{"id":25460,"date":"2025-04-16T10:07:32","date_gmt":"2025-04-16T02:07:32","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=25460"},"modified":"2025-04-16T10:07:32","modified_gmt":"2025-04-16T02:07:32","slug":"%e8%8b%b1%e4%bc%9f%e8%be%be%e5%90%ab%e9%87%8f%e4%b8%ba%e9%9b%b6%ef%bc%81%e5%8d%8e%e4%b8%ba%e5%af%86%e9%9b%86%e6%a8%a1%e5%9e%8b%e6%80%a7%e8%83%bd%e6%af%94%e8%82%a9deepseek-r1%ef%bc%8c%e7%ba%af%e6%98%87","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/04\/16\/25460\/","title":{"rendered":"\u82f1\u4f1f\u8fbe\u542b\u91cf\u4e3a\u96f6\uff01\u534e\u4e3a\u5bc6\u96c6\u6a21\u578b\u6027\u80fd\u6bd4\u80a9DeepSeek-R1\uff0c\u7eaf\u6607\u817e\u96c6\u7fa4\u8bad\u7ec3"},"content":{"rendered":"<div>\n<p><img data-dominant-color=\"9d9493\" data-has-transparency=\"false\" style=\"--dominant-color: #9d9493;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-25462\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/411b774503f898e2259463038efe4399900259-300x167-1.jpg\" width=\"300\" height=\"167\" alt=\"\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/411b774503f898e2259463038efe4399900259-300x167-1.jpg 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/411b774503f898e2259463038efe4399900259-300x167-1-150x84.jpg 150w\" sizes=\"auto, (max-width: 300px) 100vw, 300px\" \/><\/p>\n<\/div>\n<div><\/div>\n<div class=\"article-desc\">\u534e\u4e3a\u5229\u7528\u7eaf\u6607\u817e\u96c6\u7fa4\u8bad\u7ec3\u51fa\u7684\u76d8\u53e4Ultra\uff0c\u5728\u6570\u5b66\u7ade\u8d5b\u3001\u7f16\u7a0b\u7b49\u63a8\u7406\u4efb\u52a1\u5f53\u4e2d\uff0c\u548cR1\u6253\u5f97\u6709\u6765\u6709\u56de\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-1d7a5742=\"\" data-element=\"root\">\n<p>\u5bc6\u96c6\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u4e5f\u80fd\u548cDeepSeek-R1\u63b0\u624b\u8155\u4e86\uff1f<\/p>\n<p class=\"js_darkmode__2\">\u534e\u4e3a\u5229\u7528\u7eaf\u6607\u817e\u96c6\u7fa4\u8bad\u7ec3\u51fa\u7684\u76d8\u53e4Ultra\uff0c\u5728\u6570\u5b66\u7ade\u8d5b\u3001\u7f16\u7a0b\u7b49\u63a8\u7406\u4efb\u52a1\u5f53\u4e2d\uff0c\u548cR1\u6253\u5f97\u6709\u6765\u6709\u56de\u3002<\/p>\n<p class=\"js_darkmode__3\">\u5173\u952e\u662f\u6a21\u578b\u53c2\u6570\u91cf\u53ea\u6709135B\uff0c\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u96f6\u82f1\u4f1f\u8fbe\u542b\u91cf\uff0c\u800c\u4e14\u6ca1\u6709\u51fa\u73b0\u635f\u5931\u5c16\u5cf0\u3002<\/p>\n<p class=\"js_darkmode__4\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s2.51cto.com\/oss\/202504\/15\/5594c39844454e37dab4668ee464d02e58fb9d.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__5\">\u901a\u8fc7\u6539\u8fdb\u7684\u6a21\u578b\u67b6\u6784\u548c\u7cfb\u7edf\u4f18\u5316\u7b56\u7565\uff0c\u76d8\u53e4Ultra\u62e5\u6709\u4f18\u5f02\u7684\u6027\u80fd\u8868\u73b0\u548c52%\u4ee5\u4e0a\u7684\u7b97\u529b\u5229\u7528\u7387\u3002<\/p>\n<p class=\"js_darkmode__6\">\u5e76\u4e14\u6709\u7f51\u53cb\u8868\u793a\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6ca1\u6709\u51fa\u73b0\u635f\u5931\u5c16\u5cf0\u8fd9\u4e00\u7279\u5f81\uff0c\u4f3c\u4e4e\u6b64\u524d\u4ece\u672a\u5b9e\u73b0\u3002<\/p>\n<p class=\"js_darkmode__7\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/15\/89d816628702cc473ae956c084da53e7a22a6e.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h2>135B\u5bc6\u96c6\u6a21\u578b\u6bd4\u80a9DeepSeek-R1<\/h2>\n<p class=\"js_darkmode__9\">\u4f5c\u4e3a\u4e00\u4e2a\u53c2\u6570\u91cf135B\u5bc6\u96c6\u6a21\u578b\uff0c\u76d8\u53e4Ultra\u8fbe\u5230\u4e86\u540c\u5c3a\u5ea6\u5bc6\u96c6\u6a21\u578b\u7684\u6700\u4f18\u8868\u73b0\uff0c\u751a\u81f3\u53ef\u4ee5\u4e0eDeepSeek-R1\u7b49\u53c2\u6570\u91cf\u66f4\u5927\u7684MoE\u6a21\u578b\u7ade\u4e89\u3002<\/p>\n<p class=\"js_darkmode__10\">\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u6a21\u578b\u7684\u8bc4\u6d4b\u4e2d\uff0c\u76d8\u53e4Ultra\u5728\u7edd\u5927\u90e8\u5206\u82f1\u6587\u57fa\u51c6\u4efb\u52a1\u548c\u5168\u90e8\u4e2d\u6587\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u6700\u4f73\u6027\u80fd\uff0c\u4f18\u4e8eLlama 405B\u3001DeepSeek-V3\u7b49baseline\u6a21\u578b\u3002<\/p>\n<p class=\"js_darkmode__11\">\u5c24\u5176\u5728MMLU\u3001TriviaQA\u3001GSM8K\u7b49\u5177\u6709\u6311\u6218\u6027\u7684\u6570\u636e\u96c6\u4e0a\uff0c\u76d8\u53e4Ultra\u5c55\u73b0\u51fa\u4e86\u5353\u8d8a\u7684\u8bed\u8a00\u7406\u89e3\u548c\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p class=\"js_darkmode__12\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/15\/22d326253f4b54ef50760469f796df21a49f6f.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__13\">\u7ecf\u8fc7\u6307\u4ee4\u8c03\u4f18\u540e\uff0c\u76d8\u53e4Ultra\u7684\u6027\u80fd\u8fdb\u4e00\u6b65\u63d0\u5347\uff0c\u5c24\u5176\u5728AIME 2024\u3001MATH-500\u7b49\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u548cLiveCodeBench\u7b49\u7f16\u7a0b\u7ade\u8d5b\u9898\u4e0a\u8fbe\u5230\u4e86SOTA\u6c34\u5e73\u3002<\/p>\n<p class=\"js_darkmode__14\">\u7efc\u5408\u6765\u770b\uff0c\u76d8\u53e4Ultra\u8d85\u8d8a\u4e86\u5305\u62ecGPT-4o\u3001Mistral-Large 2\u7b49\u5f3a\u5927\u6a21\u578b\uff0c\u4e0eDeepSeek-R1\u7b49MoE\u6a21\u578b\u7ade\u4e89\u6fc0\u70c8\u3002<\/p>\n<p class=\"js_darkmode__15\">\u540c\u65f6\uff0c\u76d8\u53e4Ultra\u5728Arena Hard\u3001MMLU-pro\u7b49\u6db5\u76d6\u901a\u7528\u8bed\u8a00\u7406\u89e3\u548c\u63a8\u7406\u7684\u8bc4\u6d4b\u4e2d\u4e5f\u8868\u73b0\u4f18\u5f02\u3002<\/p>\n<p class=\"js_darkmode__16\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s5.51cto.com\/oss\/202504\/15\/262563231bba27b552c4125e7e717e9bbd6161.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__17\">\u90a3\u4e48\uff0c\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u6837\u7684\u6548\u679c\uff0c\u76d8\u53e4Ultra\u91c7\u7528\u4e86\u54ea\u4e9b\u5173\u952e\u6280\u672f\u5462\uff1f<\/p>\n<h2>\u201c\u4e09\u660e\u6cbb\u201d\u5c42\u5f52\u4e00\u5316\u67b6\u6784<\/h2>\n<p class=\"js_darkmode__19\">\u5982\u524d\u6587\u6240\u8ff0\uff0c\u76d8\u53e4Ultra\u662f\u4e00\u6b3e135B\u53c2\u6570\u91cf\u7684\u5bc6\u96c6\u6a21\u578b\uff0c\u4f7f\u7528\u4e8694\u5c42\u7684\u7f51\u7edc\u7ed3\u6784\u3002<\/p>\n<p class=\"js_darkmode__20\">\u76d8\u53e4Ultra\u91c7\u7528\u4e86\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff08GQA\uff09\u673a\u5236\uff0c\u5305\u542b96\u4e2a\u67e5\u8be2\u5934\uff08query head\uff09\u548c8\u4e2a\u952e\u503c\u5934\uff08key-value head\uff09\u3002<\/p>\n<p class=\"js_darkmode__21\">\u4e3a\u4e86\u89e3\u51b3\u8bad\u7ec3\u8d85\u6df1\u7f51\u7edc\u9762\u4e34\u7684\u4e0d\u7a33\u5b9a\u6027\u548c\u6536\u655b\u56f0\u96be\u7b49\u95ee\u9898\uff0c\u76d8\u53e4Ultra\u5728\u6a21\u578b\u67b6\u6784\u4e0a\u505a\u51fa\u4e86\u4e24\u4e2a\u5173\u952e\u6539\u8fdb\u2014\u2014\u6df1\u5ea6\u7f29\u653e\u7684Sandwich-Norm\u5c42\u5f52\u4e00\u5316\u548cTinyInit\u53c2\u6570\u521d\u59cb\u5316\u7b56\u7565\u3002<\/p>\n<p class=\"js_darkmode__22\">\u4f20\u7edf\u7684Transformer\u901a\u5e38\u4f7f\u7528Pre-LN\u5c42\u5f52\u4e00\u5316\uff0c\u4f46\u5728\u6df1\u5ea6\u6a21\u578b\u4e2d\uff0cPre-LN\u5bb9\u6613\u5bfc\u81f4\u6bcf\u4e2a\u5b50\u5c42\u8f93\u51fa\u5c3a\u5ea6\u7684\u6ce2\u52a8\uff0c\u5f15\u53d1\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002<\/p>\n<p class=\"js_darkmode__23\">\u76d8\u53e4Ultra\u4f7f\u7528\u7684Sandwich-Norm\u5c42\u5f52\u4e00\u5316\uff0c\u5219\u662f\u5728\u6b8b\u5dee\u8fde\u63a5\u524d\u5bf9\u6bcf\u4e2a\u5b50\u5c42\u7684\u8f93\u51fa\u505a\u5f52\u4e00\u5316\uff0c\u5e76\u6839\u636e\u7f51\u7edc\u6df1\u5ea6\u5bf9\u521d\u59cb\u5316\u503c\u8fdb\u884c\u7f29\u653e\uff0c\u4ece\u800c\u6709\u6548\u6d88\u9664\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684loss\u5c16\u5cf0\uff0c\u4f7f\u8bad\u7ec3\u8fc7\u7a0b\u66f4\u52a0\u5e73\u7a33\u3002<\/p>\n<p class=\"js_darkmode__24\">\u7528\u66f4\u5bb9\u6613\u7406\u89e3\u7684\u8bdd\u8bf4\uff0c\u4f20\u7edf\u65b9\u6cd5\u4ec5\u5728\u6bcf\u4e2a\u5b50\u5c42\u7684\u8f93\u5165\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u4f46\u8fd9\u79cd\u65b9\u6cd5\u9488\u5bf9\u8f93\u51fa\u4e5f\u8fdb\u884c\u4e86\u5f52\u4e00\u5316\uff0c\u5f62\u6210\u4e86Pre-Norm + \u5b50\u5c42 + Post-Norm\u7684\u201c\u4e09\u660e\u6cbb\u201d\u7ed3\u6784\u3002<\/p>\n<p class=\"js_darkmode__25\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s9.51cto.com\/oss\/202504\/15\/45401bf086ad28cf6ac30542e439dda4763e58.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__26\">\u4f46\u662f\uff0c\u4ec5\u4ec5\u4f7f\u7528Sandwich-Norm\u8fd8\u4e0d\u8db3\u4ee5\u5b8c\u5168\u6d88\u9664\u6df1\u5ea6\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u4e0d\u7a33\u5b9a\u6027\u2014\u2014\u968f\u7740\u7f51\u7edc\u5c42\u6570\u7684\u589e\u52a0\uff0c\u6bcf\u4e00\u5c42\u7684\u8f93\u51fa\u5c3a\u5ea6\u4ecd\u7136\u53ef\u80fd\u51fa\u73b0\u7d2f\u79ef\u6027\u7684\u6f02\u79fb\u3002<\/p>\n<p class=\"js_darkmode__27\">\u4e3a\u6b64\uff0c\u76d8\u53e4Ultra\u5728Sandwich-Norm\u7684\u57fa\u7840\u4e0a\uff0c\u8fdb\u4e00\u6b65\u5f15\u5165\u4e86\u6df1\u5ea6\u7f29\u653e\u673a\u5236\uff0c\u5bf9Post-Norm\u4e2d\u7684\u653e\u7f29\u53c2\u6570\u03b3\u8fdb\u884c\u4e86\u6df1\u5ea6\u76f8\u5173\u7684\u521d\u59cb\u5316\u3002<\/p>\n<p class=\"js_darkmode__28\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s8.51cto.com\/oss\/202504\/15\/0480fa454358d39551c3560de183c2a09dc8b4.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__29\">\u81f3\u4e8e\u6574\u4e2a\u6a21\u578b\u7684\u521d\u59cb\u5316\uff0c\u4f20\u7edf\u7684\u521d\u59cb\u5316\u901a\u5e38\u91c7\u7528\u7684Xavier\u521d\u59cb\u5316\u65b9\u6cd5\u4ec5\u8003\u8651\u6a21\u578b\u5bbd\u5ea6\uff0c\u800c\u76d8\u53e4Ultra\u91c7\u7528\u7684TinyInit\u540c\u65f6\u4f9d\u636e\u6a21\u578b\u6df1\u5ea6\u548c\u5bbd\u5ea6\u6765\u7f29\u653e\u521d\u59cb\u5316\u6743\u91cd\u7684\u6807\u51c6\u5dee\u3002<\/p>\n<p class=\"js_darkmode__30\">\u8fd9\u79cd\u521d\u59cb\u5316\u65b9\u5f0f\u6709\u52a9\u4e8e\u5728\u524d\u5411\u4f20\u64ad\u548c\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\uff0c\u7ef4\u6301\u5404\u5c42\u68af\u5ea6\u7684\u65b9\u5dee\u5728\u4e00\u4e2a\u5408\u7406\u7684\u8303\u56f4\u5185\uff0c\u907f\u514d\u4e86\u68af\u5ea6\u6d88\u5931\u6216\u7206\u70b8\u95ee\u9898\uff0c\u4f7f\u5f97\u8bad\u7ec3\u8fc7\u7a0b\u66f4\u52a0\u7a33\u5b9a\uff0c\u540c\u65f6\u4e5f\u52a0\u901f\u4e86\u6536\u655b\u3002<\/p>\n<p class=\"js_darkmode__31\">\u5b9e\u9a8c\u8868\u660e\uff0cTinyInit\u5728\u6df1\u5ea6\u6a21\u578b\u8bad\u7ec3\u4e2d\u53d6\u5f97\u4e86\u66f4\u597d\u7684\u6536\u655b\u901f\u5ea6\u548c\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\uff1b\u540c\u65f6\u9488\u5bf9embedding\u5c42\uff0c\u4fdd\u6301\u6743\u91cd\u7684\u6807\u51c6\u5dee\u63a5\u8fd11\u4e5f\u80fd\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/p>\n<p class=\"js_darkmode__32\">\u53e6\u5916\uff0c\u76d8\u53e4\u56e2\u961f\u4e5f\u9488\u5bf9Tokenizer\u8fdb\u884c\u4e86\u4f18\u5316\uff0c\u901a\u8fc7\u5728\u901a\u7528\u4e2d\u82f1\u6587\u3001\u4ee3\u7801\u3001\u6570\u5b66\u7b49\u4e0d\u540c\u9886\u57df\u5206\u522b\u8fdb\u884c\u8bcd\u9891\u7edf\u8ba1\uff0c\u518d\u5408\u5e76\u53bb\u91cd\uff0c\u6700\u7ec8\u5f97\u5230\u4e86\u4e00\u4e2a\u517c\u987e\u9886\u57df\u8986\u76d6\u548c\u7f16\u7801\u6548\u7387\u7684153376\u4e2atoken\u7684\u5e73\u8861\u8bcd\u8868\u3002<\/p>\n<p class=\"js_darkmode__33\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s4.51cto.com\/oss\/202504\/15\/b50218c8049d3a137466908fa6d36c3e7db1a0.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h2>8192\u5f20\u6607\u817eNPU\u8bad\u7ec3\u96c6\u7fa4<\/h2>\n<p class=\"js_darkmode__35\">\u76d8\u53e4Ultra\u7684\u6574\u4e2a\u8bad\u7ec3\u6d41\u7a0b\u4e3b\u8981\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5\u2014\u2014\u9884\u8bad\u7ec3\u3001\u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u548c\u6307\u4ee4\u8c03\u4f18\u3002<\/p>\n<p class=\"js_darkmode__36\">\u5176\u4e2d\u9884\u8bad\u7ec3\u53c8\u53ef\u4ee5\u5206\u4e3a\u4e09\u4e2a\u5b50\u9636\u6bb5\uff1a<\/p>\n<ul data-id=\"u738a58b-WgacajMJ\">\n<li data-id=\"l34939b2-AII8WgBm\">\u901a\u7528\u9636\u6bb5\uff1a\u4fa7\u91cd\u5efa\u7acb\u8bed\u8a00\u7406\u89e3\u548c\u77e5\u8bc6\u50a8\u5907\uff0c\u4f7f\u7528\u4e86\u5927\u91cf\u4e2d\u82f1\u6587\u901a\u7528\u8bed\u6599\uff0c\u8986\u76d6\u7f51\u9875\u3001\u4e66\u7c4d\u3001\u767e\u79d1\u7b49\u591a\u4e2a\u6765\u6e90\uff1b<\/li>\n<li data-id=\"l34939b2-G2YmjYlZ\">\u63a8\u7406\u9636\u6bb5\uff1a\u5f15\u5165\u66f4\u591a\u9ad8\u8d28\u91cf\u7684\u6570\u5b66\u548c\u4ee3\u7801\u6570\u636e\uff0c\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u540c\u65f6\u8fd8\u4f7f\u7528instruction\u6570\u636e\u6765\u5e2e\u52a9\u6a21\u578b\u5b66\u4e60\u6267\u884c\u4efb\u52a1\uff1b<\/li>\n<li data-id=\"l34939b2-16eH3mVN\">\u9000\u706b\u9636\u6bb5\uff1a\u5e2e\u52a9\u6a21\u578b\u5de9\u56fa\u77e5\u8bc6\u548c\u63a8\u7406\u80fd\u529b\uff0c\u5e76\u5f3a\u5316\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u3002\u5927\u91cf\u4f7f\u7528\u95ee\u7b54\u5bf9\u548c\u4eba\u7c7b\u53cd\u9988\u6570\u636e\u3002<\/li>\n<\/ul>\n<p class=\"js_darkmode__38\">\u7814\u7a76\u8005\u4eec\u91c7\u7528\u4e86\u57fa\u4e8e\u89c4\u5219\u548c\u6a21\u578b\u7684\u6570\u636e\u6e05\u6d17\u65b9\u6cd5\uff0c\u5e76\u8bbe\u8ba1\u4e86curriculum learning\u7b56\u7565\uff0c\u8ba9\u6a21\u578b\u5faa\u5e8f\u6e10\u8fdb\u5730\u5b66\u4e60\u4e0d\u540c\u96be\u5ea6\u7684\u6837\u672c\u3002<\/p>\n<p class=\"js_darkmode__39\">\u9884\u8bad\u7ec3\u4e2d\u4f7f\u7528\u4e86AdamW\u4f18\u5316\u5668\uff0c\u5e76\u52a8\u6001\u8c03\u6574\u8d85\u53c2\u6570\u3002<\/p>\n<p class=\"js_darkmode__40\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s8.51cto.com\/oss\/202504\/15\/325c99c67a9ac2f0712882839acffdb652088a.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__41\">\u9884\u8bad\u7ec3\u540e\uff0c\u6a21\u578b\u5728\u6700\u957f128K\u7684\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u4e0a\u8fdb\u4e00\u6b65\u8bad\u7ec3\uff0c\u901a\u8fc7\u6269\u5927RoPE\u7684\u57fa\u9891\u6765\u5b9e\u73b0\u957f\u5e8f\u5217\u5efa\u6a21\uff0c\u4ee5\u589e\u5f3a\u5904\u7406\u957f\u6587\u6863\u7684\u80fd\u529b\u3002<\/p>\n<p class=\"js_darkmode__42\">\u6700\u540e\u7684\u6307\u4ee4\u8c03\u4f18\u9636\u5219\u6bb5\u4f7f\u7528\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6765\u4f7f\u6a21\u578b\u66f4\u597d\u5730\u9002\u5e94\u4e0b\u6e38\u4efb\u52a1\uff0c\u5b66\u4f1a\u6267\u884c\u6307\u4ee4\u5e76\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u3002<\/p>\n<p class=\"js_darkmode__43\">\u8bad\u7ec3\u8bbe\u65bd\u65b9\u9762\uff0c\u76d8\u53e4Ultra\u4f7f\u7528\u4e86\u4e00\u4e2a\u75318192\u4e2a\u6607\u817eAI\u5904\u7406\u5668\u7ec4\u6210\u7684\u5927\u89c4\u6a21\u8ba1\u7b97\u96c6\u7fa4\u3002<\/p>\n<p class=\"js_darkmode__44\">\u96c6\u7fa4\u4e2d\u6bcf\u4e2a\u8282\u70b9\u5305\u542b8\u4e2aNPU\uff0c\u901a\u8fc7\u534e\u4e3a\u9ad8\u901f\u7f13\u5b58\u4e00\u81f4\u6027\u4e92\u8054HCCS\u4ee5\u5168\u4e92\u8054\u7684\u62d3\u6251\u7ed3\u6784\u8fde\u63a5\uff0c\u6bcf\u4e2aNPU\u914d\u590764GB\u5185\u5b58\uff0c\u8282\u70b9\u95f4\u5219\u901a\u8fc7200Gbps\u7684RoCE\uff08RDMA over Converged Ethernet\uff09\u7f51\u7edc\u4e92\u8054\u3002<\/p>\n<p class=\"js_darkmode__45\">\u4e3a\u4e86\u5b9e\u73b0\u76d8\u53e4Ultra\u7684\u9ad8\u6548\u8bad\u7ec3\uff0c\u7814\u7a76\u56e2\u961f\u8fd8\u91c7\u7528\u4e86\u4e00\u5957\u7cfb\u7edf\u7684\u5e76\u884c\u7b56\u7565\u548c\u4f18\u5316\u6280\u672f\u3002<\/p>\n<p class=\"js_darkmode__46\">\u5728\u5e76\u884c\u7b56\u7565\u7684\u9009\u62e9\u4e0a\uff0c\u76d8\u53e4Ultra\u7efc\u5408\u8003\u8651\u4e86\u6a21\u578b\u7684\u89c4\u6a21\u3001\u6570\u636e\u7684\u7279\u6027\u4ee5\u53ca\u786c\u4ef6\u7684\u62d3\u6251\uff0c\u6700\u7ec8\u91c7\u7528\u4e86\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u5e8f\u5217\u5e76\u884c\u548c\u6d41\u6c34\u7ebf\u5e76\u884c\u7b49\u591a\u79cd\u5e76\u884c\u65b9\u5f0f\u7684\u7ec4\u5408\uff1a<\/p>\n<ul data-id=\"u738a58b-UF1YhG90\">\n<li data-id=\"l34939b2-4V7H1IdM\">128\u8def\u6570\u636e\u5e76\u884c\uff0c\u5c06\u8bad\u7ec3\u6570\u636e\u5206\u7247\u5230\u4e0d\u540c\u8bbe\u5907\uff0c\u4fdd\u8bc1\u4e86\u6570\u636e\u541e\u5410\uff1b<\/li>\n<li data-id=\"l34939b2-hRD7ifda\">8\u8def\u5f20\u91cf\u5e76\u884c\uff0c\u5229\u7528\u8bbe\u5907\u5185\u90e8\u9ad8\u5e26\u5bbd\u5207\u5206\u5c42\u5185\u5f20\u91cf\uff0c\u5b9e\u73b0\u9ad8\u6548\u901a\u4fe1\uff1b<\/li>\n<li data-id=\"l34939b2-dpQSZQLa\">\u5e8f\u5217\u5e76\u884c\u7528\u4e8e\u5904\u7406\u8d85\u957f\u5e8f\u5217\u4ee5\u964d\u4f4e\u663e\u5b58\u538b\u529b\uff1b<\/li>\n<li data-id=\"l34939b2-7M2F4ihN\">8\u6bb5\u6d41\u6c34\u7ebf\u5e76\u884c\uff0c\u5c06\u4e0d\u540c\u5c42\u5206\u5e03\u5230\u4e0d\u540c\u8bbe\u5907\uff0c\u5f62\u6210\u9ad8\u6548\u7684\u8ba1\u7b97\u6d41\u6c34\u7ebf\u3002<\/li>\n<\/ul>\n<p class=\"js_darkmode__48\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/15\/840fc9285e0e379e7ea759b85affbcb4d43965.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__49\">\u5728\u5e76\u884c\u7b56\u7565\u7684\u57fa\u7840\u4e0a\uff0c\u76d8\u53e4Ultra\u8fd8\u4ece\u591a\u4e2a\u89d2\u5ea6\u5bf9\u8bad\u7ec3\u7cfb\u7edf\u8fdb\u884c\u4e86\u6df1\u5ea6\u4f18\u5316\u3002<\/p>\n<p class=\"js_darkmode__50\">\u4e00\u65b9\u9762\uff0c\u901a\u8fc7\u4f7f\u7528ZeRO\uff08Zero Redundancy Optimizer\uff09\u5206\u5e03\u5f0f\u4f18\u5316\u5668\uff0c\u5c06\u6a21\u578b\u72b6\u6001\u5206\u7247\u5230\u4e0d\u540c\u8bbe\u5907\uff0c\u5927\u5e45\u964d\u4f4e\u4e86\u5355\u4e2a\u8bbe\u5907\u7684\u5185\u5b58\u5360\u7528\uff0c\u5728\u63d0\u9ad8\u6570\u636e\u5e76\u884c\u5ea6\u7684\u540c\u65f6\uff0c\u786e\u4fdd\u4e86\u6bcf\u4e2a\u8bbe\u5907\u7684\u5185\u5b58\u8d1f\u62c5\u5728\u53ef\u63a5\u53d7\u8303\u56f4\u5185\u3002<\/p>\n<p class=\"js_darkmode__51\">\u53e6\u4e00\u65b9\u9762\uff0c\u7814\u7a76\u8005\u4eec\u901a\u8fc7\u5404\u79cd\u901a\u4fe1\u548c\u8ba1\u7b97\u4f18\u5316\u6280\u672f\uff0c\u6700\u5c0f\u5316\u4e86\u901a\u4fe1\u5f00\u9500\uff0c\u63d0\u5347\u4e86\u8ba1\u7b97\u6548\u7387\uff1a<\/p>\n<ul data-id=\"u738a58b-ULaeNRNf\">\n<li data-id=\"l34939b2-CQVTTK9Z\">\u901a\u8fc7\u7b97\u5b50\u878d\u5408\uff08Kernel Fusion\uff09\u5c06\u591a\u4e2a\u5c0f\u7b97\u5b50\u5408\u5e76\uff0c\u51cf\u5c11\u4e86\u5185\u5b58\u8bbf\u95ee\u548ckernel\u542f\u52a8\uff1b<\/li>\n<li data-id=\"l34939b2-pdiL0bhT\">\u901a\u8fc7\u901a\u4fe1\u8ba1\u7b97\u91cd\u53e0\uff08Communication-Computation Overlapping\uff09\u5b9e\u73b0\u901a\u4fe1\u548c\u8ba1\u7b97\u7684\u6df1\u5ea6\u4ea4\u7ec7\uff0c\u9690\u85cf\u901a\u4fe1\u5ef6\u8fdf\uff1b<\/li>\n<li data-id=\"l34939b2-rlnaaqC1\">MC^2\uff08Merged Computation &amp; Communication\uff09\u548cBOA\uff08Batch Optimization Accelerator\uff09\u5206\u522b\u5bf9\u5f20\u91cf\u5e76\u884c\u548c\u89c4\u8303\u5316\u5c42\u7684\u901a\u4fe1\u8fdb\u884c\u4e86\u4e13\u95e8\u4f18\u5316\u2026\u2026<\/li>\n<\/ul>\n<p class=\"js_darkmode__53\"><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s9.51cto.com\/oss\/202504\/15\/b12a11f5195f1a9334d734af1537d88cae9728.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p class=\"js_darkmode__54\">\u5728\u7b97\u6cd5\u3001\u5de5\u7a0b\u3001\u6570\u636e\u5404\u4e2a\u5c42\u9762\u7684\u7cbe\u7ec6\u4f18\u5316\u4e0b\uff0c\u76d8\u53e4Ultra\u5b9e\u73b0\u4e8652%\u4ee5\u4e0a\u7684\u7b97\u529b\u5229\u7528\u7387\u3002<\/p>\n<p class=\"js_darkmode__55\">\u6280\u672f\u62a5\u544a\uff1a<br \/>\nhttps:\/\/github.com\/pangu-tech\/pangu-ultra\/blob\/main\/pangu-ultra-report.pdf<\/p>\n<p>\u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_25460\" class=\"pvc_stats total_only  \" data-element-id=\"25460\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u534e\u4e3a\u5229\u7528\u7eaf\u6607\u817e\u96c6\u7fa4\u8bad\u7ec3\u51fa\u7684\u76d8\u53e4Ultra\uff0c\u5728\u6570\u5b66\u7ade\u8d5b\u3001\u7f16\u7a0b\u7b49\u63a8\u7406\u4efb\u52a1\u5f53\u4e2d\uff0c\u548cR1\u6253\u5f97\u6709\u6765\u6709\u56de\u3002 \u5bc6\u96c6\u6a21\u578b\u7684\u63a8\u7406 [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_25460\" class=\"pvc_stats total_only  \" data-element-id=\"25460\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":25462,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,17,20,80],"tags":[124,887,820],"class_list":["post-25460","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-17","category-20","category-80","tag-ai","tag-887","tag-820"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25460","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=25460"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25460\/revisions"}],"predecessor-version":[{"id":25463,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25460\/revisions\/25463"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/25462"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=25460"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=25460"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=25460"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}