{"id":6676,"date":"2025-12-26T13:31:27","date_gmt":"2025-12-26T05:31:27","guid":{"rendered":"https:\/\/nullthought.net\/?p=6676"},"modified":"2025-12-26T13:31:28","modified_gmt":"2025-12-26T05:31:28","slug":"%e9%9d%99%e9%bb%98%e6%95%b0%e6%8d%ae%e6%8d%9f%e5%9d%8f%ef%bc%88silent-data-corruption-sdc%ef%bc%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6676","title":{"rendered":"\u9759\u9ed8\u6570\u636e\u635f\u574f\uff08Silent Data Corruption, SDC\uff09"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2502.12340\" target=\"_blank\" rel=\"noreferrer noopener\">Understanding Silent Data Corruption in LLM Training<\/a><\/strong>\u7814\u7a76\u7684\u662f\u5927\u6a21\u578b\u8bad\u7ec3\u4e2d\u4e00\u79cd\u201c\u5f88\u96be\u88ab\u5bdf\u89c9\u3001\u4f46\u4f1a\u6084\u6084\u6539\u53d8\u7ed3\u679c\u201d\u7684\u786c\u4ef6\u6545\u969c\uff1a<strong>\u9759\u9ed8\u6570\u636e\u635f\u574f<\/strong>\uff08Silent Data Corruption, SDC\uff09\u3002\u5b83\u7684\u5178\u578b\u7279\u5f81\u662f\uff1a\u786c\u4ef6\u5e76\u4e0d\u4f1a\u62a5\u9519\u6216\u5d29\u6e83\uff0c\u4f46\u4f1a\u5728\u67d0\u4e9b\u65f6\u523b\u8f93\u51fa<strong>\u9519\u8bef\u7684\u8ba1\u7b97\u7ed3\u679c<\/strong>\uff0c\u4ece\u800c\u6c61\u67d3\u8bad\u7ec3\u8fc7\u7a0b\u3002\u4f5c\u8005\u5f3a\u8c03\uff0c\u968f\u7740\u8bad\u7ec3\u89c4\u6a21\u589e\u5927\uff0c\u8fd9\u7c7b\u95ee\u9898\u8d8a\u6765\u8d8a\u5e38\u89c1\uff0c\u56e0\u6b64\u9700\u8981\u7528\u66f4\u8d34\u8fd1\u771f\u5b9e\u751f\u4ea7\u73af\u5883\u7684\u65b9\u6cd5\u6765\u7406\u89e3\u5b83\u5bf9\u8bad\u7ec3\u7684\u5b9e\u9645\u5f71\u54cd\u3002<\/p>\n\n\n\n<p><strong>\u9759\u9ed8\u6570\u636e\u635f\u574fSDC<\/strong>\u5e38\u89c1\u6210\u56e0\u53ef\u4ee5\u6309\u201c\u6839\u6e90\u201d\u5206\u6210\u51e0\u5927\u7c7b\uff08\u5f88\u591a\u771f\u5b9e\u4e8b\u6545\u5f80\u5f80\u662f\u591a\u56e0\u7d20\u53e0\u52a0\uff09\uff1a<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">1) \u786c\u4ef6\u6545\u969c\uff08\u6700\u5e38\u89c1\uff09<\/h5>\n\n\n\n<p>\u8bba\u6587\u628a SDC \u7684\u4e3b\u8981\u6765\u6e90\u4e4b\u4e00\u76f4\u63a5\u5f52\u56e0\u4e8e<strong>\u786c\u4ef6\u6545\u969c<\/strong>\u3002\u5177\u4f53\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6f5c\u4f0f\/\u8fb9\u7f18\u786c\u4ef6\u7f3a\u9677\uff08latent defects \/ marginal faults\uff09<\/strong>\uff1a\u5668\u4ef6\u51fa\u5382\u6216\u88c5\u673a\u65f6\u770b\u4f3c\u6b63\u5e38\uff0c\u4f46\u5728\u67d0\u4e9b\u6761\u4ef6\u4e0b\u624d\u5f00\u59cb\u51fa\u9519\uff1b\u5728\u5927\u89c4\u6a21\u8bad\u7ec3\u8fd9\u79cd\u9ad8\u5f3a\u5ea6\u3001\u957f\u65f6\u95f4\u8d1f\u8f7d\u4e0b\u66f4\u5bb9\u6613\u66b4\u9732\u3002\n<ul class=\"wp-block-list\">\n<li>\u5178\u578b\u89e6\u53d1\u6761\u4ef6\uff1a\u9ad8\u6e29\u3001\u7279\u5b9a\u5229\u7528\u7387\/\u8d1f\u8f7d\u5f62\u6001\u3001\u957f\u65f6\u95f4\u201c\u70e4\u673a\u201d\u540e\u7684\u8001\u5316\u7b49\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u6e29\u5ea6\/\u7535\u538b\/\u65f6\u5e8f\u88d5\u91cf\u4e0d\u8db3\u5bfc\u81f4\u7684\u8ba1\u7b97\u9519\u8bef<\/strong>\uff1a\u4f8b\u5982\u9ad8\u6e29\u4f7f\u5f97\u7535\u8def\u53d8\u6162\u3001<strong>\u4f9b\u7535\u6ce2\u52a8<\/strong>\u5bfc\u81f4\u65f6\u5e8f\u8fdd\u4f8b\uff08timing violation\uff09\uff0c\u6700\u7ec8\u51fa\u73b0\u201c\u7b97\u672f\u5355\u5143\/\u77e9\u9635\u4e58\u201d\u5076\u53d1\u7b97\u9519\uff0c\u4f46\u4e0d\u4e00\u5b9a\u89e6\u53d1\u786c\u4ef6\u7684\u663e\u5f0f\u9519\u8bef\u4e0a\u62a5\u3002<\/li>\n\n\n\n<li><strong>\u5b58\u50a8\u5668\u76f8\u5173\uff08HBM\/DRAM\/\u7f13\u5b58\/\u5bc4\u5b58\u5668\uff09<\/strong>\uff1a\u4f4d\u7ffb\u8f6c\u3001\u884c\u5e72\u6270\u3001\u63a7\u5236\u5668\u8fb9\u7f18\u9519\u8bef\uff1b\u5982\u679c ECC \u672a\u8986\u76d6\u3001\u88ab\u5173\u95ed\u3001\u6216\u9519\u8bef\u843d\u5728 ECC \u96be\u4ee5\u7ea0\u6b63\/\u672a\u4e0a\u62a5\u7684\u8def\u5f84\u4e0a\uff0c\u5c31\u53ef\u80fd\u201c\u9759\u9ed8\u201d\u901a\u8fc7\u3002<\/li>\n\n\n\n<li><strong>\u4e92\u8fde\/\u7247\u4e0a\u7f51\u7edc\/PCIe\/NVLink \u7b49\u94fe\u8def\u8fb9\u7f18\u95ee\u9898<\/strong>\uff1a\u867d\u5e38\u89c1\u6709 CRC\/ECC\uff0c\u4f46\u5728\u67d0\u4e9b\u5c42\u7ea7\/\u8def\u5f84\u4e0a\u5982\u679c\u6821\u9a8c\u7f3a\u5931\u6216\u5b9e\u73b0\u7f3a\u9677\uff0c\u4e5f\u53ef\u80fd\u51fa\u73b0 silent corruption\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">2) \u73af\u5883\u56e0\u7d20\uff08\u4f8b\u5982\u8f90\u5c04\uff09<\/h5>\n\n\n\n<p>\u8bba\u6587\u4e5f\u660e\u786e\u6307\u51fa\uff1a<strong>\u8f90\u5c04\u7b49\u73af\u5883\u56e0\u7d20<\/strong>\u4f1a\u5f15\u53d1 SDC\uff08\u7ecf\u5178\u5c31\u662f\u5b87\u5b99\u5c04\u7ebf\u5bfc\u81f4\u7684\u8f6f\u9519\u8bef\/\u5355\u7c92\u5b50\u7ffb\u8f6c\uff09\u3002<br>\u8fd9\u7c7b\u5f80\u5f80\u662f<strong>\u77ac\u6001<\/strong>\u7684\uff1a\u67d0\u4e00\u6b21\u8fd0\u7b97\u6216\u67d0\u4e00\u4e2a\u6bd4\u7279\u88ab\u6270\u52a8\uff0c\u7cfb\u7edf\u4e0d\u5d29\u4f46\u7ed3\u679c\u504f\u4e86\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3) \u8f6f\u4ef6\/\u56fa\u4ef6\/\u7f16\u8bd1\u5668\/\u9a71\u52a8 Bug<\/h5>\n\n\n\n<p>SDC \u4e5f\u53ef\u80fd\u6765\u81ea<strong>\u8f6f\u4ef6\u5c42\u9762\u7684 bug<\/strong>\uff08\u5305\u62ec\u9a71\u52a8\u3001\u7f16\u8bd1\u5668\u4f18\u5316\u3001\u56fa\u4ef6\u3001\u901a\u4fe1\u5e93\u7b49\uff09\uff0c\u8868\u73b0\u4e3a\u201c\u8bed\u4e49\u9519\u8bef\u4f46\u4e0d\u62a5\u9519\u201d\u3002<br>\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7f16\u8bd1\u5668\u9519\u8bef\u4f18\u5316\u3001\u672a\u5b9a\u4e49\u884c\u4e3a\u89e6\u53d1\u3001\u8d8a\u754c\u5199\u4f46\u6ca1\u5d29\u6e83<\/li>\n\n\n\n<li>\u9a71\u52a8\/\u56fa\u4ef6\u5728\u7279\u5b9a\u5e76\u53d1\u6216\u7279\u5b9a\u6307\u4ee4\u5e8f\u5217\u4e0b\u51fa\u9519<\/li>\n\n\n\n<li>\u5206\u5e03\u5f0f\u7cfb\u7edf\u91cc\u6821\u9a8c\u4e0d\u8db3\u5bfc\u81f4\u7684 silent semantic violation\uff08\u7ed3\u679c\u770b\u4f3c\u5408\u7406\u4f46\u5176\u5b9e\u9519\u4e86\uff09<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">4) \u201c\u53ea\u5728\u7279\u5b9a\u8d1f\u8f7d\/\u6e29\u5ea6\/\u5229\u7528\u7387\u4e0b\u51fa\u73b0\u201d\u7684\u8bbe\u5907\u7279\u5f02\u6027\u7f3a\u9677\uff08\u89c4\u6a21\u5316\u8bad\u7ec3\u66f4\u5bb9\u6613\u78b0\u5230\uff09<\/h5>\n\n\n\n<p>\u5927\u89c4\u6a21\u7cfb\u7edf\u91cc\u5e38\u89c1\u4e00\u79cd\u5f88\u68d8\u624b\u7684\u60c5\u51b5\uff1a<strong>\u67d0\u4e9b\u8bbe\u5907\u53ea\u6709\u5728\u7279\u5b9a\u5229\u7528\u7387\u6216\u6e29\u5ea6\u533a\u95f4\u624d\u4f1a\u51fa\u9519<\/strong>\u3002<br>\u8fd9\u4f1a\u5bfc\u81f4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5e73\u65f6\u8dd1\u5c0f\u4efb\u52a1\u3001\u77ed\u4efb\u52a1\u6d4b\u4e0d\u51fa\u6765<\/li>\n\n\n\n<li>\u4e00\u5230\u957f\u65f6\u95f4\u6ee1\u8f7d\uff08\u5982 LLM \u8bad\u7ec3\uff09\u5c31\u5f00\u59cb\u201c\u5076\u53d1\u3001\u975e\u5747\u5300\u3001\u65f6\u6709\u65f6\u65e0\u201d\u7684\u9519\u8bef<\/li>\n<\/ul>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aJeffrey Ma, Hengzhi Pei, Leonard Lausen, George Karypis\uff0c\u6765\u81eaHarvard University\u548cAmazon Web Services\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e00\u3001\u7814\u7a76\u601d\u8def\u4e0e\u5b9e\u9a8c\u8bbe\u8ba1\uff08\u8bba\u6587\u7684\u5173\u952e\u8d21\u732e\uff09<\/h4>\n\n\n\n<p>\u4ee5\u5f80\u5f88\u591a\u5de5\u4f5c\u4f9d\u8d56\u201c\u6ce8\u5165\u6545\u969c\u201d\u53bb\u6a21\u62df\uff0c\u4f46\u6a21\u62df\u51fa\u6765\u7684\u9519\u8bef\u5206\u5e03\u4e0e\u771f\u5b9e\u751f\u4ea7\u73af\u5883\u4e0d\u4e00\u5b9a\u4e00\u81f4\u3002\u672c\u6587\u7684\u7279\u70b9\u5728\u4e8e\uff1a\u4f5c\u8005\u901a\u8fc7\u4e91\u5e73\u53f0\u62ff\u5230\u4e86\u4e00\u6279<strong>\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u88ab\u81ea\u52a8\u5316\u8fd0\u7ef4\u7cfb\u7edf\u8bc6\u522b\u5e76\u9694\u79bb\u51fa\u6765\u7684\u5f02\u5e38\u8282\u70b9<\/strong>\uff0c\u5e76\u628a\u5b83\u4eec\u4e0e\u5065\u5eb7\u8282\u70b9\u6210\u5bf9\u5bf9\u6bd4\uff0c\u5c3d\u91cf\u8ba9\u4e24\u8fb9\u8bad\u7ec3\u8fc7\u7a0b\u9664\u4e86 SDC \u4e4b\u5916\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u628a\u201cSDC \u7684\u5f71\u54cd\u201d\u4ece\u5176\u5b83\u968f\u673a\u6027\u91cc\u5265\u79bb\u51fa\u6765\uff0c\u4f5c\u8005\u505a\u4e86\u4e24\u4ef6\u5173\u952e\u4e8b\u60c5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7528\u7f16\u8bd1\u5668\u8ba9\u6267\u884c\u5c3d\u91cf<strong>\u786e\u5b9a\u6027<\/strong>\uff0c\u786e\u4fdd\u5065\u5eb7\u8282\u70b9\u4e4b\u95f4\u5728\u76f8\u540c\u8bbe\u7f6e\u4e0b\u53ef\u4ee5\u5f97\u5230\u5b8c\u5168\u4e00\u81f4\u7684\u7ed3\u679c\uff0c\u4ece\u800c\u628a\u5dee\u5f02\u5f52\u56e0\u5230 SDC\u3002<\/li>\n\n\n\n<li>\u8bbe\u8ba1\u4e86\u4e24\u79cd\u201c\u540c\u6b65\/\u7ea0\u504f\u201d\u673a\u5236\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u8ba1\u7b97\u7ea7\u540c\u6b65<\/strong>\uff1a\u5728\u6ce8\u610f\u529b\u6216\u524d\u9988\u7b49\u5b50\u6a21\u5757\u7b97\u5b8c\u540e\uff0c\u7acb\u5373\u5bf9\u6bd4\u5065\u5eb7\u4e0e\u5f02\u5e38\u8282\u70b9\u7684\u4e2d\u95f4\u7ed3\u679c\uff0c\u5e76\u7528\u5065\u5eb7\u8282\u70b9\u7684\u503c\u8986\u76d6\u5f02\u5e38\u8282\u70b9\uff0c\u907f\u514d\u9519\u8bef\u7ee7\u7eed\u6269\u6563\u5230\u540e\u7eed\u5c42\u4e0e\u540e\u7eed\u53cd\u4f20\u3002<\/li>\n\n\n\n<li><strong>\u53c2\u6570\u7ea7\u540c\u6b65<\/strong>\uff1a\u6bcf\u4e00\u6b65\u4f18\u5316\u66f4\u65b0\u540e\uff0c\u628a\u5065\u5eb7\u8282\u70b9\u7684\u53c2\u6570\u5e7f\u64ad\u8986\u76d6\u5230\u5f02\u5e38\u8282\u70b9\uff0c\u8ba9\u4e0b\u4e00\u6b65\u4ece\u540c\u4e00\u7ec4\u53c2\u6570\u51fa\u53d1\uff0c\u7528\u6765\u9694\u79bb\u201c\u5355\u6b65\u5185\u68af\u5ea6\u662f\u5426\u88ab\u6c61\u67d3\u201d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u8bba\u6587\u628a\u5206\u6790\u5206\u6210\u4e09\u5c42\u95ee\u9898\uff1a\u5b50\u6a21\u5757\u8f93\u51fa\u3001\u5355\u6b65\u68af\u5ea6\u3001\u957f\u65f6\u95f4\u8bad\u7ec3\u7d2f\u79ef\u540e\u7684\u6a21\u578b\u8d28\u91cf\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e8c\u3001\u4e3b\u8981\u53d1\u73b0\u4e00\uff1a\u5b50\u6a21\u5757\u5c42\u9762\uff0cSDC \u53d1\u751f\u201c\u5f88\u4e0d\u5747\u5300\u201d\uff0c\u4e0d\u540c\u8282\u70b9\u5dee\u5f02\u5de8\u5927<\/h4>\n\n\n\n<p>\u5728\u6ce8\u610f\u529b\u4e0e\u524d\u9988\u7684\u524d\u5411\/\u53cd\u5411\u8ba1\u7b97\u4e2d\uff0c\u4f5c\u8005\u7edf\u8ba1\u4e86\u201c\u8f93\u51fa\u4e0d\u4e00\u81f4\u51fa\u73b0\u7684\u9891\u7387\u201d\u548c\u201c\u4e0d\u4e00\u81f4\u7684\u4e25\u91cd\u7a0b\u5ea6\u201d\u3002\u7ed3\u8bba\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8282\u70b9\u4e4b\u95f4\u5dee\u5f02\u975e\u5e38\u5927<\/strong>\uff1a\u6709\u4e9b\u5f02\u5e38\u8282\u70b9\u4e0d\u4e00\u81f4\u9891\u7e41\uff0c\u6709\u4e9b\u5728\u8be5\u8bbe\u7f6e\u4e0b\u51e0\u4e4e\u4e0d\u51fa\u73b0\u3002<\/li>\n\n\n\n<li><strong>\u65f6\u95f4\u4e0a\u4e5f\u5f88\u4e0d\u5747\u5300<\/strong>\uff1a\u4e0d\u4e00\u81f4\u5e38\u5e38\u8868\u73b0\u4e3a\u201c\u957f\u65f6\u95f4\u6b63\u5e38 + \u67d0\u4e9b\u65f6\u523b\u7a81\u7136\u7206\u53d1\u5c16\u5cf0\u201d\uff0c\u751a\u81f3\u5728\u8bad\u7ec3\u521d\u59cb\u5316\u7684\u524d\u51e0\u6b65\u51fa\u73b0\u8f83\u9ad8\u5cf0\u503c\u3002<\/li>\n\n\n\n<li>\u4e25\u91cd\u65f6\uff0c\u67d0\u4e9b\u8282\u70b9\u5728\u67d0\u4e9b\u5fae\u6b65\u4e0a\u4f1a\u51fa\u73b0<strong>\u975e\u5e38\u5938\u5f20\u7684\u504f\u5dee<\/strong>\uff08\u8bba\u6587\u4e3e\u4f8b\u6709\u7684\u8282\u70b9\u504f\u5dee\u53ef\u8fbe\u767e\u500d\u91cf\u7ea7\uff09\uff0c\u8bf4\u660e SDC \u53ef\u80fd\u53ea\u5f71\u54cd\u5f88\u5c11\u91cf\u7684\u5f20\u91cf\u5143\u7d20\uff0c\u4f46\u4e00\u65e6\u547d\u4e2d\u5173\u952e\u4f4d\u7f6e\uff0c\u5c40\u90e8\u5c31\u4f1a\u975e\u5e38\u79bb\u8c31\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e09\u3001\u4e3b\u8981\u53d1\u73b0\u4e8c\uff1a\u5355\u6b65\u68af\u5ea6\u5c42\u9762\uff0c\u6574\u4f53\u201c\u566a\u58f0\u201d\u901a\u5e38\u4e0d\u7b97\u5927\uff0c\u4f46\u5e76\u975e\u7b49\u4e8e\u5b89\u5168<\/h4>\n\n\n\n<p>\u901a\u8fc7\u53c2\u6570\u7ea7\u540c\u6b65\uff0c\u4f5c\u8005\u6bd4\u8f83\u5f02\u5e38\u8282\u70b9\u4e0e\u5065\u5eb7\u8282\u70b9\u5728\u540c\u4e00\u6b65\u5f97\u5230\u7684\u68af\u5ea6\u5dee\u5f02\uff0c\u53d1\u73b0\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u68af\u5ea6\u504f\u5dee\u76f8\u5bf9\u771f\u5b9e\u68af\u5ea6\u5e76\u4e0d\u5927\uff0c\u5e76\u4e14\u968f\u7740\u8bad\u7ec3\u63a8\u8fdb\u603b\u4f53\u8fd8\u4f1a\u4e0b\u964d\uff1b\u6700\u574f\u7684\u8282\u70b9\u60c5\u5f62\u4e0b\uff0c\u68af\u5ea6\u5dee\u5f02\u5927\u81f4\u5728\u201c\u771f\u5b9e\u68af\u5ea6\u7684\u51e0\u4e2a\u767e\u5206\u70b9\u91cf\u7ea7\u201d\u3002<br>\u8fd9\u8bf4\u660e\uff1a\u5f88\u591a\u771f\u5b9e SDC \u5e76\u4e0d\u662f\u90a3\u79cd\u7acb\u523b\u628a\u68af\u5ea6\u5f7b\u5e95\u5f04\u4e71\u7684\u707e\u96be\u578b\u9519\u8bef\uff0c\u800c\u66f4\u50cf\u662f\u201c\u5076\u53d1\u7684\u5c0f\u6270\u52a8\u201d\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u56db\u3001\u4e3b\u8981\u53d1\u73b0\u4e09\uff1a\u957f\u671f\u7d2f\u79ef\u540e\uff0c\u6a21\u578b\u53ef\u80fd\u8d70\u5411\u4e0d\u540c\u7684\u6700\u4f18\u70b9\uff1b\u5fae\u8c03\u9636\u6bb5\u8fd8\u53ef\u80fd\u51fa\u73b0\u81f4\u547d\u7684\u635f\u5931\u5c16\u5cf0<\/h4>\n\n\n\n<p>\u66f4\u5173\u952e\u7684\u662f\u7d2f\u79ef\u6548\u5e94\u3002\u8bba\u6587\u5728\u4e0d\u505a\u53c2\u6570\u8986\u76d6\u7684\u60c5\u51b5\u4e0b\u7ee7\u7eed\u8bad\u7ec3\uff0c\u89c2\u5bdf\u5230\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u635f\u5931\u66f2\u7ebf\u53ef\u80fd\u770b\u8d77\u6765\u51e0\u4e4e\u4e00\u6837<\/strong>\uff0c\u4f46\u5f02\u5e38\u8282\u70b9\u4e0a\u7684\u6a21\u578b\u53c2\u6570\u4f1a\u9010\u6b65\u504f\u79bb\u5065\u5eb7\u8282\u70b9\uff0c\u610f\u5473\u7740\u5b83\u4eec\u88ab\u63a8\u5411\u4e86<strong>\u4e0d\u540c\u7684\u5c40\u90e8\u6700\u4f18<\/strong>\u3002<\/li>\n\n\n\n<li>\u5728\u4e0b\u6e38\u4efb\u52a1\u5fae\u8c03\u4e2d\uff0c\u5927\u591a\u6570\u5f02\u5e38\u8282\u70b9\u6700\u7ec8\u6548\u679c\u548c\u5065\u5eb7\u8282\u70b9\u5dee\u4e0d\u591a\uff0c\u4f46<strong>\u67d0\u4e9b\u8282\u70b9\u4f1a\u5728\u8bad\u7ec3\u4e2d\u6bb5\u6216\u672b\u6bb5\u51fa\u73b0\u660e\u663e\u7684\u635f\u5931\u5c16\u5cf0<\/strong>\uff1b\u8bba\u6587\u7ed9\u51fa\u4e00\u4e2a\u6781\u7aef\u6848\u4f8b\uff1a\u67d0\u8282\u70b9\u5728\u63a5\u8fd1\u5fae\u8c03\u7ed3\u675f\u65f6\u51fa\u73b0\u5c16\u5cf0\uff0c\u5bfc\u81f4\u6700\u7ec8\u6a21\u578b\u5728\u7279\u5b9a\u6570\u636e\u96c6\u4e0a\u8868\u73b0\u201c\u76f4\u63a5\u5d29\u6389\u201d\uff08\u6d4b\u8bd5\u51c6\u786e\u7387\u4e3a\u96f6\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4f5c\u8005\u636e\u6b64\u5f3a\u8c03\uff1a<strong>\u53ea\u76ef\u7740\u635f\u5931\u66f2\u7ebf\u5e76\u4e0d\u53ef\u9760<\/strong>\uff0c\u56e0\u4e3a SDC \u53ef\u80fd\u957f\u671f\u201c\u9759\u9ed8\u201d\uff0c\u76f4\u5230\u67d0\u6b21\u5c16\u5cf0\u624d\u9020\u6210\u4e0d\u53ef\u9006\u635f\u5bb3\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e94\u3001\u8ba8\u8bba\uff1a\u4e3a\u4f55\u68c0\u6d4b\u96be\uff0c\u4ee5\u53ca\u4e00\u4e2a\u4ee3\u8868\u6027\u5c1d\u8bd5\u4e3a\u4ec0\u4e48\u5931\u8d25<\/h4>\n\n\n\n<p>\u8bba\u6587\u8fd8\u8ba8\u8bba\u4e86\u7528\u201c\u7ed9\u77e9\u9635\u4e58\u6cd5\u52a0\u6821\u9a8c\u548c\u201d\u7684\u7ecf\u5178\u5bb9\u9519\u601d\u8def\u6765\u68c0\u6d4b SDC\uff08\u4e00\u79cd\u5e38\u89c1\u7684\u7b97\u6cd5\u7ea7\u5bb9\u9519\u65b9\u6cd5\uff09\u3002\u5b9e\u9a8c\u53d1\u73b0\uff1a\u8fd9\u79cd\u65b9\u6cd5\u5728\u591a\u6570\u5f02\u5e38\u8282\u70b9\u4e0a<strong>\u51e0\u4e4e\u68c0\u6d4b\u4e0d\u5230\u9519\u8bef<\/strong>\uff0c\u5e76\u4e14\u68c0\u6d4b\u5230\u7684\u9891\u7387\u4e0e\u4ed6\u4eec\u5728\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u4e2d\u89c2\u5bdf\u5230\u7684\u4e0d\u4e00\u81f4\u73b0\u8c61\u5e76\u4e0d\u5339\u914d\u3002<br>\u4f5c\u8005\u7ed9\u51fa\u53ef\u80fd\u539f\u56e0\uff1a\u771f\u5b9e SDC \u5f80\u5f80\u4f4e\u9891\u3001\u5bf9\u6574\u4f53\u8303\u6570\u5f71\u54cd\u5c0f\uff1b\u5f15\u5165\u68c0\u6d4b\u4f1a\u6539\u53d8\u8d1f\u8f7d\u4ece\u800c\u6539\u53d8 SDC \u53d1\u751f\u5f62\u6001\uff1b\u4ee5\u53ca\u9519\u8bef\u53ef\u80fd\u53d1\u751f\u5728\u6821\u9a8c\u4fdd\u62a4\u8303\u56f4\u4e4b\u5916\u7b49\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u516d\u3001\u5c40\u9650\u6027<\/h4>\n\n\n\n<p>\u4f5c\u8005\u4e5f\u5766\u7387\u8bf4\u660e\u4e86\u9650\u5236\uff1a\u53ef\u83b7\u53d6\u7684\u771f\u5b9e\u5f02\u5e38\u8282\u70b9\u6570\u91cf\u6709\u9650\uff1b\u5b9e\u9a8c\u4e3b\u8981\u805a\u7126\u201c\u5355\u8282\u70b9\u5185\u7684\u5f20\u91cf\u5e76\u884c\u201d\u8bad\u7ec3\uff1b\u5e76\u4e14\u4e3a\u4e86\u505a\u7ec6\u7c92\u5ea6\u5bf9\u6bd4\u800c\u5f15\u5165\u7684\u540c\u6b65\u673a\u5236\u4f1a\u5e26\u6765\u989d\u5916\u5f00\u9500\u3001\u964d\u4f4e\u5229\u7528\u7387\uff0c\u53ef\u80fd\u53cd\u8fc7\u6765\u964d\u4f4e SDC \u7684\u53d1\u751f\u9891\u7387\u2014\u2014\u5b58\u5728\u201c\u5206\u6790\u8d8a\u7ec6\u3001\u8d8a\u53ef\u80fd\u6539\u53d8\u95ee\u9898\u672c\u8eab\u201d\u7684\u6743\u8861\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e03\u3001\u4e00\u53e5\u8bdd\u603b\u7ed3<\/h4>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u7ed9\u51fa\u7684\u6838\u5fc3\u4fe1\u606f\u662f\uff1a<strong>\u771f\u5b9e\u751f\u4ea7\u73af\u5883\u91cc\u7684 SDC \u5f80\u5f80\u662f\u4f4e\u9891\u3001\u975e\u5747\u5300\u3001\u8282\u70b9\u5dee\u5f02\u6781\u5927\u7684\u201c\u5c0f\u6270\u52a8\u201d\uff0c\u77ed\u671f\u770b\u4f3c\u65e0\u4f24\u5927\u96c5\uff0c\u4f46\u957f\u671f\u4f1a\u628a\u8bad\u7ec3\u63a8\u5411\u4e0d\u540c\u7684\u6536\u655b\u7ed3\u679c\uff1b\u5728\u67d0\u4e9b\u4efb\u52a1\u548c\u65f6\u523b\uff0c\u5b83\u53c8\u53ef\u80fd\u4ee5\u635f\u5931\u5c16\u5cf0\u7684\u5f62\u5f0f\u7a81\u7136\u7206\u53d1\uff0c\u76f4\u63a5\u6bc1\u6389\u6a21\u578b\u3002<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Understanding Silent Data Corruption in LLM Training\u7814 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,86],"class_list":["post-6676","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Understanding Silent Data Corruption in LLM Training\u7814&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6676","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6676"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6676\/revisions"}],"predecessor-version":[{"id":6677,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6676\/revisions\/6677"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6676"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6676"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6676"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}