\u65b0\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u5408\u5e76\uff0c\u663e\u8457\u51cf\u5c11\u9010\u5f20\u91cf Python \u5faa\u73af\u548c kernel launch \u5f00\u9500\u3002\u4f46\u5b83\u7684\u7ec4\u7ec7\u5355\u4f4d\u4ecd\u7136\u662f\u53c2\u6570\u5217\u8868\u548c\u53c2\u6570\u7ec4\uff0c\u65e0\u6cd5\u5229\u7528 Tech-Renaissance \u8fd9\u79cd&#8221;\u540c\u8bed\u4e49\u53c2\u6570\u5728\u663e\u5b58\u4e2d\u8fde\u7eed\u6392\u5e03&#8221;\u7684 Region \u5e03\u5c40\u3002\u56e0\u6b64\uff0c\u5728\u672c\u6846\u67b6\u91cc\uff0c\u4f18\u5316\u5668\u53ef\u4ee5\u8fdb\u4e00\u6b65\u88ab\u8868\u8fbe\u6210\u6574\u533a RangeOp\uff0c\u7528\u6781\u5c11\u6570 kernel \u8986\u76d6\u6574\u7ec4\u53c2\u6570\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Tech-Renaissance \u7684\u4f18\u5316\u5668\uff1aRegion \u5373\u6279\u91cf<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u4e2d\uff0c\u4f18\u5316\u5668\u4e0d\u662f\u5bf9\u7740\u4e00\u5806\u5206\u6563\u7684\u5f20\u91cf\u505a\u5faa\u73af\uff0c\u800c\u662f\u76f4\u63a5\u5bf9\u663e\u5b58 Region \u505a\u6279\u91cf\u64cd\u4f5c\u3002\u8fd9\u662f MemoryPlan \u5206\u533a\u8bbe\u8ba1\u7684\u76f4\u63a5\u4ea7\u7269\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u5230 <code>include\/renaissance\/core\/types.h<\/code> \u91cc\u7684 Region \u679a\u4e3e\uff0c\u4e0e\u4f18\u5316\u5668\u76f4\u63a5\u76f8\u5173\u7684\u6709\u4ee5\u4e0b\u51e0\u6761\u7ebf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>W \u7cfb\u5217<\/strong>\uff1a<code>W_BN_BIAS<\/code>\u3001<code>W_BN_WEIGHT<\/code>\u3001<code>W_FC_BIAS<\/code>\u3001<code>W_FC_WEIGHT<\/code>\u3001<code>W_FIRST_CONV<\/code>\u3001<code>W_DEEP_CONV<\/code>\uff0c\u5b58\u653e FP32 \u4e3b\u6743\u91cd\uff1b<\/li>\n\n\n\n<li><strong>G \u7cfb\u5217<\/strong>\uff1a\u5bf9\u5e94\u540c\u540d\u7684\u68af\u5ea6\u533a\uff1b<\/li>\n\n\n\n<li><strong>M \u7cfb\u5217<\/strong>\uff1a\u4e00\u9636\u52a8\u91cf\u7f13\u51b2\u533a\uff1b<\/li>\n\n\n\n<li><strong>V \u7cfb\u5217<\/strong>\uff1a\u4e8c\u9636\u52a8\u91cf\u7f13\u51b2\u533a\uff08Adam\/AdamW \u4e13\u7528\uff09\uff1b<\/li>\n\n\n\n<li><strong>N \u7cfb\u5217<\/strong>\uff1aLARS \u9010\u5c42 trust ratio \u7684\u5b58\u50a8\u533a\uff1b<\/li>\n\n\n\n<li><strong>S \u7cfb\u5217<\/strong>\uff1a<code>S_SCALAR_FP32<\/code>\u3001<code>S_SCALAR_INT32<\/code>\uff0c\u5b58\u653e <code>lr<\/code>\u3001<code>wd<\/code>\u3001<code>beta<\/code>\u3001<code>scaling<\/code>\u3001<code>has_nan<\/code> \u7b49\u6807\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e\u540c\u4e00\u8bed\u4e49\u7684\u5f20\u91cf\u88ab\u8fde\u7eed\u6446\u653e\uff0c\u4f18\u5316\u5668\u53ef\u4ee5\u7528\u4e00\u6b21 <code>region_range(start, end)<\/code> \u5c31\u8986\u76d6\u5168\u90e8\u540c\u7c7b\u53c2\u6570\u3002\u4f8b\u5982\uff0c\u6240\u6709 FC \u4e0e\u5377\u79ef\u6743\u91cd\u8fde\u7eed\u653e\u5728 <code>W_FC_WEIGHT<\/code> \u5230 <code>W_DEEP_CONV<\/code> \u4e4b\u95f4\uff0c\u5b83\u4eec\u7684\u68af\u5ea6\u4e5f\u8fde\u7eed\u653e\u5728 <code>G_FC_WEIGHT<\/code> \u5230 <code>G_DEEP_CONV<\/code> \u4e4b\u95f4\uff0c\u52a8\u91cf\u7f13\u51b2\u533a\u8fde\u7eed\u653e\u5728 <code>M_FC_WEIGHT<\/code> \u5230 <code>M_DEEP_CONV<\/code> \u4e4b\u95f4\uff0c\u4e8c\u9636\u77e9\u7f13\u51b2\u533a\u8fde\u7eed\u653e\u5728 <code>V_FC_WEIGHT<\/code> \u5230 <code>V_DEEP_CONV<\/code> \u4e4b\u95f4\u3002\u8fd9\u79cd\u201c\u540c\u5f62\u540c\u533a\u201d\u7684\u5e03\u5c40\u4e0d\u662f\u5076\u7136\u7684\uff0c\u800c\u662f\u7f16\u8bd1\u5668\u548c MemoryPlan \u5171\u540c\u4fdd\u8bc1\u7684\uff1a\u5f53\u67d0\u4e00\u5c42\u88ab\u5206\u914d\u5230 <code>W_FC_WEIGHT<\/code> \u65f6\uff0c\u5b83\u7684\u68af\u5ea6\u3001\u52a8\u91cf\u3001\u4e8c\u9636\u77e9\u90fd\u4f1a\u6309\u76f8\u540c\u987a\u5e8f\u843d\u5728\u5bf9\u5e94\u533a\u57df\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u6237\u5728 <code>include\/renaissance\/algo\/optimizer.h<\/code> \u4e2d\u901a\u8fc7\u7eaf\u914d\u7f6e\u7c7b\u9009\u62e9\u4f18\u5316\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=\"\">\/\/ SGD with Momentum\ntask.optimizer(SGD().momentum(0.9f).weight_decay(5e-5f).nesterov(false));\n\n\/\/ AdamW\ntask.optimizer(AdamW().beta1(0.9f).beta2(0.999f).eps(1e-8f).weight_decay(0.01f));\n\n\/\/ LARS\ntask.optimizer(LARS().momentum(0.9f).weight_decay(5e-5f)\n                       .trust_coefficient(0.001f));<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u914d\u7f6e\u5728 <code>task.compile()<\/code> \u65f6\u88ab\u7f16\u8bd1\u5668\u8bfb\u53d6\u3002<code>plan_config_from_optimizer()<\/code> \u51fd\u6570\u5c06 <code>OptimizerKind<\/code> \u7ffb\u8bd1\u4e3a <code>PlanConfig<\/code> \u7684\u6807\u5fd7\u4f4d\uff0c\u51b3\u5b9a MemoryPlan \u662f\u5426\u5206\u914d M\u3001V\u3001N \u7cfb\u5217\u533a\u57df\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=\"\">inline PlanConfig plan_config_from_optimizer(OptimizerKind kind, bool has_ema = false) {\n    PlanConfig cfg;\n    cfg.has_ema = has_ema;\n    switch (kind) {\n        case OptimizerKind::SGD:\n            cfg.use_momentum = false;\n            cfg.use_adam     = false;\n            cfg.use_lars     = false;\n            break;\n        case OptimizerKind::SGD_MOMENTUM:\n        case OptimizerKind::SGD_NESTEROV:\n            cfg.use_momentum = true;\n            cfg.use_adam     = false;\n            cfg.use_lars     = false;\n            break;\n        case OptimizerKind::LARS:\n        case OptimizerKind::LARS_NESTEROV:\n            cfg.use_momentum = true;\n            cfg.use_adam     = false;\n            cfg.use_lars     = true;\n            break;\n        case OptimizerKind::ADAM:\n        case OptimizerKind::ADAMW:\n            cfg.use_momentum = true;\n            cfg.use_adam     = true;\n            cfg.use_lars     = false;\n            break;\n        default:\n            break;\n    }\n    return cfg;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u9009 SGD \u5c31\u4e0d\u5206\u914d M \u7cfb\u5217\uff0c\u9009 AdamW \u5c31\u591a\u5206\u914d V \u7cfb\u5217\uff0c\u6ca1\u6709\u6d6a\u8d39\u3002\u8fd9\u79cd\u8bbe\u8ba1\u4e0b\uff0c\u4f18\u5316\u5668\u7684\u9009\u62e9\u76f4\u63a5\u5f71\u54cd\u663e\u5b58\u5e03\u5c40\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f SGD\/Momentum\/Adam\/AdamW \u7684 weight \u66f4\u65b0\u53ef\u4ee5\u88ab\u8868\u8fbe\u4e3a\u4e00\u4e2a <code>RANGE_UPDATE_WEIGHT_*<\/code> \u8282\u70b9\u3002\u4ee5 AdamW \u4e3a\u4f8b\uff0c<code>src\/graph\/compiler.cpp<\/code> \u4e2d\u7684\u6784\u5efa\u903b\u8f91\u5927\u81f4\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ 1. step += 1\nGraphNode inc_node;\ninc_node.kind = GraphNode::Kind::COMPUTE;\ninc_node.compute_op = ComputeOp::SCALAR_INCREMENT;\ninc_node.input_ids  = {scalar_ids.step};\ninc_node.output_ids = {scalar_ids.step};\ntrain_cg.append(GraphId::OPTIMIZER, inc_node);\n\n\/\/ 2. bc1 = 1\/(1-beta1^step), bc2 = 1\/(1-beta2^step)\nGraphNode bc_node;\nbc_node.kind = GraphNode::Kind::COMPUTE;\nbc_node.compute_op = ComputeOp::ADAM_BIAS_CORRECTION;\nbc_node.input_ids  = {scalar_ids.step, scalar_ids.beta, scalar_ids.beta2};\nbc_node.output_ids = {scalar_ids.bias_corr1, scalar_ids.bias_corr2};\ntrain_cg.append(GraphId::OPTIMIZER, bc_node);\n\n\/\/ 3. \u6574\u533a weight \u66f4\u65b0\nMemRange w_range = memory_plan.region_range(\n    Region::W_FC_WEIGHT, Region::W_DEEP_CONV);\nMemRange g_range = memory_plan.region_range(\n    Region::G_FC_WEIGHT, Region::G_DEEP_CONV);\nMemRange m_range = memory_plan.region_range(\n    Region::M_FC_WEIGHT, Region::M_DEEP_CONV);\nMemRange v_range = memory_plan.region_range(\n    Region::V_FC_WEIGHT, Region::V_DEEP_CONV);\n\nGraphNode node;\nnode.kind = GraphNode::Kind::RANGE;\nnode.range_op = RangeOp::RANGE_UPDATE_WEIGHT_ADAMW;\nnode.input_ranges.push_back(w_range);  \/\/ W\nnode.input_ranges.push_back(g_range);  \/\/ G\nnode.input_ranges.push_back(m_range);  \/\/ M\nnode.input_ranges.push_back(v_range);  \/\/ V\nnode.output_ranges.push_back(w_range);\nnode.output_ranges.push_back(m_range);\nnode.output_ranges.push_back(v_range);\ntrain_cg.append(GraphId::OPTIMIZER, node);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8282\u70b9\u6700\u540e\u88ab\u653e\u8fdb <code>GraphId::OPTIMIZER<\/code> \u56fe\u91cc\uff0c\u8fd0\u884c\u5728 <code>StreamKind::UPDATE<\/code> \u6d41\u4e0a\u3002\u8fd0\u884c\u65f6\u53ea\u9700\u8981 launch \u4e00\u6b21\u56fe\uff0c\u5c31\u80fd\u5b8c\u6210\u6574\u4e2a\u6a21\u578b\u6240\u6709\u6743\u91cd\u7684\u66f4\u65b0\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8fd9\u4e9b Region \u7684\u8d77\u59cb\u5730\u5740\u548c\u5927\u5c0f\u5728\u7f16\u8bd1\u671f\u5c31\u5df2\u7ecf\u786e\u5b9a\uff0c<code>w_range<\/code> \u548c <code>g_range<\/code> \u4e0d\u518d\u662f\u5f20\u91cf\u63cf\u8ff0\u7b26\u5217\u8868\uff0c\u800c\u662f <code>(offset, size)<\/code> \u7684\u5185\u5b58\u533a\u95f4\u3002kernel \u542f\u52a8\u65f6\u4e0d\u9700\u8981\u904d\u5386\u53c2\u6570\u5217\u8868\u3001\u4e0d\u9700\u8981\u52a8\u6001\u8ba1\u7b97\u504f\u79fb\uff0c\u53ea\u9700\u8981\u4e24\u4e2a\u6307\u9488\u548c\u5143\u7d20\u4e2a\u6570\u3002\u8fd9\u4e3a CUDA Graph \u5168\u6355\u83b7\u521b\u9020\u4e86\u6761\u4ef6\uff1a\u56fe\u7684\u62d3\u6251\u3001\u5185\u5b58\u5730\u5740\u3001kernel grid \u5168\u90e8\u9759\u6001\u53ef\u77e5\uff0c\u8fd0\u884c\u65f6\u53ea\u662f <code>cudaGraphLaunch<\/code> \u7684\u4e00\u6b21\u91cd\u653e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001SGD\u3001Momentum\u3001AdamW \u7684 CUDA \u5185\u6838<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u7684\u5185\u6838\u5b9e\u73b0\u4f4d\u4e8e <code>src\/backend\/ops\/range\/optimizer_op.cu<\/code>\u3002\u4ee5 AdamW \u4e3a\u4f8b\uff0c\u5176\u6838\u5fc3 kernel \u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">__global__ void update_adamw_kernel(\n    float* __restrict__ w, const float* __restrict__ g,\n    float* __restrict__ m, float* __restrict__ v, size_t n,\n    const float* __restrict__ lr, const float* __restrict__ wd,\n    const float* __restrict__ b1, const float* __restrict__ b2,\n    const float* __restrict__ eps,\n    const int32_t* __restrict__ has_nan,\n    const float* __restrict__ scaling,\n    const float* __restrict__ bias_corr1,\n    const float* __restrict__ bias_corr2)\n{\n    if (*has_nan != 0) return;\n    float _lr = *lr;\n    float _wd = wd ? *wd : 0.0f;\n    float _b1 = *b1;\n    float _b2 = *b2;\n    float _eps = *eps;\n    float _inv_scaling = (scaling &amp;&amp; *scaling != 0.0f) ? (1.0f \/ *scaling) : 1.0f;\n    float _bc1 = bias_corr1 ? *bias_corr1 : 1.0f;\n    float _bc2 = bias_corr2 ? *bias_corr2 : 1.0f;\n\n    for (size_t i = blockIdx.x * blockDim.x + threadIdx.x;\n         i &lt; n; i += gridDim.x * blockDim.x) {\n        float g_i = g[i] * _inv_scaling;\n        w[i] = w[i] * (1.0f - _lr * _wd);\n        m[i] = m[i] * _b1 + (1.0f - _b1) * g_i;\n        v[i] = v[i] * _b2 + (1.0f - _b2) * g_i * g_i;\n        float m_hat = m[i] * _bc1;\n        float v_hat = v[i] * _bc2;\n        w[i] = w[i] - _lr * m_hat \/ (sqrtf(v_hat) + _eps);\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a kernel \u91cc\u878d\u5408\u4e86\u56db\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>AMP \u53cd\u7f29\u653e<\/strong>\uff1a<code>g_i = g[i] * _inv_scaling<\/code>\uff0c\u628a AMP \u9636\u6bb5\u653e\u5927\u8fc7\u7684\u68af\u5ea6\u8fd8\u539f\uff1b<\/li>\n\n\n\n<li><strong>\u6743\u91cd\u8870\u51cf<\/strong>\uff1a<code>w[i] = w[i] * (1.0f - _lr * _wd)<\/code>\uff0c\u8fd9\u662f AdamW \u7684\u89e3\u8026 weight decay\uff1b<\/li>\n\n\n\n<li><strong>\u4e00\u9636\/\u4e8c\u9636\u77e9\u66f4\u65b0<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>\u504f\u5dee\u4fee\u6b63\u540e\u7684\u53c2\u6570\u66f4\u65b0<\/strong>\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u6240\u6709\u8d85\u53c2\u6570\u90fd\u4e0d\u662f\u901a\u8fc7 kernel \u53c2\u6570\u4f20\u5165\uff0c\u800c\u662f\u901a\u8fc7<strong>\u8bbe\u5907\u6307\u9488<\/strong>\u4f20\u5165\u3002\u8fd9\u4e00\u70b9\u975e\u5e38\u5173\u952e\uff1a\u5982\u679c <code>lr<\/code>\u3001<code>wd<\/code> \u7b49\u6807\u91cf\u4ee5 kernel \u53c2\u6570\u5f62\u5f0f\u4f20\u5165\uff0c\u90a3\u4e48\u6bcf\u6b21\u5b66\u4e60\u7387\u53d8\u5316\u90fd\u9700\u8981\u91cd\u65b0\u7f16\u8bd1 CUDA Graph\uff1b\u800c\u4ee5\u8bbe\u5907\u6307\u9488\u4f20\u5165\u65f6\uff0ckernel \u8bfb\u5230\u7684\u662f\u663e\u5b58\u4e2d\u5b9e\u65f6\u53d8\u5316\u7684\u6807\u91cf\uff0c\u56fe\u7ed3\u6784\u672c\u8eab\u65e0\u9700\u6539\u53d8\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=\"\">const float* lr  = scalar_ptr&lt;0>(mp, node.input_ids.data(), ctx.rank_for_context());\nconst float* wd  = scalar_ptr&lt;1>(mp, node.input_ids.data(), ctx.rank_for_context());\n\/\/ ...<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u6807\u91cf\u5168\u90e8\u5b58\u653e\u5728 <code>S_SCALAR_FP32<\/code> \u6216 <code>S_SCALAR_INT32<\/code> \u533a\u57df\u3002\u8fd0\u884c\u65f6\u53ea\u8981\u901a\u8fc7\u4e00\u6b21 H2D \u628a\u5c0f\u6807\u91cf\u5199\u8fdb\u663e\u5b58\uff0ckernel \u8bfb\u53d6\u7684\u5c31\u662f\u6700\u65b0\u503c\uff0cCUDA Graph \u65e0\u9700\u91cd\u65b0\u6355\u83b7\u3002\u8fd9\u5bf9\u4e8e\u5b66\u4e60\u7387\u8c03\u5ea6\u81f3\u5173\u91cd\u8981\uff1a\u6bcf\u4e2a batch \u7684\u5b66\u4e60\u7387\u53ef\u4ee5\u53d8\u5316\uff0c\u4f46\u8bad\u7ec3\u56fe\u4fdd\u6301\u4e0d\u53d8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>has_nan<\/code> \u4e5f\u4ee5\u8bbe\u5907\u6307\u9488\u4f20\u5165\u3002\u4e00\u65e6 <code>RANGE_CHECK_NAN<\/code> \u68c0\u6d4b\u5230\u68af\u5ea6\u5f02\u5e38\uff0c\u6240\u6709\u4f18\u5316\u5668 kernel \u4f1a\u5728\u6700\u5f00\u59cb\u7684 <code>if (*has_nan != 0) return;<\/code> \u5904\u76f4\u63a5\u9000\u51fa\uff0c\u7b49\u4ef7\u4e8e\u8df3\u8fc7\u672c\u8f6e\u66f4\u65b0\u3002\u8fd9\u4e2a\u5224\u65ad\u5728 GPU \u5185\u90e8\u5b8c\u6210\uff0c\u4e0d\u9700\u8981 CPU \u4ecb\u5165\u6bcf\u4e2a\u53c2\u6570\u5f20\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SGD \u4e0e Momentum \u7684 kernel \u66f4\u7b80\u6d01\uff0c\u4f46\u9075\u5faa\u540c\u6837\u7684\u6a21\u5f0f\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=\"\">\/\/ SGD: w = w * (1 - lr * wd) - lr * g\n__global__ void update_sgd_kernel(...)\n{\n    if (*has_nan != 0) return;\n    float _lr = *lr;\n    float _wd = wd ? *wd : 0.0f;\n    float _inv_scaling = (scaling &amp;&amp; *scaling != 0.0f) ? (1.0f \/ *scaling) : 1.0f;\n    for (size_t i = ...; i &lt; n; i += ...) {\n        float w_i = w[i];\n        float g_i = g[i] * _inv_scaling;\n        w[i] = w_i * (1.0f - _lr * _wd) - _lr * g_i;\n    }\n}\n\n\/\/ Momentum:\n\/\/   m = beta * m + g\n\/\/   w = w * (1 - lr * wd) - lr * m\n__global__ void update_momentum_kernel(...)\n{\n    if (*has_nan != 0) return;\n    float _beta = *beta;\n    for (size_t i = ...; i &lt; n; i += ...) {\n        float g_i = g[i] * _inv_scaling;\n        m[i] = m[i] * _beta + g_i;\n        w[i] = w[i] * (1.0f - _lr * _wd) - _lr * m[i];\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u6709 kernel \u90fd\u4f7f\u7528 <code>__launch_bounds__(128, 2)<\/code> \u7f16\u8bd1\u63d0\u793a\u3002\u5b83\u544a\u8bc9\u7f16\u8bd1\u5668\uff1a\u6bcf\u5757 128 \u7ebf\u7a0b\u3001\u671f\u671b\u6bcf\u4e2a SM \u81f3\u5c11\u9a7b\u7559 2 \u4e2a block\uff08<code>minBlocksPerMultiprocessor<\/code>\uff09\uff0c\u4ece\u800c\u5f15\u5bfc\u7f16\u8bd1\u5668\u63a7\u5236\u5bc4\u5b58\u5668\u7528\u91cf\u4ee5\u4fdd\u8bc1\u5360\u7528\u7387\uff0c\u5728\u5927\u591a\u6570 GPU \u67b6\u6784\u4e0a\u662f\u4e00\u4e2a\u7a33\u5065\u7684\u5e73\u8861\u70b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">CPU fallback \u7248\u672c\u4f4d\u4e8e <code>optimizer_op.cpp<\/code> \u4e0b\u534a\u90e8\u5206\uff0c\u7528\u4e8e\u65e0 GPU \u73af\u5883\u6216\u6b63\u786e\u6027\u9a8c\u8bc1\u3002CUDA \u4e0e CPU \u7684\u6570\u5b66\u516c\u5f0f\u4fdd\u6301\u4e00\u81f4\uff0c\u8fd9\u662f Tech-Renaissance \u4fdd\u8bc1\u6570\u503c\u53ef\u590d\u73b0\u7684\u57fa\u7840\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001Bias-like \u53c2\u6570\u7684\u5206\u7ec4\u54f2\u5b66<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>compiler.cpp<\/code> \u4e2d\u6784\u5efa\u4f18\u5316\u5668\u56fe\u65f6\uff0cTech-Renaissance \u628a\u6240\u6709\u53ef\u8bad\u7ec3\u53c2\u6570\u5206\u6210\u4e86\u4e24\u7ec4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Weight \u7ec4<\/strong>\uff1a<code>W_FC_WEIGHT<\/code>\u3001<code>W_FIRST_CONV<\/code>\u3001<code>W_DEEP_CONV<\/code>\uff0c\u9700\u8981 weight decay\uff1b<\/li>\n\n\n\n<li><strong>Bias-like \u7ec4<\/strong>\uff1a<code>W_BN_BIAS<\/code>\u3001<code>W_BN_WEIGHT<\/code>\u3001<code>W_FC_BIAS<\/code>\uff0c<strong>\u4e0d\u9700\u8981 weight decay<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u5206\u7ec4\u4e0d\u662f\u968f\u610f\u5b9a\u7684\uff0c\u800c\u662f\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4e2d\u7684\u4e00\u4e2a\u5e38\u89c1\u7ea6\u5b9a\uff1aBatchNorm \u7684 scale\/bias \u548c\u5168\u8fde\u63a5\u5c42\u7684 bias \u901a\u5e38\u4e0d\u53c2\u4e0e L2 \u6b63\u5219\u5316\u3002\u5982\u679c\u5f3a\u884c\u628a BN \u7684 gamma \u4e5f\u505a weight decay\uff0c\u5728\u5f88\u591a\u6807\u51c6\u8bad\u7ec3\u914d\u7f6e\u91cc\u4f1a\u88ab\u8ba4\u4e3a\u7834\u574f\u4e86\u5f52\u4e00\u5316\u5c42\u7684\u5c3a\u5ea6\u4e0d\u53d8\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee3\u7801\u91cc\u901a\u8fc7 <code>static_assert<\/code> \u5f3a\u5236\u8981\u6c42\u8fd9\u4e24\u7ec4\u5728 Region \u679a\u4e3e\u4e2d\u8fde\u7eed\u6392\u5217\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_assert(\n    static_cast&lt;int>(Region::W_BN_BIAS) + 1 == static_cast&lt;int>(Region::W_BN_WEIGHT) &amp;&amp;\n    static_cast&lt;int>(Region::W_BN_WEIGHT) + 1 == static_cast&lt;int>(Region::W_FC_BIAS),\n    \"W_BN_BIAS, W_BN_WEIGHT, W_FC_BIAS must be consecutive in Region enum\"\n);\nstatic_assert(\n    static_cast&lt;int>(Region::G_BN_BIAS) + 1 == static_cast&lt;int>(Region::G_BN_WEIGHT) &amp;&amp;\n    static_cast&lt;int>(Region::G_BN_WEIGHT) + 1 == static_cast&lt;int>(Region::G_FC_BIAS),\n    \"G_BN_BIAS, G_BN_WEIGHT, G_FC_BIAS must be consecutive in Region enum\"\n);\nstatic_assert(\n    static_cast&lt;int>(Region::M_BN_BIAS) + 1 == static_cast&lt;int>(Region::M_BN_WEIGHT) &amp;&amp;\n    static_cast&lt;int>(Region::M_BN_WEIGHT) + 1 == static_cast&lt;int>(Region::M_FC_BIAS),\n    \"M_BN_BIAS, M_BN_WEIGHT, M_FC_BIAS must be consecutive in Region enum\"\n);\nstatic_assert(\n    static_cast&lt;int>(Region::V_BN_BIAS) + 1 == static_cast&lt;int>(Region::V_BN_WEIGHT) &amp;&amp;\n    static_cast&lt;int>(Region::V_BN_WEIGHT) + 1 == static_cast&lt;int>(Region::V_FC_BIAS),\n    \"V_BN_BIAS, V_BN_WEIGHT, V_FC_BIAS must be consecutive in Region enum\"\n);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f Bias-like \u66f4\u65b0\u53ef\u4ee5\u7528\u4e00\u4e2a <code>region_range(W_BN_BIAS, W_FC_BIAS)<\/code> \u8986\u76d6\u5168\u90e8\uff0c\u5bf9\u5e94 <code>RANGE_UPDATE_BIAS_*<\/code> \u7cfb\u5217 RangeOp\u3002Bias \u8def\u5f84\u7684 kernel \u4e0d\u4f20 <code>wd<\/code> \u6307\u9488\uff0c\u5185\u90e8 <code>_wd = 0.0f<\/code>\uff0c\u56e0\u6b64\u65e0\u8bba SGD\u3001Momentum\u3001Nesterov \u8fd8\u662f Adam\/AdamW\uff0cBias-like \u53c2\u6570\u90fd\u81ea\u7136\u5f97\u5230\u6b63\u786e\u7684\u884c\u4e3a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Adam \u4e0e AdamW \u7684 Bias \u8def\u5f84\u5171\u7528\u540c\u4e00\u4e2a <code>RANGE_UPDATE_BIAS_ADAM<\/code>\uff0c\u56e0\u4e3a\u5f53 <code>wd = nullptr<\/code> \u65f6\uff0c<code>update_adam_kernel<\/code> \u548c <code>update_adamw_kernel<\/code> \u6570\u5b66\u7b49\u4ef7\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=\"\">\/\/ Bias \u8def\u5f84 wd \u56fa\u5b9a\u4f20 nullptr\uff0c_wd = 0.0f\nfloat _wd = wd ? *wd : 0.0f;\n\/\/ Adam \u7684 g_i = g * inv_scaling + _wd * w = g * inv_scaling\n\/\/ AdamW \u7684 g_i = g * inv_scaling\n\/\/ \u5f53 _wd == 0 \u65f6\u4e24\u8005\u5b8c\u5168\u76f8\u540c<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u201c\u7528\u5206\u7ec4\u4ee3\u66ff\u9010\u5c42\u5224\u65ad\u201d\u7684\u601d\u8def\uff0c\u5927\u5e45\u7b80\u5316\u4e86\u7f16\u8bd1\u671f\u56fe\u6784\u5efa\uff0c\u4e5f\u51cf\u5c11\u4e86\u8fd0\u884c\u65f6\u7684 kernel \u6570\u91cf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001Adam \u7684\u504f\u5dee\u4fee\u6b63\uff1a\u4e24\u4e2a\u6807\u91cf kernel<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Adam\/AdamW \u8fd8\u9700\u8981\u5728\u6bcf\u6b65\u66f4\u65b0\u524d\u8ba1\u7b97\u504f\u5dee\u4fee\u6b63\u7cfb\u6570 <code>bc1<\/code> \u548c <code>bc2<\/code>\u3002Tech-Renaissance \u6ca1\u6709\u628a\u8fd9\u4ef6\u4e8b\u653e\u5230\u4e3b\u66f4\u65b0 kernel \u91cc\uff0c\u800c\u662f\u5355\u72ec\u653e\u5728 <code>GraphId::OPTIMIZER<\/code> \u56fe\u7684\u6700\u524d\u9762\uff0c\u7528\u4e24\u4e2a\u6781\u5c0f\u7684 kernel \u5b8c\u6210\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=\"\">\/\/ SCALAR_INCREMENT: step += 1\n__global__ void scalar_increment_kernel(int32_t* value) {\n    if (threadIdx.x == 0 &amp;&amp; blockIdx.x == 0) {\n        *value += 1;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u9700\u8981\u8bf4\u660e\u7684\u662f\uff0c\u8fd9\u4e2a <code>step += 1<\/code> \u7684 kernel \u4e0e\u540e\u7eed\u6240\u6709\u53c2\u6570\u66f4\u65b0 kernel \u540c\u5904 <code>GraphId::OPTIMIZER<\/code> \u56fe\uff1b\u5f53 <code>has_nan<\/code> \u6807\u5fd7\u4e3a\u771f\u65f6\uff0c\u6574\u6bb5 <code>OPTIMIZER<\/code> \u56fe\u4e0d\u4f1a\u88ab launch\uff08\u89c1 <code>src\/backend\/graph_executor.cpp<\/code>\uff09\uff0c\u56e0\u6b64 step \u4e0d\u4f1a\u5728 NaN\/Inf \u6b65\u9519\u8bef\u63a8\u8fdb\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=\"\">\/\/ ADAM_BIAS_CORRECTION:\n\/\/   bc1 = 1 \/ (1 - beta1^step)\n\/\/   bc2 = 1 \/ (1 - beta2^step)\n__global__ void adam_bias_correction_kernel(\n    const int32_t* __restrict__ step,\n    const float*   __restrict__ beta1,\n    const float*   __restrict__ beta2,\n    float*         __restrict__ bias_corr1,\n    float*         __restrict__ bias_corr2)\n{\n    if (threadIdx.x != 0 || blockIdx.x != 0) return;\n    int32_t t = *step;\n    float b1 = *beta1;\n    float b2 = *beta2;\n    *bias_corr1 = 1.0f \/ (1.0f - powf(b1, static_cast&lt;float>(t)));\n    *bias_corr2 = 1.0f \/ (1.0f - powf(b2, static_cast&lt;float>(t)));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e24\u4e2a kernel \u90fd\u8fd0\u884c\u5728 <code>UPDATE<\/code> \u6d41\u4e0a\uff0c\u4e0e\u540e\u7eed\u7684\u5927\u89c4\u6a21\u6743\u91cd\u66f4\u65b0 kernel \u5904\u4e8e\u540c\u4e00\u6761\u6d41\uff0c\u5929\u7136\u6309\u987a\u5e8f\u6267\u884c\u3002\u5b83\u4eec\u7684\u8ba1\u7b97\u91cf\u6781\u5c0f\uff0c\u4f46\u628a\u5b83\u4eec\u72ec\u7acb\u51fa\u6765\u7684\u597d\u5904\u662f\uff1a<strong>\u504f\u5dee\u4fee\u6b63\u53ea\u9700\u8981\u505a\u4e00\u6b21<\/strong>\uff0c\u968f\u540e\u88ab\u6240\u6709 weight \u548c bias \u66f4\u65b0 kernel \u5171\u4eab\u3002\u6ca1\u6709\u91cd\u590d\u8ba1\u7b97\uff0c\u4e5f\u6ca1\u6709\u628a\u6807\u91cf\u903b\u8f91\u8026\u5408\u8fdb\u5927\u6570\u636e kernel\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\uff1a\u5728\u8bad\u7ec3\u6b65\u6570\u6781\u5927\u7684\u751f\u4ea7\u573a\u666f\u4e2d\uff0c\u4e3a\u4e86\u907f\u514d\u5355\u7cbe\u5ea6 <code>powf<\/code> \u7684\u820d\u5165\/\u4e0b\u6ea2\u8bef\u5dee\uff0c\u4e5f\u53ef\u5c06\u8fd9\u4e00\u6b65\u6539\u5728\u6838\u5185\u7528 <code>double<\/code> \u8ba1\u7b97\uff0c\u6216\u6539\u4e3a <code>beta^t<\/code> \u7684\u8fed\u4ee3\u7d2f\u4e58\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u80fd\u4f1a\u95ee\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u5728\u4e3b\u66f4\u65b0 kernel \u91cc\u76f4\u63a5\u7b97 <code>powf(beta, t)<\/code>\uff1f\u6280\u672f\u4e0a\u5f53\u7136\u53ef\u4ee5\uff0c\u4f46\u90a3\u6837\u6bcf\u4e2a\u7ebf\u7a0b\u90fd\u8981\u91cd\u590d\u4e00\u6b21\u6307\u6570\u8fd0\u7b97\uff0c\u800c\u6307\u6570\u8fd0\u7b97\u5728 GPU \u4e0a\u5e76\u4e0d\u4fbf\u5b9c\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u628a\u5b83\u72ec\u7acb\u6210\u6807\u91cf kernel \u540e\uff0c<code>bc1<\/code> \u548c <code>bc2<\/code> \u53ef\u4ee5\u88ab\u591a\u4e2a RangeOp \u8282\u70b9\u5171\u4eab\u2014\u2014weight \u66f4\u65b0\u7528\u4e00\u6b21\uff0cbias \u66f4\u65b0\u4e5f\u7528\u4e00\u6b21\uff0c\u786e\u4fdd\u6240\u6709\u53c2\u6570\u7684\u504f\u5dee\u4fee\u6b63\u5728\u540c\u4e00\u6b65\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u5916\uff0c\u628a <code>step<\/code> \u7684\u9012\u589e\u4e5f\u505a\u6210\u663e\u5f0f\u8282\u70b9\uff0c\u8ba9\u8bad\u7ec3\u6b65\u6570\u8fd9\u4e2a\u5168\u5c40\u72b6\u6001\u5b8c\u5168\u5728 GPU \u4fa7\u7ef4\u62a4\u3002CPU \u53ea\u8d1f\u8d23\u5728\u6bcf\u4e2a epoch \u6216\u6bcf\u4e2a batch \u5f00\u59cb\u65f6\u628a\u5b66\u4e60\u7387\u7b49\u5c0f\u6807\u91cf H2D \u5199\u5165\u663e\u5b58\uff0c\u4e0d\u53c2\u4e0e\u6b65\u6570\u8ba1\u6570\u3002\u8fd9\u79cd\u201cCPU \u53ea\u505a\u51b3\u7b56\u3001GPU \u505a\u6267\u884c\u201d\u7684\u5206\u5de5\uff0c\u662f\u9759\u6001\u56fe\u6846\u67b6\u5b9e\u73b0\u6781\u81f4\u4f4e\u5ef6\u8fdf\u7684\u5173\u952e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001LARS\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u8d70 RangeOp\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">LARS \u662f\u4e09\u79cd\u4f18\u5316\u5668\u91cc\u6700\u7279\u6b8a\u7684\u4e00\u4e2a\u3002SGD\u3001AdamW \u7684\u66f4\u65b0\u53ea\u4f9d\u8d56\u5143\u7d20\u7ea7\u522b\u7684\u8bfb\u5199\uff0c\u53ef\u4ee5\u6279\u91cf\u8986\u76d6\u6574\u4e2a Region\uff1b\u4f46 LARS \u9700\u8981\u4e3a<strong>\u6bcf\u4e00\u5c42<\/strong>\u8ba1\u7b97\u4e00\u4e2a trust ratio\uff0c\u800c\u8fd9\u4e2a trust ratio \u662f <code>||w_l||<\/code> \u548c <code>||g_l||<\/code> \u4e24\u4e2a L2 \u8303\u6570\u7684\u6bd4\u503c\uff0c\u672c\u8d28\u4e0a\u662f<strong>\u5bf9\u5355\u4e2a\u5f20\u91cf\u7684\u5f52\u7ea6\u64cd\u4f5c<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">RangeOp \u7684\u8bbe\u8ba1\u76ee\u6807\u662f\u201c\u65e0\u89c6\u5f20\u91cf\u8fb9\u754c\uff0c\u4e00\u6b21\u6027\u5904\u7406\u540c\u7c7b\u6570\u636e\u201d\u3002\u4e00\u65e6\u9700\u8981 per-tensor \u7684\u5f52\u7ea6\u7ed3\u679c\uff0cRangeOp \u5c31\u4e0d\u518d\u9002\u7528\uff0c\u56e0\u4e3a\u6574\u4e2a Region \u88ab\u63c9\u5728\u4e00\u8d77\u540e\uff0c\u4f60\u65e0\u6cd5\u533a\u5206\u54ea\u4e00\u6bb5\u5c5e\u4e8e\u54ea\u4e00\u5c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64 Tech-Renaissance \u4e3a LARS \u5355\u72ec\u8bbe\u8ba1\u4e86\u4e00\u5957 ComputeOp\uff0c\u7f16\u8bd1\u5668\u6309\u5c42\u6784\u5efa\u8282\u70b9\u3002\u5728 <code>src\/graph\/compiler.cpp<\/code> \u4e2d\uff0cLARS \u7684\u56fe\u88ab\u5206\u914d\u5230\u4e09\u4e2a\u72ec\u7acb\u7684 <code>GraphId<\/code>\uff1a<code>LARS_FC_OPT<\/code>\u3001<code>LARS_FIRST_CONV_OPT<\/code>\u3001<code>LARS_DEEP_CONV_OPT<\/code>\uff0c\u5206\u522b\u5bf9\u5e94 FC \u6743\u91cd\u3001\u9996\u5c42\u5377\u79ef\u3001\u6df1\u5c42\u5377\u79ef\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=\"\">auto build_lars_pair = [&amp;](Region w_region, GraphId gid,\n                            ComputeOp trust_op, ComputeOp update_op,\n                            int32_t temp_partial_id) {\n    Region g_region = paired_grad_region(w_region);\n    Region m_region = paired_momentum_region(w_region);\n    Region n_region = paired_norm_region(w_region);\n\n    const auto&amp; w_ids = memory_plan.get_ids_by_region(w_region);\n    const auto&amp; g_ids = memory_plan.get_ids_by_region(g_region);\n    const auto&amp; m_ids = memory_plan.get_ids_by_region(m_region);\n    const auto&amp; n_ids = memory_plan.get_ids_by_region(n_region);\n\n    for (size_t i = 0; i &lt; w_ids.size(); ++i) {\n        \/\/ Step 1: \u8ba1\u7b97 trust ratio\n        GraphNode trust_node;\n        trust_node.kind = GraphNode::Kind::COMPUTE;\n        trust_node.compute_op = trust_op;\n        trust_node.input_ids  = {w_ids[i], g_ids[i], scalar_ids.tc,\n                                  scalar_ids.wd, scalar_ids.eps,\n                                  scalar_ids.scaling, scalar_ids.has_nan,\n                                  temp_partial_id};\n        trust_node.output_ids = {n_ids[i]};\n        train_cg.append(gid, trust_node);\n\n        \/\/ Step 2: \u7528 trust ratio \u66f4\u65b0\u6743\u91cd\n        GraphNode update_node;\n        update_node.kind = GraphNode::Kind::COMPUTE;\n        update_node.compute_op = update_op;\n        update_node.input_ids  = {w_ids[i], g_ids[i], m_ids[i], n_ids[i],\n                                   scalar_ids.lr, scalar_ids.beta,\n                                   scalar_ids.wd, scalar_ids.scaling,\n                                   scalar_ids.has_nan};\n        update_node.output_ids = {w_ids[i], m_ids[i]};\n        train_cg.append(gid, update_node);\n    }\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684\u5173\u952e\u662f<strong>\u4e09\u5f20 LARS \u56fe\u3001\u4e09\u6761\u8ba1\u7b97\u6d41<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>LARS_FC_OPT<\/code> \u2192 <code>COMP_1<\/code><\/li>\n\n\n\n<li><code>LARS_FIRST_CONV_OPT<\/code> \u2192 <code>COMP_2<\/code><\/li>\n\n\n\n<li><code>LARS_DEEP_CONV_OPT<\/code> \u2192 <code>COMP_3<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">FC\u3001\u9996\u5c42\u5377\u79ef\u3001\u6df1\u5c42\u5377\u79ef\u7684 trust ratio \u5f52\u7ea6\u548c\u6743\u91cd\u66f4\u65b0\u5206\u522b\u843d\u5728\u4e09\u6761\u72ec\u7acb\u7684\u8ba1\u7b97\u6d41\u4e0a\uff0c\u5929\u7136\u5e76\u884c\u3002<code>src\/backend\/op_stream_policy.cpp<\/code> \u4e2d\u7684\u6d41\u5206\u914d\u7b56\u7565\u786e\u8ba4\u4e86\u8fd9\u4e00\u70b9\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">case ComputeOp::LARS_COMPUTE_TRUST_RATIO_FC:\ncase ComputeOp::LARS_UPDATE_FC:\ncase ComputeOp::LARS_NESTEROV_UPDATE_FC:\n    return StreamKind::COMP_1;\n\ncase ComputeOp::LARS_COMPUTE_TRUST_RATIO_FIRST:\ncase ComputeOp::LARS_UPDATE_FIRST:\ncase ComputeOp::LARS_NESTEROV_UPDATE_FIRST:\n    return StreamKind::COMP_2;\n\ncase ComputeOp::LARS_COMPUTE_TRUST_RATIO_DEEP:\ncase ComputeOp::LARS_UPDATE_DEEP:\ncase ComputeOp::LARS_NESTEROV_UPDATE_DEEP:\n    return StreamKind::COMP_3;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>lars_op.cu<\/code> \u4e2d\u7684 trust ratio kernel \u91c7\u7528\u4e24\u9636\u6bb5 reduce\uff1aphase1 \u7528\u591a\u4e2a block \u5e76\u884c\u7d2f\u52a0\u5c40\u90e8 <code>sum_w2<\/code>\/<code>sum_g2<\/code>\uff0cphase2 \u7528\u5355\u4e2a\u7ebf\u7a0b\u5b8c\u6210\u6700\u7ec8\u5f52\u7ea6\u5e76\u8ba1\u7b97 <code>eta<\/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=\"\">\/\/ Phase 1: \u6bcf\u4e2a block \u5f52\u7ea6\u4e00\u6bb5\u6570\u636e\uff0c\u8f93\u51fa\u5c40\u90e8 sum_w2 \/ sum_g2\n__global__ void lars_trust_ratio_phase1_kernel(...)\n{\n    float local_w2 = 0.0f, local_g2 = 0.0f;\n    for (size_t i = blockIdx.x * blockDim.x + threadIdx.x;\n         i &lt; n; i += gridDim.x * blockDim.x) {\n        float wv = w[i];\n        float gv = g[i];\n        local_w2 += wv * wv;\n        local_g2 += gv * gv;\n    }\n    \/\/ warp shuffle + shared memory \u5b8c\u6210 block \u5185\u5f52\u7ea6\n    \/\/ \u6700\u7ec8\u5199\u5165 out_w2[blockIdx.x], out_g2[blockIdx.x]\n}\n\n\/\/ Phase 2: \u5355\u7ebf\u7a0b\u6c47\u603b\u6240\u6709 block \u7ed3\u679c\uff0c\u8ba1\u7b97 eta\n__global__ void lars_trust_ratio_phase2_kernel(...)\n{\n    float sum_w2 = 0.0f, sum_g2 = 0.0f;\n    for (int i = 0; i &lt; num_blocks; ++i) {\n        sum_w2 += in_w2[i];\n        sum_g2 += in_g2[i];\n    }\n    float w_norm = sqrtf(sum_w2);\n    float g_norm = sqrtf(sum_g2) * inv_scaling;\n    float eta = 1.0f;\n    if (w_norm >= 1e-12f &amp;&amp; g_norm >= 1e-12f) {\n        eta = tc * w_norm \/ (g_norm + wd * w_norm + eps);\n        if (eta > 100.0f) eta = 100.0f;  \/\/ \u6570\u503c\u7a33\u5b9a\u6027\u94b3\u5236\n    }\n    *out_eta = eta;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e34\u65f6\u7f13\u51b2\u533a <code>T_TEMP_FP32<\/code> \u7684\u5206\u914d\u987a\u5e8f\u662f FC \u2192 FirstConv \u2192 DeepConv\uff0c\u7d22\u5f15\u4ece 0 \u5f00\u59cb\uff0c\u6bcf\u4e2a LARS \u5c42\u5bf9\u5360\u7528 <code>2 \u00d7 kLarsMaxPartial<\/code> \u4e2a float\u3002<code>kLarsMaxPartial<\/code> \u5b9a\u4e49\u4e3a 65535\uff0c\u56e0\u6b64\u6bcf\u4e2a\u5c42\u5bf9\u7ea6 512 KB\uff0c\u8fdc\u5c0f\u4e8e\u4e00\u6b21\u4e2d\u95f4\u7279\u5f81\u56fe\u7684\u5185\u5b58\u5360\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e24\u9636\u6bb5\u5f52\u7ea6\u7684\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e<strong>\u5e76\u884c\u5ea6<\/strong>\u3002\u5355 block \u5f52\u7ea6\u6700\u591a\u53ea\u80fd\u5229\u7528 1024 \u4e2a\u7ebf\u7a0b\uff0c\u4f46 Phase 1 \u53ef\u4ee5\u542f\u52a8\u591a\u8fbe 65535 \u4e2a block\uff0c\u6bcf\u4e2a block 256 \u4e2a\u7ebf\u7a0b\u2014\u2014\u5bf9\u4e8e\u5927\u578b\u5168\u8fde\u63a5\u5c42\uff08\u6570\u767e\u4e07\u53c2\u6570\uff09\u7684\u8303\u6570\u8ba1\u7b97\uff0c\u8fd9\u80fd\u5145\u5206\u5229\u7528 GPU \u7684\u8ba1\u7b97\u5355\u5143\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LARS \u7684\u6743\u91cd\u66f4\u65b0 kernel \u4e5f\u6709\u5176\u72ec\u7279\u4e4b\u5904\uff1a\u5b83\u540c\u65f6\u63a5\u53d7 <code>eta<\/code>\uff08per-layer trust ratio\uff09\u548c <code>lr<\/code>\uff08\u5168\u5c40\u5b66\u4e60\u7387\uff09\uff0c\u5c06\u4e24\u8005\u76f8\u4e58\u4f5c\u4e3a\u5b9e\u9645\u5b66\u4e60\u7387\uff0c\u5e76\u5728\u4e00\u6b21\u904d\u5386\u4e2d\u5b8c\u6210\u52a8\u91cf\u66f4\u65b0\u548c\u6743\u91cd\u8870\u51cf\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=\"\">\/\/ LARS_UPDATE:\n\/\/   gp      = g + wd * w\n\/\/   m_new   = beta * m + lr * eta * gp\n\/\/   w_new   = w - m_new\n\/\/   m       = m_new\nfor (size_t i = ...; i &lt; n; i += ...) {\n    float wv = w[i];\n    float gv = g[i] * inv_scaling;\n    float gp = gv + _wd * wv;\n    float m_new = _beta * m[i] + _lr * _eta * gp;\n    w[i] = wv - m_new;\n    m[i] = m_new;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f LARS \u7684 weight decay \u662f\u5d4c\u5165\u5728\u68af\u5ea6\u66f4\u65b0\u4e2d\u7684\uff08<code>gp = gv + _wd * wv<\/code>\uff09\uff0c\u8fd9\u4e0e SGD\/Adam \u7684 weight decay \u5d4c\u5165\u65b9\u5f0f\u4e0d\u540c\uff0c\u662f LARS \u539f\u59cb\u8bba\u6587\u4e2d\u5b9a\u4e49\u7684\u516c\u5f0f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">LARS \u7684 Bias-like \u53c2\u6570\u5219\u4e0d\u53c2\u4e0e trust ratio\uff0c\u76f4\u63a5\u9000\u5316\u4e3a\u5e26 Momentum \u6216 Nesterov \u7684\u6807\u51c6 RangeOp \u66f4\u65b0\uff0c\u4e0e SGD_MOMENTUM \/ SGD_NESTEROV \u5171\u7528\u540c\u4e00\u6761\u8def\u5f84\u3002\u8fd9\u518d\u6b21\u4f53\u73b0\u4e86\u201c\u6309\u884c\u4e3a\u5206\u7ec4\u201d\u7684\u8bbe\u8ba1\u54f2\u5b66\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001NaN \u4fdd\u62a4\u3001\u68af\u5ea6\u7f29\u653e\u4e0e CUDA Graph \u517c\u5bb9\u6027<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f18\u5316\u5668\u66f4\u65b0\u4e0d\u662f\u5b64\u7acb\u6267\u884c\u7684\u2014\u2014\u5b83\u5904\u5728\u6574\u4e2a\u8bad\u7ec3\u5faa\u73af\u7684\u6700\u540e\u9636\u6bb5\uff0c\u524d\u9762\u662f\u68af\u5ea6\u8ba1\u7b97\u3001\u68af\u5ea6\u7f29\u653e\uff08AMP \u7684\u53cd\u5411\u8fc7\u7a0b\uff09\u3001NaN \u68c0\u6d4b\u3002\u8fd9\u4e9b\u73af\u8282\u7684\u4fe1\u606f\u901a\u8fc7\u4e00\u4e2a <code>has_nan<\/code> \u6807\u5fd7\u4f20\u9012\u7ed9\u4f18\u5316\u5668\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>optimizer_op.cpp<\/code> \u7684 launcher \u4e2d\uff0c\u6bcf\u4e2a\u4f18\u5316\u5668\u5728\u542f\u52a8 kernel \u524d\u90fd\u4f1a\u89e3\u6790 <code>has_nan<\/code> \u6807\u5fd7\u7684\u6307\u9488\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=\"\">const int32_t* has_nan = static_cast&lt;const int32_t*>(\n    ArenaKeeper::instance().ptr_at(ctx.rank_for_context(),\n                                   mp.get_dtensor(node.input_ids.back()).offset()));<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">kernel \u5185\u90e8\u7684\u7b2c\u4e00\u6761\u6307\u4ee4\u5c31\u662f\u68c0\u67e5\u8fd9\u4e2a\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=\"\">if (*has_nan != 0) return;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8bbe\u8ba1\u7684\u7cbe\u5999\u4e4b\u5904\u5728\u4e8e\uff1a<strong>\u5373\u4f7f kernel \u4ec0\u4e48\u90fd\u4e0d\u505a\uff0cCUDA Graph \u7684\u91cd\u653e\u4f9d\u7136\u6709\u6548<\/strong>\u3002\u56e0\u4e3a <code>has_nan<\/code> \u662f\u4e00\u4e2a\u53ef\u66f4\u65b0\u7684 GPU \u6807\u91cf\uff0cGraph \u91cd\u653e\u524d\u53ef\u4ee5\u901a\u8fc7 host \u7aef\u66f4\u65b0\u5176\u503c\uff0ckernel \u8bfb\u53d6\u540e\u6839\u636e\u503c\u51b3\u5b9a\u662f\u5426\u6267\u884c\u5b9e\u9645\u8ba1\u7b97\u3002\u8fd9\u4fdd\u8bc1\u4e86\u786e\u5b9a\u6027\u8bad\u7ec3\u573a\u666f\u4e0b\u7684\u56fe\u7a33\u5b9a\u6027\u2014\u2014\u4e0d\u4f1a\u56e0\u4e3a\u67d0\u6b65\u51fa\u73b0 NaN \u800c\u5bfc\u81f4\u56fe\u7ed3\u6784\u53d8\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u68af\u5ea6\u7f29\u653e\uff08gradient scaling\uff09\u7684\u5904\u7406\u540c\u6837\u88ab\u878d\u5408\u8fdb\u4e86\u4f18\u5316\u5668 kernel \u5185\u90e8\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=\"\">float _inv_scaling = (scaling &amp;&amp; *scaling != 0.0f) ? (1.0f \/ *scaling) : 1.0f;\n\/\/ ...\nfloat g_i = g[i] * _inv_scaling;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u907f\u514d\u4e86\u989d\u5916\u7684\u4e00\u6b21\u5168\u5c40\u663e\u5b58\u904d\u5386\uff0c\u5c06 unscale \u7684\u5f00\u9500\u5b8c\u5168\u9690\u85cf\u5728\u4e86\u4f18\u5316\u5668\u7684\u8ba1\u7b97\u4e2d\u3002\u4e00\u4e2a\u503c\u5f97\u4e0e PyTorch \u5bf9\u6bd4\u7684\u7ec6\u8282\u662f\uff1aPyTorch \u7684 AMP \u901a\u5e38\u9700\u8981\u663e\u5f0f\u8c03\u7528 <code>scaler.unscale_(optimizer)<\/code> \u6765\u8fd8\u539f\u68af\u5ea6\uff0c\u7136\u540e\u518d\u8c03\u7528 <code>optimizer.step()<\/code>\u3002\u8fd9\u4e24\u4e2a\u6b65\u9aa4\u662f\u5206\u79bb\u7684\u2014\u2014unscale \u904d\u5386\u4e00\u6b21\u6240\u6709\u68af\u5ea6\uff0cstep \u518d\u904d\u5386\u4e00\u6b21\u6240\u6709\u53c2\u6570\u3002\u5728\u6211\u4eec\u7684\u878d\u5408\u8bbe\u8ba1\u4e2d\uff0cunscale \u548c update \u5728\u540c\u4e00\u4e2a kernel \u7684\u540c\u4e00\u4e2a\u5faa\u73af\u4e2d\u5b8c\u6210\uff0c\u51cf\u5c11\u4e86\u663e\u5b58\u8bbf\u95ee\u6b21\u6570\u3002\u5bf9\u4e8e memory-bound \u7684\u4f18\u5316\u5668\u66f4\u65b0\u6765\u8bf4\uff0c\u8fd9\u662f\u4e00\u9879\u5b9e\u8d28\u6027\u7684\u6027\u80fd\u6539\u8fdb\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\uff1a\u4e00\u4e2a\u5c0f\u6807\u91cf H2D \u5c31\u591f\u4e86<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u4e2d\uff0c\u5b66\u4e60\u7387\u4e0d\u662f\u4f18\u5316\u5668\u72b6\u6001\u7684\u4e00\u90e8\u5206\uff0c\u800c\u662f\u7531 Scheduler \u8ba1\u7b97\u3001\u901a\u8fc7\u4e00\u6b21\u6781\u5c0f\u7684 H2D \u62f7\u8d1d\u5199\u5165 <code>S_SCALAR_FP32<\/code> \u533a\u57df\u7684\u5168\u5c40\u6807\u91cf\u3002\u7531\u4e8e\u4f18\u5316\u5668 kernel \u4ee5\u8bbe\u5907\u6307\u9488\u8bfb\u53d6 <code>lr<\/code>\uff0c\u5b66\u4e60\u7387\u7684\u53d8\u5316\u4e0d\u4f1a\u5bfc\u81f4 CUDA Graph \u91cd\u65b0\u6355\u83b7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u8bbe\u8ba1\u5728\u8fd0\u884c\u65f6\u7684 <code>GraphExecutor::update_lr_scalar()<\/code> \u4e2d\u4f53\u73b0\uff1a\u5b83\u53ea\u628a\u4e00\u4e2a float \u503c copy \u5230\u663e\u5b58\uff0c\u8017\u65f6\u901a\u5e38\u5728\u5fae\u79d2\u7ea7\uff0c\u5e76\u4e14\u53ef\u4ee5\u548c\u540e\u7eed\u8ba1\u7b97\u6d41\u91cd\u53e0\u3002\u5982\u679c\u662f step-by-batch \u7684\u8c03\u5ea6\u7b56\u7565\uff0c\u6bcf\u4e2a batch \u90fd\u4f1a\u6709\u4e00\u6b21\u8fd9\u6837\u7684\u5c0f H2D\uff0c\u4f46\u56e0\u4e3a\u5b83\u53ea\u5199\u5355\u4e2a\u6807\u91cf\uff0c\u5f00\u9500\u51e0\u4e4e\u53ef\u4ee5\u5ffd\u7565\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5982\u679c\u4f18\u5316\u5668 step \u662f\u4ee5 kernel \u53c2\u6570\u5f62\u5f0f\u63a5\u6536\u5b66\u4e60\u7387\uff0c\u90a3\u4e48\u5b66\u4e60\u7387\u6bcf\u53d8\u4e00\u6b21\u5c31\u9700\u8981\u91cd\u6784\u56fe\uff0cstep-by-batch \u7684\u8c03\u5ea6\u6210\u672c\u5c31\u4f1a\u9ad8\u5f97\u591a\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u4e00\u4e0b\u672c\u6846\u67b6\u878d\u5408\u4f18\u5316\u5668\u7684\u6838\u5fc3\u8bbe\u8ba1\u601d\u8def\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4ece\u201c\u9010\u53c2\u6570\u5faa\u73af\u201d\u5230\u201c\u6574\u533a\u6279\u91cf\u66f4\u65b0\u201d<\/strong>\uff1a\u5229\u7528 MemoryPlan \u7684 Region \u8fde\u7eed\u5e03\u5c40\uff0c\u5c06\u6a21\u578b\u6240\u6709\u53c2\u6570\u5206\u4e3a Weight \u7ec4\u548c Bias \u7ec4\uff0c\u6bcf\u7ec4\u53ea\u9700\u4e00\u4e2a kernel \u5b8c\u6210\u5168\u90e8\u66f4\u65b0\u3002\u65e0\u8bba\u6a21\u578b\u6709\u591a\u5c11\u5c42\uff0c\u4f18\u5316\u5668\u66f4\u65b0\u6c38\u8fdc\u662f 1~2 \u4e2a kernel\u3002<\/li>\n\n\n\n<li><strong>Weight \u4e0e Bias \u7684\u5206\u7ec4\u8bbe\u8ba1<\/strong>\uff1a\u57fa\u4e8e weight decay \u7684\u8bed\u4e49\u5dee\u5f02\u2014\u2014Weight \u7ec4\u65bd\u52a0 weight decay\uff0cBias \u7ec4 weight decay \u6052\u4e3a\u96f6\u3002\u8fd9\u79cd\u5206\u7ec4\u4e0d\u662f\u666e\u901a\u7684\u6027\u80fd\u6280\u5de7\uff0c\u800c\u662f\u6df1\u5ea6\u5b66\u4e60\u754c\u957f\u671f\u9a8c\u8bc1\u7684\u6536\u655b\u5b9e\u8df5\uff08\u9632\u6b62\u6b63\u5219\u5316\u8fc7\u5ea6\u62b9\u53bb\u4eff\u5c04\u53d8\u6362\u7684\u5e73\u79fb\u8bed\u4e49\uff09\u3002<\/li>\n\n\n\n<li><strong>\u4e94\u79cd\u4f18\u5316\u5668\u3001\u7edf\u4e00\u6846\u67b6<\/strong>\uff1aSGD\u3001Momentum\u3001Nesterov\u3001Adam\u3001AdamW \u901a\u8fc7\u7edf\u4e00\u7684 RangeOp \u8def\u5f84\u5b9e\u73b0\uff0cLARS \u901a\u8fc7 ComputeOp \u8def\u5f84\u5b9e\u73b0\u3002\u4e24\u8005\u5171\u4eab\u76f8\u540c\u7684\u6807\u91cf\u53c2\u6570\u7ba1\u7406\u3001NaN \u4fdd\u62a4\u548c\u56fe\u517c\u5bb9\u673a\u5236\u3002<\/li>\n\n\n\n<li><strong>LARS \u7684\u4e09\u6d41\u5e76\u884c<\/strong>\uff1a\u5c06 FC \u5c42\u3001\u9996\u5c42\u5377\u79ef\u3001\u6df1\u5c42\u5377\u79ef\u7684 trust ratio \u8ba1\u7b97\u548c\u66f4\u65b0\u5206\u522b\u6620\u5c04\u5230\u4e09\u6761\u8ba1\u7b97\u6d41\u4e0a\u5e76\u884c\u6267\u884c\uff0c\u914d\u5408\u4e24\u9636\u6bb5\u5f52\u7ea6\uff0c\u5c06 LARS \u7684\u989d\u5916\u5f00\u9500\u964d\u5230\u6700\u4f4e\u3002<\/li>\n\n\n\n<li><strong>CUDA Graph \u5168\u517c\u5bb9<\/strong>\uff1a\u6240\u6709\u6807\u91cf\u53c2\u6570\u4ee5\u6307\u9488\u4f20\u5165\uff0c\u5b66\u4e60\u7387\u3001\u52a8\u91cf\u7cfb\u6570\u3001bias correction \u7b49\u53ef\u4ee5\u5728 Graph \u91cd\u653e\u524d\u52a8\u6001\u66f4\u65b0\uff0c\u65e0\u9700\u91cd\u65b0\u6355\u83b7\u56fe\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bb0\u4f4f\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u4f18\u5316\u5668\u662f<strong>\u6574\u533a\u6279\u91cf<\/strong>\u7684\uff1a\u6240\u6709\u540c\u884c\u4e3a\u7684\u53c2\u6570\u5728\u5185\u5b58\u4e2d\u8fde\u7eed\uff0c\u4e00\u6b21 kernel launch \u8986\u76d6\u5168\u90e8\u3002SGD\/Momentum\/AdamW \u7684 weight \u66f4\u65b0\u901a\u5e38\u53ea\u9700\u8981\u4e00\u4e2a <code>RANGE_UPDATE_WEIGHT_*<\/code> \u8282\u70b9\uff0cbias \u66f4\u65b0\u53ea\u9700\u8981\u4e00\u4e2a <code>RANGE_UPDATE_BIAS_*<\/code> \u8282\u70b9\uff0c\u518d\u52a0\u4e0a Adam \u7684\u4e24\u4e2a\u6807\u91cf kernel\u3002\u65e0\u8bba\u6a21\u578b\u662f 10 \u5c42\u8fd8\u662f 100 \u5c42\uff0c\u4f18\u5316\u5668 step \u7684 kernel \u6570\u91cf\u662f\u6052\u5b9a\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u8bbe\u8ba1\u4e5f\u5e26\u6765\u4e86\u66f4\u597d\u7684<strong>\u663e\u5b58\u8bbf\u95ee\u5c40\u90e8\u6027<\/strong>\u3002\u4e3b\u6743\u91cd\u3001\u68af\u5ea6\u3001\u52a8\u91cf\u7f13\u51b2\u533a\u6309\u76f8\u540c\u987a\u5e8f\u8fde\u7eed\u5b58\u653e\uff0ckernel \u5728 grid-stride loop \u4e2d\u987a\u5e8f\u8bfb\u5199\uff0c\u80fd\u591f\u66f4\u597d\u5730\u5229\u7528\u5408\u5e76\u8bbf\u95ee\u548c\u7f13\u5b58\u884c\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u9010\u5f20\u91cf\u66f4\u65b0\u867d\u7136\u6bcf\u4e2a\u5f20\u91cf\u5185\u90e8\u4e5f\u662f\u987a\u5e8f\u8bbf\u95ee\uff0c\u4f46\u5f20\u91cf\u4e4b\u95f4\u7684\u5730\u5740\u4e0d\u8fde\u7eed\uff0c\u5bb9\u6613\u5728 TLB \u548c cache \u4e0a\u4ea7\u751f\u66f4\u591a\u6296\u52a8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u7136\uff0c\u8fd9\u79cd\u6574\u533a\u6279\u91cf\u7684\u524d\u63d0\u662f\uff1a\u6240\u6709\u88ab\u5408\u5e76\u7684\u53c2\u6570\u5fc5\u987b\u5171\u4eab\u76f8\u540c\u7684\u8d85\u53c2\u6570\u548c\u884c\u4e3a\u3002Tech-Renaissance \u76ee\u524d\u628a BN bias\/weight\u3001FC bias \u5206\u4e3a\u4e00\u7ec4\uff0cFC weight\u3001Conv weight \u5206\u4e3a\u53e6\u4e00\u7ec4\uff0c\u6b63\u662f\u56e0\u4e3a\u5b83\u4eec\u7684 weight decay \u884c\u4e3a\u4e00\u81f4\u3002\u5982\u679c\u4f60\u9700\u8981\u7ed9\u4e0d\u540c\u5c42\u8bbe\u7f6e\u5b8c\u5168\u4e0d\u540c\u7684\u8d85\u53c2\u6570\uff0c\u8fd9\u79cd\u5f3a\u6279\u91cf\u8bbe\u8ba1\u5c31\u4e0d\u592a\u9002\u7528\u2014\u2014\u4e0d\u8fc7\u5bf9\u6807\u51c6 CNN \u8bad\u7ec3\u6765\u8bf4\uff0c\u8fd9\u79cd\u5206\u7ec4\u5df2\u7ecf\u8986\u76d6\u4e86\u7edd\u5927\u591a\u6570\u573a\u666f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f18\u5316\u5668\u878d\u5408\u662f\u663e\u5b58\u5206\u533a\u8bbe\u8ba1\u5e26\u6765\u7684\u76f4\u63a5\u7ea2\u5229\u2014\u2014\u5206\u533a\u8ba9\u201c\u6279\u91cf\u201d\u6210\u4e3a\u53ef\u80fd\uff0c\u878d\u5408\u8ba9\u201c\u6279\u91cf\u201d\u53d1\u6325\u51fa\u6027\u80fd\u4ef7\u503c\u3002\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u635f\u5931\u51fd\u6570\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u4e0e\u53c2\u6570\u521d\u59cb\u5316\uff0c\u770b\u770b\u6846\u67b6\u7684\u8bad\u7ec3\u7b97\u6cd5\u914d\u7f6e\u5c42\u662f\u5982\u4f55\u8bbe\u8ba1\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341\u4e00 \u8bad\u7ec3\u795e\u7ecf\u7f51\u7edc\uff0c\u672c\u8d28\u4e0a\u5c31\u662f\u65e0 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":651,"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-547","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\/547","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=547"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/547\/revisions"}],"predecessor-version":[{"id":708,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/547\/revisions\/708"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/651"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=547"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=547"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=547"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}