{"id":154,"date":"2024-07-23T15:17:38","date_gmt":"2024-07-23T07:17:38","guid":{"rendered":"https:\/\/www.onetts.com\/ai\/?p=154"},"modified":"2024-07-25T21:17:29","modified_gmt":"2024-07-25T13:17:29","slug":"deepseek-v2","status":"publish","type":"post","link":"https:\/\/www.onetts.com\/ai\/models\/deepseek-v2\/","title":{"rendered":"DeepSeek-V2"},"content":{"rendered":"<p>DeepSeek-V2 \u662f\u7531 DeepSeek-AI \u5f00\u53d1\u7684\u4e00\u6b3e\u5148\u8fdb\u7684\u57fa\u4e8e Mixture-of-Experts (MoE) \u7684\u6df7\u5408\u4e13\u5bb6\u8bed\u8a00\u6a21\u578b\uff0c\u62e5\u6709 236B\uff082.36 \u4e07\u4ebf\uff09\u7684\u603b\u53c2\u6570\u91cf\uff0c\u5176\u4e2d\u6bcf\u6b21\u6fc0\u6d3b 21B\uff082100 \u4ebf\uff09\u53c2\u6570\u3002\u652f\u6301\u957f\u8fbe 128K \u4e2a token \u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u80fd\u591f\u5904\u7406\u590d\u6742\u7684\u8bed\u8a00\u4efb\u52a1\u3002<\/p>\n<p>\u4e0e\u524d\u4ee3\u6a21\u578b DeepSeek 67B \u76f8\u6bd4\uff0cDeepSeek-V2 \u5728\u6027\u80fd\u4e0a\u6709\u4e86\u663e\u8457\u63d0\u5347\uff0c\u540c\u65f6\u5728\u8bad\u7ec3\u6210\u672c\u4e0a\u8282\u7701\u4e86 42.5%\uff0c\u51cf\u5c11\u4e86 93.3% \u7684\u952e\u503c\uff08KV\uff09\u7f13\u5b58\uff0c\u5e76\u63d0\u9ad8\u4e86\u6700\u5927\u751f\u6210\u541e\u5410\u91cf\u81f3 5.76 \u500d\u3002<\/p>\n<p>DeepSeek-V2 \u5728\u4e00\u4e2a\u5305\u542b 8.1 \u4e07\u4ebf tokens \u7684\u591a\u6837\u5316\u548c\u9ad8\u8d28\u91cf\u8bed\u6599\u5e93\u4e0a\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\uff0c\u5e76\u901a\u8fc7\u76d1\u7763\u5f0f\u5fae\u8c03\uff08Supervised Fine-Tuning\uff0c\u7b80\u79f0 SFT\uff09\u548c\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning\uff0c\u7b80\u79f0 RL\uff09\u8fdb\u4e00\u6b65\u4f18\u5316\u4e86\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n<p>DeepSeek-V2 \u91c7\u7528\u4e86\u521b\u65b0\u7684\u67b6\u6784\u8bbe\u8ba1\uff0c\u5305\u62ec Multi-head Latent Attention (MLA) \u548c DeepSeekMoE\uff0c\u8fd9\u4e9b\u6280\u672f\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u7684\u63a8\u7406\u6548\u7387\u548c\u8bad\u7ec3\u7ecf\u6d4e\u6027\u3002<\/p>\n<ul>\n<li><strong>Multi-head Latent Attention (MLA)<\/strong>\uff1a\u901a\u8fc7\u4f4e\u79e9\u952e\u503c\u8054\u5408\u538b\u7f29\u6280\u672f\uff0cMLA \u51cf\u5c11\u4e86\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u952e\u503c\u7f13\u5b58\u9700\u6c42\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u63a8\u7406\u6548\u7387\u3002<\/li>\n<li><strong>DeepSeekMoE<\/strong>\uff1a\u901a\u8fc7\u7a00\u758f\u8ba1\u7b97\uff0cDeepSeekMoE \u4f7f\u5f97\u5728\u7ecf\u6d4e\u6210\u672c\u4e0b\u8bad\u7ec3\u51fa\u5f3a\u5927\u7684\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002<\/li>\n<\/ul>\n<h2>\u6a21\u578b\u8bc4\u6d4b<\/h2>\n<p>DeepSeek-V2 \u5728\u591a\u4e2a\u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\uff1a<\/p>\n<ul>\n<li>MMLU\uff1a\u5728\u82f1\u8bed\u9886\u57df\uff0cDeepSeek-V2 \u7684\u5f97\u5206\u4e3a 78.5\uff0c\u8d85\u8fc7\u4e86 LLaMA3 70B \u548c Mixtral 8x22B\u3002<\/li>\n<li>BBH\uff1a\u5728\u82f1\u8bed\u9886\u57df\uff0cDeepSeek-V2 \u7684\u5f97\u5206\u4e3a 78.9\uff0c\u540c\u6837\u8868\u73b0\u4f18\u5f02\u3002<\/li>\n<li>C-Eval\uff1a\u5728\u4e2d\u6587\u9886\u57df\uff0cDeepSeek-V2 \u7684\u5f97\u5206\u4e3a 81.7\uff0c\u663e\u8457\u9ad8\u4e8e\u5176\u4ed6\u6a21\u578b\u3002<\/li>\n<li>CMMLU\uff1a\u5728\u4e2d\u6587\u9886\u57df\uff0cDeepSeek-V2 \u7684\u5f97\u5206\u4e3a 84.0\uff0c\u7ee7\u7eed\u4fdd\u6301\u9886\u5148\u3002<\/li>\n<li>HumanEval\uff1a\u5728\u4ee3\u7801\u751f\u6210\u9886\u57df\uff0cDeepSeek-V2 \u7684\u5f97\u5206\u4e3a 48.8\uff0c\u663e\u793a\u51fa\u5176\u5728\u7f16\u7a0b\u8bed\u8a00\u5904\u7406\u4e0a\u7684\u80fd\u529b\u3002<\/li>\n<li>MBPP\uff1a\u5728\u4ee3\u7801\u751f\u6210\u9886\u57df\uff0cDeepSeek-V2 \u7684\u5f97\u5206\u4e3a 66.6\uff0c\u8868\u73b0\u4f18\u5f02\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0cDeepSeek-V2 \u5728 &#8220;Needle In A Haystack&#8221;\uff08NIAH\uff09\u6d4b\u8bd5\u4e2d\u4e5f\u8868\u73b0\u51fa\u8272\uff0c\u5c24\u5176\u662f\u5728 128K \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u957f\u5ea6\u4e0a\u3002<\/p>\n<h2>\u90e8\u7f72\u4f7f\u7528<\/h2>\n<p>\u90e8\u7f72 DeepSeek-V2 \u6d89\u53ca\u4ee5\u4e0b\u6b65\u9aa4\uff1a<\/p>\n<ol>\n<li><strong>\u73af\u5883\u51c6\u5907<\/strong>\uff1a\u786e\u4fdd\u6709\u8db3\u591f\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u5982\u9ad8\u6027\u80fd GPU\u3002<\/li>\n<li><strong>\u6a21\u578b\u4e0b\u8f7d<\/strong>\uff1a\u4ece\u00a0<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V2\" target=\"_blank\" rel=\"noopener\">DeepSeek-V2 GitHub<\/a>\u00a0\u4e0b\u8f7d\u6a21\u578b\u68c0\u67e5\u70b9\u3002<\/li>\n<li><strong>\u4f9d\u8d56\u5b89\u88c5<\/strong>\uff1a\u5b89\u88c5\u5fc5\u8981\u7684\u5e93\u548c\u6846\u67b6\uff0c\u5982 PyTorch \u6216 TensorFlow\u3002<\/li>\n<li><strong>\u6a21\u578b\u52a0\u8f7d<\/strong>\uff1a\u52a0\u8f7d\u6a21\u578b\u53c2\u6570\uff0c\u5e76\u8fdb\u884c\u5fc5\u8981\u7684\u9884\u5904\u7406\u3002<\/li>\n<li><strong>\u63a8\u7406\u6267\u884c<\/strong>\uff1a\u901a\u8fc7\u6a21\u578b\u6267\u884c\u63a8\u7406\u4efb\u52a1\uff0c\u5982\u6587\u672c\u751f\u6210\u6216\u8bed\u8a00\u7406\u89e3\u3002<\/li>\n<li><strong>\u6027\u80fd\u4f18\u5316<\/strong>\uff1a\u6839\u636e\u5b9e\u9645\u9700\u6c42\u8c03\u6574\u6a21\u578b\u53c2\u6570\u548c\u786c\u4ef6\u8d44\u6e90\uff0c\u4ee5\u4f18\u5316\u6027\u80fd\u3002<\/li>\n<\/ol>\n<h2>\u5e38\u89c1\u95ee\u9898<\/h2>\n<p><strong>Q: \u5982\u4f55\u4f18\u5316 DeepSeek-V2 \u7684\u63a8\u7406\u6548\u7387\uff1f<\/strong><\/p>\n<p>A: \u53ef\u4ee5\u901a\u8fc7\u51cf\u5c11\u6a21\u578b\u7684\u53c2\u6570\u91cf\u3001\u4f7f\u7528\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\uff08\u5982 FP8\uff09\u548c\u5b9e\u65bd\u952e\u503c\u7f13\u5b58\u91cf\u5316\u6765\u4f18\u5316\u63a8\u7406\u6548\u7387\u3002<\/p>\n<p><strong>Q: DeepSeek-V2 \u662f\u5426\u652f\u6301\u591a\u8bed\u8a00\uff1f<\/strong><\/p>\n<p>A: DeepSeek-V2 \u4e3b\u8981\u9488\u5bf9\u4e2d\u6587\u548c\u82f1\u6587\u8fdb\u884c\u4e86\u4f18\u5316\uff0c\u4f46\u5728\u5176\u4ed6\u8bed\u8a00\u4e0a\u53ef\u80fd\u8868\u73b0\u6709\u9650\u3002<\/p>\n<p><strong>Q: \u5982\u4f55\u5904\u7406\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u4e0d\u5e73\u8861\u8d1f\u8f7d\u95ee\u9898\uff1f<\/strong><\/p>\n<p>A: \u53ef\u4ee5\u901a\u8fc7\u8f85\u52a9\u635f\u5931\u51fd\u6570\uff08\u5982\u4e13\u5bb6\u7ea7\u5e73\u8861\u635f\u5931\u3001\u8bbe\u5907\u7ea7\u5e73\u8861\u635f\u5931\u548c\u901a\u4fe1\u5e73\u8861\u635f\u5931\uff09\u6765\u81ea\u52a8\u5b66\u4e60\u8def\u7531\u7b56\u7565\uff0c\u786e\u4fdd\u8d1f\u8f7d\u5e73\u8861\u3002<\/p>\n<h2>\u76f8\u5173\u8d44\u6e90<\/h2>\n<ul>\n<li><strong>\u6a21\u578b\u4e0b\u8f7d<\/strong>\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/www.modelscope.cn\/models\/deepseek-ai\/DeepSeek-V2\" target=\"_blank\" rel=\"noopener\">DeepSeek-V2 \u6a21\u578b\u4e0b\u8f7d<\/a><\/li>\n<li><strong>\u5b98\u65b9\u6587\u6863<\/strong>\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V2\" target=\"_blank\" rel=\"noopener\">DeepSeek-V2 \u5b98\u65b9\u6587\u6863<\/a><\/li>\n<li><strong>API \u5e73\u53f0<\/strong>\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/platform.deepseek.com\/\" target=\"_blank\" rel=\"noopener\">DeepSeek Platform<\/a><\/li>\n<li><strong>\u804a\u5929\u7f51\u7ad9<\/strong>\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/chat.deepseek.com\/\" target=\"_blank\" rel=\"noopener\">DeepSeek Chat<\/a><\/li>\n<li><strong>\u8bb8\u53ef\u8bc1<\/strong>\uff1aDeepSeek-V2 \u7684\u4ee3\u7801\u5e93\u91c7\u7528 MIT \u8bb8\u53ef\u8bc1\uff0c\u6a21\u578b\u7684\u4f7f\u7528\u53d7\u6a21\u578b\u8bb8\u53ef\u8bc1\u7684\u7ea6\u675f\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>DeepSeek-V2 \u662f\u7531 DeepSeek-AI \u5f00\u53d1\u7684\u4e00\u6b3e\u5148\u8fdb\u7684\u57fa\u4e8e Mixture-of-Experts (MoE) \u7684\u6df7\u5408\u4e13\u5bb6\u8bed\u8a00\u6a21\u578b\uff0c\u62e5\u6709 236B\uff082.36 \u4e07\u4ebf\uff09\u7684\u603b\u53c2\u6570\u91cf\uff0c\u5176\u4e2d\u6bcf\u6b21\u6fc0\u6d3b 21B\uff082100 \u4ebf\uff09\u53c2\u6570\u3002\u652f\u6301\u957f\u8fbe 128K \u4e2a token \u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u80fd\u591f\u5904\u7406\u590d\u6742\u7684\u8bed\u8a00\u4efb\u52a1\u3002 \u4e0e\u524d\u4ee3\u6a21\u578b DeepSeek 67B \u76f8\u6bd4\uff0cDeepSeek-V2 \u5728\u6027\u80fd\u4e0a\u6709\u4e86\u663e\u8457\u63d0\u5347\uff0c\u540c\u65f6\u5728\u8bad\u7ec3\u6210\u672c\u4e0a\u8282\u7701\u4e86 42.5%\uff0c\u51cf\u5c11\u4e86 93.3% \u7684\u952e\u503c\uff08KV\uff09\u7f13\u5b58\uff0c\u5e76\u63d0\u9ad8\u4e86\u6700\u5927\u751f\u6210\u541e\u5410\u91cf\u81f3 5.76 \u500d\u3002 De<\/p>\n","protected":false},"author":1,"featured_media":155,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[43],"collection":[],"company":[24],"rank":[],"class_list":["post-154","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-llm","tag-moe-llm","company-deepseek"],"_links":{"self":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/154","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=154"}],"version-history":[{"count":3,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/154\/revisions"}],"predecessor-version":[{"id":158,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/154\/revisions\/158"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/media\/155"}],"wp:attachment":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/media?parent=154"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/categories?post=154"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/tags?post=154"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/collection?post=154"},{"taxonomy":"company","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/company?post=154"},{"taxonomy":"rank","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/rank?post=154"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}