{"id":1114,"date":"2025-02-06T18:10:04","date_gmt":"2025-02-06T10:10:04","guid":{"rendered":"https:\/\/lingbo.online\/?p=1114"},"modified":"2025-11-06T09:04:18","modified_gmt":"2025-11-06T01:04:18","slug":"deepseekv3-technical-report","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/learning_note\/deepseekv3-technical-report\/","title":{"rendered":"Deepseek V3 Technical Report"},"content":{"rendered":"<h3>MetaData<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4 2024.12.27<\/li>\n<li>\u4f5c\u8005\uff1aDeepSeek-AI<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2412.19437v1\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/abs\/2412.19437v1<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/deepseek-ai\/DeepSeek-V3<\/a><\/li>\n<\/ul>\n<h3>\u6458\u8981<\/h3>\n<p>DeepSeek-V3\u662f\u4e00\u4e2a\u529f\u80fd\u5f3a\u5927\u7684\u4e13\u5bb6\u6df7\u5408\uff08Mixture-of-Experts\uff0cMoE\uff09\u8bed\u8a00\u6a21\u578b\uff0c\u5177\u5907\u4ee5\u4e0b\u663e\u8457\u7279\u70b9\uff1a<\/p>\n<ul>\n<li>\u53c2\u6570\u89c4\u6a21\u4e0e\u6fc0\u6d3b\u673a\u5236\n<ul>\n<li>\u53c2\u6570\u603b\u91cf\uff1a\u5305\u542b6710\u4ebf\u4e2a\u53c2\u6570\u3002<\/li>\n<li>\u53c2\u6570\u6fc0\u6d3b\uff1a\u6bcf\u4e2a\u7b26\u5143\u6fc0\u6d3b370\u4ebf\u4e2a\u53c2\u6570\uff0c\u5b9e\u73b0\u9ad8\u6548\u7684\u4fe1\u606f\u5904\u7406\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u67b6\u6784\u521b\u65b0\n<ul>\n<li>\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08Multi-head Latent Attention\uff0cMLA\uff09\uff1a\u91c7\u7528\u6b64\u67b6\u6784\u4ee5\u4f18\u5316\u63a8\u7406\u6548\u7387\u3002<\/li>\n<li>DeepSeekMoE\u67b6\u6784\uff1a\u8be5\u67b6\u6784\u5728DeepSeek-V2\u4e2d\u5df2\u5f97\u5230\u5145\u5206\u9a8c\u8bc1\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u6a21\u578b\u7684\u6027\u80fd\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u8bad\u7ec3\u7b56\u7565\u4e0e\u6027\u80fd\u63d0\u5347\n<ul>\n<li>\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff1a\u5f00\u521b\u6027\u7b56\u7565\uff0c\u786e\u4fdd\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u7a33\u5b9a\u6027\u3002<\/li>\n<li>\u591aToken\u9884\u6d4b\u8bad\u7ec3\u76ee\u6807\uff1a\u8bbe\u5b9a\u6b64\u76ee\u6807\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u6027\u80fd\u3002<\/li>\n<li>\u9884\u8bad\u7ec3\u4e0e\u5fae\u8c03\uff1a\u572814.8\u4e07\u4ebf\u4e2a\u591a\u6837\u5316\u4e14\u9ad8\u8d28\u91cf\u7684Token\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u968f\u540e\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\u548c\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\uff0c\u5145\u5206\u53d1\u6325\u6a21\u578b\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u6027\u80fd\u8bc4\u4f30\u4e0e\u8bad\u7ec3\u6548\u7387\n<ul>\n<li>\u6027\u80fd\u5353\u8d8a\uff1a\u7efc\u5408\u8bc4\u4f30\u663e\u793a\uff0cDeepSeek-V3\u7684\u6027\u80fd\u4f18\u4e8e\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\uff0c\u8fbe\u5230\u4e0e\u9886\u5148\u7684\u95ed\u6e90\u6a21\u578b\u76f8\u5f53\u7684\u6c34\u5e73\u3002<\/li>\n<li>\u8bad\u7ec3\u6210\u672c\u4f4e\uff1a\u5b8c\u6574\u8bad\u7ec3\u4ec5\u9700278.8\u4e07H800GPU\u5c0f\u65f6\uff0c\u4f53\u73b0\u4e86\u9ad8\u6548\u7684\u8bad\u7ec3\u6548\u7387\u3002<\/li>\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u7a33\u5b9a\uff1a\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u672a\u9047\u5230\u4efb\u4f55\u4e0d\u53ef\u6062\u590d\u7684\u635f\u5931\u5cf0\u503c\uff0c\u65e0\u9700\u8fdb\u884c\u56de\u6eda\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u6a21\u578b\u83b7\u53d6\n<ul>\n<li>\u6a21\u578b\u68c0\u67e5\u70b9\uff1a\u53ef\u5728<a href=\"https:\/\/github.com\/deepseek-ai\/DeepSeek-V3\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/deepseek-ai\/DeepSeek-V3<\/a>\u83b7\u53d6\uff0c\u4fbf\u4e8e\u7814\u7a76\u4eba\u5458\u548c\u5f00\u53d1\u8005\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u7814\u7a76\u4e0e\u5e94\u7528\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3>1 \u5f15\u8a00<\/h3>\n<p>\u8fd1\u5e74\u6765\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u6b63\u5728\u7ecf\u5386\u5feb\u901f\u8fed\u4ee3\u548c\u53d1\u5c55\uff0c\u9010\u6e10\u7f29\u5c0f\u4e0e\u4eba\u5de5\u901a\u7528\u667a\u80fd\uff08AGI\uff09\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u9664\u4e86\u95ed\u6e90\u6a21\u578b\u4e4b\u5916\uff0c\u5f00\u6e90\u6a21\u578b\u4e5f\u5728\u53d6\u5f97\u663e\u8457\u8fdb\u5c55\uff0c\u52aa\u529b\u7f29\u5c0f\u4e0e\u95ed\u6e90\u6a21\u578b\u7684\u5dee\u8ddd\u3002\u5176\u4e2d\u5305\u62ecDeepSeek\u7cfb\u5217\u3001LLaMA\u7cfb\u5217\u3001Qwen\u7cfb\u5217\u548cMistral\u7cfb\u5217\u3002\u4e3a\u8fdb\u4e00\u6b65\u7a81\u7834\u5f00\u6e90\u6a21\u578b\u7684\u80fd\u529b\u8fb9\u754c\uff0c\u6211\u4eec\u6269\u5c55\u4e86\u6211\u4eec\u7684\u6a21\u578b\uff0c\u5e76\u5f15\u5165\u4e86DeepSeek-V3\uff0c\u8fd9\u662f\u4e00\u4e2a\u62e5\u67096710\u4ebf\u53c2\u6570\u7684\u5927\u578b\u4e13\u5bb6\u6df7\u5408\u6a21\u578b\uff08MoE\uff09\uff0c\u5176\u4e2d\u6bcf\u4e2a\u7b26\u5143\u6fc0\u6d3b370\u4ebf\u53c2\u6570\u3002<br \/>\u6211\u4eec\u59cb\u7ec8\u81f4\u529b\u4e8e\u5728\u5f3a\u5927\u7684\u6a21\u578b\u6027\u80fd\u548c\u7ecf\u6d4e\u6210\u672c\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002\u5728\u67b6\u6784\u65b9\u9762\uff0cDeepSeek-V3\u7ee7\u7eed\u91c7\u7528\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u4ee5\u5b9e\u73b0\u9ad8\u6548\u63a8\u7406\uff0c\u5e76\u91c7\u7528DeepSeekMoE\u4ee5\u5b9e\u73b0\u7ecf\u6d4e\u9ad8\u6548\u7684\u8bad\u7ec3\u3002\u8fd9\u4e24\u79cd\u67b6\u6784\u5df2\u5728DeepSeek-V2\u4e2d\u5f97\u5230\u9a8c\u8bc1\uff0c\u8bc1\u660e\u4e86\u5b83\u4eec\u5728\u4fdd\u6301\u5f3a\u5927\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\uff0c\u80fd\u591f\u5b9e\u73b0\u9ad8\u6548\u8bad\u7ec3\u548c\u63a8\u7406\u3002\u9664\u4e86\u57fa\u672c\u67b6\u6784\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u5f15\u5165\u4e86\u4e24\u79cd\u989d\u5916\u7684\u7b56\u7565\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u80fd\u529b\u3002\u9996\u5148\uff0cDeepSeek-V3\u5f00\u521b\u4e86\u4e00\u79cd\u65e0\u8f85\u52a9\u635f\u5931\uff08auxiliary-loss-free\uff09\u7684\u7b56\u7565\u7528\u4e8e\u8d1f\u8f7d\u5747\u8861\uff0c\u65e8\u5728\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u56e0\u9f13\u52b1\u8d1f\u8f7d\u5747\u8861\u800c\u5bf9\u6a21\u578b\u6027\u80fd\u9020\u6210\u7684\u8d1f\u9762\u5f71\u54cd\u3002\u5176\u6b21\uff0cDeepSeek-V3\u91c7\u7528\u591a\u7b26\u5143\uff08multi-token\uff09\u9884\u6d4b\u8bad\u7ec3\u76ee\u6807\uff0c\u8fd9\u6709\u52a9\u4e8e\u63d0\u9ad8\u5728\u8bc4\u4f30\u57fa\u51c6\u4e0a\u7684\u6574\u4f53\u6027\u80fd\u3002<br \/>\u4e3a\u4e86\u5b9e\u73b0\u9ad8\u6548\u8bad\u7ec3\uff0c\u6211\u4eec\u652f\u6301FP8\u6df7\u5408\u7cbe\u5ea6\uff08FP8 mixed precision training\uff09\u8bad\u7ec3\uff0c\u5e76\u5bf9\u8bad\u7ec3\u6846\u67b6\u8fdb\u884c\u4e86\u5168\u9762\u4f18\u5316\u3002\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u5df2\u6210\u4e3a\u4e00\u79cd\u5f88\u6709\u524d\u666f\u7684\u9ad8\u6548\u8bad\u7ec3\u89e3\u51b3\u65b9\u6848\uff0c\u5176\u53d1\u5c55\u4e0e\u786c\u4ef6\u80fd\u529b\u7684\u8fdb\u6b65\u5bc6\u5207\u76f8\u5173\u3002\u5728\u672c\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2aFP8\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6\uff0c\u5e76\u9996\u6b21\u5728\u4e00\u4e2a\u6781\u5927\u89c4\u6a21\u7684\u6a21\u578b\u4e0a\u9a8c\u8bc1\u4e86\u5176\u6709\u6548\u6027\u3002\u901a\u8fc7\u652f\u6301FP8\u8ba1\u7b97\u548c\u5b58\u50a8\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u8bad\u7ec3\u52a0\u901f\u548cGPU\u5185\u5b58\u4f7f\u7528\u7684\u51cf\u5c11\u3002\u81f3\u4e8e\u8bad\u7ec3\u6846\u67b6\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86DualPipe\u7b97\u6cd5\u4ee5\u5b9e\u73b0\u9ad8\u6548\u7684\u6d41\u6c34\u7ebf\u5e76\u884c\u3002\u8be5\u7b97\u6cd5\u5177\u6709\u66f4\u5c11\u7684\u6d41\u6c34\u7ebf\u6c14\u6ce1\uff0c\u5e76\u901a\u8fc7\u8ba1\u7b97-\u901a\u4fe1\u91cd\u53e0\u9690\u85cf\u4e86\u8bad\u7ec3\u671f\u95f4\u7684\u5927\u90e8\u5206\u901a\u4fe1\u3002\u8fd9\u79cd\u91cd\u53e0\u786e\u4fdd\uff0c\u968f\u7740\u6a21\u578b\u8fdb\u4e00\u6b65\u6269\u5c55\uff0c\u53ea\u8981\u6211\u4eec\u4fdd\u6301\u6052\u5b9a\u7684\u8ba1\u7b97\u4e0e\u901a\u4fe1\u6bd4\u7387\uff0c\u6211\u4eec\u4ecd\u7136\u53ef\u4ee5\u5728\u8282\u70b9\u95f4\u4f7f\u7528\u7ec6\u7c92\u5ea6\u7684\u4e13\u5bb6\uff0c\u540c\u65f6\u5b9e\u73b0\u63a5\u8fd1\u4e8e\u96f6\u7684\u5168\u5bf9\u5168\u901a\u4fe1\u5f00\u9500\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5f00\u53d1\u4e86\u9ad8\u6548\u7684\u8de8\u8282\u70b9\u5168\u5bf9\u5168\u901a\u4fe1\u5185\u6838\uff0c\u4ee5\u5145\u5206\u5229\u7528InfiniBand\uff08IB\uff09\u548cNVLink\u5e26\u5bbd\u3002\u6211\u4eec\u8fd8\u7cbe\u5fc3\u4f18\u5316\u4e86\u5185\u5b58\u5360\u7528\uff0c\u4f7f\u5f97\u65e0\u9700\u4f7f\u7528\u4ee3\u4ef7\u9ad8\u6602\u7684\u5f20\u91cf\u5e76\u884c\u5c31\u80fd\u8bad\u7ec3DeepSeek-V3\u3002\u901a\u8fc7\u8fd9\u4e9b\u52aa\u529b\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u9ad8\u8bad\u7ec3\u6548\u7387\u3002<br \/>\u5728\u9884\u8bad\u7ec3\u671f\u95f4\uff0c\u6211\u4eec\u572814.8T\u9ad8\u8d28\u91cf\u548c\u591a\u6837\u5316\u7684\u7b26\u5143\u4e0a\u8bad\u7ec3DeepSeek-V3\uff0c\u6574\u4e2a\u9884\u8bad\u7ec3\u8fc7\u7a0b\u975e\u5e38\u7a33\u5b9a\u3002\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u6ca1\u6709\u9047\u5230\u4efb\u4f55\u4e0d\u53ef\u6062\u590d\u7684\u635f\u5931\u5cf0\u503c\uff0c\u4e5f\u65e0\u9700\u56de\u6eda\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5bf9DeepSeek-V3\u8fdb\u884c\u4e86\u4e24\u9636\u6bb5\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u3002\u5728\u7b2c\u4e00\u9636\u6bb5\uff0c\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u523032K\uff0c\u7b2c\u4e8c\u9636\u6bb5\u8fdb\u4e00\u6b65\u6269\u5c55\u5230128K\u3002\u6b64\u540e\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u540e\u8bad\u7ec3\uff0c\u5305\u62ec\u5728DeepSeek-V3\u7684\u57fa\u7840\u6a21\u578b\u4e0a\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\uff0c\u4ee5\u4f7f\u5176\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\uff0c\u5e76\u8fdb\u4e00\u6b65\u91ca\u653e\u5176\u6f5c\u529b\u3002\u5728\u540e\u671f\u8bad\u7ec3\u9636\u6bb5\uff0c\u6211\u4eec\u4eceDeepSeek-R1\u7cfb\u5217\u6a21\u578b\u4e2d\u63d0\u53d6\u63a8\u7406\u80fd\u529b\uff0c\u540c\u65f6\u4ed4\u7ec6\u4fdd\u6301\u6a21\u578b\u7cbe\u5ea6\u548c\u751f\u6210\u957f\u5ea6\u4e4b\u95f4\u7684\u5e73\u8861\u3002<br \/>\u6211\u4eec\u5728\u5168\u9762\u7684\u57fa\u51c6\u6d4b\u8bd5\u96c6\u4e0a\u8bc4\u4f30\u4e86DeepSeek-V3\u3002\u5c3d\u7ba1\u5176\u8bad\u7ec3\u6210\u672c\u7ecf\u6d4e\uff0c\u4f46\u7efc\u5408\u8bc4\u4f30\u663e\u793a\uff0cDeepSeek-V3-Base\u5df2\u6210\u4e3a\u76ee\u524d\u53ef\u7528\u7684\u6700\u5f3a\u5f00\u6e90\u57fa\u7840\u6a21\u578b\uff0c\u5c24\u5176\u662f\u5728\u4ee3\u7801\u548c\u6570\u5b66\u65b9\u9762\u3002\u5176\u804a\u5929\u7248\u672c\u4e5f\u4f18\u4e8e\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\uff0c\u5e76\u5728\u4e00\u7cfb\u5217\u6807\u51c6\u548c\u5f00\u653e\u5f0f\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5b9e\u73b0\u4e86\u4e0e\u9886\u5148\u7684\u95ed\u6e90\u6a21\u578b\uff08\u5305\u62ecGPT-4o\u548cClaude-3.5-Sonnet\uff09\u76f8\u5f53\u7684\u6027\u80fd\u3002<br \/>\u6700\u540e\uff0c\u6211\u4eec\u518d\u6b21\u5f3a\u8c03DeepSeek-V3\u7684\u7ecf\u6d4e\u9ad8\u6548\u7684\u8bad\u7ec3\u6210\u672c\u3002\u5982\u88681\u6240\u793a\uff0c\u8fd9\u4e00\u6210\u672c\u662f\u901a\u8fc7\u6211\u4eec\u5bf9\u7b97\u6cd5\u3001\u6846\u67b6\u548c\u786c\u4ef6\u7684\u4f18\u5316\u534f\u540c\u8bbe\u8ba1\u5b9e\u73b0\u7684\u3002\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u8bad\u7ec3DeepSeek-V3\u6bcf\u4e07\u4ebf\u4e2a\u7b26\u5143\u4ec5\u970018\u4e07\u4e2aH800GPU\u5c0f\u65f6\uff0c\u5373\u5728\u6211\u4eec\u62e5\u67092048\u4e2aH800GPU\u7684\u96c6\u7fa4\u4e0a\u8fd0\u884c3.7\u5929\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u9636\u6bb5\u5728\u4e0d\u5230\u4e24\u4e2a\u6708\u5185\u5b8c\u6210\uff0c\u8017\u8d39\u4e86266.4\u4e07\u4e2aGPU\u5c0f\u65f6\u3002\u52a0\u4e0a\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u768411.9\u4e07\u4e2aGPU\u5c0f\u65f6\u548c\u540e\u671f\u8bad\u7ec3\u76845000\u4e2aGPU\u5c0f\u65f6\uff0cDeepSeek-V3\u7684\u5168\u90e8\u8bad\u7ec3\u4ec5\u9700278.8\u4e07\u4e2aGPU\u5c0f\u65f6\u3002\u5047\u8bbeH800GPU\u7684\u79df\u8d41\u4ef7\u683c\u4e3a\u6bcfGPU\u5c0f\u65f62\u7f8e\u5143\uff0c\u6211\u4eec\u7684\u603b\u8bad\u7ec3\u6210\u672c\u4ec5\u4e3a557.6\u4e07\u7f8e\u5143\u3002\u8bf7\u6ce8\u610f\uff0c\u4e0a\u8ff0\u6210\u672c\u4ec5\u5305\u62ecDeepSeek-V3\u7684\u6b63\u5f0f\u8bad\u7ec3\uff0c\u4e0d\u5305\u62ec\u4e0e\u5148\u524d\u5173\u4e8e\u67b6\u6784\u3001\u7b97\u6cd5\u6216\u6570\u636e\u7684\u7814\u7a76\u548c\u6d88\u878d\u5b9e\u9a8c\u76f8\u5173\u7684\u6210\u672c\u3002<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250206111918706.jpg\" alt=\"\" \/><\/p>\n<p>\u88681 DeepSeek-V3\u7684\u8bad\u7ec3\u6210\u672c\uff0c\u5047\u8bbeH800\u7684\u79df\u8d41\u4ef7\u683c\u4e3a\u6bcfGPU\u5c0f\u65f62\u7f8e\u5143<\/p>\n<p>\u6211\u4eec\u7684\u4e3b\u8981\u8d21\u732e\u5305\u62ec\uff1a<\/p>\n<p><strong>\u67b6\u6784\uff1a\u521b\u65b0\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\u548c\u8bad\u7ec3\u76ee\u6807<\/strong><\/p>\n<ul>\n<li>\u5728DeepSeek-V2\u9ad8\u6548\u67b6\u6784\u7684\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u5f00\u521b\u4e86\u4e00\u79cd\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff0c\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u4e86\u56e0\u9f13\u52b1\u8d1f\u8f7d\u5747\u8861\u800c\u9020\u6210\u7684\u6027\u80fd\u4e0b\u964d\u3002<\/li>\n<li>\u6211\u4eec\u7814\u7a76\u4e86\u4e00\u79cd\u591a\u7b26\u5143\u9884\u6d4b\uff08MTP\uff09\u76ee\u6807\uff0c\u5e76\u8bc1\u660e\u5b83\u6709\u5229\u4e8e\u63d0\u5347\u6a21\u578b\u6027\u80fd\u3002\u5b83\u8fd8\u53ef\u4ee5\u7528\u4e8e\u63a8\u6d4b\u6027\u89e3\u7801\uff0c\u4ee5\u52a0\u901f\u63a8\u7406\u8fc7\u7a0b\u3002<\/li>\n<\/ul>\n<p><strong>\u9884\u8bad\u7ec3\uff1a\u8fc8\u5411\u6781\u81f4\u8bad\u7ec3\u6548\u7387<\/strong><\/p>\n<ul>\n<li>\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2aFP8\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6\uff0c\u5e76\u9996\u6b21\u9a8c\u8bc1\u4e86FP8\u8bad\u7ec3\u5728\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u4e0a\u7684\u53ef\u884c\u6027\u548c\u6709\u6548\u6027\u3002<\/li>\n<li>\u901a\u8fc7\u7b97\u6cd5\u3001\u6846\u67b6\u548c\u786c\u4ef6\u7684\u534f\u540c\u8bbe\u8ba1\uff0c\u6211\u4eec\u514b\u670d\u4e86\u8de8\u8282\u70b9MoE\u8bad\u7ec3\u4e2d\u7684\u901a\u4fe1\u74f6\u9888\uff0c\u5b9e\u73b0\u4e86\u63a5\u8fd1\u5b8c\u5168\u7684\u8ba1\u7b97\u901a\u4fe1\u91cd\u53e0\u3002\u8fd9\u663e\u8457\u63d0\u9ad8\u4e86\u6211\u4eec\u7684\u8bad\u7ec3\u6548\u7387\u5e76\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u65e0\u9700\u989d\u5916\u5f00\u9500\u7684\u60c5\u51b5\u4e0b\u8fdb\u4e00\u6b65\u6269\u5927\u6a21\u578b\u89c4\u6a21\u3002<\/li>\n<li>\u6211\u4eec\u4ec5\u4ee5 266.4 \u4e07 H800 GPU \u5c0f\u65f6\u7684\u7ecf\u6d4e\u6210\u672c\uff0c\u5728 14.8 \u4e07\u4ebf\u7b26\u5143\u4e0a\u5b8c\u6210\u4e86 DeepSeek-V3 \u7684\u9884\u8bad\u7ec3\uff0c\u4ece\u800c\u4ea7\u751f\u4e86\u76ee\u524d\u6700\u5f3a\u5927\u7684\u5f00\u6e90\u57fa\u7840\u6a21\u578b\u3002\u9884\u8bad\u7ec3\u540e\u7684\u540e\u7eed\u8bad\u7ec3\u9636\u6bb5\u4ec5\u9700 0.1 \u4e07 GPU \u5c0f\u65f6\u3002<\/li>\n<\/ul>\n<p><strong>\u8bad\u7ec3\u540e\u9636\u6bb5\uff1a\u6765\u81ea DeepSeek-R1 \u7684\u77e5\u8bc6\u84b8\u998f<\/strong><\/p>\n<ul>\n<li>\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u521b\u65b0\u7684\u65b9\u6cd5\uff0c\u5c06\u63a8\u7406\u80fd\u529b\uff08\u7279\u522b\u662f\u6765\u81eaDeepSeek R1\u7cfb\u5217\u6a21\u578b\u4e4b\u4e00\u7684\uff09\u4ece\u957f\u94fe\u601d\u7ef4\uff08CoT\uff09\u6a21\u578b\u4e2d\u84b8\u998f\u5230\u6807\u51c6\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4e2d\uff0c\u7279\u522b\u662fDeepSeek-V3\u3002\u6211\u4eec\u7684\u6d41\u7a0b\u5de7\u5999\u5730\u5c06R1\u7684\u9a8c\u8bc1\u548c\u53cd\u601d\u6a21\u5f0f\uff08verification and reflection patterns\uff09\u878d\u5165DeepSeek-V3\uff0c\u5e76\u663e\u8457\u63d0\u9ad8\u4e86\u5176\u63a8\u7406\u6027\u80fd\u3002\u540c\u65f6\uff0c\u6211\u4eec\u4e5f\u4fdd\u6301\u4e86\u5bf9DeepSeek-V3\u8f93\u51fa\u98ce\u683c\u548c\u957f\u5ea6\u7684\u63a7\u5236\u3002<\/li>\n<\/ul>\n<p><strong>\u6838\u5fc3\u8bc4\u4f30\u7ed3\u679c\u603b\u7ed3<\/strong><\/p>\n<ul>\n<li><strong>\u77e5\u8bc6\uff1a<\/strong>\n<ul>\n<li>\u5728MMLU\u3001MMLU-Pro\u548cGPQA\u7b49\u6559\u80b2\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cDeepSeek-V3\u7684\u6027\u80fd\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\uff0c\u5728MMLU\u4e0a\u8fbe\u523088.5\uff0c\u5728MMLU-Pro\u4e0a\u8fbe\u523075.9\uff0c\u5728GPQA\u4e0a\u8fbe\u523059.1\u3002\u5176\u6027\u80fd\u4e0eGPT-4o\u548cClaude-Sonnet-3.5\u7b49\u9886\u5148\u7684\u95ed\u6e90\u6a21\u578b\u76f8\u5f53\uff0c\u7f29\u5c0f\u4e86\u8be5\u9886\u57df\u5f00\u6e90\u6a21\u578b\u548c\u95ed\u6e90\u6a21\u578b\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002<\/li>\n<li>\u5bf9\u4e8e\u4e8b\u5b9e\u6027\u57fa\u51c6\u6d4b\u8bd5\uff0cDeepSeek-V3\u5728SimpleQA\u548c\u4e2d\u6587SimpleQA\u4e0a\u90fd\u5c55\u73b0\u51fa\u4f18\u4e8e\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\u7684\u6027\u80fd\u3002\u867d\u7136\u5728\u82f1\u6587\u4e8b\u5b9e\u77e5\u8bc6\uff08SimpleQA\uff09\u65b9\u9762\u843d\u540e\u4e8eGPT-4o\u548cClaude-Sonnet-3.5\uff0c\u4f46\u5728\u4e2d\u6587\u4e8b\u5b9e\u77e5\u8bc6\uff08\u4e2d\u6587 SimpleQA\uff09\u65b9\u9762\u5374\u8d85\u8fc7\u4e86\u8fd9\u4e9b\u6a21\u578b\uff0c\u7a81\u663e\u4e86\u5176\u5728\u4e2d\u6587\u4e8b\u5b9e\u77e5\u8bc6\u65b9\u9762\u7684\u4f18\u52bf\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4ee3\u7801\u3001\u6570\u5b66\u548c\u63a8\u7406\uff1a<\/strong>\n<ul>\n<li>\u5728\u6240\u6709\u975e\u957f\u94fe\u601d\u7ef4\uff08CoT\uff09\u5f00\u6e90\u548c\u95ed\u6e90\u6a21\u578b\u4e2d\uff0cDeepSeek-V3 \u5728\u4e0e\u6570\u5b66\u76f8\u5173\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5b83\u751a\u81f3\u5728\u67d0\u4e9b\u57fa\u51c6\u6d4b\u8bd5\uff08\u4f8b\u5982 MATH-500\uff09\u4e0a\u8d85\u8d8a\u4e86 o1-preview\uff0c\u5c55\u73b0\u4e86\u5176\u5f3a\u5927\u7684\u6570\u5b66\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<li>\u5728\u4e0e\u7f16\u7801\u76f8\u5173\u7684\u4efb\u52a1\u4e2d\uff0cDeepSeek-V3\u6210\u4e3a\u7f16\u7801\u7ade\u8d5b\u57fa\u51c6\u6d4b\u8bd5\uff08\u4f8b\u5982 LiveCodeBench\uff09\u4e2d\u6027\u80fd\u6700\u9ad8\u7684\u6a21\u578b\uff0c\u5de9\u56fa\u4e86\u5176\u5728\u8be5\u9886\u57df\u7684\u9886\u5148\u5730\u4f4d\u3002\u5bf9\u4e8e\u5de5\u7a0b\u76f8\u5173\u4efb\u52a1\uff0c\u867d\u7136DeepSeek-V3\u7684\u6027\u80fd\u7565\u4f4e\u4e8eClaude-Sonnet-3.5\uff0c\u4f46\u5b83\u4ecd\u7136\u663e\u8457\u8d85\u8d8a\u5176\u4ed6\u6240\u6709\u6a21\u578b\uff0c\u8bc1\u660e\u4e86\u5176\u5728\u5404\u79cd\u6280\u672f\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u7ade\u4e89\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u5728\u672c\u6587\u7684\u5176\u4f59\u90e8\u5206\uff0c\u6211\u4eec\u5c06\u9996\u5148\u8be6\u7ec6\u4ecb\u7ecdDeepSeek-V3\u6a21\u578b\u67b6\u6784\uff08\u7b2c2\u8282\uff09\u3002\u968f\u540e\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u6211\u4eec\u7684\u57fa\u7840\u8bbe\u65bd\uff0c\u5305\u62ec\u6211\u4eec\u7684\u8ba1\u7b97\u96c6\u7fa4\u3001\u8bad\u7ec3\u6846\u67b6\u3001\u5bf9FP8\u8bad\u7ec3\u7684\u652f\u6301\u3001\u63a8\u7406\u90e8\u7f72\u7b56\u7565\u4ee5\u53ca\u6211\u4eec\u5bf9\u672a\u6765\u786c\u4ef6\u8bbe\u8ba1\u7684\u5efa\u8bae\uff08\u7b2c3\u8282\uff09\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u63cf\u8ff0\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u5305\u62ec\u8bad\u7ec3\u6570\u636e\u7684\u6784\u5efa\u3001\u8d85\u53c2\u6570\u8bbe\u7f6e\u3001\u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u6280\u672f\u3001\u76f8\u5173\u7684\u8bc4\u4f30\u4ee5\u53ca\u4e00\u4e9b\u8ba8\u8bba\uff08\u7b2c4\u8282\uff09\u3002\u6b64\u540e\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u6211\u4eec\u5728\u8bad\u7ec3\u540e\u7684\u52aa\u529b\uff0c\u5305\u62ec\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u3001\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u3001\u76f8\u5e94\u7684\u8bc4\u4f30\u548c\u8ba8\u8bba\uff08\u7b2c5\u8282\uff09\u3002\u6700\u540e\uff0c\u6211\u4eec\u5c06\u603b\u7ed3\u8fd9\u9879\u5de5\u4f5c\uff0c\u8ba8\u8bbaDeepSeek-V3\u73b0\u6709\u7684\u5c40\u9650\u6027\uff0c\u5e76\u4e3a\u672a\u6765\u7684\u7814\u7a76\u63d0\u51fa\u6f5c\u5728\u7684\u65b9\u5411\uff08\u7b2c6\u8282\uff09\u3002<\/p>\n<h3>2 \u67b6\u6784<\/h3>\n<p>\u6211\u4eec\u9996\u5148\u4ecb\u7ecdDeepSeek-V3\u7684\u57fa\u672c\u67b6\u6784\u3002\u5176\u6838\u5fc3\u7279\u70b9\u662f\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\uff0c\u7528\u4e8e\u5b9e\u73b0\u9ad8\u6548\u63a8\u7406\uff0c\u4ee5\u53caDeepSeekMoE\uff0c\u7528\u4e8e\u5b9e\u73b0\u7ecf\u6d4e\u9ad8\u6548\u7684\u8bad\u7ec3\u3002\u968f\u540e\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u591a\u7b26\u5143\u9884\u6d4b\uff08MTP\uff09\u8bad\u7ec3\u76ee\u6807\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u8fd9\u4e00\u76ee\u6807\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u6a21\u578b\u5728\u8bc4\u4f30\u57fa\u51c6\u4e0a\u7684\u6574\u4f53\u6027\u80fd\u3002\u5bf9\u4e8e\u5176\u4ed6\u672a\u660e\u786e\u63d0\u53ca\u7684\u7ec6\u8282\uff0cDeepSeek-V3\u5747\u9075\u5faaDeepSeek-V2\u7684\u8bbe\u7f6e\u3002<\/p>\n<h4>2.1 \u57fa\u672c\u67b6\u6784<\/h4>\n<p>DeepSeek-V3\u7684\u57fa\u672c\u67b6\u6784\u4ecd\u7136\u57fa\u4e8eTransformer\u6846\u67b6\u3002\u4e3a\u4e86\u5b9e\u73b0\u9ad8\u6548\u7684\u63a8\u7406\u548c\u7ecf\u6d4e\u7684\u8bad\u7ec3\uff0cDeepSeek-V3\u91c7\u7528\u4e86\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u548cDeepSeekMoE\uff0c\u8fd9\u4e9b\u6280\u672f\u5df2\u7ecf\u5728DeepSeek-V2\u4e2d\u5f97\u5230\u4e86\u5145\u5206\u9a8c\u8bc1\u3002\u4e0eDeepSeek-V2\u76f8\u6bd4\uff0c\u552f\u4e00\u7684\u4f8b\u5916\u662f\uff0c\u6211\u4eec\u989d\u5916\u5f15\u5165\u4e86\u4e00\u79cd\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff0c\u7528\u4e8eDeepSeekMoE\uff0c\u4ee5\u51cf\u8f7b\u56e0\u9700\u8981\u4fdd\u8bc1\u8d1f\u8f7d\u5747\u8861\u800c\u5bfc\u81f4\u7684\u6027\u80fd\u4e0b\u964d\u3002\u56fe2\u5c55\u793a\u4e86DeepSeek-V3\u7684\u57fa\u672c\u67b6\u6784\uff0c\u6211\u4eec\u5c06\u5728\u672c\u8282\u4e2d\u7b80\u8981\u56de\u987eMLA\u548cDeepSeekMoE\u7684\u7ec6\u8282\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250206114129122.png\" alt=\"\" \/><\/p>\n<p>\u56fe2 DeepSeek-V3\u57fa\u672c\u67b6\u6784\u56fe\u793a\u3002\u9075\u5faaDeepSeek-V2\uff0c\u6211\u4eec\u91c7\u7528MLA\u548cDeepSeekMoE\u6765\u5b9e\u73b0\u9ad8\u6548\u7684\u63a8\u7406\u548c\u7ecf\u6d4e\u9ad8\u6548\u7684\u8bad\u7ec3\u3002<\/p>\n<h5>2.1.1 \u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b<\/h5>\n<p>\u5bf9\u4e8e\u6ce8\u610f\u529b\u673a\u5236\uff0cDeepSeek-V3 \u91c7\u7528\u4e86\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09\u67b6\u6784\u3002\u8bbe<span class=\"katex-eq\" data-katex-display=\"false\">d<\/span>\u8868\u793a\u5d4c\u5165\u7ef4\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">n_h<\/span>\u8868\u793a\u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">d_h<\/span>\u8868\u793a\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">h_t\\in\\mathbb{R}^d<\/span>\u8868\u793a\u7ed9\u5b9a\u6ce8\u610f\u529b\u5c42\u4e2d\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\u4e2a\u7b26\u5143\u7684\u6ce8\u610f\u529b\u8f93\u5165\u3002MLA\u7684\u6838\u5fc3\u662f\u6ce8\u610f\u529b\u952e\u548c\u503c\u7684\u4f4e\u79e9\u8054\u5408\u538b\u7f29(low-rank joint compression)\uff0c\u4ee5\u51cf\u5c11\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u952e\u503c\uff08KV\uff09\u7f13\u5b58\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250206114907448.jpg\" alt=\"\" \/><br \/>\u5176\u4e2d\uff1a<br \/><span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{c}_{t}^{kv} \\in \\mathbb{R}^{d_c}<\/span>\u662f\u952e\u548c\u503c\u7684\u538b\u7f29\u6f5c\u5728\u5411\u91cf\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">d_{c} (\\ll d_hn_h)<\/span>\u8868\u793aKV\u538b\u7f29\u7ef4\u5ea6\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">W^{UK},W^{UV} \\in \\mathbb{R}^{d_hn_h\u00d7d_c}<\/span>\u5206\u522b\u662f\u952e\u548c\u503c\u7684\u964d\u7ef4\u548c\u5347\u7ef4\u6295\u5f71\u77e9\u9635\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}^{KR} \\in \\mathbb{R}^{d_h^R\u00d7d}<\/span> \u662f\u7528\u4e8e\u751f\u6210\u643a\u5e26\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165\uff08RoPE\uff09\u7684\u89e3\u8026\u952e\u7684\u77e9\u9635\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">\\text{RoPE}(\\cdot)<\/span>\u8868\u793a\u5e94\u7528RoPE\u77e9\u9635\u7684\u64cd\u4f5c\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">[\\cdot; \\cdot]<\/span>\u8868\u793a\u5411\u91cf\u8fde\u63a5\u3002<br \/>\u8bf7\u6ce8\u610f\uff0c\u5bf9\u4e8eMLA\uff0c\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u53ea\u9700\u8981\u7f13\u5b58\u84dd\u8272\u65b9\u6846\u5185\u7684\u5411\u91cf\uff08\u5373<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{c}_{t}^{KV}<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{k}_{t}^{R}<\/span>\uff09\uff0c\u8fd9\u5728\u4fdd\u6301\u4e0e\u6807\u51c6\u591a\u5934\u6ce8\u610f\u529b\uff08Multi-Head Attention,MHA\uff09\u76f8\u5f53\u7684\u6027\u80fd\u7684\u540c\u65f6\uff0c\u663e\u8457\u51cf\u5c11\u4e86KV\u7f13\u5b58\u3002<br \/>\u5bf9\u4e8e\u6ce8\u610f\u529b\u67e5\u8be2\uff0c\u6211\u4eec\u4e5f\u6267\u884c\u4f4e\u79e9\u538b\u7f29\uff0c\u4ee5\u51cf\u5c11\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u6fc0\u6d3b\u5185\u5b58\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250206121220460.jpg\" alt=\"\" \/><br \/>\u5176\u4e2d\uff1a<br \/><span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{c}_{t}^{Q} \\in \\mathbb{R}^{d_{c}&#039;}<\/span>\u662f\u67e5\u8be2\u7684\u538b\u7f29\u6f5c\u5728\u5411\u91cf\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">d_{c}&#039;(\\ll d_hn_h)<\/span>\u8868\u793a\u67e5\u8be2\u538b\u7f29\u7ef4\u5ea6\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}^{DQ} \\in \\mathbb{R}^{d_{c}&#039; \\times d}<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}^{UQ} \\in \\mathbb{R}^{d_hn_h \\times d_{q}&#039;}<\/span>\u5206\u522b\u662f\u67e5\u8be2\u7684\u964d\u7ef4\u548c\u5347\u7ef4\u6295\u5f71\u77e9\u9635\uff1b<br \/><span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}^{QR} \\in \\mathbb{R}^{d_h^Rn_h \\times d{q}&#039;}<\/span>\u662f\u7528\u4e8e\u751f\u6210\u643a\u5e26RoPE\u7684\u89e3\u8026\u67e5\u8be2\u7684\u77e9\u9635\u3002<br \/>\u6700\u7ec8\uff0c\u6ce8\u610f\u529b\u67e5\u8be2(<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{q}_{t,i}<\/span>)\u3001\u952e(<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{k}_{j,i}<\/span>)\u548c\u503c(<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{v}_{j,i}^{C}<\/span>)\u7ec4\u5408\u8d77\u6765\u4ea7\u751f\u6700\u7ec8\u7684\u6ce8\u610f\u529b\u8f93\u51fa<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{u}_{t}<\/span>\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250206122521585.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}_{O} \\in \\mathbb{R}^{d \\times d_hn_h}<\/span>\u8868\u793a\u8f93\u51fa\u6295\u5f71\u77e9\u9635\u3002<\/p>\n<h5>2.1.2 \u5177\u6709\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861\u7684DeepSeekMoE<\/h5>\n<h6>DeepSeekMoE\u7684\u57fa\u672c\u67b6\u6784<\/h6>\n<p>\u5bf9\u4e8e\u524d\u9988\u7f51\u7edc\uff08FFNs\uff09\uff0cDeepSeek-V3 \u91c7\u7528\u4e86 DeepSeekMoE \u67b6\u6784\u3002\u4e0e GShard \u7b49\u4f20\u7edf\u7684 MoE \u67b6\u6784\u76f8\u6bd4\uff0cDeepSeekMoE \u4f7f\u7528\u66f4\u7ec6\u7c92\u5ea6\u7684\u4e13\u5bb6\uff0c\u5e76\u5c06\u4e00\u4e9b\u4e13\u5bb6\u9694\u79bb\u4e3a\u5171\u4eab\u4e13\u5bb6\u3002\u8bbe<span class=\"katex-eq\" data-katex-display=\"false\">u_t<\/span>\u8868\u793a\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\u4e2a\u7b26\u5143\u7684FFN\u8f93\u5165\uff0c\u6211\u4eec\u6309\u4ee5\u4e0b\u65b9\u5f0f\u8ba1\u7b97FFN\u8f93\u51fa<span class=\"katex-eq\" data-katex-display=\"false\">h_t&#039;<\/span>:<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250210081537743.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">N_s<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">N_r<\/span>\u5206\u522b\u8868\u793a\u5171\u4eab\u4e13\u5bb6\u548c\u8def\u7531\u4e13\u5bb6\u7684\u6570\u91cf\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">FNN_{i}^{(s)}(\\cdot)<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">FNN_{i}^{(r)}(\\cdot)<\/span>\u5206\u522b\u8868\u793a\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u5171\u4eab\u4e13\u5bb6\u548c\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u8def\u7531\u4e13\u5bb6\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">K_r<\/span>\u8868\u793a\u6fc0\u6d3b\u7684\u8def\u7531\u4e13\u5bb6\u7684\u6570\u91cf\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">g_{i,t}<\/span>\u662f\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u4e13\u5bb6\u7684\u95e8\u63a7\u503c\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">s_{i,t}<\/span>\u662f\u7b26\u5143\u5230\u4e13\u5bb6\u7684\u4eb2\u548c\u529b\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">e_i<\/span>\u662f\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u8def\u7531\u4e13\u5bb6\u7684\u8d28\u5fc3\u5411\u91cf\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">Topk(\u22c5,K)<\/span>\u8868\u793a\u5305\u542b\u9488\u5bf9\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u7b26\u5143\u548c\u6240\u6709\u8def\u7531\u4e13\u5bb6\u8ba1\u7b97\u51fa\u7684\u4eb2\u548c\u529b\u5206\u6570\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">K<\/span>\u4e2a\u6700\u9ad8\u5206\u6570\u7684\u96c6\u5408\u3002\u4e0eDeepSeek-V2\u7565\u6709\u4e0d\u540c\u7684\u662f\uff0cDeepSeek-V3\u4f7f\u7528Sigmoid\u51fd\u6570\u8ba1\u7b97\u4eb2\u548c\u529b\u5206\u6570\uff0c\u5e76\u5728\u6240\u6709\u9009\u62e9\u7684\u4eb2\u548c\u529b\u5206\u6570\u4e2d\u5e94\u7528\u5f52\u4e00\u5316\u6765\u751f\u6210\u95e8\u63a7\u503c\u3002<\/p>\n<h6>\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861<\/h6>\n<p>\u5bf9\u4e8e MoE \u6a21\u578b\uff0c\u4e0d\u5747\u8861\u7684\u4e13\u5bb6\u8d1f\u8f7d\u4f1a\u5bfc\u81f4\u8def\u7531\u5d29\u6e83\uff0c\u5e76\u5728\u4e13\u5bb6\u5e76\u884c\u573a\u666f\u4e2d\u964d\u4f4e\u8ba1\u7b97\u6548\u7387\u3002\u5e38\u89c4\u89e3\u51b3\u65b9\u6848\u901a\u5e38\u4f9d\u8d56\u4e8e\u8f85\u52a9\u635f\u5931\u6765\u907f\u514d\u4e0d\u5747\u8861\u8d1f\u8f7d\u3002\u7136\u800c\uff0c\u8fc7\u5927\u7684\u8f85\u52a9\u635f\u5931\u4f1a\u635f\u5bb3\u6a21\u578b\u6027\u80fd\u3002\u4e3a\u4e86\u5728\u8d1f\u8f7d\u5747\u8861\u548c\u6a21\u578b\u6027\u80fd\u4e4b\u95f4\u53d6\u5f97\u66f4\u597d\u7684\u5e73\u8861\uff0c\u6211\u4eec\u5f00\u521b\u4e86\u4e00\u79cd\u65e0\u8f85\u52a9\u635f\u5931\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\u6765\u786e\u4fdd\u8d1f\u8f7d\u5747\u8861\u3002<br \/>\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u4e3a\u6bcf\u4e2a\u4e13\u5bb6\u5f15\u5165\u4e00\u4e2a\u504f\u5dee\u9879<span class=\"katex-eq\" data-katex-display=\"false\">b_i<\/span>\uff0c\u5e76\u5c06\u5176\u6dfb\u52a0\u5230\u76f8\u5e94\u7684\u4eb2\u548c\u529b\u5206\u6570<span class=\"katex-eq\" data-katex-display=\"false\">s_{i,t}<\/span>\u4e2d\u4ee5\u786e\u5b9atop-K\u8def\u7531\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250210090254695.jpg\" alt=\"\" \/><br \/>\u6ce8\u610f\uff0c\u504f\u5dee\u9879\u4ec5\u7528\u4e8e\u8def\u7531\u3002\u95e8\u63a7\u503c\u5c06\u4e58\u4ee5FFN\u8f93\u51fa\uff0c\u4ecd\u7136\u6765\u81ea\u539f\u59cb\u4eb2\u548c\u529b\u5206\u6570<span class=\"katex-eq\" data-katex-display=\"false\">s_{i,t}<\/span>\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u6301\u7eed\u76d1\u63a7\u6bcf\u4e2a\u8bad\u7ec3\u6b65\u9aa4\u6574\u4e2a\u6279\u6b21\u7684\u4e13\u5bb6\u8d1f\u8f7d\u3002\u5728\u6bcf\u4e2a\u6b65\u9aa4\u7ed3\u675f\u65f6\uff0c\u5982\u679c\u5176\u5bf9\u5e94\u7684\u4e13\u5bb6\u8fc7\u8f7d\uff0c\u6211\u4eec\u5c06\u504f\u5dee\u9879\u51cf\u5c11<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\uff1b\u5982\u679c\u5176\u5bf9\u5e94\u7684\u4e13\u5bb6\u8d1f\u8f7d\u4e0d\u8db3\uff0c\u6211\u4eec\u5c06\u504f\u5dee\u9879\u589e\u52a0<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\uff0c\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\u662f\u4e00\u4e2a\u79f0\u4e3a\u504f\u5dee\u66f4\u65b0\u901f\u5ea6\u7684\u8d85\u53c2\u6570\u3002\u901a\u8fc7\u52a8\u6001\u8c03\u6574\uff0cDeepSeek-V3\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u5747\u8861\u7684\u4e13\u5bb6\u8d1f\u8f7d\uff0c\u5e76\u53d6\u5f97\u6bd4\u901a\u8fc7\u7eaf\u8f85\u52a9\u635f\u5931\u6765\u9f13\u52b1\u8d1f\u8f7d\u5747\u8861\u7684\u6a21\u578b\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n<h6>\u8865\u5145\u7684\u5e8f\u5217\u7ea7\u8f85\u52a9\u635f\u5931<\/h6>\n<p>\u5c3d\u7ba1DeepSeek-V3\u4e3b\u8981\u4f9d\u8d56\u4e8e\u65e0\u8f85\u52a9\u635f\u5931\u7684\u7b56\u7565\u6765\u8fdb\u884c\u8d1f\u8f7d\u5747\u8861\uff0c\u4f46\u4e3a\u4e86\u9632\u6b62\u4efb\u4f55\u5355\u4e2a\u5e8f\u5217\u5185\u7684\u6781\u7aef\u4e0d\u5e73\u8861\uff0c\u6211\u4eec\u8fd8\u91c7\u7528\u4e86\u4e00\u79cd\u8865\u5145\u7684\u5e8f\u5217\u7ea7\u5e73\u8861\u635f\u5931\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211020050779.jpg\" alt=\"\" \/><br \/>\u5176\u4e2d\uff0c\u5e73\u8861\u56e0\u5b50<span class=\"katex-eq\" data-katex-display=\"false\">\\alpha<\/span>\u662f\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u5bf9\u4e8eDeepSeek-V3\uff0c\u5b83\u5c06\u88ab\u8d4b\u4e88\u4e00\u4e2a\u6781\u5c0f\u7684\u503c\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbb{1}(\u22c5)<\/span>\u8868\u793a\u6307\u793a\u51fd\u6570\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">T<\/span>\u8868\u793a\u5e8f\u5217\u4e2d\u7b26\u5143\u7684\u6570\u91cf\u3002\u57fa\u4e8e\u5e8f\u5217\u7684\u5e73\u8861\u635f\u5931\u9f13\u52b1\u6bcf\u4e2a\u5e8f\u5217\u4e0a\u7684\u4e13\u5bb6\u8d1f\u8f7d\u4fdd\u6301\u5e73\u8861\u3002<\/p>\n<h6>\u8282\u70b9\u9650\u5236\u8def\u7531<\/h6>\n<p>\u4e0eDeepSeek-V2\u4f7f\u7528\u7684\u8bbe\u5907\u9650\u5236\u8def\u7531\u7c7b\u4f3c\uff0cDeepSeek-V3\u4e5f\u4f7f\u7528\u53d7\u9650\u8def\u7531\u673a\u5236\u6765\u9650\u5236\u8bad\u7ec3\u671f\u95f4\u7684\u901a\u4fe1\u6210\u672c\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u6211\u4eec\u786e\u4fdd\u6bcf\u4e2a\u7b26\u5143\u6700\u591a\u4f1a\u88ab\u53d1\u9001\u5230<span class=\"katex-eq\" data-katex-display=\"false\">M<\/span>\u4e2a\u8282\u70b9\uff0c\u8fd9\u4e9b\u8282\u70b9\u6839\u636e\u6bcf\u4e2a\u8282\u70b9\u4e0a\u5206\u5e03\u7684\u4e13\u5bb6\u6700\u9ad8<span class=\"katex-eq\" data-katex-display=\"false\">\\frac{K_r}{M}<\/span>\u4e2a\u4eb2\u548c\u529b\u5206\u6570\u4e4b\u548c\u6765\u9009\u62e9\u3002\u5728\u6b64\u7ea6\u675f\u4e0b\uff0c\u6211\u4eec\u7684MoE\u8bad\u7ec3\u6846\u67b6\u51e0\u4e4e\u53ef\u4ee5\u5b9e\u73b0\u5b8c\u5168\u7684\u8ba1\u7b97-\u901a\u4fe1\u91cd\u53e0\u3002<\/p>\n<h6>\u4e0d\u4e22\u5f03\u7b26\u5143<\/h6>\n<p>\u7531\u4e8e\u6709\u6548\u7684\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff0cDeepSeek-V3\u5728\u5176\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u826f\u597d\u7684\u8d1f\u8f7d\u5e73\u8861\u3002\u56e0\u6b64\uff0cDeepSeek-V3 \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0d\u4f1a\u4e22\u5f03\u4efb\u4f55\u7b26\u5143\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u5b9e\u73b0\u4e86\u7279\u5b9a\u7684\u90e8\u7f72\u7b56\u7565\u4ee5\u786e\u4fdd\u63a8\u7406\u8d1f\u8f7d\u5e73\u8861\uff0c\u56e0\u6b64 DeepSeek-V3\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4e5f\u4e0d\u4f1a\u4e22\u5f03\u7b26\u5143\u3002<\/p>\n<h4>2.2 \u591a\u7b26\u5143\u9884\u6d4b<\/h4>\n<p>\u53d7\u76f8\u5173\u7814\u7a76\u7684\u542f\u53d1\uff0c\u6211\u4eec\u7814\u7a76\u5e76\u4e3aDeepSeek-V3\u8bbe\u7f6e\u4e86\u4e00\u4e2a\u591a\u7b26\u5143\u9884\u6d4b\uff08Multi-Token Prediction\uff0cMTP\uff09\u76ee\u6807\uff0c\u8be5\u76ee\u6807\u5c06\u9884\u6d4b\u8303\u56f4\u6269\u5c55\u5230\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u591a\u4e2a\u672a\u6765\u7b26\u5143\u3002\u4e00\u65b9\u9762\uff0cMTP\u76ee\u6807\u4f7f\u8bad\u7ec3\u4fe1\u53f7\u66f4\u52a0\u5bc6\u96c6\uff0c\u5e76\u53ef\u80fd\u63d0\u9ad8\u6570\u636e\u6548\u7387\u3002\u53e6\u4e00\u65b9\u9762\uff0cMTP\u53ef\u4ee5\u4f7f\u6a21\u578b\u9884\u5148\u89c4\u5212\u5176\u8868\u793a\uff0c\u4ee5\u4fbf\u66f4\u597d\u5730\u9884\u6d4b\u672a\u6765\u7684\u7b26\u5143\u3002\u56fe3\u8bf4\u660e\u4e86\u6211\u4eec\u5bf9MTP\u7684\u5b9e\u73b0\u3002\u4e0e\u67d0\u4e9b\u7814\u7a76\u4f7f\u7528\u72ec\u7acb\u7684\u8f93\u51fa\u5934\u5e76\u884c\u9884\u6d4b\u591a\u4e2a\u9644\u52a0\u7b26\u5143\u4e0d\u540c\uff0c\u6211\u4eec\u987a\u5e8f\u9884\u6d4b\u9644\u52a0\u7b26\u5143\uff0c\u5e76\u5728\u6bcf\u4e2a\u9884\u6d4b\u6df1\u5ea6\u4fdd\u6301\u5b8c\u6574\u7684\u56e0\u679c\u94fe\u3002\u6211\u4eec\u5728\u672c\u8282\u4ecb\u7ecd\u4e86MTP\u5b9e\u73b0\u7684\u7ec6\u8282\u3002<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211021401225.png\" alt=\"\" \/><\/p>\n<p>\u56fe3 \u6211\u4eec\u591a\u7b26\u5143\u9884\u6d4b (MTP) \u5b9e\u73b0\u7684\u793a\u610f\u56fe\u3002 \u6211\u4eec\u4fdd\u7559\u6bcf\u4e2a\u6df1\u5ea6\u6bcf\u4e2a\u7b26\u5143\u9884\u6d4b\u7684\u5b8c\u6574\u56e0\u679c\u94fe\u3002<\/p>\n<h6>MTP\u6a21\u5757<\/h6>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u7684MTP\u5b9e\u73b0\u4f7f\u7528<span class=\"katex-eq\" data-katex-display=\"false\">D<\/span>\u4e2a\u987a\u5e8f\u6a21\u5757\u6765\u9884\u6d4b<span class=\"katex-eq\" data-katex-display=\"false\">D<\/span>\u4e2a\u9644\u52a0\u7b26\u5143\u3002\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u4e2aMTP\u6a21\u5757\u7531\u4e00\u4e2a\u5171\u4eab\u5d4c\u5165\u5c42<span class=\"katex-eq\" data-katex-display=\"false\">Emb(\u22c5)<\/span>\u3001\u4e00\u4e2a\u5171\u4eab\u8f93\u51fa\u5934<span class=\"katex-eq\" data-katex-display=\"false\">OutHead(\u22c5)<\/span>\u3001\u4e00\u4e2aTransformer\u5757 <span class=\"katex-eq\" data-katex-display=\"false\">TRM_k(\u22c5)<\/span>\u548c\u4e00\u4e2a\u6295\u5f71\u77e9\u9635<span class=\"katex-eq\" data-katex-display=\"false\">M_k\\in\\mathbb{R}^{d\u00d72d}<\/span>\u7ec4\u6210\u3002\u5bf9\u4e8e\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u8f93\u5165\u7b26\u5143<span class=\"katex-eq\" data-katex-display=\"false\">t_i<\/span>\uff0c\u5728\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u4e2a\u9884\u6d4b\u6df1\u5ea6\uff0c\u6211\u4eec\u9996\u5148\u7ed3\u5408\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u7b26\u5143\u5728\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">(k-1)<\/span>\u4e2a\u6df1\u5ea6<span class=\"katex-eq\" data-katex-display=\"false\">h_i^{k-1}\\in\\mathbb{R}^d<\/span>\u7684\u8868\u793a\u548c\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">(i+k)<\/span>\u4e2a\u7b26\u5143<span class=\"katex-eq\" data-katex-display=\"false\">Emb(t_{i+k})<\/span>\u7684\u5d4c\u5165\uff0c\u8fdb\u884c\u7ebf\u6027\u6295\u5f71\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211022111504.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">[\\cdot; \\cdot]<\/span>\u8868\u793a\u8fde\u63a5\u3002\u7279\u522b\u662f\uff0c\u5f53<span class=\"katex-eq\" data-katex-display=\"false\">k=1<\/span>\u65f6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{h}_i^{k-1}<\/span>\u6307\u7684\u662f\u4e3b\u6a21\u578b\u7ed9\u51fa\u7684\u8868\u793a\u3002\u8bf7\u6ce8\u610f\uff0c\u5bf9\u4e8e\u6bcf\u4e2aMTP\u6a21\u5757\uff0c\u5176\u5d4c\u5165\u5c42\u4e0e\u4e3b\u6a21\u578b\u5171\u4eab\u3002\u7ec4\u5408\u7684<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{h}_i^{k&#039;}<\/span>\u4f5c\u4e3a\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u4e2a\u6df1\u5ea6\u7684Transformer\u5757\u7684\u8f93\u5165\uff0c\u4ee5\u4ea7\u751f\u5f53\u524d\u6df1\u5ea6<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{h}_i^k<\/span>\u7684\u8f93\u51fa\u8868\u793a\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211022926559.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">T<\/span>\u8868\u793a\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">i:j<\/span>\u8868\u793a\u5207\u7247\u64cd\u4f5c\uff08\u5305\u542b\u5de6\u53f3\u8fb9\u754c\uff09\u3002\u6700\u540e\uff0c\u4ee5<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{h}i^k<\/span>\u4f5c\u4e3a\u8f93\u5165\uff0c\u5171\u4eab\u8f93\u51fa\u5934\u5c06\u8ba1\u7b97\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u4e2a\u9644\u52a0\u9884\u6d4b\u7b26\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">P_{i+1+k}^{k} \\in \\mathbb{R}^V<\/span>\u7684\u6982\u7387\u5206\u5e03\uff0c\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">V<\/span>\u662f\u8bcd\u6c47\u91cf\u5927\u5c0f\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211023220701.jpg\" alt=\"\" \/><\/p>\n<p>\u8f93\u51fa\u5934<span class=\"katex-eq\" data-katex-display=\"false\">\\text{OutHead}(\\cdot)<\/span>\u5c06\u8868\u793a\u7ebf\u6027\u6620\u5c04\u5230logits\uff0c\u968f\u540e\u5e94\u7528<span class=\"katex-eq\" data-katex-display=\"false\">\\text{Softmax}(\\cdot)<\/span>\u51fd\u6570\u6765\u8ba1\u7b97\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u4e2a\u9644\u52a0\u7b26\u5143\u7684\u9884\u6d4b\u6982\u7387\u3002\u6b64\u5916\uff0c\u5bf9\u4e8e\u6bcf\u4e2aMTP\u6a21\u5757\uff0c\u5176\u8f93\u51fa\u5934\u4e0e\u4e3b\u6a21\u578b\u5171\u4eab\u3002\u6211\u4eec\u4fdd\u6301\u9884\u6d4b\u56e0\u679c\u94fe\u7684\u539f\u5219\u7c7b\u4f3c\u4e8e\u67d0\u4e9b\u7814\u7a76\uff0c\u4f46\u5176\u4e3b\u8981\u76ee\u6807\u662f\u63a8\u6d4b\u6027\u89e3\u7801\uff0c\u800c\u6211\u4eec\u5229\u7528MTP\u6539\u8fdb\u8bad\u7ec3\u3002<\/p>\n<h6>MTP\u8bad\u7ec3\u76ee\u6807<\/h6>\n<p>\u5bf9\u4e8e\u6bcf\u4e2a\u9884\u6d4b\u6df1\u5ea6\uff0c\u6211\u4eec\u8ba1\u7b97\u4ea4\u53c9\u71b5\u635f\u5931<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}_{MTP}^k<\/span>\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211024105375.jpg\" alt=\"\" \/><\/p>\n<p>\u5176\u4e2d\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">T<\/span>\u8868\u793a\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">t_i<\/span>\u8868\u793a\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u4f4d\u7f6e\u7684\u771f\u5b9e\u7b26\u5143\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">P_i^k[t_i]<\/span>\u8868\u793a\u7531\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u4e2aMTP\u6a21\u5757\u7ed9\u51fa\u7684<span class=\"katex-eq\" data-katex-display=\"false\">t_i<\/span>\u7684\u5bf9\u5e94\u9884\u6d4b\u6982\u7387\u3002\u6700\u540e\uff0c\u6211\u4eec\u5c06\u6240\u6709\u6df1\u5ea6\u7684MTP\u635f\u5931\u53d6\u5e73\u5747\u503c\uff0c\u5e76\u4e58\u4ee5\u6743\u91cd\u56e0\u5b50<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\uff0c\u5f97\u5230\u6574\u4f53MTP\u635f\u5931<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}_{MTP}<\/span>\uff0c\u8fd9\u4f5c\u4e3aDeepSeek-V3\u7684\u9644\u52a0\u8bad\u7ec3\u76ee\u6807\uff1a<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211024449377.jpg\" alt=\"\" \/><\/p>\n<h6>\u63a8\u7406\u4e2d\u7684MTP<\/h6>\n<p>\u6211\u4eec\u7684MTP\u7b56\u7565\u4e3b\u8981\u65e8\u5728\u63d0\u9ad8\u4e3b\u6a21\u578b\u7684\u6027\u80fd\uff0c\u56e0\u6b64\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u4e22\u5f03MTP\u6a21\u5757\uff0c\u4e3b\u6a21\u578b\u53ef\u4ee5\u72ec\u7acb\u6b63\u5e38\u8fd0\u884c\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u53ef\u4ee5\u5c06\u8fd9\u4e9bMTP\u6a21\u5757\u91cd\u65b0\u7528\u4e8e\u63a8\u6d4b\u6027\u89e3\u7801\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u751f\u6210\u5ef6\u8fdf\u3002<\/p>\n<h3>3 \u57fa\u7840\u8bbe\u65bd<\/h3>\n<h4>3.1 \u8ba1\u7b97\u96c6\u7fa4<\/h4>\n<p>DeepSeek-V3\u7684\u8bad\u7ec3\u662f\u5728\u4e00\u4e2a\u914d\u59072048\u4e2aNVIDIA H800 GPU\u7684\u96c6\u7fa4\u4e0a\u5b8c\u6210\u7684\u3002\u8be5\u96c6\u7fa4\u4e2d\u7684\u6bcf\u4e2a\u8282\u70b9\u5305\u542b8\u4e2aGPU\uff0c\u8fd9\u4e9bGPU\u901a\u8fc7\u8282\u70b9\u5185\u7684NVLink\u548cNVSwitch\u8fdb\u884c\u8fde\u63a5\u3002\u5728\u4e0d\u540c\u8282\u70b9\u4e4b\u95f4\uff0c\u5229\u7528InfiniBand(IB)\u4e92\u8fde\u6765\u4fc3\u8fdb\u901a\u4fe1\u3002<\/p>\n<h4>3.2 \u8bad\u7ec3\u6846\u67b6<\/h4>\n<p>DeepSeek-V3\u7684\u8bad\u7ec3\u7531HAI-LLM\u6846\u67b6\u652f\u6301\uff0c\u8fd9\u662f\u4e00\u4e2a\u7531\u6211\u4eec\u7684\u5de5\u7a0b\u5e08\u4ece\u5934\u5f00\u59cb\u6784\u5efa\u7684\u9ad8\u6548\u8f7b\u91cf\u7ea7\u8bad\u7ec3\u6846\u67b6\u3002\u603b\u4f53\u800c\u8a00DeepSeek-V3\u5e94\u7528\u4e8616\u8def\u6d41\u6c34\u7ebf\u5e76\u884c\uff08Pipeline Parallelism,PP\uff09\uff0c\u8de8\u8d8a8\u4e2a\u8282\u70b9\u768464\u8def\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallelism,EP\uff09\uff0c\u4ee5\u53caZeRO-1\u6570\u636e\u5e76\u884c\uff08Data Parallelism,DP\uff09\u3002<br \/>\u4e3a\u4e86\u4fc3\u8fdbDeepSeek-V3\u7684\u9ad8\u6548\u8bad\u7ec3\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u7ec6\u81f4\u7684\u5de5\u7a0b\u4f18\u5316\u3002\u9996\u5148\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u7528\u4e8e\u9ad8\u6548\u6d41\u6c34\u7ebf\u5e76\u884c\u7684DualPipe\u7b97\u6cd5\u3002\u4e0e\u73b0\u6709\u7684\u6d41\u6c34\u7ebf\u5e76\u884c\u65b9\u6cd5\u76f8\u6bd4\uff0cDualPipe\u7684\u6d41\u6c34\u7ebf\u6c14\u6ce1\u66f4\u5c11\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u91cd\u53e0\u4e86\u524d\u5411\u548c\u540e\u5411\u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u9636\u6bb5\uff0c\u4ece\u800c\u89e3\u51b3\u4e86\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\u5f15\u5165\u7684\u6c89\u91cd\u901a\u4fe1\u5f00\u9500\u7684\u6311\u6218\u3002\u5176\u6b21\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u9ad8\u6548\u7684\u8de8\u8282\u70b9\u5168\u5bf9\u5168\u901a\u4fe1\u5185\u6838\uff0c\u4ee5\u5145\u5206\u5229\u7528IB\u548cNVLink\u5e26\u5bbd\uff0c\u5e76\u8282\u7701\u4e13\u95e8\u7528\u4e8e\u901a\u4fe1\u7684\u6d41\u591a\u5904\u7406\u5668(Streaming Multiprocessors,SM)\u3002\u6700\u540e\uff0c\u6211\u4eec\u4ed4\u7ec6\u4f18\u5316\u4e86\u8bad\u7ec3\u671f\u95f4\u7684\u5185\u5b58\u5360\u7528\uff0c\u4ece\u800c\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u4e0d\u4f7f\u7528\u4ee3\u4ef7\u9ad8\u6602\u7684\u5f20\u91cf\u5e76\u884c(Tensor Parallelism,TP)\u7684\u60c5\u51b5\u4e0b\u8bad\u7ec3DeepSeek-V3\u3002<\/p>\n<h5>3.2.1 DualPipe\u548c\u8ba1\u7b97-\u901a\u4fe1\u91cd\u53e0<\/h5>\n<p>\u5bf9\u4e8eDeepSeek-V3\uff0c\u8de8\u8282\u70b9\u4e13\u5bb6\u5e76\u884c\u5f15\u5165\u7684\u901a\u4fe1\u5f00\u9500\u5bfc\u81f4\u8ba1\u7b97\u4e0e\u901a\u4fe1\u6bd4\u7387\u4f4e\u6548\uff0c\u7ea6\u4e3a1:1\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u540d\u4e3aDualPipe\u7684\u521b\u65b0\u6d41\u6c34\u7ebf\u5e76\u884c\u7b97\u6cd5\u3002\u8be5\u7b97\u6cd5\u4e0d\u4ec5\u901a\u8fc7\u6709\u6548\u5730\u91cd\u53e0\u524d\u5411\u548c\u540e\u5411\u8ba1\u7b97-\u901a\u4fe1\u9636\u6bb5\u6765\u52a0\u901f\u6a21\u578b\u8bad\u7ec3\uff0c\u8fd8\u51cf\u5c11\u4e86\u6d41\u6c34\u7ebf\u6c14\u6ce1\u3002<br \/>DualPipe\u7684\u6838\u5fc3\u601d\u60f3\u662f\u91cd\u53e0\u4e00\u5bf9\u5355\u72ec\u7684\u524d\u5411\u548c\u540e\u5411\u5757\u4e2d\u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u6bcf\u4e2a\u5757\u5212\u5206\u4e3a\u56db\u4e2a\u4e3b\u8981\u7ec4\u4ef6\uff1a<br \/>\u6ce8\u610f\u529b\uff08Attention\uff09\u3001\u5168\u5bf9\u5168\u8c03\u5ea6\uff08All-to-All Sharding\uff09\u3001MLP\uff08\u591a\u5c42\u611f\u77e5\u673a\uff09\u3001\u5168\u5bf9\u5168\u7ec4\u5408\uff08All-to-All Combining\uff09\u3002\u7279\u522b\u5730\uff0c\u5bf9\u4e8e\u540e\u5411\u5757\uff0c\u6ce8\u610f\u529b\u548cMLP\u8fdb\u4e00\u6b65\u88ab\u5206\u6210\u4e24\u90e8\u5206\uff1a\u8f93\u5165\u7684\u53cd\u5411\u548c\u6743\u91cd\u7684\u53cd\u5411\uff0c\u7c7b\u4f3c\u4e8eZeroBubble\u4e2d\u7684\u65b9\u6cd5\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u6709\u4e00\u4e2aPP\u901a\u4fe1\u7ec4\u4ef6\u3002 \u5982\u56fe4\u6240\u793a\uff0c\u5bf9\u4e8e\u4e00\u5bf9\u524d\u5411\u548c\u540e\u5411\u5757\uff0c\u6211\u4eec\u91cd\u65b0\u6392\u5217\u8fd9\u4e9b\u7ec4\u4ef6\u5e76\u624b\u52a8\u8c03\u6574\u4e13\u7528\u4e8e\u901a\u4fe1\u4e0e\u8ba1\u7b97\u7684GPU SM\u6bd4\u4f8b\u3002\u5728\u8fd9\u79cd\u91cd\u53e0\u7b56\u7565\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u786e\u4fdd\u5728\u6267\u884c\u8fc7\u7a0b\u4e2d\u6240\u6709\u5230\u6240\u6709\u548cPP\u901a\u4fe1\u90fd\u53ef\u4ee5\u5b8c\u5168\u9690\u85cf\u3002\u9274\u4e8e\u9ad8\u6548\u7684\u91cd\u53e0\u7b56\u7565\uff0c\u5b8c\u6574\u7684DualPipe\u8c03\u5ea6\u5982\u56fe5\u6240\u793a\u3002\u5b83\u91c7\u7528\u53cc\u5411\u6d41\u6c34\u7ebf\u8c03\u5ea6\uff0c\u540c\u65f6\u4ece\u6d41\u6c34\u7ebf\u7684\u4e24\u7aef\u9988\u9001\u5fae\u6279\u6b21\uff0c\u5e76\u4e14\u53ef\u4ee5\u5b8c\u5168\u91cd\u53e0\u5f88\u5927\u4e00\u90e8\u5206\u901a\u4fe1\u3002\u8fd9\u79cd\u91cd\u53e0\u8fd8\u786e\u4fdd\uff0c\u968f\u7740\u6a21\u578b\u8fdb\u4e00\u6b65\u6269\u5c55\uff0c\u53ea\u8981\u6211\u4eec\u4fdd\u6301\u6052\u5b9a\u7684\u8ba1\u7b97\u4e0e\u901a\u4fe1\u6bd4\u7387\uff0c\u6211\u4eec\u4ecd\u7136\u53ef\u4ee5\u5728\u8282\u70b9\u4e4b\u95f4\u4f7f\u7528\u7ec6\u7c92\u5ea6\u7684\u4e13\u5bb6\uff0c\u540c\u65f6\u5b9e\u73b0\u63a5\u8fd1\u96f6\u7684\u5168\u90e8\u5230\u5168\u90e8\u901a\u4fe1\u5f00\u9500\u3002<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211025717844.png\" alt=\"\" \/><\/p>\n<p>\u56fe4\uff1a\u4e00\u5bf9\u5355\u72ec\u7684\u524d\u5411\u548c\u540e\u5411\u5757\u7684\u91cd\u53e0\u7b56\u7565\uff08Transformer\u5757\u7684\u8fb9\u754c\u672a\u5bf9\u9f50\uff09\u3002\u6a59\u8272\u8868\u793a\u524d\u5411\uff0c\u7eff\u8272\u8868\u793a\u201c\u8f93\u5165\u7684\u53cd\u5411\u201d\uff0c\u84dd\u8272\u8868\u793a\u201c\u6743\u91cd\u7684\u53cd\u5411\u201d\uff0c\u7d2b\u8272\u8868\u793a\u6d41\u6c34\u7ebf\u5e76\u884c\u901a\u4fe1\uff0c\u7ea2\u8272\u8868\u793a\u5c4f\u969c\u3002\u5168\u5bf9\u5168\u901a\u4fe1\u548c\u6d41\u6c34\u7ebf\u5e76\u884c\u901a\u4fe1\u90fd\u53ef\u4ee5\u5b8c\u5168\u9690\u85cf\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211025741533.png\" alt=\"\" \/><\/p>\n<p>\u56fe5\uff1a8\u4e2aPP\u79e9\u548c\u4e24\u4e2a\u65b9\u5411\u4e0a20\u4e2a\u5fae\u6279\u6b21\u7684DualPipe\u8c03\u5ea6\u793a\u4f8b\u3002\u53cd\u5411\u7684\u5fae\u6279\u6b21\u4e0e\u6b63\u5411\u7684\u5fae\u6279\u6b21\u5bf9\u79f0\uff0c\u4e3a\u4e86\u7b80\u5316\u8bf4\u660e\uff0c\u6211\u4eec\u7701\u7565\u4e86\u5b83\u4eec\u7684\u6279\u6b21ID\u3002\u7531\u5171\u4eab\u9ed1\u8fb9\u5305\u56f4\u7684\u4e24\u4e2a\u5355\u5143\u683c\u5177\u6709\u76f8\u4e92\u91cd\u53e0\u7684\u8ba1\u7b97\u548c\u901a\u4fe1\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u5373\u4f7f\u5728\u6ca1\u6709\u5927\u91cf\u901a\u4fe1\u8d1f\u62c5\u7684\u66f4\u4e00\u822c\u60c5\u51b5\u4e0b\uff0cDualPipe\u4ecd\u7136\u8868\u73b0\u51fa\u6548\u7387\u4f18\u52bf\u3002\u5728\u88682\u4e2d\uff0c\u6211\u4eec\u603b\u7ed3\u4e86\u4e0d\u540c\u6d41\u6c34\u7ebf\u5e76\u884c\uff08PP\uff09\u65b9\u6cd5\u7684\u6d41\u6c34\u7ebf\u6c14\u6ce1\u548c\u5185\u5b58\u4f7f\u7528\u60c5\u51b5\u3002\u5982\u8868\u6240\u793a\uff0c\u4e0eZB1P\u548c1F1B\u76f8\u6bd4\uff0cDualPipe\u663e\u8457\u51cf\u5c11\u4e86\u6d41\u6c34\u7ebf\u6c14\u6ce1\uff0c\u800c\u5cf0\u503c\u6fc0\u6d3b\u5185\u5b58\u4ec5\u589e\u52a0\u4e86<span class=\"katex-eq\" data-katex-display=\"false\">\\frac{1}{PP}<\/span>\u500d\u3002\u5c3d\u7ba1DualPipe\u9700\u8981\u4fdd\u7559\u6a21\u578b\u53c2\u6570\u7684\u4e24\u4e2a\u526f\u672c\uff0c\u4f46\u8fd9\u4e0d\u4f1a\u663e\u8457\u589e\u52a0\u5185\u5b58\u6d88\u8017\uff0c\u56e0\u4e3a\u6211\u4eec\u5728\u8bad\u7ec3\u671f\u95f4\u4f7f\u7528\u4e86\u8f83\u5927\u7684\u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u5927\u5c0f\u3002\u4e0eChimera\u76f8\u6bd4\uff0cDualPipe\u53ea\u9700\u8981\u6d41\u6c34\u7ebf\u9636\u6bb5\u548c\u5fae\u6279\u6b21\u80fd\u591f\u88ab2\u6574\u9664\uff0c\u800c\u4e0d\u9700\u8981\u5fae\u6279\u6b21\u80fd\u591f\u88ab\u6d41\u6c34\u7ebf\u9636\u6bb5\u6574\u9664\u3002\u6b64\u5916\uff0c\u5bf9\u4e8eDualPipe\uff0c\u6c14\u6ce1\u548c\u6fc0\u6d3b\u5185\u5b58\u90fd\u4e0d\u4f1a\u968f\u7740\u5fae\u6279\u6b21\u6570\u91cf\u7684\u589e\u52a0\u800c\u589e\u52a0\u3002<br \/><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250211030233571.jpg\" alt=\"\" \/><\/p>\n<p>\u88682\uff1a \u4e0d\u540c\u6d41\u6c34\u7ebf\u5e76\u884c\u65b9\u6cd5\u7684\u6d41\u6c34\u7ebf\u963b\u585e\u548c\u5185\u5b58\u4f7f\u7528\u60c5\u51b5\u6bd4\u8f83\u3002<span class=\"katex-eq\" data-katex-display=\"false\">F<\/span>\u8868\u793a\u524d\u5411\u5757\u7684\u6267\u884c\u65f6\u95f4\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">B<\/span>\u8868\u793a\u5b8c\u6574\u540e\u5411\u5757\u7684\u6267\u884c\u65f6\u95f4\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">W<\/span>\u8868\u793a\u201c\u6743\u91cd\u540e\u5411\u201d\u5757\u7684\u6267\u884c\u65f6\u95f4\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">F&amp;B<\/span>\u8868\u793a\u4e24\u4e2a\u76f8\u4e92\u91cd\u53e0\u7684\u524d\u5411\u548c\u540e\u5411\u5757\u7684\u6267\u884c\u65f6\u95f4\u3002<\/p>\n<h5>3.2.2 \u9ad8\u6548\u5b9e\u73b0\u8de8\u8282\u70b9\u5168\u5bf9\u5168\u901a\u4fe1<\/h5>\n<p>\u4e3a\u4e86\u786e\u4fddDualPipe\u5177\u6709\u8db3\u591f\u7684\u8ba1\u7b97\u6027\u80fd\uff0c\u6211\u4eec\u5b9a\u5236\u4e86\u9ad8\u6548\u7684\u8de8\u8282\u70b9\u5168\u5bf9\u5168\u901a\u4fe1\u5185\u6838\uff08\u5305\u62ec\u8c03\u5ea6\u548c\u7ec4\u5408\uff09\uff0c\u4ee5\u8282\u7701\u4e13\u7528\u4e8e\u901a\u4fe1\u7684\u6d41\u591a\u5904\u7406\u5668\uff08Streaming Multiprocessors\uff0cSM\uff09\u6570\u91cf\u3002\u5185\u6838\u7684\u5b9e\u73b0\u4e0eMoE\u95e8\u63a7\u7b97\u6cd5\u548c\u6211\u4eec\u96c6\u7fa4\u7684\u7f51\u7edc\u62d3\u6251\u7ed3\u6784\u5171\u540c\u8bbe\u8ba1\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5728\u6211\u4eec\u7684\u96c6\u7fa4\u4e2d\uff0c\u8de8\u8282\u70b9GPU\u901a\u8fc7InfiniBand(IB)\u5b8c\u5168\u4e92\u8fde\uff0c\u8282\u70b9\u5185\u901a\u4fe1\u901a\u8fc7NVLink\u5904\u7406\u3002NVLink\u63d0\u4f9b160 GB\/s\u7684\u5e26\u5bbd\uff0c\u5927\u7ea6\u662fIB (50 GB\/s)\u76843.2\u500d\u3002<br \/>\u4e3a\u4e86\u6709\u6548\u5229\u7528IB\u548cNVLink\u7684\u4e0d\u540c\u5e26\u5bbd\uff0c\u6211\u4eec\u5c06\u6bcf\u4e2a\u7b26\u5143\u6700\u591a\u8c03\u5ea6\u52304\u4e2a\u8282\u70b9\uff0c\u4ece\u800c\u51cf\u5c11IB\u901a\u4fe1\u91cf\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u7b26\u5143\uff0c\u5728\u5176\u8def\u7531\u51b3\u7b56\u505a\u51fa\u540e\uff0c\u5b83\u5c06\u9996\u5148\u901a\u8fc7IB\u4f20\u8f93\u5230\u76ee\u6807\u8282\u70b9\u4e0a\u5177\u6709\u76f8\u540c\u8282\u70b9\u5185\u7d22\u5f15\u7684GPU\u3002\u4e00\u65e6\u5230\u8fbe\u76ee\u6807\u8282\u70b9\uff0c\u6211\u4eec\u5c06\u52aa\u529b\u786e\u4fdd\u5b83\u901a\u8fc7NVLink\u5373\u65f6\u8f6c\u53d1\u5230\u627f\u8f7d\u5176\u76ee\u6807\u4e13\u5bb6\u7684\u7279\u5b9aGPU\uff0c\u800c\u4e0d\u4f1a\u88ab\u968f\u540e\u5230\u8fbe\u7684\u7b26\u5143\u963b\u585e\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u901a\u8fc7IB\u548cNVLink\u7684\u901a\u4fe1\u5b8c\u5168\u91cd\u53e0\uff0c\u6bcf\u4e2a\u7b26\u5143\u53ef\u4ee5\u6709\u6548\u5730\u9009\u62e9\u6bcf\u4e2a\u8282\u70b9\u5e73\u57473.2\u4e2a\u4e13\u5bb6\uff0c\u800c\u4e0d\u4f1a\u4ea7\u751fNVLink\u7684\u989d\u5916\u5f00\u9500\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5c3d\u7ba1DeepSeek-V3\u5728\u5b9e\u8df5\u4e2d\u53ea\u9009\u62e98\u4e2a\u8def\u7531\u4e13\u5bb6\uff0c\u4f46\u5b83\u53ef\u4ee5\u5c06\u6b64\u6570\u91cf\u6700\u591a\u6269\u5c55\u523013\u4e2a\u4e13\u5bb6\uff084\u4e2a\u8282\u70b9<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>3.2\u4e2a\u4e13\u5bb6\/\u8282\u70b9\uff09\uff0c\u540c\u65f6\u4fdd\u6301\u76f8\u540c\u7684\u901a\u4fe1\u6210\u672c\u3002\u603b\u4f53\u800c\u8a00\uff0c\u5728\u8fd9\u79cd\u901a\u4fe1\u7b56\u7565\u4e0b\uff0c\u53ea\u9700\u898120\u4e2aSM\u5c31\u80fd\u5145\u5206\u5229\u7528IB\u548cNVLink\u7684\u5e26\u5bbd\u3002<br \/>\u8be6\u7ec6\u6765\u8bf4\uff0c\u6211\u4eec\u91c7\u7528warp\u4e13\u4e1a\u5316\u6280\u672f\uff0c\u5e76\u5c0620\u4e2aSM\u5206\u621010\u4e2a\u901a\u4fe1\u901a\u9053\u3002\u5728\u8c03\u5ea6\u8fc7\u7a0b\u4e2d\uff0c(1)IB\u53d1\u9001\uff0c(2)IB\u5230NVLink\u8f6c\u53d1\uff0c\u4ee5\u53ca(3)NVLink\u63a5\u6536\u5206\u522b\u7531\u5404\u81ea\u7684warp\u5904\u7406\u3002\u5206\u914d\u7ed9\u6bcf\u4e2a\u901a\u4fe1\u4efb\u52a1\u7684warp\u6570\u4f1a\u6839\u636e\u6240\u6709SM\u4e0a\u7684\u5b9e\u9645\u5de5\u4f5c\u8d1f\u8f7d\u52a8\u6001\u8c03\u6574\u3002\u540c\u6837\uff0c\u5728\u5408\u5e76\u8fc7\u7a0b\u4e2d\uff0c(1)NVLink\u53d1\u9001\uff0c(2)NVLink\u5230IB\u7684\u8f6c\u53d1\u548c\u7d2f\u79ef\uff0c\u4ee5\u53ca(3)IB\u63a5\u6536\u548c\u7d2f\u79ef\u4e5f\u7531\u52a8\u6001\u8c03\u6574\u7684warp\u5904\u7406\u3002\u6b64\u5916\uff0c\u8c03\u5ea6\u548c\u5408\u5e76\u5185\u6838\u90fd\u4e0e\u8ba1\u7b97\u6d41\u91cd\u53e0\uff0c\u56e0\u6b64\u6211\u4eec\u4e5f\u8003\u8651\u5b83\u4eec\u5bf9\u5176\u4ed6SM\u8ba1\u7b97\u5185\u6838\u7684\u5f71\u54cd\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u91c7\u7528\u5b9a\u5236\u7684PTX\uff08Parallel Thread Execution\uff0c\u5e76\u884c\u7ebf\u7a0b\u6267\u884c\uff09\u6307\u4ee4\u5e76\u81ea\u52a8\u8c03\u6574\u901a\u4fe1\u5757\u5927\u5c0f\uff0c\u8fd9\u663e\u8457\u51cf\u5c11\u4e86L2\u7f13\u5b58\u7684\u4f7f\u7528\u4ee5\u53ca\u5bf9\u5176\u4ed6SM\u7684\u5e72\u6270\u3002<\/p>\n<h5>3.2.3 \u6781\u4f4e\u7684\u5185\u5b58\u5360\u7528\uff0c\u51e0\u4e4e\u6ca1\u6709\u989d\u5916\u5f00\u9500<\/h5>\n<p>\u4e3a\u4e86\u51cf\u5c11\u8bad\u7ec3\u671f\u95f4\u7684\u5185\u5b58\u5360\u7528\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u4ee5\u4e0b\u6280\u672f\uff1a<\/p>\n<h6>RMSNorm\u548cMLA\u4e0a\u6295\u5f71\u7684\u91cd\u65b0\u8ba1\u7b97<\/h6>\n<p>\u6211\u4eec\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u91cd\u65b0\u8ba1\u7b97\u6240\u6709RMSNorm\u64cd\u4f5c\u548cMLA\u4e0a\u6295\u5f71\uff0c\u4ece\u800c\u65e0\u9700\u6301\u4e45\u5b58\u50a8\u5b83\u4eec\u7684\u8f93\u51fa\u6fc0\u6d3b\u503c\u3002\u901a\u8fc7\u5c11\u91cf\u989d\u5916\u5f00\u9500\uff0c\u6b64\u7b56\u7565\u663e\u8457\u51cf\u5c11\u4e86\u5b58\u50a8\u6fc0\u6d3b\u503c\u6240\u9700\u7684\u5185\u5b58\u3002<\/p>\n<h6>CPU \u4e2d\u7684\u6307\u6570\u79fb\u52a8\u5e73\u5747<\/h6>\n<p>\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u6211\u4eec\u4fdd\u7559\u6a21\u578b\u53c2\u6570\u7684\u6307\u6570\u79fb\u52a8\u5e73\u5747\uff08Exponential Moving Average, EMA\uff09\uff0c\u4ee5\u4fbf\u5728\u5b66\u4e60\u7387\u8870\u51cf\u540e\u5c3d\u65e9\u4f30\u8ba1\u6a21\u578b\u6027\u80fd\u3002EMA\u53c2\u6570\u5b58\u50a8\u5728CPU\u5185\u5b58\u4e2d\uff0c\u5e76\u5728\u6bcf\u4e2a\u8bad\u7ec3\u6b65\u9aa4\u4e4b\u540e\u5f02\u6b65\u66f4\u65b0\u3002\u6b64\u65b9\u6cd5\u4f7f\u6211\u4eec\u80fd\u591f\u5728\u4e0d\u4ea7\u751f\u989d\u5916\u5185\u5b58\u6216\u65f6\u95f4\u5f00\u9500\u7684\u60c5\u51b5\u4e0b\u7ef4\u62a4EMA\u53c2\u6570\u3002<\/p>\n<h6>\u7528\u4e8e\u591a\u7b26\u5143\u9884\u6d4b\u7684\u5171\u4eab\u5d4c\u5165\u548c\u8f93\u51fa\u5934<\/h6>\n<p>\u4f7f\u7528DualPipe\u7b56\u7565\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u4e2d\u6700\u6d45\u7684\u5c42\uff08\u5305\u62ec\u5d4c\u5165\u5c42\uff09\u548c\u6700\u6df1\u7684\u5c42\uff08\u5305\u62ec\u8f93\u51fa\u5934\uff09\u90e8\u7f72\u5728\u540c\u4e00\u4e2a\u6d41\u6c34\u7ebf\u5e76\u884c\uff08PP\uff09\u79e9\u4e0a\u3002\u6b64\u5b89\u6392\u4f7f\u5f97MTP\u6a21\u5757\u548c\u4e3b\u6a21\u578b\u4e4b\u95f4\u53ef\u4ee5\u7269\u7406\u5171\u4eab\u5171\u4eab\u5d4c\u5165\u548c\u8f93\u51fa\u5934\u7684\u53c2\u6570\u548c\u68af\u5ea6\u3002\u8fd9\u79cd\u7269\u7406\u5171\u4eab\u673a\u5236\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u6211\u4eec\u7684\u5185\u5b58\u6548\u7387\u3002<\/p>\n<h4>3.3 FP8\u8bad\u7ec3<\/h4>\n<p>\u53d7\u8fd1\u671f\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u8fdb\u5c55\u7684\u542f\u53d1\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u5229\u7528FP8\u6570\u636e\u683c\u5f0f\u5bf9DeepSeek-V3\u8fdb\u884c\u8bad\u7ec3\u7684\u7ec6\u7c92\u5ea6\u6df7\u5408\u7cbe\u5ea6\u6846\u67b6\u3002\u5c3d\u7ba1\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u524d\u666f\u5e7f\u9614\uff0c\u4f46\u5176\u5f80\u5f80\u53d7\u9650\u4e8e\u6fc0\u6d3b\u503c\u3001\u6743\u91cd\u548c\u68af\u5ea6\u4e2d\u5f02\u5e38\u503c\u7684\u5b58\u5728\u3002\u5c3d\u7ba1\u5728\u63a8\u7406\u91cf\u5316\u65b9\u9762\u5df2\u53d6\u5f97\u663e\u8457\u8fdb\u5c55\uff0c\u4f46\u5728\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u4e2d\u6210\u529f\u5e94\u7528\u4f4e\u7cbe\u5ea6\u6280\u672f\u7684\u7814\u7a76\u76f8\u5bf9\u8f83\u5c11\u3002\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\u5e76\u6709\u6548\u6269\u5c55 FP8 \u683c\u5f0f\u7684\u52a8\u6001\u8303\u56f4\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u7ec6\u7c92\u5ea6\u91cf\u5316\u7b56\u7565\uff1a\u91c7\u7528 <span class=\"katex-eq\" data-katex-display=\"false\">1 \\times N_c<\/span> \u5143\u7d20\u7684\u5206\u5757\u5206\u7ec4\u6216 <span class=\"katex-eq\" data-katex-display=\"false\">N_c \\times N_c<\/span> \u5143\u7d20\u7684\u5757\u5206\u7ec4\u3002\u5728\u6211\u4eec\u63d0\u9ad8\u7cbe\u5ea6\u7684\u7d2f\u52a0\u673a\u5236\u4e0b\uff0c\u76f8\u5173\u7684\u53bb\u91cf\u5316\u5f00\u9500\u5f97\u5230\u4e86\u5f88\u5927\u7a0b\u5ea6\u7684\u7f13\u89e3\uff0c\u8fd9\u662f\u5b9e\u73b0\u51c6\u786e\u7684 FP8 \u901a\u7528\u77e9\u9635\u4e58\u6cd5\uff08GEMM\uff09\u7684\u5173\u952e\u65b9\u9762\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u8fdb\u4e00\u6b65\u964d\u4f4e MoE \u8bad\u7ec3\u4e2d\u7684\u5185\u5b58\u548c\u901a\u4fe1\u5f00\u9500\uff0c\u6211\u4eec\u5728 FP8 \u4e2d\u7f13\u5b58\u548c\u5206\u53d1\u6fc0\u6d3b\u503c\uff0c\u540c\u65f6\u4ee5 BF16 \u683c\u5f0f\u5b58\u50a8\u4f4e\u7cbe\u5ea6\u4f18\u5316\u5668\u72b6\u6001\u3002\u6211\u4eec\u5728\u4e24\u4e2a\u7c7b\u4f3c\u4e8e DeepSeek-V2-Lite \u548c DeepSeek-V2 \u7684\u6a21\u578b\u89c4\u6a21\u4e0a\u9a8c\u8bc1\u4e86\u6240\u63d0\u51fa\u7684 FP8 \u6df7\u5408\u7cbe\u5ea6\u6846\u67b6\uff0c\u8bad\u7ec3\u4e86\u5927\u7ea6 1 \u4e07\u4ebf\u4e2a\u6807\u8bb0\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4e0e BF16 \u57fa\u7ebf\u76f8\u6bd4\uff0c\u6211\u4eec\u7684 FP8 \u8bad\u7ec3\u6a21\u578b\u7684\u76f8\u5bf9\u635f\u5931\u8bef\u5dee\u59cb\u7ec8\u4f4e\u4e8e 0.25%\uff0c\u8fd9\u4e00\u6c34\u5e73\u5b8c\u5168\u5728\u8bad\u7ec3\u968f\u673a\u6027\u7684\u53ef\u63a5\u53d7\u8303\u56f4\u5185\u3002<\/p>\n<h5>3.3.1 \u6df7\u5408\u7cbe\u5ea6\u6846\u67b6<\/h5>\n<div class=\"markdown-container\" data-v-010c3103=\"\" data-v-b341a789=\"\">\n<div class=\"markdown\" data-v-53cd9be5=\"\" data-v-010c3103=\"\">\n<div class=\"paragraph\">\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7528\u4e8eFP8\u8bad\u7ec3\u7684\u6df7\u5408\u7cbe\u5ea6\u6846\u67b6\u3002\u8be5\u6846\u67b6\u5728\u5927\u591a\u6570\u8ba1\u7b97\u5bc6\u96c6\u578b\u64cd\u4f5c\u4e2d\u91c7\u7528FP8\u7cbe\u5ea6\uff0c\u540c\u65f6\u6218\u7565\u6027\u5730\u4fdd\u7559\u90e8\u5206\u5173\u952e\u64cd\u4f5c\u7684\u539f\u59cb\u6570\u636e\u683c\u5f0f\uff08\u4f8b\u5982BF16\u6216FP32\uff09\uff0c\u4ee5\u5e73\u8861\u8bad\u7ec3\u6548\u7387\u4e0e\u6570\u503c\u7a33\u5b9a\u6027\u3002\u56fe6\u5c55\u793a\u4e86\u6846\u67b6\u7684\u6574\u4f53\u7ed3\u6784\u3002<\/div>\n<div><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-1140\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213111429986.png\" alt=\"\" width=\"1024\" height=\"295\" srcset=\"http:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213111429986.png 1024w, http:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213111429986-300x86.png 300w, http:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213111429986-768x221.png 768w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div>\n<div style=\"text-align: center;\">\u56fe6 <span id=\"yiyi-407\" class=\"yiyi-st\">\u4f7f\u7528FP8\u6570\u636e\u683c\u5f0f\u7684\u6574\u4f53\u6df7\u5408\u7cbe\u5ea6\u6846\u67b6\u3002<\/span><span id=\"yiyi-408\" class=\"yiyi-st\">\u4e3a\u6e05\u695a\u8d77\u89c1\uff0c\u4ec5\u8bf4\u660e\u4e86<span id=\"S3.F6.2.1\" class=\"ltx_text ltx_font_typewriter\">Linear<\/span>\u8fd0\u7b97\u7b26\u3002<\/span><\/div>\n<div>\u00a0<\/div>\n<div class=\"paragraph\">\u9996\u5148\uff0c\u4e3a\u4e86\u52a0\u901f\u6a21\u578b\u8bad\u7ec3\uff0c\u6838\u5fc3\u8ba1\u7b97\u5185\u6838\uff08\u5373GEMM\u8fd0\u7b97\uff09\u5927\u591a\u4ee5FP8\u7cbe\u5ea6\u5b9e\u73b0\u3002\u8fd9\u4e9bGEMM\u8fd0\u7b97\u4ee5FP8\u5f20\u91cf\u4e3a\u8f93\u5165\uff0c\u5e76\u8f93\u51faBF16\u6216FP32\u7684\u7ed3\u679c\u3002\u5982\u56fe6\u6240\u793a\uff0c\u4e0eLinear\u7b97\u5b50\u76f8\u5173\u7684\u4e09\u4e2aGEMM\u8fd0\u7b97\u2014\u2014Fprop\uff08\u524d\u5411\u4f20\u64ad\uff09\u3001Dgrad\uff08\u6fc0\u6d3b\u53cd\u5411\u4f20\u64ad\uff09\u548cWgrad\uff08\u6743\u91cd\u53cd\u5411\u4f20\u64ad\uff09\u2014\u2014\u5747\u4ee5FP8\u7cbe\u5ea6\u6267\u884c\u3002\u8fd9\u79cd\u8bbe\u8ba1\u7406\u8bba\u4e0a\u53ef\u5c06\u8ba1\u7b97\u901f\u5ea6\u63d0\u5347\u4e00\u500d\uff0c\u76f8\u8f83\u4e8e\u539f\u6709\u7684BF16\u65b9\u6cd5\u3002\u6b64\u5916\uff0cFP8 Wgrad GEMM\u5141\u8bb8\u4ee5FP8\u683c\u5f0f\u5b58\u50a8\u6fc0\u6d3b\u503c\uff0c\u4f9b\u53cd\u5411\u4f20\u64ad\u4f7f\u7528\uff0c\u4ece\u800c\u663e\u8457\u51cf\u5c11\u5185\u5b58\u6d88\u8017\u3002<\/div>\n<div class=\"paragraph\">\u5c3d\u7ba1FP8\u683c\u5f0f\u5728\u6548\u7387\u4e0a\u5177\u6709\u663e\u8457\u4f18\u52bf\uff0c\u4f46\u67d0\u4e9b\u7b97\u5b50\u7531\u4e8e\u5bf9\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u8f83\u4e3a\u654f\u611f\uff0c\u4ecd\u9700\u66f4\u9ad8\u7684\u7cbe\u5ea6\u3002\u6b64\u5916\uff0c\u4e00\u4e9b\u8ba1\u7b97\u6210\u672c\u8f83\u4f4e\u7684\u7b97\u5b50\u53ef\u4ee5\u4f7f\u7528\u66f4\u9ad8\u7cbe\u5ea6\uff0c\u800c\u4e0d\u4f1a\u5bf9\u6574\u4f53\u8bad\u7ec3\u6210\u672c\u4ea7\u751f\u663e\u8457\u5f71\u54cd\u3002\u56e0\u6b64\uff0c\u7ecf\u8fc7\u7ec6\u81f4\u7814\u7a76\uff0c\u6211\u4eec\u4e3a\u4ee5\u4e0b\u7ec4\u4ef6\u4fdd\u7559\u4e86\u539f\u59cb\u7cbe\u5ea6\uff08\u4f8b\u5982\uff0cBF16\u6216FP32\uff09\uff1a\u5d4c\u5165\u6a21\u5757\u3001\u8f93\u51fa\u5934\u3001MoE\u95e8\u63a7\u6a21\u5757\u3001\u5f52\u4e00\u5316\u7b97\u5b50\u4ee5\u53ca\u6ce8\u610f\u529b\u7b97\u5b50\u3002\u8fd9\u4e9b\u7ec4\u4ef6\u6709\u9488\u5bf9\u6027\u5730\u4fdd\u7559\u9ad8\u7cbe\u5ea6\uff0c\u786e\u4fdd\u4e86\u6a21\u578b\u7684\u7a33\u5b9a\u8bad\u7ec3\u52a8\u6001\u3002\u4e3a\u8fdb\u4e00\u6b65\u4fdd\u969c\u6570\u503c\u7a33\u5b9a\u6027\uff0c\u6211\u4eec\u5c06\u4e3b\u6743\u91cd\u3001\u6743\u91cd\u68af\u5ea6\u548c\u4f18\u5316\u5668\u72b6\u6001\u4ee5\u66f4\u9ad8\u7cbe\u5ea6\u5b58\u50a8\u3002\u867d\u7136\u8fd9\u4e9b\u9ad8\u7cbe\u5ea6\u7ec4\u4ef6\u4f1a\u5e26\u6765\u4e00\u5b9a\u7684\u5185\u5b58\u5f00\u9500\uff0c\u4f46\u53ef\u901a\u8fc7\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u7cfb\u7edf\u4e2d\u8de8\u591a\u4e2a\u6570\u636e\u5e76\u884c\uff08DP\uff09\u8282\u70b9\u8fdb\u884c\u9ad8\u6548\u5206\u7247\uff0c\u4ece\u800c\u6700\u5927\u9650\u5ea6\u5730\u964d\u4f4e\u5176\u5f71\u54cd\u3002<\/div>\n<div>\n<div class=\"markdown-container\" data-v-010c3103=\"\" data-v-b341a789=\"\">\n<div class=\"markdown\" data-v-53cd9be5=\"\" data-v-010c3103=\"\">\n<h5>3.3.2 \u91cf\u5316\u4e0e\u4e58\u6cd5\u5e26\u6765\u7684\u7cbe\u5ea6\u6539\u8fdb<\/h5>\n<div class=\"paragraph\">\u57fa\u4e8e\u6df7\u5408\u7cbe\u5ea6FP8\u6846\u67b6\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u591a\u79cd\u7b56\u7565\u4ee5\u63d0\u9ad8\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u7684\u7cbe\u5ea6\uff0c\u91cd\u70b9\u5173\u6ce8\u91cf\u5316\u65b9\u6cd5\u548c\u4e58\u6cd5\u8fc7\u7a0b\u3002<\/div>\n<h6>\u7ec6\u7c92\u5ea6\u91cf\u5316<\/h6>\n<div class=\"paragraph\">\u5728\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u6846\u67b6\u4e2d\uff0c\u7531\u4e8eFP8\u683c\u5f0f\u7684\u52a8\u6001\u8303\u56f4\u6709\u9650\uff08\u53d7\u5176\u51cf\u5c11\u7684\u6307\u6570\u4f4d\u9650\u5236\uff09\uff0c\u4e0a\u6ea2\u548c\u4e0b\u6ea2\u662f\u5e38\u89c1\u95ee\u9898\u3002\u4f20\u7edf\u65b9\u6cd5\u901a\u8fc7\u5c06\u8f93\u5165\u5f20\u91cf\u7684\u6700\u5927\u7edd\u5bf9\u503c\u7f29\u653e\u81f3FP8\u7684\u6700\u5927\u53ef\u8868\u793a\u503c\uff0c\u4ee5\u4f7f\u8f93\u5165\u5206\u5e03\u4e0eFP8\u683c\u5f0f\u7684\u53ef\u8868\u793a\u8303\u56f4\u5bf9\u9f50\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u5bf9\u6fc0\u6d3b\u5f02\u5e38\u503c\u9ad8\u5ea6\u654f\u611f\uff0c\u4f1a\u663e\u8457\u964d\u4f4e\u91cf\u5316\u7cbe\u5ea6\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7ec6\u7c92\u5ea6\u91cf\u5316\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u5728\u66f4\u7ec6\u7c92\u5ea6\u7684\u7ea7\u522b\u5e94\u7528\u7f29\u653e\u3002\u5982\u56fe7(a)\u6240\u793a\uff1a(1) \u5bf9\u4e8e\u6fc0\u6d3b\uff0c\u6211\u4eec\u57fa\u4e8e1\u00d7128\u7684\u5757\uff08\u5373\u6bcf\u4e2a\u7b26\u5143\u7684128\u4e2a\u901a\u9053\uff09\u5bf9\u5143\u7d20\u8fdb\u884c\u5206\u7ec4\u548c\u7f29\u653e\uff1b(2) \u5bf9\u4e8e\u6743\u91cd\uff0c\u6211\u4eec\u57fa\u4e8e128\u00d7128\u7684\u5757\uff08\u5373\u6bcf\u4e2a128\u4e2a\u8f93\u5165\u901a\u9053\u5bf9\u5e94128\u4e2a\u8f93\u51fa\u901a\u9053\uff09\u5bf9\u5143\u7d20\u8fdb\u884c\u5206\u7ec4\u548c\u7f29\u653e\u3002\u8be5\u65b9\u6cd5\u901a\u8fc7\u6839\u636e\u8f83\u5c0f\u7684\u5143\u7d20\u7ec4\u8c03\u6574\u6bd4\u4f8b\uff0c\u786e\u4fdd\u91cf\u5316\u8fc7\u7a0b\u80fd\u591f\u66f4\u597d\u5730\u9002\u5e94\u5f02\u5e38\u503c\u3002\u5728\u9644\u5f55B.2\u4e2d\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u8ba8\u8bba\u4e86\u5728\u6fc0\u6d3b\u91cf\u5316\u4e2d\u5e94\u7528\u4e0e\u6743\u91cd\u7c7b\u4f3c\u7684\u5757\u7ea7\u5206\u7ec4\u548c\u7f29\u653e\u65f6\u7684\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u6027\u3002<\/div>\n<div><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-1141\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213113238949.png\" alt=\"\" width=\"1024\" height=\"519\" srcset=\"http:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213113238949.png 1024w, http:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213113238949-300x152.png 300w, http:\/\/youpaiyun.lingbo.online\/2025\/02\/20250213113238949-768x389.png 768w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div>\n<div style=\"text-align: center;\"><span class=\"ltx_tag ltx_tag_figure\"><span id=\"yiyi-433\" class=\"yiyi-st\">\u56fe 7\uff1a<\/span><\/span><span id=\"yiyi-434\" class=\"yiyi-st\">(a) \u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7ec6\u7c92\u5ea6\u91cf\u5316\u65b9\u6cd5\u6765\u51cf\u8f7b\u7531\u7279\u5f81\u5f02\u5e38\u503c\u5f15\u8d77\u7684\u91cf\u5316\u8bef\u5dee\uff1b\u4e3a\u7b80\u5355\u8d77\u89c1\uff0c\u4ec5\u8bf4\u660e\u4e86<span id=\"S3.F7.4.1\" class=\"ltx_text ltx_font_typewriter\">Fprop<\/span>\u3002<\/span><span id=\"yiyi-435\" class=\"yiyi-st\">(b) \u7ed3\u5408\u6211\u4eec\u7684\u91cf\u5316\u7b56\u7565\uff0c\u6211\u4eec\u901a\u8fc7\u4ee5[letax]N_c=128[\/latex]\u4e2a\u5143\u7d20 MMA \u7684\u95f4\u9694\u63d0\u5347\u5230 CUDA \u5185\u6838\u6765\u63d0\u9ad8 FP8 GEMM \u7cbe\u5ea6\uff0c\u4ee5\u5b9e\u73b0\u9ad8\u7cbe\u5ea6\u7d2f\u79ef\u3002<\/span><\/div>\n<div class=\"paragraph\">\u6211\u4eec\u65b9\u6cd5\u7684\u4e00\u4e2a\u5173\u952e\u6539\u8fdb\u662f\u5728GEMM\u8fd0\u7b97\u7684\u5185\u7ef4\u4e2d\u5f15\u5165\u6bcf\u7ec4\u7f29\u653e\u56e0\u5b50\u3002\u6807\u51c6FP8 GEMM\u8fd0\u7b97\u4e0d\u652f\u6301\u8fd9\u4e00\u529f\u80fd\uff0c\u4f46\u7ed3\u5408\u6211\u4eec\u7cbe\u786e\u7684FP32\u7d2f\u52a0\u7b56\u7565\uff0c\u53ef\u4ee5\u6709\u6548\u5730\u5b9e\u73b0\u5b83\u3002<\/div>\n<div class=\"paragraph\">\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6211\u4eec\u7684\u7ec6\u7c92\u5ea6\u91cf\u5316\u7b56\u7565\u4e0e\u5fae\u7f29\u653e\u683c\u5f0f\u7684\u7406\u5ff5\u9ad8\u5ea6\u4e00\u81f4\uff0c\u800cNVIDIA\u4e0b\u4e00\u4ee3GPU\uff08Blackwell\u7cfb\u5217\uff09\u7684Tensor Cores\u5df2\u5ba3\u5e03\u652f\u6301\u5177\u6709\u66f4\u5c0f\u91cf\u5316\u7c92\u5ea6\u7684\u5fae\u7f29\u653e\u683c\u5f0f\u3002\u6211\u4eec\u5e0c\u671b\u6211\u4eec\u7684\u8bbe\u8ba1\u53ef\u4ee5\u4f5c\u4e3a\u672a\u6765\u5de5\u4f5c\u7684\u53c2\u8003\uff0c\u4ee5\u8ddf\u4e0a\u6700\u65b0\u7684GPU\u67b6\u6784\u3002<\/div>\n<h6>\u63d0\u9ad8\u7d2f\u52a0\u7cbe\u5ea6<\/h6>\n<div class=\"paragraph\">\u4f4e\u7cbe\u5ea6GEMM\u8fd0\u7b97\u901a\u5e38\u4f1a\u9047\u5230\u4e0b\u6ea2\u95ee\u9898\uff0c\u5176\u7cbe\u5ea6\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u9ad8\u7cbe\u5ea6\u7d2f\u52a0\uff0c\u901a\u5e38\u4ee5FP32\u7cbe\u5ea6\u6267\u884c\u3002\u7136\u800c\uff0c\u6211\u4eec\u89c2\u5bdf\u5230NVIDIA H800 GPU\u4e0aFP8 GEMM\u7684\u7d2f\u52a0\u7cbe\u5ea6\u4ec5\u9650\u4e8e\u4fdd\u7559\u7ea614\u4f4d\uff0c\u8fdc\u4f4e\u4e8eFP32\u7d2f\u52a0\u7cbe\u5ea6\u3002\u5f53\u5185\u90e8\u7ef4\u5ea6K\u8f83\u5927\u65f6\uff0c\u8fd9\u4e00\u95ee\u9898\u5c06\u66f4\u52a0\u7a81\u51fa\uff0c\u8fd9\u5728\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3\u4e2d\u662f\u4e00\u4e2a\u5178\u578b\u573a\u666f\uff0c\u5176\u4e2d\u6279\u91cf\u5927\u5c0f\u548c\u6a21\u578b\u5bbd\u5ea6\u90fd\u4f1a\u589e\u52a0\u3002\u4ee5K=4096\u7684\u4e24\u4e2a\u968f\u673a\u77e9\u9635\u7684GEMM\u8fd0\u7b97\u4e3a\u4f8b\uff0c\u5728\u6211\u4eec\u7684\u9884\u5907\u6d4b\u8bd5\u4e2d\uff0cTensor Core\u4e2d\u6709\u9650\u7684\u7d2f\u52a0\u7cbe\u5ea6\u5bfc\u81f4\u6700\u5927\u76f8\u5bf9\u8bef\u5dee\u63a5\u8fd12%\u3002\u5c3d\u7ba1\u5b58\u5728\u8fd9\u4e9b\u95ee\u9898\uff0c\u4f46\u5728\u4e00\u4e9bFP8\u6846\u67b6\u4e2d\uff0c\u6709\u9650\u7684\u7d2f\u52a0\u7cbe\u5ea6\u4ecd\u7136\u662f\u9ed8\u8ba4\u9009\u9879\uff0c\u4e25\u91cd\u9650\u5236\u4e86\u8bad\u7ec3\u7cbe\u5ea6\u3002<\/div>\n<div class=\"paragraph\">\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u91c7\u7528\u5c06\u8ba1\u7b97\u63d0\u5347\u5230CUDA Core\u4ee5\u83b7\u5f97\u66f4\u9ad8\u7cbe\u5ea6\u7684\u65b9\u6cd5\u3002\u8be5\u8fc7\u7a0b\u5982\u56fe7(b)\u6240\u793a\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5728Tensor Core\u4e0a\u6267\u884cMMA\uff08\u77e9\u9635\u4e58\u6cd5\u7d2f\u52a0\uff09\u671f\u95f4\uff0c\u4f7f\u7528\u6709\u9650\u7684\u4f4d\u5bbd\u7d2f\u52a0\u4e2d\u95f4\u7ed3\u679c\u3002\u4e00\u65e6\u8fbe\u5230<span class=\"katex-eq\" data-katex-display=\"false\">N_c<\/span>\u95f4\u9694\uff0c\u8fd9\u4e9b\u90e8\u5206\u7ed3\u679c\u5c06\u88ab\u590d\u5236\u5230CUDA Core\u4e0a\u7684FP32\u5bc4\u5b58\u5668\u4e2d\uff0c\u5728\u90a3\u91cc\u6267\u884c\u5168\u7cbe\u5ea6FP32\u7d2f\u52a0\u3002\u5982\u524d\u6240\u8ff0\uff0c\u6211\u4eec\u7684\u7ec6\u7c92\u5ea6\u91cf\u5316\u6cbf\u5185\u90e8\u7ef4\u5ea6K\u5e94\u7528\u6bcf\u4e2a\u7ec4\u7684\u6bd4\u4f8b\u56e0\u5b50\u3002\u8fd9\u4e9b\u6bd4\u4f8b\u56e0\u5b50\u53ef\u4ee5\u5728CUDA Core\u4e0a\u9ad8\u6548\u5730\u4e58\u4ee5\uff0c\u4f5c\u4e3a\u53cd\u91cf\u5316\u8fc7\u7a0b\uff0c\u5e76\u5177\u6709\u6700\u5c0f\u7684\u989d\u5916\u8ba1\u7b97\u6210\u672c\u3002<\/div>\n<div class=\"paragraph\">\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u79cd\u4fee\u6539\u964d\u4f4e\u4e86\u5355\u4e2a\u7ebf\u7a0b\u5757\u7684WGMMA\uff08\u7ebf\u7a0b\u7ec4\u7ea7\u77e9\u9635\u4e58\u6cd5\u7d2f\u52a0\uff09\u6307\u4ee4\u53d1\u51fa\u7387\u3002\u7136\u800c\uff0c\u5728H800\u67b6\u6784\u4e0a\uff0c\u901a\u5e38\u6709\u4e24\u4e2aWGMMA\u540c\u65f6\u5b58\u5728\uff1a\u5f53\u4e00\u4e2a\u7ebf\u7a0b\u5757\u6267\u884c\u63d0\u5347\u64cd\u4f5c\u65f6\uff0c\u53e6\u4e00\u4e2a\u7ebf\u7a0b\u5757\u80fd\u591f\u6267\u884cMMA\u64cd\u4f5c\u3002\u8fd9\u79cd\u8bbe\u8ba1\u80fd\u591f\u4f7f\u8fd9\u4e24\u4e2a\u64cd\u4f5c\u91cd\u53e0\uff0c\u4fdd\u6301Tensor Core\u7684\u9ad8\u5229\u7528\u7387\u3002\u6839\u636e\u6211\u4eec\u7684\u5b9e\u9a8c\uff0c\u8bbe\u7f6e\u95f4\u9694\u4e3a128\u4e2a\u5143\u7d20\uff0c\u76f8\u5f53\u4e8e4\u4e2aWGMMAs\uff0c\u4ee3\u8868\u4e86\u53ef\u4ee5\u5728\u4e0d\u5f15\u5165\u5927\u91cf\u989d\u5916\u5f00\u9500\u7684\u60c5\u51b5\u4e0b\u663e\u8457\u63d0\u9ad8\u7cbe\u5ea6\u7684\u6700\u5c0f\u7d2f\u52a0\u95f4\u9694\u3002<\/div>\n<h6>\u5c3e\u6570\u6bd4\u6307\u6570<\/h6>\n<div class=\"paragraph\">\u4e0e\u5148\u524d\u5de5\u4f5c\u91c7\u7528\u7684\u6df7\u5408FP8\u683c\u5f0f\u5f62\u6210\u5bf9\u6bd4\uff0c\u8be5\u683c\u5f0f\u5728Fprop\u4e2d\u4f7f\u7528E4M3\uff084\u4f4d\u6307\u6570\u548c3\u4f4d\u5c3e\u6570\uff09\uff0c\u5728Dgrad\u548cWgrad\u4e2d\u4f7f\u7528E5M2\uff085\u4f4d\u6307\u6570\u548c2\u4f4d\u5c3e\u6570\uff09\uff0c\u6211\u4eec\u5bf9\u6240\u6709\u5f20\u91cf\u91c7\u7528E4M3\u683c\u5f0f\u4ee5\u83b7\u5f97\u66f4\u9ad8\u7684\u7cbe\u5ea6\u3002\u6211\u4eec\u5c06\u8fd9\u79cd\u65b9\u6cd5\u7684\u53ef\u884c\u6027\u5f52\u56e0\u4e8e\u6211\u4eec\u7ec6\u7c92\u5ea6\u7684\u91cf\u5316\u7b56\u7565\uff0c\u5373\u57fa\u4e8e\u5757\u7684\u7f29\u653e\u3002\u901a\u8fc7\u5bf9\u8f83\u5c0f\u7684\u5143\u7d20\u7ec4\u8fdb\u884c\u64cd\u4f5c\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u6709\u6548\u5730\u5728\u8fd9\u4e9b\u5206\u7ec4\u5143\u7d20\u4e4b\u95f4\u5171\u4eab\u6307\u6570\u4f4d\uff0c\u4ece\u800c\u51cf\u8f7b\u4e86\u6709\u9650\u52a8\u6001\u8303\u56f4\u7684\u5f71\u54cd\u3002<\/div>\n<h6>\u5728\u7ebf\u91cf\u5316<\/h6>\n<div class=\"paragraph\">\u5ef6\u8fdf\u91cf\u5316\u88ab\u7528\u4e8e\u5f20\u91cf\u91cf\u5316\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u7ef4\u62a4\u5148\u524d\u8fed\u4ee3\u4e2d\u6700\u5927\u7edd\u5bf9\u503c\u7684\u8bb0\u5f55\u4ee5\u63a8\u65ad\u5f53\u524d\u503c\u3002\u4e3a\u4e86\u786e\u4fdd\u7cbe\u786e\u7684\u6bd4\u4f8b\u5e76\u7b80\u5316\u6846\u67b6\uff0c\u6211\u4eec\u9488\u5bf9\u6bcf\u4e2a1\u00d7128\u6fc0\u6d3b\u5757\u6216128\u00d7128\u6743\u91cd\u5757\u5728\u7ebf\u8ba1\u7b97\u6700\u5927\u7edd\u5bf9\u503c\u3002\u57fa\u4e8e\u6b64\uff0c\u6211\u4eec\u63a8\u5bfc\u51fa\u7f29\u653e\u56e0\u5b50\uff0c\u7136\u540e\u5c06\u6fc0\u6d3b\u6216\u6743\u91cd\u5728\u7ebf\u91cf\u5316\u4e3aFP8\u683c\u5f0f\u3002<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>MetaData \u53d1\u8868\u65f6\u95f4 2024.12.27 \u4f5c\u8005\uff1aDeepSeek-AI \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/ab &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[60,41],"tags":[66,65],"class_list":["post-1114","post","type-post","status-publish","format-standard","hentry","category-paper_read","category-learning_note","tag-deepseek-v3","tag-65"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1114","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1114"}],"version-history":[{"count":8,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1114\/revisions"}],"predecessor-version":[{"id":1148,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1114\/revisions\/1148"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1114"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1114"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1114"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}