{"id":1057,"date":"2025-08-11T13:38:00","date_gmt":"2025-08-11T05:38:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1057"},"modified":"2025-09-17T13:39:37","modified_gmt":"2025-09-17T05:39:37","slug":"%e4%bc%98%e5%8c%96%e5%99%a8%e3%80%8c%e8%af%b8%e7%a5%9e%e4%b9%8b%e6%88%98%e3%80%8d%ef%bc%9fadamw-%e5%87%ad%e3%80%8c%e7%a8%b3%e5%ae%9a%e3%80%8d%e8%83%9c%e5%87%ba","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/08\/11\/%e4%bc%98%e5%8c%96%e5%99%a8%e3%80%8c%e8%af%b8%e7%a5%9e%e4%b9%8b%e6%88%98%e3%80%8d%ef%bc%9fadamw-%e5%87%ad%e3%80%8c%e7%a8%b3%e5%ae%9a%e3%80%8d%e8%83%9c%e5%87%ba\/","title":{"rendered":"\u4f18\u5316\u5668\u300c\u8bf8\u795e\u4e4b\u6218\u300d\uff1fAdamW \u51ed\u300c\u7a33\u5b9a\u300d\u80dc\u51fa"},"content":{"rendered":"\n<p>\u81ea 2014 \u5e74\u63d0\u51fa\u4ee5\u6765\uff0cAdam \u53ca\u5176\u6539\u8fdb\u7248 AdamW \u957f\u671f\u5360\u636e\u5f00\u653e\u6743\u91cd\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u7684\u4e3b\u5bfc\u5730\u4f4d\uff0c\u5e2e\u52a9\u6a21\u578b\u5728\u6d77\u91cf\u6570\u636e\u4e0b\u4fdd\u6301\u7a33\u5b9a\u5e76\u5b9e\u73b0\u8f83\u5feb\u6536\u655b\u3002<\/p>\n\n\n\n<p>\u968f\u7740\u6a21\u578b\u89c4\u6a21\u8fc5\u901f\u6269\u5927\uff0c\u9884\u8bad\u7ec3\u5df2\u6210\u4e3a\u8ba1\u7b97\u5bc6\u96c6\u578b\u4efb\u52a1\u7684\u5178\u578b\u4ee3\u8868\uff0c\u5728\u5927\u6a21\u578b\u7814\u53d1\u4e2d\u5f80\u5f80\u662f\u6700\u4e3b\u8981\u7684\u8ba1\u7b97\u5f00\u9500\u3002\u5728\u8fd9\u79cd\u80cc\u666f\u4e0b\uff0c\u4f18\u5316\u5668\u7684\u8bbe\u8ba1\u76f4\u63a5\u5173\u7cfb\u5230\u6536\u655b\u901f\u5ea6\u4e0e\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u8005\u4eec\u63a2\u7d22\u4e86\u591a\u79cd\u6539\u8fdb\u65b9\u5411\uff0c\u5176\u4e2d\u6700\u5feb\u7684\u4f18\u5316\u5668\u5f80\u5f80\u91c7\u7528\u77e9\u9635\u578b\u9884\u6761\u4ef6\u5b50\uff08\u5982 Muon\u3001Soap\u3001Kron\uff09\uff0c\u76f8\u8f83\u4e8e\u7ecf\u8fc7\u4e25\u683c\u8c03\u4f18\u7684 AdamW\uff0c\u53ef\u4ee5\u5e26\u6765\u7ea6 30\u201340% \u7684\u8fed\u4ee3\u7ea7\u522b\u52a0\u901f\u3002<\/p>\n\n\n\n<p>\u65af\u5766\u798f\u5927\u5b66 Percy Liang \u56e2\u961f\u7684\u7814\u7a76\u6307\u51fa\uff0c\u5c3d\u7ba1\u5b58\u5728\u8bb8\u591a\u58f0\u79f0\u80fd\u63d0\u4f9b\u663e\u8457\u52a0\u901f\uff081.4 \u81f3 2 \u500d\uff09\u7684\u66ff\u4ee3\u65b9\u6848\uff0c<strong>AdamW \u4f9d\u7136\u662f\u9884\u8bad\u7ec3\u7684\u7a33\u5065\u9996\u9009\uff0c\u4f46\u77e9\u9635\u578b\u65b9\u6cd5\u5728\u7279\u5b9a\u6570\u636e\u2013\u6a21\u578b\u6bd4\u4f8b\u4e0b\u5c55\u73b0\u51fa\u660e\u663e\u4f18\u52bf<\/strong>\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bba\u6587\u6807\u9898\uff1aFantastic Pretraining Optimizers and \u00a0Where to Find Them<\/li>\n\n\n\n<li>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/www.arxiv.org\/pdf\/2509.02046v1<\/li>\n\n\n\n<li>Github\uff1ahttps:\/\/github.com\/marin-community\/marin\/issues\/1290<\/li>\n\n\n\n<li>\u535a\u5ba2\uff1ahttps:\/\/wandb.ai\/marin-community\/marin\/reports\/Fantastic-Optimizers-and-Where-to-Find-Them&#8211;VmlldzoxMjgzMzQ2NQ<\/li>\n<\/ul>\n\n\n\n<p>\u7814\u7a76\u8005\u8ba4\u4e3a\uff0c\u8fd9\u79cd\u73b0\u8c61\u53ef\u80fd\u6e90\u4e8e\u4e24\u4e2a\u5173\u952e\u7684\u65b9\u6cd5\u8bba\u7f3a\u9677\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u95ee\u9898 1\uff1a\u4e0d\u516c\u5e73\u7684\u8d85\u53c2\u6570\u8c03\u4f18\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u57fa\u7ebf\u6a21\u578b\u901a\u5e38\u8c03\u4f18\u4e0d\u8db3<\/strong>\uff1a\u5728\u5e38\u7528\u7684 AdamW \u57fa\u7ebf\u4e2d\uff0c\u4ec5\u4ec5\u662f\u8c03\u4f18\u5b66\u4e60\u7387\u8fd9\u4e00\u4e2a\u53c2\u6570\uff0c\u5c31\u80fd\u5728 1.3 \u4ebf\u53c2\u6570\u89c4\u6a21\u7684\u6a21\u578b\u4e0a\u5b9e\u73b0 2 \u500d\u7684\u52a0\u901f\u3002<\/p>\n\n\n\n<p><strong>\u56fa\u5b9a\u5171\u4eab\u7684\u8d85\u53c2\u6570\u5e76\u4e0d\u80fd\u4fdd\u8bc1\u6bd4\u8f83\u7684\u516c\u5e73\u6027<\/strong>\uff1a\u4f8b\u5982\uff0c\u4e0e\u6807\u51c6\u7684\u6743\u91cd\u8870\u51cf\u503c 0.1 \u76f8\u6bd4\uff0cLion \u4f18\u5316\u5668\u66f4\u504f\u597d\u8f83\u9ad8\u7684\u6743\u91cd\u8870\u51cf\u503c\uff08\u5982 0.6\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"394\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-336-1024x394.jpg\" alt=\"\" class=\"wp-image-1058\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-336-1024x394.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-336-300x115.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-336-768x295.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-336.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>\u5de6\uff1a\u5e38\u7528\u7684 AdamW \u57fa\u7ebf\u5b58\u5728\u8c03\u4f18\u4e0d\u8db3\u7684\u95ee\u9898\u3002 \u5728 Brown \u7b49\u4eba [2020] \u63d0\u51fa\u3001\u5e76\u88ab\u540e\u7eed\u591a\u9879\u7814\u7a76\u91c7\u7528\u7684 GPT-3 \u8bad\u7ec3\u65b9\u6848\u4e2d\uff0c\u4ec5\u4ec5\u9488\u5bf9\u4e00\u4e2a 1 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u8c03\u6574\u5b66\u4e60\u7387\u8fd9\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u4fbf\u53ef\u5b9e\u73b0\u9ad8\u8fbe 2 \u500d\u7684\u52a0\u901f\uff0c\u8fd9\u51f8\u663e\u4e86\u8fdb\u884c\u6070\u5f53\u8d85\u53c2\u6570\u4f18\u5316\u7684\u91cd\u8981\u6027\u3002\u53f3\uff1a\u5728\u4e0d\u540c\u4f18\u5316\u5668\u4e4b\u95f4\u56fa\u5b9a\u8d85\u53c2\u6570\u5e76\u4e0d\u80fd\u4fdd\u8bc1\u6bd4\u8f83\u7684\u516c\u5e73\u6027\u3002 \u5728\u4ee5\u5f80\u7684\u7814\u7a76\u4e2d\uff0c\u50cf\u5b66\u4e60\u7387\u548c\u6743\u91cd\u8870\u51cf\u8fd9\u7c7b\u5171\u4eab\u8d85\u53c2\u6570\u901a\u5e38\u88ab\u8bbe\u4e3a\u5e38\u91cf\u3002\u7136\u800c\uff0c\u5373\u4f7f\u662f\u6982\u5ff5\u4e0a\u76f8\u4f3c\u7684\u4f18\u5316\u5668\uff0c\u5176\u5bf9\u5e94\u7684\u6700\u4f18\u8d85\u53c2\u6570\u4e5f\u53ef\u80fd\u5927\u76f8\u5f84\u5ead\u3002<\/sup><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u95ee\u9898 2\uff1a\u6d4b\u8bd5\u89c4\u6a21\u4e0d\u8db3<\/li>\n<\/ul>\n\n\n\n<p>\u5927\u591a\u6570\u6d4b\u8bd5\u4ec5\u4f7f\u7528\u5c0f\u578b\u6a21\u578b\uff08\u53c2\u6570\u8fdc\u5c0f\u4e8e 10 \u4ebf\uff09\u6216\u9075\u5faa Chinchilla \u8bba\u6587\u63d0\u51fa\u7684 1 \u500d\u6570\u636e\u914d\u6bd4\u3002\u90a3\u4e48\uff0c\u5728\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u6216\u66f4\u9ad8\u7684\u6570\u636e\u914d\u6bd4\u4e0b\uff0c\u7ed3\u679c\u4f1a\u5982\u4f55\u5462\uff1f<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u8bad\u7ec3\u65e9\u671f\u7684\u68c0\u67e5\u70b9\u4e5f\u53ef\u80fd\u4ea7\u751f\u8bef\u5bfc\uff0c\u5728\u5b66\u4e60\u7387\u8870\u51cf\u9636\u6bb5\uff0c\u4e0d\u540c\u65b9\u6cd5\u7684\u635f\u5931\u66f2\u7ebf\u53ef\u80fd\u4f1a\u53d1\u751f\u4ea4\u53c9\uff0c\u4ece\u800c\u5bfc\u81f4\u6700\u7ec8\u6392\u540d\u53cd\u8f6c\u3002\u56e0\u6b64\uff0c\u5fc5\u987b\u5728\uff08\u4e0d\u540c\u7684\uff09\u8bbe\u5b9a\u4e0b\u8fdb\u884c\u8bad\u7ec3\u7ed3\u675f\u65f6\u7684\u6700\u7ec8\u8bc4\u4f30\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"373\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-338-1024x373.jpg\" alt=\"\" class=\"wp-image-1060\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-338-1024x373.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-338-300x109.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-338-768x280.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-338.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>\u5de6\uff1a\u52a0\u901f\u6548\u679c\u968f\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\u800c\u8870\u51cf\u3002 \u5c3d\u7ba1\u4e00\u4e9b\u4f18\u5316\u5668\u5728\u53c2\u6570\u91cf\u5c0f\u4e8e 10 \u4ebf\u7684\u6a21\u578b\u4e0a\u76f8\u6bd4 AdamW \u80fd\u5c55\u73b0\u51fa\u8f83\u9ad8\u7684\u52a0\u901f\u6bd4\uff081.3-1.4 \u500d\uff09\uff0c\u4f46\u5f53\u6a21\u578b\u89c4\u6a21\u589e\u81f3 12 \u4ebf\u53c2\u6570\u65f6\uff0c\u5176\u52a0\u901f\u6bd4\u4f1a\u8870\u51cf\u81f3\u4ec5 1.1 \u500d\u3002\u53f3\uff1a\u57fa\u4e8e\u77e9\u9635\u7684\u4f18\u5316\u5668\u6027\u80fd\u7a33\u5b9a\u4f18\u4e8e\u57fa\u4e8e\u6807\u91cf\u7684\u4f18\u5316\u5668\u3002 \u8be5\u56fe\u5c55\u793a\u4e86\u4e09\u79cd\u57fa\u4e8e\u6807\u91cf\u7684\u4f18\u5316\u5668\uff08AdamW, Nesterov AdamW, Mars\uff09\u548c\u4e09\u79cd\u57fa\u4e8e\u77e9\u9635\u7684\u4f18\u5316\u5668\uff08Kron, Soap, Muon\uff09\u5728\u4e0d\u540c Chinchilla \u6570\u636e\u914d\u6bd4\u4e0b\u8bad\u7ec3\u65f6\u7684\u635f\u5931\u66f2\u7ebf\u3002\u57fa\u4e8e\u77e9\u9635\u7684\u4f18\u5316\u5668\u76f8\u6bd4\u57fa\u4e8e\u6807\u91cf\u7684\u4f18\u5316\u5668\u5b9e\u73b0\u4e86\u4e00\u81f4\u7684\u52a0\u901f\u6548\u679c\u3002\u6b64\u5916\uff0c\u5728\u8fc7\u8bad\u7ec3\uff08overtrained\uff09\u7684\u60c5\u51b5\u4e0b\uff0c\u8fd9\u4e09\u79cd\u57fa\u4e8e\u77e9\u9635\u7684\u4f18\u5316\u5668\u6700\u7ec8\u4f1a\u6536\u655b\u5230\u76f8\u4f3c\u7684\u635f\u5931\u503c\u3002<\/sup><\/p>\n\n\n\n<p>\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e00\u5047\u8bbe\uff0c\u7814\u7a76\u4eba\u5458\u8fdb\u884c\u4e86\u7cfb\u7edf\u6027\u7684\u6bd4\u8f83\u7814\u7a76\uff0c\u6db5\u76d6\u4e86\u5341\u4e00\u79cd\u4e0d\u540c\u7684\u6df1\u5ea6\u5b66\u4e60\u4f18\u5316\u5668\u3002\u4ed6\u4eec\u5728\u591a\u79cd\u6a21\u578b\u89c4\u6a21\uff08\u4ece 1 \u4ebf\u5230 12 \u4ebf\u53c2\u6570\uff09\u548c\u6570\u636e\u2013\u6a21\u578b\u6bd4\u4f8b\uff08\u53c2\u7167 Chinchilla \u6700\u4f18\u6bd4\u4f8b\u7684 1 \u500d\u81f3 8 \u500d\uff09\u4e0b\uff0c\u4e3a\u6bcf\u4e00\u79cd\u4f18\u5316\u5668\u90fd\u8fdb\u884c\u4e86\u4e25\u8c28\u3001\u72ec\u7acb\u7684\u8d85\u53c2\u6570\u8c03\u4f18\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"736\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-341-1024x736.jpg\" alt=\"\" class=\"wp-image-1063\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-341-1024x736.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-341-300x216.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-341-768x552.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-341.jpg 1107w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; \u672c\u7814\u7a76\u6240\u4f7f\u7528\u7684\u4f18\u5316\u5668\u3002<\/sup><\/p>\n\n\n\n<p>\u7814\u7a76\u53d1\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u72ec\u7acb\u8c03\u4f18\u81f3\u5173\u91cd\u8981<\/strong>\uff1a\u4e00\u4e2a\u4f18\u5316\u5668\u7684\u6700\u4f18\u8d85\u53c2\u6570\u914d\u7f6e\u5f80\u5f80\u65e0\u6cd5\u76f4\u63a5\u8fc1\u79fb\u5230\u53e6\u4e00\u79cd\u4f18\u5316\u5668\u4e0a\u3002\u5982\u679c\u7f3a\u4e4f\u72ec\u7acb\u8c03\u4f18\uff0c\u4e0d\u4ec5\u6bd4\u8f83\u7ed3\u679c\u7f3a\u4e4f\u516c\u5e73\u6027\uff0c\u800c\u4e14\u65b0\u4f18\u5316\u5668\u76f8\u8f83\u4e8e\u7cbe\u5fc3\u8c03\u4f18\u8fc7\u7684 AdamW\uff0c\u5b9e\u9645\u52a0\u901f\u6548\u679c\u8fdc\u4f4e\u4e8e\u5176\u58f0\u79f0\u7684\u6570\u503c\u3002<\/li>\n\n\n\n<li><strong>\u77ed\u671f\u8bc4\u4f30\u5177\u6709\u8bef\u5bfc\u6027<\/strong>\uff1a\u4ec5\u5728\u77ed\u65f6\u95f4\u8bad\u7ec3\u7a97\u53e3\u5185\u8bc4\u4f30\u4f18\u5316\u5668\u6027\u80fd\u662f\u4e0d\u53ef\u9760\u7684\u3002\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\u548c\u5b66\u4e60\u7387\u8870\u51cf\uff0c\u4e0d\u540c\u4f18\u5316\u5668\u7684\u6027\u80fd\u6392\u540d\u53ef\u80fd\u4f1a\u53d1\u751f\u9006\u8f6c\uff0c\u5176\u635f\u5931\u66f2\u7ebf\u751a\u81f3\u4f1a\u591a\u6b21\u4ea4\u53c9\u3002<\/li>\n\n\n\n<li><strong>\u77e9\u9635\u65b9\u6cd5\u6027\u80fd\u9886\u5148<\/strong>\uff1a\u6240\u6709\u901f\u5ea6\u6700\u5feb\u7684\u4f18\u5316\u5668\u90fd\u91c7\u7528\u4e86\u57fa\u4e8e\u77e9\u9635\u7684\u9884\u6761\u4ef6\u5b50\uff0c\u800c\u975e\u4f20\u7edf\u7684\u9010\u5143\u7d20\u6807\u91cf\u7f29\u653e\u3002Muon\u3001Soap \u548c Kron \u7b49\u65b9\u6cd5\uff0c\u76f8\u6bd4\u4e25\u683c\u8c03\u4f18\u540e\u7684 AdamW\uff0c\u80fd\u591f\u5b9e\u73b0 30\u201340% \u7684\u5355\u6b65\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6709\u8da3\u7684\u662f\uff0c\u6700\u4f18\u9009\u62e9\u4e5f\u4e0e\u5177\u4f53\u573a\u666f\u76f8\u5173\uff1a\u5728\u6807\u51c6 Chinchilla \u6570\u636e\u6bd4\u4f8b\u4e0b\uff0cMuon \u8868\u73b0\u6700\u4f73\uff1b\u800c\u5f53\u6570\u636e\u91cf\u76f8\u5bf9\u4e8e\u6a21\u578b\u89c4\u6a21\u7684\u6bd4\u4f8b\u63d0\u5347\u81f3 8 \u500d\u4ee5\u4e0a\u65f6\uff0cSoap \u5219\u6210\u4e3a\u66f4\u4f18\u7684\u9009\u62e9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"677\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-340-1024x677.jpg\" alt=\"\" class=\"wp-image-1062\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-340-1024x677.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-340-300x198.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-340-768x507.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-340.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u65b9\u6cd5<\/strong><\/p>\n\n\n\n<p>\u7814\u7a76\u8bbe\u8ba1\u4e86\u4e00\u5957\u4e25\u8c28\u7684\u65b9\u6cd5\u8bba\u6765\u8bc4\u4f30\u8fd9\u4e9b\u4f18\u5316\u5668\uff0c\u8be5\u65b9\u6cd5\u5206\u4e3a\u4e09\u4e2a\u4e3b\u8981\u9636\u6bb5\u3002\u9996\u5148\u662f\u901a\u7528\u8bbe\u7f6e\u9636\u6bb5\uff0c\u660e\u786e\u4e86\u5b9e\u9a8c\u73af\u5883\u3002\u7814\u7a76\u4f7f\u7528\u4e86\u56db\u79cd\u4e0d\u540c\u89c4\u6a21\u7684 Transformer \u6a21\u578b\uff0c\u53c2\u6570\u91cf\u4ece 130M \u5230 1.2B\uff0c\u5e8f\u5217\u957f\u5ea6\u5747\u4e3a 4096\uff0c\u5e76\u8be6\u7ec6\u5217\u4e3e\u4e86\u5404\u6a21\u578b\u5c42\u6570\u3001\u9690\u85cf\u7ef4\u5ea6\u7b49\u5177\u4f53\u914d\u7f6e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"230\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-337-1024x230.jpg\" alt=\"\" class=\"wp-image-1059\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-337-1024x230.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-337-300x67.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-337-768x173.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-337.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>&nbsp; &nbsp; &nbsp; \u6240\u7814\u7a76\u7684\u5404\u4e2a\u6a21\u578b\u89c4\u6a21\u7684\u8be6\u7ec6\u67b6\u6784\u8d85\u53c2\u6570\u3002<\/sup><\/p>\n\n\n\n<p>\u6570\u636e\u65b9\u9762\uff0c\u7814\u7a76\u6df7\u5408\u4f7f\u7528\u4e86 DCLM-baseline\u3001StarCoder V2 \u548c ProofPile 2 \u6570\u636e\u96c6\uff0c\u5e76\u4f7f\u7528 LLaMA-3 \u5206\u8bcd\u5668\u8fdb\u884c\u5206\u8bcd\uff0c\u786e\u4fdd\u4e86\u8bad\u7ec3\u6570\u636e\u7684\u4e30\u5bcc\u6027\u3002\u8bc4\u4f30\u7684\u4f18\u5316\u5668\u6db5\u76d6\u4e86 AdamW\u3001NAdamW\u3001Mars\u3001Cautious\u3001Lion\u3001Adam-mini\u3001Muon\u3001Scion\u3001Kron (PSGD) \u3001Soap \u548c Sophia\uff0c\u4ee3\u8868\u4e86\u5f53\u524d\u6df1\u5ea6\u5b66\u4e60\u4f18\u5316\u9886\u57df\u7684\u4e3b\u6d41\u548c\u524d\u6cbf\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u9636\u6bb5 I: \u5168\u9762\u53c2\u6570\u626b\u63cf<\/p>\n\n\n\n<p>\u7814\u7a76\u65e8\u5728\u89e3\u51b3\u57fa\u7ebf\u4f18\u5316\u5668\u8d85\u53c2\u6570\u8c03\u6574\u4e0d\u5f53\u5bfc\u81f4\u5176\u6027\u80fd\u88ab\u4f4e\u4f30\u7684\u95ee\u9898\u3002\u7814\u7a76\u91c7\u7528\u4e86\u5750\u6807\u4e0b\u964d\u6cd5\uff0c\u5bf9\u6240\u6709\u4f18\u5316\u5668\u7684\u8d85\u53c2\u6570\uff08\u5305\u62ec\u5b66\u4e60\u7387\u3001\u6743\u91cd\u8870\u51cf\u3001\u9884\u70ed\u6b65\u6570\u3001\u03b2\u2081\u3001\u03b2\u2082\u3001\u03b5\u3001\u6700\u5927\u68af\u5ea6\u8303\u6570\u548c\u6279\u6b21\u5927\u5c0f\uff09\u5728\u9884\u8bbe\u7f51\u683c\u4e0a\u8fdb\u884c\u4e86\u8be6\u5c3d\u641c\u7d22\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e00\u9636\u6bb5\u7684\u5b9e\u9a8c\u8bbe\u7f6e\u6db5\u76d6\u4e86 130M\u3001300M \u548c 500M \u6a21\u578b\u5728 1 \u500d Chinchilla \u6570\u636e\u91cf\u4e0b\u7684\u8bad\u7ec3\uff0c\u4ee5\u53ca 130M \u6a21\u578b\u5728 2 \u500d\u30014 \u500d\u30018 \u500d Chinchilla \u6570\u636e\u91cf\u4e0b\u7684\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u53d1\u73b0\uff0c\u5bf9\u6bcf\u4e2a\u4f18\u5316\u5668\u8fdb\u884c\u4e25\u683c\u7684\u8d85\u53c2\u6570\u8c03\u6574\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u4e0d\u540c\u4f18\u5316\u5668\u4e4b\u95f4\u7684\u6700\u4f18\u8d85\u53c2\u6570\u914d\u7f6e\u5dee\u5f02\u663e\u8457\uff0c\u76f2\u76ee\u8fc1\u79fb\u8d85\u53c2\u6570\u4f1a\u5bfc\u81f4\u4e0d\u516c\u5e73\u7684\u6bd4\u8f83\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u7814\u7a76\u4e5f\u89c2\u5bdf\u5230\uff0c\u4e0e\u7ecf\u8fc7\u7cbe\u5fc3\u8c03\u6574\u7684\u57fa\u7ebf AdamW \u76f8\u6bd4\uff0c\u5b9e\u9645\u7684\u52a0\u901f\u6548\u679c\u666e\u904d\u4f4e\u4e8e\u6b64\u524d\u4e00\u4e9b\u7814\u7a76\u6240\u58f0\u79f0\u7684\u6c34\u5e73\u3002<\/p>\n\n\n\n<p>\u9636\u6bb5 II: \u654f\u611f\u8d85\u53c2\u6570\u8bc6\u522b<\/p>\n\n\n\n<p>\u7814\u7a76\u6839\u636e\u7b2c\u4e00\u9636\u6bb5\u7684\u7ed3\u679c\uff0c\u8bc6\u522b\u51fa\u90a3\u4e9b\u6700\u4f18\u503c\u4f1a\u968f\u6a21\u578b\u89c4\u6a21\u53d8\u5316\u7684\u654f\u611f\u8d85\u53c2\u6570\uff0c\u4f8b\u5982\u5b66\u4e60\u7387\u548c\u9884\u70ed\u957f\u5ea6\u3002\u968f\u540e\uff0c\u8fd9\u4e9b\u654f\u611f\u8d85\u53c2\u6570\u5728 300M \u548c 500M \u6a21\u578b\u4ee5\u53ca 2 \u500d\u30014 \u500d\u30018 \u500d Chinchilla \u6570\u636e\u91cf\u4e0b\u8fdb\u884c\u4e86\u8fdb\u4e00\u6b65\u7684\u7f51\u683c\u641c\u7d22\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"498\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-343-1024x498.jpg\" alt=\"\" class=\"wp-image-1065\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-343-1024x498.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-343-300x146.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-343-768x374.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-343.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>\u7b2c\u4e00\u9636\u6bb5\u4e0e\u7b2c\u4e8c\u9636\u6bb5\u7684\u4e3b\u8981\u7ed3\u679c\u3002\u4e0a\u56fe\uff1a \u6211\u4eec\u7ed8\u5236\u4e86\u7b2c\u4e00\u9636\u6bb5\u548c\u7b2c\u4e8c\u9636\u6bb5\u5b9e\u9a8c\u4e2d\uff0c\u6a21\u578b\u5728 C4\/EN \u6570\u636e\u96c6\u4e0a\u7684\u9a8c\u8bc1\u96c6\u635f\u5931\u3002\u56fe\u4e2d\u7684\u6bcf\u4e00\u4e2a\u70b9\u90fd\u5bf9\u5e94\u4e8e\u6bcf\u79cd\u4f18\u5316\u5668\u5728\u76f8\u5e94\u7684 Chinchilla \u6570\u636e\u914d\u6bd4\u4e0b\u6240\u80fd\u8fbe\u5230\u7684\u6700\u4f18\u635f\u5931\u503c\u3002\u4e0b\u56fe\uff1a \u6211\u4eec\u9488\u5bf9\u90e8\u5206\u4f18\u5316\u5668\uff0c\u7ed8\u5236\u4e86\u5b83\u4eec\u5728 HellaSwag \u57fa\u51c6\u4e0a\u7684\u6027\u80fd\u3002\u8fd9\u4e9b\u4f18\u5316\u5668\u5305\u62ec\uff1aAdamW \u57fa\u7ebf\u3001\u6027\u80fd\u6392\u540d\u524d 2 \u7684\u57fa\u4e8e\u6807\u91cf\u7684\u4f18\u5316\u5668\uff0c\u4ee5\u53ca\u6027\u80fd\u6392\u540d\u524d 3 \u7684\u57fa\u4e8e\u77e9\u9635\u7684\u4f18\u5316\u5668\u3002\u6027\u80fd\u6570\u636e\u6765\u81ea\u4e8e\u5b83\u4eec\u5404\u81ea\u6700\u4f18\u7684\u8fd0\u884c\u6279\u6b21\u3002<\/sup><\/p>\n\n\n\n<p>\u901a\u8fc7\u7ed3\u5408\u524d\u4e24\u4e2a\u9636\u6bb5\u7684\u7ed3\u679c\uff0c\u7814\u7a76\u83b7\u5f97\u4e86 12 \u79cd\u4e0d\u540c\u8bbe\u7f6e\u4e0b\u7684\u8fd1\u4e4e\u6700\u4f18\u8d85\u53c2\u6570\u96c6\u53ca\u5176\u5bf9\u5e94\u7684\u635f\u5931\u3002\u4e3a\u4e86\u91cf\u5316\u4e0d\u540c\u4f18\u5316\u5668\u76f8\u5bf9\u4e8e AdamW \u7684\u52a0\u901f\u6548\u679c\uff0c\u7814\u7a76\u62df\u5408\u4e86 AdamW \u635f\u5931\u968f\u6570\u636e\u9884\u7b97\u53d8\u5316\u7684\u7f29\u653e\u5b9a\u5f8b\uff0c\u5e76\u4ee5\u6b64\u8ba1\u7b97\u51fa\u8fbe\u5230\u76f8\u540c\u635f\u5931\u6240\u9700\u7684 AdamW \u6570\u636e\u91cf\u4e0e\u4f18\u5316\u5668\u5b9e\u9645\u6240\u9700\u6570\u636e\u91cf\u4e4b\u6bd4\uff0c\u4f5c\u4e3a\u52a0\u901f\u6bd4\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u53d1\u73b0\uff0c\u57fa\u4e8e\u77e9\u9635\u7684\u4f18\u5316\u5668\u867d\u7136\u8868\u73b0\u666e\u904d\u4f18\u4e8e\u57fa\u4e8e\u6807\u91cf\u7684\u4f18\u5316\u5668\uff0c\u4f46\u5176\u52a0\u901f\u6bd4\u5728\u5b9e\u9645\u6d4b\u8bd5\u4e2d\u5747\u672a\u8d85\u8fc7 1.4 \u500d\u3002\u8bb8\u591a\u66ff\u4ee3\u4f18\u5316\u5668\u5728\u5c0f\u89c4\u6a21\u6a21\u578b\u6216\u6709\u9650\u6570\u636e\u6bd4\u4f8b\u4e0b\u770b\u4f3c\u5177\u6709\u4f18\u52bf\uff0c\u4f46\u968f\u7740\u6a21\u578b\u89c4\u6a21\u6269\u5927\uff0c\u8fd9\u4e9b\u52a0\u901f\u4f18\u52bf\u9010\u6e10\u6d88\u5931\u751a\u81f3\u53cd\u8f6c\uff0cAdamW \u4f9d\u7136\u662f\u6700\u7a33\u5065\u7684\u9884\u8bad\u7ec3\u9996\u9009\u3002<\/p>\n\n\n\n<p><strong>\u9636\u6bb5 III: \u6848\u4f8b\u7814\u7a76<\/strong><\/p>\n\n\n\n<p>\u8be5\u9636\u6bb5\u65e8\u5728\u5bf9\u66f4\u5927\u89c4\u6a21\u7684\u5b9e\u9a8c\u8fdb\u884c\u6df1\u5165\u63a2\u7d22\u3002\u7814\u7a76\u9996\u5148\u68c0\u9a8c\u4e86\u8d85\u53c2\u6570\u7684\u62df\u5408\u7a0b\u5ea6\uff0c\u901a\u8fc7\u62df\u5408\u5f62\u5f0f\u4e3a<img loading=\"lazy\" decoding=\"async\" width=\"315\" height=\"48\" src=\"blob:http:\/\/www.applicationsofllm.com\/6db28406-8124-442e-a060-4ae85ee6d4ff\" alt=\"\u56fe\u7247\">&nbsp;\u7684\u5e73\u6ed1\u5b9a\u5f8b\uff0c\u9884\u6d4b\u4e86\u5728\u6a21\u578b\u89c4\u6a21 N \u548c\u6570\u636e\u89c4\u6a21 D \u4e0b\u7684\u6700\u4f18\u8bbe\u7f6e\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e9b\u7f29\u653e\u5b9a\u5f8b\uff0c\u7814\u7a76\u5bf9 1.2B \u6a21\u578b\u5728 1 \u500d Chinchilla \u6570\u636e\u91cf\u4e0b\u8fdb\u884c\u4e86\u5168\u9762\u626b\u63cf\uff0c\u7ed3\u679c\u663e\u793a\u9884\u6d4b\u7684\u914d\u7f6e\u4e0e\u5b9e\u9645\u6700\u4f18\u914d\u7f6e\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u5f02\u6781\u5c0f\uff0c\u8bc1\u660e\u4e86\u9884\u6d4b\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p>\u968f\u540e\uff0c\u7814\u7a76\u8fdb\u884c\u4e86\u4e24\u9879\u6848\u4f8b\u7814\u7a76\uff1a\u4e00\u662f\u8bad\u7ec3 1.2B \u6a21\u578b\u5728 1 \u81f3 8 \u500d Chinchilla \u6570\u636e\u91cf\u4e0b\uff0c\u4ee5\u68c0\u9a8c\u4f18\u5316\u5668\u52a0\u901f\u6548\u679c\u968f\u6a21\u578b\u89c4\u6a21\u6269\u5c55\u7684\u53d8\u5316\uff1b\u4e8c\u662f\u5728 16 \u500d Chinchilla \u6570\u636e\u91cf\u4e0b\u8bad\u7ec3 130M \u548c 300M \u6a21\u578b\uff0c\u4ee5\u89c2\u5bdf\u5728\u6781\u7aef\u6570\u636e\u91cf\u4e0e\u6a21\u578b\u6bd4\u4f8b\u4e0b\u7684\u4f18\u5316\u5668\u8868\u73b0\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"259\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-342-1024x259.jpg\" alt=\"\" class=\"wp-image-1064\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-342-1024x259.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-342-300x76.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-342-768x194.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-342.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>\u6848\u4f8b\u5206\u6790\u3002\u5de6\u56fe\uff1a \u5728 12 \u4ebf\u53c2\u6570\u6a21\u578b\u4e0a\uff0cAdamW\u3001NAdamW\u3001Muon \u548c Soap \u56db\u79cd\u4f18\u5316\u5668\u7684\u9a8c\u8bc1\u96c6\u635f\u5931\u7f29\u653e\u60c5\u51b5\u3002\u7ed3\u679c\u663e\u793a\uff0cMuon \u548c Soap \u76f8\u6bd4 AdamW \u4ecd\u6709\u663e\u8457\u7684\u52a0\u901f\u6548\u679c\uff0c\u4f46\u76f8\u6bd4 NAdamW \u5df2\u65e0\u660e\u663e\u52a0\u901f\u4f18\u52bf\u3002\u4e2d\u56fe\uff1a \u91c7\u7528\u4e0e\u56fe 3 \u76f8\u540c\u7684\u65b9\u6cd5\u4f30\u7b97\u52a0\u901f\u6bd4\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0cMuon \u548c Soap \u7684\u52a0\u901f\u6bd4\u968f\u6a21\u578b\u89c4\u6a21\u589e\u5927\u800c\u8870\u51cf\uff0c\u6700\u7ec8\u964d\u81f3\u4ec5 1.1 \u500d\u3002\u53f3\u56fe\uff1a \u5728 3 \u4ebf\u53c2\u6570\u6a21\u578b\u548c 16 \u500d Chinchilla \u6570\u636e\u914d\u6bd4\u7684\u8bbe\u5b9a\u4e0b\uff0c\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5f53\u6570\u636e\u4e0e\u6a21\u578b\u7684\u6bd4\u4f8b\u8fdb\u4e00\u6b65\u589e\u5927\u65f6\uff0cSoap \u7684\u6027\u80fd\u4f18\u4e8e Muon\u3002<\/sup><\/p>\n\n\n\n<p>\u8fd9\u4e00\u9636\u6bb5\u7684\u7ed3\u679c\u8fdb\u4e00\u6b65\u63ed\u793a\u4e86 Muon \u4f18\u5316\u5668\u7684\u6f5c\u5728\u5c40\u9650\u6027\uff1a\u5c3d\u7ba1 Muon \u5bf9\u9ad8\u8fbe 1.2B \u53c2\u6570\u7684\u6a21\u578b\u4ecd\u6709\u52a0\u901f\u6548\u679c\uff0c\u4f46\u52a0\u901f\u6bd4\u4f1a\u4e0b\u964d\u5230 1.2 \u500d\u4ee5\u4e0b\u3002\u5728\u9ad8\u6570\u636e\u4e0e\u6a21\u578b\u6bd4\u4f8b\uff08\u5982 16 \u500d Chinchilla\uff09\u4e0b\uff0cNAdamW \u548c Soap \u5728 130M \u6a21\u578b\u4e0a\u8d85\u8d8a\u4e86 Muon\uff0c\u4e14 Soap \u5728 300M \u6a21\u578b\u4e0a\u4e5f\u8d85\u8fc7\u4e86 Muon\u3002\u7814\u7a76\u63a8\u6d4b\uff0c\u5728\u6570\u636e\u4e0e\u6a21\u578b\u6bd4\u4f8b\u5f88\u9ad8\u65f6\uff0cSoap \u548c Kron \u6240\u7ef4\u6301\u7684\u4e8c\u9636\u52a8\u91cf\u53d8\u5f97\u66f4\u4e3a\u6709\u6548\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u81ea 2014 \u5e74\u63d0\u51fa\u4ee5\u6765\uff0cAdam \u53ca\u5176\u6539\u8fdb\u7248 AdamW \u957f\u671f\u5360\u636e\u5f00\u653e\u6743\u91cd\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u7684\u4e3b\u5bfc\u5730\u4f4d\uff0c\u5e2e\u52a9\u6a21\u578b [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":1061,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-1057","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-10"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1057","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1057"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1057\/revisions"}],"predecessor-version":[{"id":1066,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1057\/revisions\/1066"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1061"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1057"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1057"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1057"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}