{"id":88,"date":"2024-07-21T18:06:27","date_gmt":"2024-07-21T10:06:27","guid":{"rendered":"https:\/\/www.onetts.com\/ai\/?p=88"},"modified":"2024-07-21T20:09:36","modified_gmt":"2024-07-21T12:09:36","slug":"qwen2-57b-a14b","status":"publish","type":"post","link":"https:\/\/www.onetts.com\/ai\/models\/qwen2-57b-a14b\/","title":{"rendered":"Qwen2-57B-A14B"},"content":{"rendered":"<p data-spm-anchor-id=\"a2c6h.13066512.0.i4.763536afrXJ4Mz\">Qwen2-57B-A14B \u662f Qwen2 \u7cfb\u5217\u4e2d\u7684\u4e00\u6b3e\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u5c5e\u4e8e Mixture-of-Experts\uff08MoE\uff09\u6a21\u578b\u3002\u8be5\u6a21\u578b\u62e5\u6709 57 \u4ebf\u53c2\u6570\uff0c\u5176\u4e2d\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u65f6\u6fc0\u6d3b\u7684\u53c2\u6570\u4e3a 14 \u4ebf\u3002Qwen2-57B-A14B \u57fa\u4e8e Transformer \u67b6\u6784\uff0c\u91c7\u7528 SwiGLU \u6fc0\u6d3b\u3001\u6ce8\u610f\u529b QKV \u504f\u7f6e\u3001\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u7b49\u6280\u672f\u3002\u6b64\u5916\uff0c\u5b83\u8fd8\u914d\u5907\u4e86\u4e00\u4e2a\u6539\u8fdb\u7684\u5206\u8bcd\u5668\uff0c\u80fd\u591f\u9002\u5e94\u591a\u79cd\u81ea\u7136\u8bed\u8a00\u548c\u4ee3\u7801\u3002<\/p>\n<h2 id=\"-\">\u6a21\u578b\u8bc4\u6d4b<\/h2>\n<p>Qwen2-57B-A14B \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u6db5\u76d6\u4e86\u8bed\u8a00\u7406\u89e3\u3001\u8bed\u8a00\u751f\u6210\u3001\u7f16\u7a0b\u3001\u6570\u5b66\u3001\u79d1\u5b66\u77e5\u8bc6\u3001\u63a8\u7406\u548c\u591a\u8bed\u8a00\u80fd\u529b\u7b49\u65b9\u9762\u3002\u4ee5\u4e0b\u662f\u4e00\u4e9b\u5173\u952e\u7684\u8bc4\u6d4b\u6570\u636e\uff1a<\/p>\n<ul>\n<li><strong>\u82f1\u8bed\u4efb\u52a1<\/strong>\uff1aMMLU\uff085-shot\uff09\u3001MMLU-Pro\uff085-shot\uff09\u3001GPQA\uff085-shot\uff09\u3001\u5b9a\u7406 QA\uff085-shot\uff09\u3001BBH\uff083-shot\uff09\u3001HellaSwag\uff0810-shot\uff09\u3001Winogrande\uff085-shot\uff09\u3001TruthfulQA\uff080-shot\uff09\u3001ARC-C\uff0825-shot\uff09<\/li>\n<li><strong>\u7f16\u7a0b\u4efb\u52a1<\/strong>\uff1aEvalPlus\uff080-shot\uff09\uff08HumanEval, MBPP, HumanEval+, MBPP+\uff09\u3001MultiPL-E\uff080-shot\uff09\uff08Python, C++, JAVA, PHP, TypeScript, C#, Bash, JavaScript\uff09<\/li>\n<li><strong>\u6570\u5b66\u4efb\u52a1<\/strong>\uff1aGSM8K\uff084-shot\uff09\u3001MATH\uff084-shot\uff09<\/li>\n<li><strong>\u4e2d\u6587\u4efb\u52a1<\/strong>\uff1aC-Eval\uff085-shot\uff09\u3001CMMLU\uff085-shot\uff09<\/li>\n<li><strong>\u591a\u8bed\u8a00\u4efb\u52a1<\/strong>\uff1aMulti-Exam\uff08M3Exam 5-shot, IndoMMLU 3-shot, ruMMLU 5-shot, mMMLU 5-shot\uff09\u3001Multi-Understanding\uff08BELEBELE 5-shot, XCOPA 5-shot, XWinograd 5-shot, XStoryCloze 0-shot, PAWS-X 5-shot\uff09\u3001Multi-Mathematics\uff08MGSM 8-shot\uff09\u3001Multi-Translation\uff08Flores-101 5-shot\uff09<\/li>\n<\/ul>\n<p>\u5728\u8fd9\u4e9b\u4efb\u52a1\u4e2d\uff0cQwen2-57B-A14B \u663e\u793a\u51fa\u6bd4\u5148\u524d\u53d1\u5e03\u7684 Qwen1.5 \u7b49\u6a21\u578b\u66f4\u9ad8\u7684\u6027\u80fd\u3002<\/p>\n<h2 id=\"-\">\u90e8\u7f72\u4f7f\u7528<\/h2>\n<h4 id=\"-\">\u90e8\u7f72\u6b65\u9aa4<\/h4>\n<ol>\n<li><strong>\u5b89\u88c5\u4f9d\u8d56<\/strong>\uff1a\u9996\u5148\u9700\u8981\u5b89\u88c5 Hugging Face \u7684\u00a0<code>transformers<\/code>\u00a0\u5e93\uff0c\u5efa\u8bae\u7248\u672c\u4e3a\u00a0<code>transformers&gt;=4.40.0<\/code>\u3002\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u547d\u4ee4\u5b89\u88c5\uff1a\n<pre><code class=\"lang-bash\">pip <span class=\"token function\">install<\/span> transformers\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u52a0\u8f7d\u6a21\u578b<\/strong>\uff1a\u4f7f\u7528\u00a0<code>transformers<\/code>\u00a0\u5e93\u52a0\u8f7d Qwen2-57B-A14B \u6a21\u578b\uff1a\n<pre><code class=\"lang-python\"><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> AutoModelForSeq2SeqLM\r\n\r\nmodel <span class=\"token operator\">=<\/span> AutoModelForSeq2SeqLM<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span><span class=\"token string\">\"qwen2-57b-a14b\"<\/span><span class=\"token punctuation\">)<\/span>\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5206\u8bcd\u5668<\/strong>\uff1a\u52a0\u8f7d\u76f8\u5e94\u7684\u5206\u8bcd\u5668\u4ee5\u5904\u7406\u8f93\u5165\u6587\u672c\uff1a\n<pre><code class=\"lang-python\"><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> AutoTokenizer\r\n\r\ntokenizer <span class=\"token operator\">=<\/span> AutoTokenizer<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span><span class=\"token string\">\"qwen2-57b-a14b\"<\/span><span class=\"token punctuation\">)<\/span>\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u6587\u672c\u751f\u6210<\/strong>\uff1a\u4f7f\u7528\u6a21\u578b\u548c\u5206\u8bcd\u5668\u751f\u6210\u6587\u672c\uff1a\n<pre><code class=\"lang-python\">input_text <span class=\"token operator\">=<\/span> <span class=\"token string\">\"Hello, what is your name?\"<\/span>\r\ninputs <span class=\"token operator\">=<\/span> tokenizer<span class=\"token punctuation\">(<\/span>input_text<span class=\"token punctuation\">,<\/span> return_tensors<span class=\"token operator\">=<\/span><span class=\"token string\">\"pt\"<\/span><span class=\"token punctuation\">)<\/span>\r\noutputs <span class=\"token operator\">=<\/span> model<span class=\"token punctuation\">.<\/span>generate<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>inputs<span class=\"token punctuation\">)<\/span>\r\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>tokenizer<span class=\"token punctuation\">.<\/span>decode<span class=\"token punctuation\">(<\/span>outputs<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> skip_special_tokens<span class=\"token operator\">=<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span>\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u6a21\u578b\u5fae\u8c03<\/strong>\uff1a\u6839\u636e\u7279\u5b9a\u4efb\u52a1\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u4f8b\u5982\u4f7f\u7528 SFT\uff08Supervised Fine-Tuning\uff09\u3001RLHF\uff08Reinforcement Learning from Human Feedback\uff09\u6216\u7ee7\u7eed\u9884\u8bad\u7ec3\u3002<\/li>\n<\/ol>\n<h2 id=\"-\">\u5e38\u89c1\u95ee\u9898<\/h2>\n<ul>\n<li><strong>Q: \u5982\u4f55\u5904\u7406\u591a\u8bed\u8a00\u8f93\u5165\uff1f<\/strong>\n<ul>\n<li>A: \u4f7f\u7528 Qwen2-57B-A14B \u7684\u6539\u8fdb\u5206\u8bcd\u5668\uff0c\u53ef\u4ee5\u5904\u7406\u591a\u79cd\u81ea\u7136\u8bed\u8a00\u548c\u4ee3\u7801\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>Q: \u5982\u4f55\u4f18\u5316\u6a21\u578b\u6027\u80fd\uff1f<\/strong>\n<ul>\n<li>A: \u53ef\u4ee5\u901a\u8fc7\u5fae\u8c03\u6a21\u578b\uff08\u5982 SFT\u3001RLHF\uff09\u6216\u7ee7\u7eed\u9884\u8bad\u7ec3\u6765\u4f18\u5316\u6027\u80fd\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>Q: \u5982\u4f55\u5904\u7406\u6a21\u578b\u7684\u5185\u5b58\u5360\u7528\uff1f<\/strong>\n<ul>\n<li>A: MoE \u67b6\u6784\u5141\u8bb8\u5728\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u4e2d\u53ea\u6fc0\u6d3b\u90e8\u5206\u53c2\u6570\uff0c\u4ece\u800c\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h2 id=\"-\">\u76f8\u5173\u8d44\u6e90<\/h2>\n<ul>\n<li><strong>ModelScope \u6a21\u578b\u9875\u9762<\/strong>\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/modelscope.cn\/models\/qwen\/Qwen2-57B-A14B\">Qwen2-57B-A14B<\/a><\/li>\n<li><strong>GitHub \u4ed3\u5e93<\/strong>\uff1a\u53ef\u4ee5\u8bbf\u95ee Qwen2 \u7684 GitHub \u4ed3\u5e93\u83b7\u53d6\u66f4\u591a\u4ee3\u7801\u548c\u6587\u6863\u3002<\/li>\n<li><strong>\u6280\u672f\u535a\u5ba2<\/strong>\uff1aQwen2 \u7684\u6280\u672f\u535a\u5ba2\u63d0\u4f9b\u4e86\u8be6\u7ec6\u7684\u6a21\u578b\u4ecb\u7ecd\u548c\u4f7f\u7528\u6307\u5357\u3002<\/li>\n<\/ul>\n<h3 id=\"-\">\u5f15\u7528<\/h3>\n<p>\u5982\u679c\u60a8\u5728\u7814\u7a76\u4e2d\u4f7f\u7528\u4e86 Qwen2-57B-A14B \u6a21\u578b\uff0c\u8bf7\u5f15\u7528\u4ee5\u4e0b\u6587\u732e\uff1a<\/p>\n<pre><code>@article{qwen2,\r\n  title={Qwen2 Technical Report},\r\n  year={2024}\r\n}\r\n<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>Qwen2-57B-A14B \u662f Qwen2 \u7cfb\u5217\u4e2d\u7684\u4e00\u6b3e\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u5c5e\u4e8e Mixture-of-Experts\uff08MoE\uff09\u6a21\u578b\u3002\u8be5\u6a21\u578b\u62e5\u6709 57 \u4ebf\u53c2\u6570\uff0c\u5176\u4e2d\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u65f6\u6fc0\u6d3b\u7684\u53c2\u6570\u4e3a 14 \u4ebf\u3002Qwen2-57B-A14B \u57fa\u4e8e Transformer \u67b6\u6784\uff0c\u91c7\u7528 SwiGLU \u6fc0\u6d3b\u3001\u6ce8\u610f\u529b QKV \u504f\u7f6e\u3001\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u7b49\u6280\u672f\u3002\u6b64\u5916\uff0c\u5b83\u8fd8\u914d\u5907\u4e86\u4e00\u4e2a\u6539\u8fdb\u7684\u5206\u8bcd\u5668\uff0c\u80fd\u591f\u9002\u5e94\u591a\u79cd\u81ea\u7136\u8bed\u8a00\u548c\u4ee3\u7801\u3002 \u6a21\u578b\u8bc4\u6d4b Qwen2-57B-A14B \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u6db5\u76d6\u4e86\u8bed\u8a00\u7406\u89e3\u3001\u8bed\u8a00\u751f\u6210\u3001\u7f16\u7a0b\u3001\u6570\u5b66\u3001\u79d1\u5b66\u77e5\u8bc6\u3001\u63a8\u7406\u548c\u591a<\/p>\n","protected":false},"author":1,"featured_media":89,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[43],"collection":[44],"company":[6],"rank":[],"class_list":["post-88","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-llm","tag-moe-llm","collection-qwen2","company-alibaba"],"_links":{"self":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/88","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/comments?post=88"}],"version-history":[{"count":1,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/88\/revisions"}],"predecessor-version":[{"id":90,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/88\/revisions\/90"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/media\/89"}],"wp:attachment":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/media?parent=88"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/categories?post=88"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/tags?post=88"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/collection?post=88"},{"taxonomy":"company","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/company?post=88"},{"taxonomy":"rank","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/rank?post=88"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}