{"id":415304,"date":"2024-06-13T11:32:01","date_gmt":"2024-06-13T03:32:01","guid":{"rendered":"https:\/\/www.idc.net\/help\/?p=415304"},"modified":"2024-06-13T11:32:01","modified_gmt":"2024-06-13T03:32:01","slug":"%e7%9f%a5%e5%90%8dai%e7%a0%94%e7%a9%b6%e8%80%85%e6%b7%b1%e6%8c%96%e8%b0%b7%e6%ad%8cgemma%ef%bc%9a%e5%8f%82%e6%95%b0%e4%b8%8d%e6%ad%a270%e4%ba%bf%ef%bc%8c%e8%ae%be%e8%ae%a1%e5%8e%9f%e5%88%99%e5%be%88","status":"publish","type":"post","link":"https:\/\/idc.net\/help\/415304\/","title":{"rendered":"\u77e5\u540dAI\u7814\u7a76\u8005\u6df1\u6316\u8c37\u6b4cGemma\uff1a\u53c2\u6570\u4e0d\u6b6270\u4ebf\uff0c\u8bbe\u8ba1\u539f\u5219\u5f88\u72ec\u7279"},"content":{"rendered":"<p>\u5c31\u5728\u51e0\u5929\u524d\uff0c\u5f00\u6e90\u5927\u6a21\u578b\u9886\u57df\u8fce\u6765\u4e86\u91cd\u78c5\u65b0\u73a9\u5bb6\uff1a\u8c37\u6b4c\u63a8\u51fa\u4e86\u5168\u65b0\u7684\u5f00\u6e90\u6a21\u578b\u7cfb\u5217\u300cGemma\u300d\u3002\u76f8\u6bd4 Gemini\uff0cGemma \u66f4\u52a0\u8f7b\u91cf\uff0c\u540c\u65f6\u4fdd\u6301\u514d\u8d39\u53ef\u7528\uff0c\u6a21\u578b\u6743\u91cd\u4e5f\u4e00\u5e76\u5f00\u6e90\u4e86\uff0c\u4e14\u5141\u8bb8\u5546\u7528\u3002<\/p>\n<p>\u8c37\u6b4c\u53d1\u5e03\u4e86\u5305\u542b\u4e24\u79cd\u6743\u91cd\u89c4\u6a21\u7684\u6a21\u578b\uff1aGemma 2B \u548c Gemma 7B\u3002\u5c3d\u7ba1\u4f53\u91cf\u8f83\u5c0f\uff0c\u4f46 Gemma \u5df2\u7ecf\u300c\u5728\u5173\u952e\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u660e\u663e\u8d85\u8d8a\u4e86\u66f4\u5927\u7684\u6a21\u578b\u300d\uff0c\u5305\u62ec Llama-2 7B \u548c 13B\uff0c\u4ee5\u53ca\u98ce\u5934\u6b63\u52b2\u7684 Mistral 7B\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5173\u4e8e Gemma \u7684\u6280\u672f\u62a5\u544a\u4e5f\u4e00\u5e76\u653e\u51fa\u3002<\/p>\n<p>\u76f8\u4fe1\u5927\u5bb6\u5df2\u7ecf\u5bf9 Gemma \u7684\u76f8\u5173\u5185\u5bb9\u8fdb\u884c\u4e86\u7cfb\u7edf\u7814\u7a76\uff0c\u672c\u6587\u77e5\u540d\u673a\u5668\u5b66\u4e60\u4e0e AI \u7814\u7a76\u8005 Sebastian Raschka \u5411\u6211\u4eec\u4ecb\u7ecd\u4e86 Gemma \u76f8\u6bd4\u4e8e\u5176\u4ed6 LLM \u7684\u4e00\u4e9b\u72ec\u7279\u8bbe\u8ba1\u539f\u5219\u3002<\/p>\n<p>Raschka \u9996\u5148\u4ece\u6a21\u578b\u6027\u80fd\u5c55\u5f00\uff0c\u4ed6\u8868\u793a\u770b\u8fc7\u6280\u672f\u62a5\u544a\u7684\u5c0f\u4f19\u4f34\u53ef\u80fd\u90fd\u6709\u4e00\u4e2a\u7591\u95ee\uff0c\u662f\u4ec0\u4e48\u8ba9 Gemma \u8868\u73b0\u5982\u6b64\u51fa\u8272\uff1f\u8bba\u6587\u4e2d\u6ca1\u6709\u660e\u786e\u8bf4\u660e\u539f\u56e0\uff0cSebastian Raschka \u8ba4\u4e3a\u53ef\u4ee5\u4ece\u4e0b\u9762\u4e24\u70b9\u5f97\u51fa\u7ed3\u8bba\uff1a<\/p>\n<ul>\n<li>\u9996\u5148\u662f\u8bcd\u6c47\u91cf\u5927\uff0cGemma \u8bcd\u6c47\u91cf\u8fbe\u5230 256000 \u4e2a\u5355\u8bcd\uff0c\u76f8\u6bd4\u4e4b\u4e0b\uff0cLlama \u7684\u8bcd\u6c47\u91cf\u4e3a 32000 \u4e2a\u5355\u8bcd\uff1b<\/li>\n<li>\u5176\u6b21\u662f\u8bad\u7ec3\u6570\u636e\u96c6\u8fbe 6 \u4e07\u4ebf token\uff0c\u4f5c\u4e3a\u5bf9\u6bd4\uff0cLlama \u4ec5\u63a5\u53d7\u4e86\u5176\u4e2d\u4e09\u5206\u4e4b\u4e00\u7684\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<p>\u5728\u67b6\u6784\u65b9\u9762\uff0cRaschka \u5217\u4e3e\u4e86 Gemma \u4e0e LLama 2 7B \u548c OLMo 7B \u7684\u67b6\u6784\u6982\u89c8\u3002<\/p>\n<p>\u5728\u6a21\u578b\u5927\u5c0f\u4e0a\uff0cRaschka \u8868\u793a Gemma 2B \u6709\u591a\u67e5\u8be2\u6ce8\u610f\u529b\uff0c\u800c Gemma 7B \u6ca1\u6709\u3002\u53e6\u5916\uff0c\u4e0e Llama 2 \u76f8\u6bd4\uff0cGemma 7B \u5177\u6709\u76f8\u5bf9\u8f83\u5927\u7684\u524d\u9988\u5c42\uff0c\u5c3d\u7ba1\u5176\u5c42\u6570\u8f83\u5c11\uff0828 VS 32\uff09\uff0c\u4f46 Gemma \u4e2d\u7684\u53c2\u6570\u6570\u91cf\u5374\u76f8\u5f53\u5927\u3002<\/p>\n<p>Raschka \u731c\u6d4b Gemma 7B \u5b9e\u9645\u4e0a\u603b\u5171\u6709 93 \u4ebf\u4e2a\u53c2\u6570\uff0c\u5982\u679c\u8003\u8651\u5230\u6743\u91cd\u5171\u4eab\uff08Weight tying\uff09\u7684\u8bdd\uff0c\u5219\u6709 85 \u4ebf\u4e2a\u53c2\u6570\u3002\u6743\u91cd\u5171\u4eab\u610f\u5473\u7740\u6a21\u578b\u5728\u8f93\u5165\u5d4c\u5165\u548c\u8f93\u51fa\u6295\u5f71\u5c42\u4e2d\u5171\u4eab\u76f8\u540c\u7684\u6743\u91cd\uff0c\u7c7b\u4f3c\u4e8e GPT-2 \u548c OLMo 1B\uff08OLMO 7B \u7684\u8bad\u7ec3\u6ca1\u6709\u6743\u91cd\u5171\u4eab\uff09\u3002<\/p>\n<h4>\u5f52\u4e00\u5316\u5c42<\/h4>\n<p>\u53e6\u4e00\u4e2a\u5f15\u4eba\u6ce8\u76ee\u7684\u7ec6\u8282\u662f\u4ee5\u4e0b\u51fa\u81ea Gemma \u8bba\u6587\u4e2d\u7684\u6bb5\u843d\u3002<\/p>\n<p>\u5f52\u4e00\u5316\u4f4d\u7f6e\u3002\u8c37\u6b4c\u5bf9\u6bcf\u4e2a transformer \u5b50\u5c42\u7684\u8f93\u5165\u548c\u8f93\u51fa\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u8fd9\u4e0e\u5355\u72ec\u5f52\u4e00\u5316\u8f93\u5165\u6216\u8f93\u51fa\u7684\u6807\u51c6\u505a\u6cd5\u4e0d\u540c\u3002\u8c37\u6b4c\u4f7f\u7528 RMSNorm \u4f5c\u4e3a\u5f52\u4e00\u5316\u5c42\u3002<\/p>\n<p>\u4e4d\u4e00\u770b\uff0c\u770b\u8d77\u6765\u50cf Gemma \u5728\u6bcf\u4e2a transformer \u5757\u4e4b\u540e\u90fd\u6709\u4e00\u4e2a\u989d\u5916\u7684 RMSNorm \u5c42\u3002\u4f46\u662f\uff0c\u901a\u8fc7\u67e5\u770b\u300ckeras-nlp\u300d\u9879\u76ee\u7684\u5b98\u65b9\u4ee3\u7801\u5b9e\u73b0\uff0c\u539f\u6765 Gemma \u4ec5\u4ec5\u4f7f\u7528\u4e86 GPT-2\u3001Llama 2 \u7b49\u5176\u4ed6 LLM \u4f7f\u7528\u7684\u5e38\u89c4\u9884\u5f52\u4e00\u5316\u65b9\u6848\uff0c\u5177\u4f53\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<p>GPT\u3001Llama 2 \u548c\u5176\u4ed6 LLM \u4e2d\u5178\u578b\u7684\u5c42\u5f52\u4e00\u5316\u4f4d\u7f6e\uff0cGemma \u4e2d\u6ca1\u6709\u4ec0\u4e48\u65b0\u4e1c\u897f\u3002\u6765\u6e90\uff1ahttps:\/\/github.com\/rasbt\/LLMs-from-scratch<\/p>\n<h4>GeGLU \u6fc0\u6d3b<\/h4>\n<p>Gemma \u4e0e\u5176\u4ed6\u67b6\u6784\u4e4b\u95f4\u7684\u4e00\u5927\u533a\u522b\u662f\u5b83\u4f7f\u7528\u4e86 GeGLU \u6fc0\u6d3b\uff0c\u800c GeGLU \u6fc0\u6d3b\u662f\u5728 2020 \u5e74\u7684\u8c37\u6b4c\u8bba\u6587\u300aGLU Variants Improve Transformer\u300b\u4e2d\u63d0\u51fa\u7684\u3002<\/p>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2002.05202.pdf<\/p>\n<p>GeLU \u5168\u79f0\u4e3a\u9ad8\u65af\u8bef\u5dee\u7ebf\u6027\u5355\u5143\uff08Gaussian Error Linear Unit\uff09\uff0c\u5b83\u662f\u4e00\u4e2a\u6fc0\u6d3b\u51fd\u6570\uff0c\u8d8a\u6765\u8d8a\u591a\u5730\u88ab\u4f5c\u4e3a\u4f20\u7edf ReLU \u7684\u66ff\u4ee3\u65b9\u6848\u3002GeLU \u7684\u6d41\u884c\u5f97\u76ca\u4e8e\u5b83\u6709\u80fd\u529b\u5f15\u5165\u975e\u7ebf\u6027\u7279\u5f81\uff0c\u5e76\u5141\u8bb8\u4e3a\u8d1f\u8f93\u5165\u503c\u6267\u884c\u68af\u5ea6\u4f20\u64ad\uff0c\u8fd9\u89e3\u51b3\u4e86 ReLU \u7684\u4e00\u5927\u5c40\u9650\uff0c\u5b8c\u5168\u963b\u65ad\u4e86\u8d1f\u503c\u3002<\/p>\n<p>\u73b0\u5728\uff0c\u4f5c\u4e3a GeLU \u7684\u95e8\u7ebf\u6027\u5355\u5143\u53d8\u4f53\uff0cGeGLU \u7684\u6fc0\u6d3b\u88ab\u5206\u5272\u4e3a\u4e24\u90e8\u5206\uff0c\u5206\u522b\u662f sigmoid \u5355\u5143\u548c\u7ebf\u6027\u6620\u5c04\u5355\u5143\uff08\u5b83\u4e0e sigmoid \u5355\u5143\u7684\u8f93\u51fa\u9010\u5143\u7d20\u76f8\u4e58\uff09\uff0c\u5177\u4f53\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<p>GeLU \u4e0e ReLU \u6fc0\u6d3b\u51fd\u6570\u56fe\u793a\u6bd4\u8f83\uff0c\u6765\u6e90\uff1ahttps:\/\/github.com\/rasbt\/LLMs-from-scratch<\/p>\n<p>\u540c\u65f6\uff0cGeGLU \u4e0e Llama 2\u3001Mistral \u7b49\u5176\u4ed6 LLM \u4f7f\u7528\u7684 SwiGLU \u6fc0\u6d3b\u7c7b\u4f3c\u3002\u552f\u4e00\u7684\u533a\u522b\u662f GeGLU \u4f7f\u7528\u7684\u57fa\u7840\u6fc0\u6d3b\u662f GeLU \u800c\u4e0d\u662f Swish\u3002<\/p>\n<p>\u4e0b\u56fe\u5c55\u793a\u4e86 GeLU\uff08GPT-2\uff09\u3001SwiGLU\uff08Llama 2\uff09\u548c GeGLU\uff08Gemma\uff09\u7684\u4f2a\u4ee3\u7801\u3002<\/p>\n<p>\u9700\u8981\u6ce8\u610f\uff0c\u4e0e\u4f7f\u7528 GeLU\uff08\u4ec5\u7ebf\u6027\uff09\u7684\u5e38\u89c4\u524d\u9988\u6a21\u5757\u76f8\u6bd4\uff0c\u4f7f\u7528 SwiGLU \u548c GeGLU \u7684\u524d\u9988\u6a21\u5757\u5404\u591a\u4e86\u4e00\u4e2a\u7ebf\u6027\u5c42\uff08\u5206\u522b\u662f linear_1 \u548c linear_2\uff09\u3002\u4e0d\u8fc7\uff0c\u5728 SwiGLU \u548c GeGLU \u524d\u9988\u6a21\u5757\u4e2d\uff0clinear_1 \u548c linear_2 \u901a\u5e38\u901a\u8fc7\u5c06\u5355\u4e2a\u7ebf\u6027\u5c42\u5206\u5272\u4e3a\u4e24\u90e8\u5206\u800c\u83b7\u5f97\uff0c\u56e0\u6b64\u4e0d\u4f1a\u589e\u52a0\u53c2\u6570\u89c4\u6a21\u3002<\/p>\n<p>\u90a3\u662f\u5426 GeGLU \u5c31\u6bd4 SwiGLU \u5f3a\u5462\uff1f\u5e76\u6ca1\u6709\u6d88\u878d\u5b9e\u9a8c\u6765\u8bc1\u5b9e\u8fd9\u4e00\u70b9\u3002Raschka \u731c\u6d4b\u8c37\u6b4c\u9009\u62e9\u4f7f\u7528 GeGLU\uff0c\u53ea\u662f\u4e3a\u4e86\u8ba9 Gemma \u4e0e Llama 2 \u7565\u6709\u4e0d\u540c\u3002<\/p>\n<p>\u4e3e\u4f8b\u800c\u8a00\uff0cGemma \u4e3a RMSNorm \u5c42\u6dfb\u52a0\u4e86 + 1 \u7684\u504f\u79fb\u91cf\uff0c\u5e76\u901a\u8fc7\u9690\u85cf\u5c42\u7ef4\u6570\u7684\u5f00\u7acb\u65b9\u6839\u6765\u5f52\u4e00\u5316\u5d4c\u5165\u3002Gemma \u8bba\u6587\u4e2d\u6ca1\u6709\u63d0\u53ca\u6216\u8ba8\u8bba\u8fd9\u4e9b\u7ec6\u8282\uff0c\u6240\u4ee5\u5b83\u4eec\u7684\u91cd\u8981\u6027\u4e5f\u4e0d\u6e05\u695a\u3002<\/p>\n<h4>\u7ed3\u8bba<\/h4>\n<p>\u5bf9\u4e8e\u5f00\u6e90 LLM \u800c\u8a00\uff0cGemma \u505a\u51fa\u4e86\u975e\u5e38\u68d2\u7684\u8d21\u732e\uff0c\u5c55\u793a\u4e86 7B \u53c2\u6570\u89c4\u6a21\u4e5f\u80fd\u6210\u5c31\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u5e76\u6709\u6f5c\u529b\u5728\u771f\u5b9e\u4e16\u754c\u7684\u7528\u4f8b\u4e2d\u53d6\u4ee3 Llama 2 \u548c Mistral\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u76ee\u524d 7B \u5927\u5c0f\u89c4\u6a21\u7684\u5f00\u6e90\u6a21\u578b\u5df2\u7ecf\u6709\u5f88\u591a\u4e86\uff0c\u56e0\u6b64 Gemma 2B \u66f4\u52a0\u6709\u8da3\uff0c\u5b83\u53ef\u4ee5\u8f7b\u677e\u5730\u5728\u5355\u4e2a GPU \u4e0a\u8fd0\u884c\u3002\u5f53\u7136\uff0cGemma 2B \u4e0e 2.7B \u5927\u5c0f\u7684 phi-2 \u4e4b\u95f4\u7684\u5bf9\u6bd4\u4e5f\u5c06\u4f1a\u5f88\u6709\u8da3\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5c31\u5728\u51e0\u5929\u524d\uff0c\u5f00\u6e90\u5927\u6a21\u578b\u9886\u57df\u8fce\u6765\u4e86\u91cd\u78c5\u65b0\u73a9\u5bb6\uff1a\u8c37\u6b4c\u63a8\u51fa\u4e86\u5168\u65b0\u7684\u5f00\u6e90\u6a21\u578b\u7cfb\u5217\u300cGemma\u300d\u3002\u76f8\u6bd4 Gemini\uff0cG [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7023],"tags":[],"class_list":["post-415304","post","type-post","status-publish","format-standard","hentry","category-chatgpt"],"_links":{"self":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415304","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/comments?post=415304"}],"version-history":[{"count":0,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415304\/revisions"}],"wp:attachment":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/media?parent=415304"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/categories?post=415304"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/tags?post=415304"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}