{"id":5463,"date":"2025-01-03T11:31:10","date_gmt":"2025-01-03T03:31:10","guid":{"rendered":"https:\/\/nullthought.net\/?p=5463"},"modified":"2025-02-24T21:38:02","modified_gmt":"2025-02-24T13:38:02","slug":"deepseek-v3%e6%8a%80%e6%9c%af%e6%8a%a5%e5%91%8a%e6%a6%82%e8%bf%b0","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5463","title":{"rendered":"DeepSeek-V3\u6280\u672f\u62a5\u544a\u6982\u8ff0"},"content":{"rendered":"\n<p><strong>\u4e00\u3001\u7b80\u4ecb<\/strong><\/p>\n\n\n\n<p>\u8fd1\u5e74\u6765\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u53d6\u5f97\u4e86\u98de\u901f\u53d1\u5c55\uff0c\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u8fc8\u8fdb\u7684\u6b65\u4f10\u8d8a\u6765\u8d8a\u5feb\u3002<strong><a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-V3<\/a><\/strong>\u4f5c\u4e3a\u4e00\u6b3e\u5927\u89c4\u6a21\u7684\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08MoE\uff09\uff0c\u5177\u67096710\u4ebf\u53c2\u6570\uff0c\u5176\u4e2d\u6bcf\u4e2a\u4ee4\u724c\u6fc0\u6d3b37\u4ebf\u53c2\u6570\u3002\u4e3a\u4e86\u5b9e\u73b0\u9ad8\u6548\u7684\u63a8\u7406\u4e0e\u7ecf\u6d4e\u7684\u8bad\u7ec3\uff0cDeepSeek-V3\u91c7\u7528\u4e86\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08Multi-head<br>Latent Attention, MLA\uff09\u548cDeepSeekMoE\u67b6\u6784\uff0c\u8fd9\u4e9b\u67b6\u6784\u7ecf\u8fc7<a href=\"https:\/\/arxiv.org\/abs\/2405.04434\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-V2<\/a>\u7684\u9a8c\u8bc1\uff0c\u5c55\u793a\u4e86\u5728\u9ad8\u6548\u63a8\u7406\u548c<strong>\u4f4e\u6210\u672c\u8bad\u7ec3<\/strong>\u65b9\u9762\u7684\u826f\u597d\u6548\u679c\u3002<\/p>\n\n\n\n<p>\u5728\u67b6\u6784\u8bbe\u8ba1\u4e0a\uff0cDeepSeek-V3\u521b\u65b0\u6027\u5730\u5f15\u5165\u4e86\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff0c\u4ee5\u53ca\u57fa\u4e8e\u591a\u4ee4\u724c\u9884\u6d4b\u7684\u8bad\u7ec3\u76ee\u6807\uff0c\u4ece\u800c\u5728\u591a\u4e2a\u6807\u51c6\u8bc4\u6d4b\u57fa\u51c6\u4e0a\u663e\u8457\u63d0\u9ad8\u4e86\u6a21\u578b\u6027\u80fd\u3002\u540c\u65f6\uff0c\u4e3a\u4e86\u51cf\u5c11\u8bad\u7ec3\u7684\u786c\u4ef6\u9700\u6c42\uff0cDeepSeek-V3\u5f15\u5165\u4e86FP8\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6\uff0c\u51cf\u5c11\u4e86GPU\u5185\u5b58\u4f7f\u7528\u91cf\uff0c\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n\n\n\n<p>DeepSeek-V3\u901a\u8fc7\u5bf914.8\u4e07\u4ebf\u9ad8\u8d28\u91cf\u7684\u591a\u6837\u5316\u6570\u636e\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u5e76\u5728\u6b64\u57fa\u7840\u4e0a\u8fdb\u884c\u4e86\u8d85\u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u3002\u901a\u8fc7\u540e\u7eed\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u9636\u6bb5\uff0c\u4f7f\u6a21\u578b\u5145\u5206\u53d1\u6325\u5176\u6f5c\u529b\u3002\u8bc4\u4f30\u7ed3\u679c\u8868\u660e\uff0cDeepSeek-V3\u5728\u591a\u4e2a\u5f00\u6e90\u548c\u5c01\u95ed\u6e90\u6a21\u578b\u4e2d\u90fd\u8868\u73b0\u51fa\u8272\uff0c\u5c24\u5176\u5728\u6570\u5b66\u3001\u4ee3\u7801\u548c\u63a8\u7406\u4efb\u52a1\u4e0a\u5177\u6709\u663e\u8457\u7684\u4f18\u52bf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"891\" height=\"296\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/01\/image-2.png\" alt=\"\" class=\"wp-image-5464\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/01\/image-2.png 891w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/01\/image-2-300x100.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/01\/image-2-768x255.png 768w\" sizes=\"auto, (max-width: 891px) 100vw, 891px\" \/><figcaption class=\"wp-element-caption\"><a href=\"https:\/\/www.deepseek.com\/\" target=\"_blank\" rel=\"noreferrer noopener\">deepseek.com<\/a><\/figcaption><\/figure>\n\n\n\n<p><strong>\u4e8c\u3001\u67b6\u6784<\/strong><\/p>\n\n\n\n<p>DeepSeek-V3\u7684\u67b6\u6784\u57fa\u4e8eTransformer\u6846\u67b6\uff0c\u5e76\u8fdb\u884c\u4e86\u591a\u9879\u4f18\u5316\uff0c\u4e3b\u8981\u5305\u62ec\u4ee5\u4e0b\u51e0\u90e8\u5206\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\uff1a<\/strong>DeepSeek-V3\u91c7\u7528\u4e86\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u67b6\u6784\uff0c\u76f8\u6bd4\u4e8e\u4f20\u7edf\u7684\u591a\u5934\u6ce8\u610f\u529b\uff08MHA\uff09\u673a\u5236\uff0cMLA\u901a\u8fc7\u4f4e\u79e9\u8054\u5408\u538b\u7f29\u663e\u8457\u51cf\u5c11\u4e86\u952e\u503c\u5bf9\uff08KV\uff09\u7f13\u5b58\u7684\u5185\u5b58\u5360\u7528\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u4e0e\u6807\u51c6MHA\u673a\u5236\u76f8\u5f53\u7684\u6027\u80fd\u3002\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0cMLA\u4f7f\u5f97\u7f13\u5b58\u7684\u952e\u503c\u5bf9\u5927\u5927\u51cf\u5c11\uff0c\u4ece\u800c\u663e\u8457\u51cf\u5c11\u4e86\u5185\u5b58\u4f7f\u7528\uff0c\u5e76\u4e14\u4fdd\u6301\u4e86\u8f83\u9ad8\u7684\u63a8\u7406\u6548\u7387\u3002\n<ul class=\"wp-block-list\">\n<li>\u901a\u8fc7\u4f4e\u79e9\u538b\u7f29\uff0cMLA\u51cf\u5c11\u4e86\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u5bf9\u6bcf\u4e2a\u4ee4\u724c\u7684\u5185\u5b58\u9700\u6c42\uff0c\u4ece\u800c\u80fd\u591f\u5728\u66f4\u5927\u7684\u6a21\u578b\u548c\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\u8fdb\u884c\u8bad\u7ec3\u548c\u63a8\u7406\u3002<\/li>\n\n\n\n<li>\u91c7\u7528\u4e86\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff08RoPE\uff09\uff0c\u6709\u6548\u5904\u7406\u4e86\u5728\u5927\u89c4\u6a21\u6a21\u578b\u4e2d\u5904\u7406\u957f\u5e8f\u5217\u65f6\u7684\u4f4d\u7f6e\u4f9d\u8d56\u95ee\u9898\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>DeepSeekMoE<\/strong>\uff1aDeepSeek-V3\u91c7\u7528\u4e86DeepSeekMoE\u67b6\u6784\uff0c\u8be5\u67b6\u6784\u4e0e\u4f20\u7edf\u7684MoE\u67b6\u6784\u4e0d\u540c\uff0c\u5177\u6709\u66f4\u7ec6\u7c92\u5ea6\u7684\u4e13\u5bb6\uff0c\u5e76\u901a\u8fc7\u5171\u4eab\u4e13\u5bb6\u51cf\u5c11\u4e86\u6a21\u578b\u7684\u8ba1\u7b97\u91cf\u3002\u4e0eDeepSeek-V2\u4e0d\u540c\uff0cDeepSeek-V3\u5f15\u5165\u4e86\u4e00\u79cd\u521b\u65b0\u7684\u65e0\u8f85\u52a9\u635f\u5931\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff0c\u907f\u514d\u4e86\u7531\u4e8e\u8d1f\u8f7d\u4e0d\u5747\u800c\u5bfc\u81f4\u7684\u6a21\u578b\u6027\u80fd\u4e0b\u964d\u3002\n<ul class=\"wp-block-list\">\n<li><strong>\u65e0\u8f85\u52a9\u635f\u5931\u8d1f\u8f7d\u5747\u8861<\/strong>\uff1a\u4f20\u7edfMoE\u6a21\u578b\u901a\u5e38\u4f9d\u8d56\u8f85\u52a9\u635f\u5931\u6765\u5b9e\u73b0\u8d1f\u8f7d\u5747\u8861\uff0c\u4f46\u8fd9\u79cd\u65b9\u6cd5\u53ef\u80fd\u4f1a\u964d\u4f4e\u6a21\u578b\u7684\u8868\u73b0\u3002DeepSeek-V3\u5f15\u5165\u4e86\u4e00\u4e2a\u8f85\u52a9\u635f\u5931\u6d88\u9664\u7b56\u7565\uff0c\u901a\u8fc7\u52a8\u6001\u8c03\u6574\u4e13\u5bb6\u7684\u8d1f\u8f7d\u6765\u786e\u4fdd\u5747\u8861\uff0c\u540c\u65f6\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u8d1f\u9762\u5f71\u54cd\u3002<\/li>\n\n\n\n<li><strong>\u7ec6\u7c92\u5ea6\u4e13\u5bb6<\/strong>\uff1aDeepSeekMoE\u67b6\u6784\u4f7f\u7528\u4e86\u66f4\u7ec6\u7c92\u5ea6\u7684\u4e13\u5bb6\uff0c\u51cf\u5c11\u4e86\u6bcf\u4e2a\u4e13\u5bb6\u7684\u8ba1\u7b97\u8d1f\u62c5\uff0c\u4ece\u800c\u5728\u591a\u4e2a\u4e13\u5bb6\u4e4b\u95f4\u66f4\u6709\u6548\u5730\u5206\u914d\u8ba1\u7b97\u4efb\u52a1\uff0c\u63d0\u9ad8\u4e86\u8ba1\u7b97\u6548\u7387\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u591a\u4ee4\u724c\u9884\u6d4b\uff08MTP\uff09<\/strong>\uff1aDeepSeek-V3\u91c7\u7528\u4e86\u591a\u4ee4\u724c\u9884\u6d4b\uff08MTP\uff09\u8bad\u7ec3\u76ee\u6807\uff0c\u8fd9\u4e00\u76ee\u6807\u5141\u8bb8\u6a21\u578b\u5728\u6bcf\u4e2a\u4f4d\u7f6e\u540c\u65f6\u9884\u6d4b\u591a\u4e2a\u672a\u6765\u7684\u4ee4\u724c\u3002MTP\u7684\u5b9e\u65bd\u80fd\u591f\u7a20\u5bc6\u5316\u8bad\u7ec3\u4fe1\u53f7\uff0c\u63d0\u9ad8\u6570\u636e\u5229\u7528\u6548\u7387\uff0c\u5e76\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u89c4\u5212\u5176\u8868\u793a\u4ee5\u8fdb\u884c\u672a\u6765\u4ee4\u724c\u7684\u9884\u6d4b\u3002\n<ul class=\"wp-block-list\">\n<li><strong>\u591a\u4ee4\u724c\u9884\u6d4b\u6a21\u5757<\/strong>\uff1a\u6bcf\u4e2aMTP\u6a21\u5757\u7531\u5171\u4eab\u7684\u5d4c\u5165\u5c42\u3001\u5171\u4eab\u7684\u8f93\u51fa\u5934\u3001Transformer\u5757\u4ee5\u53ca\u6295\u5f71\u77e9\u9635\u7ec4\u6210\u3002\u4e0e\u4f20\u7edf\u7684\u591a\u4ee4\u724c\u9884\u6d4b\u65b9\u6cd5\u4e0d\u540c\uff0cDeepSeek-V3\u7684MTP\u6a21\u5757\u91c7\u7528\u4e86\u987a\u5e8f\u9884\u6d4b\u7684\u65b9\u5f0f\uff0c\u786e\u4fdd\u6bcf\u6b21\u9884\u6d4b\u65f6\u90fd\u4fdd\u6301\u5b8c\u6574\u7684\u56e0\u679c\u94fe\u3002<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u76ee\u6807<\/strong>\uff1a\u5728\u6bcf\u4e2a\u9884\u6d4b\u6df1\u5ea6\u4e0a\uff0cDeepSeek-V3\u8ba1\u7b97\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u5e76\u5bf9\u6240\u6709\u6df1\u5ea6\u7684MTP\u635f\u5931\u8fdb\u884c\u5e73\u5747\uff0c\u6700\u7ec8\u901a\u8fc7\u52a0\u6743\u7cfb\u6570\u6765\u5f97\u5230\u6574\u4f53\u7684MTP\u635f\u5931\u3002\u8fd9\u4e00\u8bad\u7ec3\u76ee\u6807\u5728\u63d0\u5347\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\uff0c\u4e5f\u4e3a\u63a8\u7406\u52a0\u901f\u63d0\u4f9b\u4e86\u7406\u8bba\u652f\u6301\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p><strong>\u4e09\u3001\u57fa\u7840\u8bbe\u65bd<\/strong><\/p>\n\n\n\n<p>DeepSeek-V3\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u4f9d\u8d56\u4e8e\u5f3a\u5927\u7684\u8ba1\u7b97\u57fa\u7840\u8bbe\u65bd\uff0c\u4e3b\u8981\u7531\u4ee5\u4e0b\u51e0\u90e8\u5206\u6784\u6210\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u8ba1\u7b97\u96c6\u7fa4<\/strong>\uff1aDeepSeek-V3\u7684\u8bad\u7ec3\u57fa\u4e8e2048\u53f0NVIDIA H800 GPU\u7684\u8ba1\u7b97\u96c6\u7fa4\u3002\u6bcf\u4e2a\u8282\u70b9\u5305\u542b8\u5f20GPU\uff0c\u901a\u8fc7NVLink\u548cNVSwitch\u8fdb\u884c\u8fde\u63a5\uff0c\u4e0d\u540c\u8282\u70b9\u4e4b\u95f4\u901a\u8fc7InfiniBand\u8fdb\u884c\u901a\u4fe1\u3002<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u6846\u67b6\u4f18\u5316<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>DualPipe\u7b97\u6cd5<\/strong>\uff1aDeepSeek-V3\u91c7\u7528\u4e86DualPipe\u7b97\u6cd5\u6765\u4f18\u5316\u6d41\u6c34\u7ebf\u5e76\u884c\u6027\uff0c\u89e3\u51b3\u4e86\u5728\u8de8\u8282\u70b9\u8bad\u7ec3\u65f6\u7531\u901a\u4fe1\u5f15\u8d77\u7684\u6548\u7387\u74f6\u9888\u3002\u901a\u8fc7\u91cd\u53e0\u524d\u5411\u548c\u540e\u5411\u8ba1\u7b97\uff0c\u51cf\u5c11\u4e86\u6d41\u6c34\u7ebf\u4e2d\u7684\u7a7a\u95f2\u65f6\u95f4\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\u3002<\/li>\n\n\n\n<li><strong>\u8de8\u8282\u70b9\u901a\u4fe1\u4f18\u5316<\/strong>\uff1a\u4e3a\u4e86\u5145\u5206\u5229\u7528\u786c\u4ef6\u5e26\u5bbd\uff0cDeepSeek-V3\u8bbe\u8ba1\u4e86\u9ad8\u6548\u7684\u8de8\u8282\u70b9\u901a\u4fe1\u5185\u6838\uff0c\u786e\u4fdd\u901a\u4fe1\u8fc7\u7a0b\u4e2d\u7684\u6570\u636e\u4f20\u8f93\u9ad8\u6548\u65e0\u963b\uff0c\u5e76\u4e14\u6700\u5927\u9650\u5ea6\u51cf\u5c11\u4e86\u901a\u4fe1\u5f00\u9500\u3002<\/li>\n\n\n\n<li><strong>\u5185\u5b58\u4f18\u5316<\/strong>\uff1aDeepSeek-V3\u901a\u8fc7\u591a\u9879\u5185\u5b58\u4f18\u5316\u6280\u672f\uff0c\u51cf\u5c11\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5185\u5b58\u7684\u5360\u7528\u3002\u6bd4\u5982\u901a\u8fc7\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u91cd\u65b0\u8ba1\u7b97RMSNorm\u64cd\u4f5c\u548cMLA\u7684\u4e0a\u6295\u5f71\u6765\u51cf\u5c11\u5b58\u50a8\u9700\u6c42\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>FP8\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>FP8\u8bad\u7ec3<\/strong>\uff1a\u4e3a\u4e86\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\uff0cDeepSeek-V3\u91c7\u7528\u4e86FP8\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6\u3002\u5728\u8fd9\u4e00\u6846\u67b6\u4e0b\uff0c\u5927\u591a\u6570\u8ba1\u7b97\u5bc6\u96c6\u578b\u64cd\u4f5c\u90fd\u5728FP8\u7cbe\u5ea6\u4e0b\u8fdb\u884c\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u8ba1\u7b97\u901f\u5ea6\uff0c\u540c\u65f6\u51cf\u5c11\u4e86\u5185\u5b58\u9700\u6c42\u3002<\/li>\n\n\n\n<li><strong>\u7ec6\u7c92\u5ea6\u91cf\u5316<\/strong>\uff1a\u4e3a\u4e86\u7f13\u89e3\u7531\u4e8eFP8\u7cbe\u5ea6\u5f15\u8d77\u7684\u91cf\u5316\u8bef\u5dee\uff0cDeepSeek-V3\u91c7\u7528\u4e86\u7ec6\u7c92\u5ea6\u7684\u91cf\u5316\u65b9\u6cd5\uff0c\u5728\u66f4\u5c0f\u7684\u5143\u7d20\u7ec4\u4e0a\u8fdb\u884c\u91cf\u5316\uff0c\u4f7f\u5f97\u6bcf\u4e2a\u64cd\u4f5c\u7684\u7cbe\u5ea6\u66f4\u52a0\u7a33\u5b9a\u3002<\/li>\n\n\n\n<li><strong>\u7d2f\u79ef\u7cbe\u5ea6\u63d0\u5347<\/strong>\uff1a\u5bf9\u4e8e\u4f4e\u7cbe\u5ea6\u7684GEMM\u64cd\u4f5c\uff0cDeepSeek-V3\u91c7\u7528\u4e86\u5728CUDA\u6838\u5fc3\u4e0a\u8fdb\u884c\u9ad8\u7cbe\u5ea6\u7d2f\u79ef\u7684\u7b56\u7565\uff0c\u4ece\u800c\u786e\u4fdd\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u6570\u503c\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p><strong>\u56db\u3001\u9884\u8bad\u7ec3<\/strong><\/p>\n\n\n\n<p>DeepSeek-V3\u7684\u9884\u8bad\u7ec3\u8fc7\u7a0b\u5305\u62ec\u4ee5\u4e0b\u51e0\u4e2a\u91cd\u8981\u90e8\u5206\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6570\u636e\u6784\u5efa\u4e0e\u53c2\u6570\u8bbe\u7f6e<\/strong>\n<ul class=\"wp-block-list\">\n<li>DeepSeek-V3\u4f7f\u7528\u4e8614.8\u4e07\u4ebf\u4e2a\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u5316\u7684\u4ee4\u724c\u6570\u636e\u8fdb\u884c\u9884\u8bad\u7ec3\u3002\u8fd9\u4e9b\u6570\u636e\u6765\u81ea\u591a\u4e2a\u9886\u57df\uff0c\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5230\u4e30\u5bcc\u7684\u77e5\u8bc6\u8868\u793a\u3002<\/li>\n\n\n\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u91c7\u7528\u4e8632K\u548c128K\u7684\u4e0a\u4e0b\u6587\u6269\u5c55\u7b56\u7565\uff0c\u9010\u6b65\u589e\u52a0\u6a21\u578b\u80fd\u591f\u5904\u7406\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u6a21\u578b\u5bf9\u957f\u6587\u6863\u7684\u7406\u89e3\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u7a33\u5b9a\u6027<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u5728\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cDeepSeek-V3\u8868\u73b0\u51fa\u4e86\u6781\u9ad8\u7684\u7a33\u5b9a\u6027\u3002\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6ca1\u6709\u51fa\u73b0\u65e0\u6cd5\u6062\u590d\u7684\u635f\u5931\u5c16\u5cf0\uff0c\u4e14\u8bad\u7ec3\u8fc7\u7a0b\u987a\u5229\u8fdb\u884c\uff0c\u6ca1\u6709\u51fa\u73b0\u8fc7\u56de\u6eda\u73b0\u8c61\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u6548\u7387<\/strong>\n<ul class=\"wp-block-list\">\n<li>DeepSeek-V3\u7684\u9884\u8bad\u7ec3\u9636\u6bb5\u603b\u5171\u6d88\u8017\u4e862788\u4e07H800 GPU\u5c0f\u65f6\uff0c\u8bad\u7ec3\u6210\u672c\u4e3a557.6\u4e07\u7f8e\u5143\u3002\u901a\u8fc7\u9ad8\u6548\u7684\u8bad\u7ec3\u6846\u67b6\u548c\u4f18\u5316\u7b56\u7565\uff0cDeepSeek-V3\u80fd\u591f\u5728\u4e24\u4e2a\u6708\u5185\u5b8c\u621014.8\u4e07\u4ebf\u4ee4\u724c\u7684\u9884\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p><strong>\u4e94\u3001\u540e\u8bad\u7ec3<\/strong><\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u5728\u9884\u8bad\u7ec3\u4e4b\u540e\uff0cDeepSeek-V3\u901a\u8fc7\u76d1\u7763\u5fae\u8c03\u6765\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6a21\u578b\u7684\u6027\u80fd\u3002\u901a\u8fc7\u5bf9\u8bdd\u6570\u636e\u548c\u4e13\u4e1a\u9886\u57df\u6570\u636e\u7684\u5fae\u8c03\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u9002\u5e94\u5b9e\u9645\u5e94\u7528\u573a\u666f\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\u5305\u62ec\u4e86\u5956\u52b1\u6a21\u578b\u548c\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08Group Relative Policy Optimization\uff09\u3002\u8fd9\u4e9b\u6280\u672f\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u8fdb\u884c\u63a8\u7406\u4efb\u52a1\uff0c\u5e76\u4f18\u5316\u5176\u751f\u6210\u7ed3\u679c\u7684\u8d28\u91cf\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u8bc4\u4f30\u7ed3\u679c<\/strong>\n<ul class=\"wp-block-list\">\n<li>DeepSeek-V3\u5728\u591a\u4e2a\u8bc4\u6d4b\u57fa\u51c6\u4e0a\u8868\u73b0\u5353\u8d8a\uff0c\u5c24\u5176\u5728\u6570\u5b66\u3001\u4ee3\u7801\u3001\u63a8\u7406\u7b49\u4efb\u52a1\u4e0a\u8d85\u8d8a\u4e86\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\uff0c\u5e76\u4e14\u5728\u4e00\u4e9b\u4efb\u52a1\u4e0a\u4e0e\u9886\u5148\u7684\u5c01\u95ed\u6e90\u6a21\u578b\u5982GPT-4\u548cClaude-3.5\u6301\u5e73\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p><strong>\u516d\u3001\u7ed3\u8bba\u3001\u5c40\u9650\u6027\u4e0e\u672a\u6765\u65b9\u5411<\/strong><\/p>\n\n\n\n<p>DeepSeek-V3\u5728\u67b6\u6784\u521b\u65b0\u3001\u8bad\u7ec3\u6548\u7387\u548c\u63a8\u7406\u6027\u80fd\u65b9\u9762\u5c55\u73b0\u4e86\u5de8\u5927\u7684\u6f5c\u529b\uff0c\u5c24\u5176\u5728\u6210\u672c\u548c\u6027\u80fd\u7684\u5e73\u8861\u65b9\u9762\u505a\u51fa\u4e86\u91cd\u8981\u8d21\u732e\u3002\u7136\u800c\uff0c\u4ecd\u6709\u8bb8\u591a\u6311\u6218\u9700\u8981\u89e3\u51b3\uff0c\u5982\u5982\u4f55\u8fdb\u4e00\u6b65\u6269\u5c55\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001\u4f18\u5316\u591a\u6a21\u6001\u6570\u636e\u5904\u7406\u7b49\u3002<\/p>\n\n\n\n<p>\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u5305\u62ec\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u901f\u5ea6\u3001\u66f4\u9ad8\u6548\u7684\u786c\u4ef6\u67b6\u6784\u8bbe\u8ba1\u3001\u4ee5\u53ca\u591a\u6a21\u6001\u5b66\u4e60\u548c\u751f\u6210\u80fd\u529b\u7684\u589e\u5f3a\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u76f8\u5173\u9605\u8bfb\uff1a<strong><a href=\"https:\/\/nullthought.net\/?p=5614\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-R1\u6280\u672f\u62a5\u544a\u6982\u8ff0<\/a><\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u7b80\u4ecb \u8fd1\u5e74\u6765\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u53d6\u5f97\u4e86\u98de\u901f\u53d1\u5c55\uff0c\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u8fc8\u8fdb\u7684\u6b65\u4f10\u8d8a\u6765\u8d8a\u5feb\u3002DeepS [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,96,95,86],"class_list":["post-5463","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-deepseek","tag-transformer","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u4e00\u3001\u7b80\u4ecb \u8fd1\u5e74\u6765\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u53d6\u5f97\u4e86\u98de\u901f\u53d1\u5c55\uff0c\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u8fc8\u8fdb\u7684\u6b65\u4f10\u8d8a\u6765\u8d8a\u5feb\u3002DeepS&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5463","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5463"}],"version-history":[{"count":3,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5463\/revisions"}],"predecessor-version":[{"id":5619,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5463\/revisions\/5619"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5463"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5463"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5463"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}