{"id":31242,"date":"2025-04-21T09:04:08","date_gmt":"2025-04-21T01:04:08","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/31242.html"},"modified":"2025-04-21T09:04:08","modified_gmt":"2025-04-21T01:04:08","slug":"%e4%b8%80%e6%96%87%e6%b7%b1%e5%85%a5%e4%ba%86%e8%a7%a3deepseek-r1%ef%bc%9a%e6%a8%a1%e5%9e%8b%e6%9e%b6%e6%9e%84","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/31242.html","title":{"rendered":"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784"},"content":{"rendered":"<p>\u672c\u6587\u6df1\u5165\u63a2\u8ba8\u4e86 DeepSeek-R1 \u6a21\u578b\u67b6\u6784\u3002\u8ba9\u6211\u4eec\u4ece\u8f93\u5165\u5230\u8f93\u51fa\u8ffd\u8e2a DeepSeek-R1 \u6a21\u578b&#xff0c;\u4ee5\u627e\u5230\u67b6\u6784\u4e2d\u7684\u65b0\u53d1\u5c55\u548c\u5173\u952e\u90e8\u5206\u3002DeepSeek-R1 \u57fa\u4e8e DeepSeek-V3-Base \u6a21\u578b\u67b6\u6784\u3002\u672c\u6587\u65e8\u5728\u6db5\u76d6\u5176\u8bbe\u8ba1\u7684\u6240\u6709\u91cd\u8981\u65b9\u9762\u3002<\/p>\n<h3>&#x1f4dd; 1. \u8f93\u5165\u4e0a\u4e0b\u6587\u957f\u5ea6<\/h3>\n<p>DeepSeek-R1\u7684\u8f93\u5165\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a128K\u3002<\/p>\n<p>DeepSeek-R1 \u4ece\u5176\u57fa\u7840\u6a21\u578b DeepSeek-V3-Base \u7ee7\u627f\u4e86 128K \u4e0a\u4e0b\u6587\u957f\u5ea6\u3002\u6700\u521d&#xff0c;DeepSeek-V3 \u4f7f\u7528 4K \u4e0a\u4e0b\u6587\u957f\u5ea6\u8fdb\u884c\u9884\u8bad\u7ec3\u3002\u7136\u540e&#xff0c;\u5229\u7528 YaRN \u6280\u672f&#xff0c;\u4e24\u9636\u6bb5\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u9996\u5148\u5c06\u5176\u589e\u52a0\u5230 32K&#xff0c;\u7136\u540e\u589e\u52a0\u5230 128K\u3002<\/p>\n<p>YaRN&#xff08;\u53e6\u4e00\u79cd RoPE \u6269\u5c55\u65b9\u6cd5&#xff09;\u662f\u4e00\u79cd\u65e8\u5728\u6709\u6548\u6269\u5c55\u4f7f\u7528\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165 (RoPE)\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u6280\u672f\u3002RoPE \u4f7f\u7528\u65cb\u8f6c\u77e9\u9635\u5bf9\u4f4d\u7f6e\u4fe1\u606f\u8fdb\u884c\u7f16\u7801&#xff0c;\u800c YaRN \u5219\u4fee\u6539\u8fd9\u4e9b\u65cb\u8f6c\u9891\u7387\u7684\u7f29\u653e\u65b9\u5f0f\u3002\u5b83\u4e0d\u662f\u7b80\u5355\u5730\u63a8\u65ad\u9891\u7387&#xff08;\u8fd9\u901a\u5e38\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d&#xff09;&#xff0c;\u800c\u662f\u5e73\u6ed1\u5730\u63d2\u5165\u548c\u8c03\u6574\u8fd9\u4e9b\u9891\u7387&#xff0c;\u4ece\u800c\u80fd\u591f\u66f4\u597d\u5730\u63a8\u5e7f\u5230\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u3002\u5b83\u5728\u8ba1\u7b97\u4e0a\u662f\u9ad8\u6548\u7684&#xff0c;\u5e76\u4e14\u65e0\u9700\u5927\u91cf\u91cd\u65b0\u8bad\u7ec3\u5373\u53ef\u6269\u5c55\u6a21\u578b\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/p>\n<h3>&#x1f3db; 2. \u603b\u5c42\u6570<\/h3>\n<p>DeepSeek-R1 \u7531\u4e00\u4e2a\u5d4c\u5165\u5c42\u3001\u5176\u540e\u7684 61 \u4e2a\u53d8\u6362\u5668\u5c42\u4ee5\u53ca\u8f93\u51fa\u9636\u6bb5\u7684\u591a\u4e2a\u9884\u6d4b\u5934\u7ec4\u6210\u3002<\/p>\n<p>DeepSeek-R1 \u5728\u6240\u6709 Transformer \u5c42\u4e0a\u91c7\u7528\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA) \u5c42&#xff0c;\u800c\u4e0d\u662f\u6807\u51c6\u591a\u5934\u6ce8\u610f\u529b\u3002\u524d\u4e09\u4e2a Transformer \u5c42\u4e0e\u5176\u4ed6\u5c42\u4e0d\u540c&#xff0c;\u4f7f\u7528\u6807\u51c6\u524d\u9988\u7f51\u7edc (FFN) \u5c42\u3002\u4ece\u7b2c 4 \u5c42\u5230\u7b2c 61 \u5c42&#xff0c;\u6df7\u5408\u4e13\u5bb6 (MoE) \u5c42\u53d6\u4ee3\u4e86 FFN \u5c42\u3002MLA \u548c MoE \u7684\u7ec6\u8282\u5c06\u5728\u4ee5\u4e0b\u90e8\u5206\u4e2d\u63a2\u8ba8\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010400-68059980d42ec.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u5e26\u6709\u7ef4\u5ea6\u7684\u5b8c\u6574\u6a21\u578b\u67b6\u6784\u63cf\u8ff0&#xff1a;<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010401-68059981386a0.png\" alt=\"\" \/><\/p>\n<p>DeepSeek-V3\u4f7f\u7528\u591a\u6807\u8bb0\u9884\u6d4b (MTP) \u6280\u672f&#xff0c;\u5229\u7528\u6700\u540e\u4e24\u4e2a\u9884\u6d4b\u5934\u9884\u6d4b\u63a5\u4e0b\u6765\u76842 \u4e2a\u6807\u8bb0\u3002\u7b2c\u4e8c\u4e2a\u9884\u6d4b\u6807\u8bb0\u7684\u63a5\u53d7\u7387\u4ecb\u4e8e85% \u548c 90%\u4e4b\u95f4&#xff0c;\u8868\u660e\u5728\u5404\u4e2a\u751f\u6210\u4e3b\u9898\u4e2d\u5747\u5177\u6709\u5f88\u9ad8\u7684\u53ef\u9760\u6027\u3002 DeepSeek-R1 (DeepSeek-V3) \u603b\u5171\u5305\u542b 671B \u4e2a\u53c2\u6570&#xff0c;\u5176\u4e2d\u6bcf\u4e2a\u6807\u8bb0\u6fc0\u6d3b 37B \u4e2a\u3002\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0<\/p>\n<h3>&#x1f52c; 3. \u524d 3 \u4e2a DeepSeek-R1 \u5c42<\/h3>\n<p>\u524d 3 \u5c42\u7531\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA) \u548c\u6807\u51c6 FFN \u5c42\u7ec4\u6210\u3002\u8fd9\u4e9b\u901a\u5e38\u88ab\u79f0\u4e3a\u201c\u5bc6\u96c6 LLM \u5c42\u201d&#xff0c;\u56e0\u4e3a FFN \u5c42\u4e0d\u4f1a\u88ab MoE \u5c42\u53d6\u4ee3&#xff0c;\u76f8\u6bd4\u4e4b\u4e0b MoE \u5c42\u88ab\u8ba4\u4e3a\u66f4\u7a00\u758f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010401-680599817c347.png\" alt=\"\" \/><\/p>\n<p>DeepSeek-R1 \u4e2d\u7684\u524d 3 \u4e2a Transformer \u5c42<\/p>\n<h3>&#x1f9e9; 4. DeepSeek-R1 \u7684\u7b2c 4 \u5c42\u81f3\u7b2c 61 \u5c42<\/h3>\n<p>\u8fd9\u4e9b\u5c42\u7531 MLA \u5c42\u548c MoE \u5c42\u7ec4\u6210\u3002\u6211\u4eec\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u90e8\u5206\u4e2d\u4e86\u89e3\u4ec0\u4e48\u662f MLA \u5c42\u548c MoE \u5c42\u4ee5\u53ca\u5b83\u4eec\u5982\u4f55\u5de5\u4f5c\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010401-68059981aa409.png\" alt=\"\" \/><\/p>\n<p>MoE Transformer \u5c42<\/p>\n<h3>&#x1f9e0; 5.\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b&#xff08;MLA&#xff09;<\/h3>\n<p>\u73b0\u5728\u8ba9\u6211\u4eec\u4e86\u89e3\u4ec0\u4e48\u662f MLA\u3002<\/p>\n<p>MLA \u9996\u6b21\u5728 DeepSeek-V2 \u4e2d\u5f15\u5165&#xff0c;\u5e76\u5df2\u5ef6\u7eed\u5230 DeepSeek-V3 \u548c DeepSeek-R1\u3002<\/p>\n<p>\u4e3a\u4ec0\u4e48\u8981\u5f00\u53d1MLA&#xff1f;<\/p>\n<p>\u4ee5\u4e0b\u662f DeepSeek-V2 \u8bba\u6587\/\u6280\u672f\u62a5\u544a\u4e2d\u7684\u4e00\u4efd\u58f0\u660e&#xff0c;\u5b83\u6e05\u695a\u5730\u9610\u8ff0\u4e86\u5f00\u53d1 MLA \u80cc\u540e\u7684\u539f\u56e0\u3002<\/p>\n<p>\u201c\u4f20\u7edf\u7684Transformer\u6a21\u578b\u901a\u5e38\u91c7\u7528\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236&#xff08;MHA&#xff09;&#xff0c;\u4f46\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d&#xff0c;\u5176\u7e41\u91cd\u7684\u952e\u503c&#xff08;KV&#xff09;\u7f13\u5b58\u4f1a\u6210\u4e3a\u9650\u5236\u63a8\u7406\u6548\u7387\u7684\u74f6\u9888\u3002\u4e3a\u4e86\u51cf\u5c11KV\u7f13\u5b58&#xff0c;\u63d0\u51fa\u4e86\u591a\u67e5\u8be2\u6ce8\u610f\u529b\u673a\u5236&#xff08;MQA&#xff09;\u548c\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u673a\u5236&#xff08;GQA&#xff09;\u3002\u5b83\u4eec\u9700\u8981\u7684KV\u7f13\u5b58\u91cf\u7ea7\u8f83\u5c0f&#xff0c;\u4f46\u6027\u80fd\u4e0d\u5982MHA\u3002<\/p>\n<p>\u5bf9\u4e8e DeepSeek-V2&#xff0c;\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u521b\u65b0\u7684\u6ce8\u610f\u529b\u673a\u5236&#xff0c;\u79f0\u4e3a\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA)\u3002MLA \u914d\u5907\u4e86\u4f4e\u79e9\u952e\u503c\u8054\u5408\u538b\u7f29&#xff0c;\u6027\u80fd\u4f18\u4e8e MHA&#xff0c;\u4f46\u6240\u9700\u7684 KV \u7f13\u5b58\u91cf\u5374\u5c11\u5f97\u591a\u3002\u201d<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010402-68059982543d4.png\" alt=\"\" \/><\/p>\n<p>MLA \u4e0e MHA\u3001GQA \u548c MQA \u7684\u6bd4\u8f83&#xff1a;\u6765\u81ea DeepSeek-V2 \u7684\u56fe\u8868<\/p>\n<p>MLA \u5982\u4f55\u5b9e\u73b0\u51cf\u5c11 KV \u7f13\u5b58\u4ee5\u52a0\u5feb\u63a8\u7406\u901f\u5ea6&#xff1f;<\/p>\n<p>\u201cMLA \u7684\u6838\u5fc3\u662f\u5bf9\u6ce8\u610f\u529b\u952e\u548c\u503c\u8fdb\u884c\u4f4e\u79e9\u8054\u5408\u538b\u7f29&#xff0c;\u4ee5\u51cf\u5c11\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u952e\u503c (KV) \u7f13\u5b58\u3002\u201d \u2014 DeepSeek-V2<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010402-68059982a66b0.png\" alt=\"\" \/><\/p>\n<p>\u8ba9\u6211\u4eec\u4e00\u6b65\u4e00\u6b65\u5730\u7406\u89e3\u8fd9\u4e2a\u56fe&#xff1a;<\/p>\n<p>\u6b65\u9aa4 1&#xff1a;Q\u3001K \u548c V \u7684\u4e0b\u6295\u5f71<\/p>\n<p>MLA \u5c42\u7684\u8f93\u5165\u662f<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>         \u200b <\/p>\n<p>        h_t\u200b <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mord\">\u200b<\/span><\/span><\/span><\/span><\/span>\u3002\u4e3a\u6e05\u695a\u8d77\u89c1&#xff0c;\u6211\u4eec\u5047\u8bbe <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u7684\u5f62\u72b6\u4e3a (input_sequence_length\u00d72000)\u3002<\/p>\n<p>\u5728\u4f20\u7edf\u7684 Transformer \u5c42\u4e2d&#xff0c;\u6743\u91cd\u77e9\u9635\u7528\u4e8e\u5c06 h_t\u200b \u6295\u5f71\u5230\u67e5\u8be2 (Q)\u3001\u952e (K) \u548c\u503c (V) \u8868\u793a\u4e2d\u3002\u8fd9\u4e9b\u8868\u793a\u4e2d\u7684\u6bcf\u4e00\u4e2a\u901a\u5e38\u90fd\u4fdd\u7559\u4e0e\u8f93\u5165\u76f8\u540c\u7684\u9690\u85cf\u7ef4\u5ea6&#xff0c;\u4ece\u800c\u5bfc\u81f4 Q\u3001K\u3001V \u5177\u6709 (input_sequence_length\u00d72000) \u7684\u5f62\u72b6\u3002<\/p>\n<p>MLA \u5c42\u7684\u8f93\u5165\u662f<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b\u3002\u4e3a\u6e05\u695a\u8d77\u89c1&#xff0c;\u6211\u4eec\u5047\u8bbe<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u7684\u5f62\u72b6\u4e3a (input_sequence_length\u00d72000)\u3002<\/p>\n<p>\u5728\u4f20\u7edf\u7684 Transformer \u5c42\u4e2d&#xff0c;\u6743\u91cd\u77e9\u9635\u7528\u4e8e\u5c06 h_t\u200b \u6295\u5f71\u5230\u67e5\u8be2 (Q)\u3001\u952e (K) \u548c\u503c (V) \u8868\u793a\u4e2d\u3002\u8fd9\u4e9b\u8868\u793a\u4e2d\u7684\u6bcf\u4e00\u4e2a\u901a\u5e38\u90fd\u4fdd\u7559\u4e0e\u8f93\u5165\u76f8\u540c\u7684\u9690\u85cf\u7ef4\u5ea6&#xff0c;\u4ece\u800c\u5bfc\u81f4 Q\u3001K\u3001V \u5177\u6709 (input_sequence_length\u00d72000) \u7684\u5f62\u72b6\u3002<\/p>\n<p>\u7136\u800c&#xff0c;\u5728\u5177\u6709\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b (MLA) \u7684 Transformer \u5c42\u4e2d&#xff0c;\u6743\u91cd\u77e9\u9635\u751f\u6210\u7684 Q\u3001K \u548c V \u7684\u5c3a\u5bf8\u660e\u663e\u5c0f\u4e8e\u8f93\u5165\u3002MLA \u4e0d\u4f1a\u4fdd\u7559\u5b8c\u6574\u7684\u9690\u85cf\u7ef4\u5ea6&#xff0c;\u800c\u662f\u4f1a\u51cf\u5c0f\u5b83\u4eec\u7684\u5c3a\u5bf8\u3002\u4f8b\u5982&#xff0c;\u5982\u679c\u8f93\u5165<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b \u7684\u5f62\u72b6\u4e3a (\u5e8f\u5217\u957f\u5ea6 \u00d7 2000)&#xff0c;\u5219\u751f\u6210\u7684 Q\u3001K \u548c V \u7684\u5f62\u72b6\u53ef\u80fd\u4e3a (\u5e8f\u5217\u957f\u5ea6 \u00d7 100)\u3002<\/p>\n<p>\u5728\u5b9e\u65bd\u8fc7\u7a0b\u4e2d&#xff0c;\u901a\u5e38\u4f1a\u878d\u5408 Q\u3001K \u548c V \u7684\u6743\u91cd\u77e9\u9635&#xff0c;\u4ee5\u63d0\u9ad8GPU \u4e0a\u7684\u8ba1\u7b97\u548c\u5185\u5b58\u6548\u7387\u3002\u4e0d\u662f\u5e94\u7528\u5355\u72ec\u7684\u6295\u5f71&#xff0c;\u800c\u662f\u4f7f\u7528\u7ec4\u5408\u6743\u91cd\u77e9\u9635\u6765\u4f18\u5316\u64cd\u4f5c\u3002\u5728 MLA \u4e2d&#xff0c; K\u548cV\u7684\u751f\u6210\u9075\u5faa\u8fd9\u4e00\u539f\u5219\u3002\u5177\u4f53\u800c\u8a00&#xff0c;\u5728\u65b9\u7a0b\u4e2d\u4f7f\u7528\u5355\u4e2a\u6743\u91cd\u77e9\u9635&#xff0c;\u8868\u793a\u4e3a<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          W <\/p>\n<p>           D <\/p>\n<p>           K <\/p>\n<p>           V <\/p>\n<p>        W^{DKV} <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8413em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">W<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0715em\">DK<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.2222em\">V<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u200b\u3002\u8fd9\u91cc&#xff0c; <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          W <\/p>\n<p>           D <\/p>\n<p>           K <\/p>\n<p>           V <\/p>\n<p>        W^{DKV} <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8413em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">W<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0715em\">DK<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.2222em\">V<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b \u4e2d\u7684\u201cD\u201d\u4ee3\u8868\u4e0b\u6295\u5f71\u6743\u91cd\u77e9\u9635&#xff0c;\u53cd\u6620\u4e86\u5176\u5728\u964d\u4f4e\u7ef4\u5ea6\u4ee5\u5b9e\u73b0\u9ad8\u6548\u6ce8\u610f\u529b\u8ba1\u7b97\u65b9\u9762\u7684\u4f5c\u7528\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010403-6805998332404.png\" alt=\"\" \/><\/p>\n<p>\u6f5c\u5728 K \u548c V \u5d4c\u5165<\/p>\n<p>\u6b64\u6295\u5f71\u7684\u8f93\u51fa\u662f\u5305\u542b K \u548c V \u7684\u8fde\u63a5\u8868\u793a\u3002\u53ef\u4ee5\u4f7f\u7528\u7b80\u5355\u7684\u5207\u7247\u673a\u5236\u8f7b\u677e\u63d0\u53d6\u5b83\u4eec\u3002\u7ed3\u679c\u8f93\u51fa\u7684\u5f62\u72b6\u4e3a (\u5e8f\u5217\u957f\u5ea6 \u00d7 200)&#xff0c;\u5176\u4e2d\u7b2c\u4e00\u4e2a (\u5e8f\u5217\u957f\u5ea6 \u00d7 100) \u5bf9\u5e94\u4e8e K&#xff0c;\u5176\u4f59 (\u5e8f\u5217\u957f\u5ea6 \u00d7 100) \u5bf9\u5e94\u4e8e V\u3002<\/p>\n<p>\u538b\u7f29\u7684 K \u548c V \u8f93\u51fa\u5728\u63a8\u7406\u671f\u95f4\u88ab\u7f13\u5b58&#xff0c;\u4ece\u800c\u663e\u8457\u51cf\u5c11\u4e86 KV \u7f13\u5b58\u7684\u5185\u5b58\u5360\u7528\u3002<\/p>\n<p>\u7c7b\u4f3c\u5730&#xff0c;Q \u4e5f\u5728MLA\u4e2d\u88ab\u538b\u7f29\u3002Q \u7684\u7ed3\u679c\u5f62\u72b6\u4e3a(\u5e8f\u5217\u957f\u5ea6 \u00d7 100)\u3002<\/p>\n<p>\u7b2c 2 \u6b65&#xff1a;Q\u3001K \u548c V \u7684\u5411\u4e0a\u6295\u5f71<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010403-6805998358a04.png\" alt=\"\" \/><\/p>\n<p>\u538b\u7f29\u540e&#xff0c;Q\u3001K \u548c V\u88ab\u5411\u4e0a\u6295\u5f71\u56de\u66f4\u5927\u7684\u5c3a\u5bf8\u4ee5\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u8fd9\u4e2a\u66f4\u5927\u7684\u5c3a\u5bf8\u53ef\u4ee5\u5339\u914d\u539f\u59cb\u8f93\u5165<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u200b&#xff0c;\u4e5f\u53ef\u4ee5\u9075\u5faa\u57fa\u4e8e\u6ce8\u610f\u529b\u5934\u914d\u7f6e\u7684\u7ed3\u6784\u3002<\/p>\n<p>\u4f8b\u5982&#xff0c;\u5411\u4e0a\u6295\u5f71\u7684\u5f62\u72b6\u53ef\u4ee5\u662f&#xff1a;<\/p>\n<ul>\n<li>(\u5e8f\u5217\u957f\u5ea6\u00d72000)&#xff0c;\u4e0e\u8f93\u5165\u5927\u5c0f\u5339\u914d\u3002<\/li>\n<li>\u5e8f\u5217\u957f\u5ea6\u00d73200&#xff09;&#xff0c;\u5176\u4e2d3200\u6765\u81ea\u4e8e64\u00d750&#xff08;\u670964 \u4e2a\u6ce8\u610f\u529b\u5934&#xff0c;\u6bcf\u4e2a\u5934\u6709 50 \u4e2a\u7ef4\u5ea6&#xff09;\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010403-6805998390ca1.png\" alt=\"\" \/><\/p>\n<p>K \u548c V \u5411\u4e0a\u6295\u5f71<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010403-68059983b57b7.png\" alt=\"\" \/><\/p>\n<p>Q \u5411\u4e0a\u6295\u5f71<\/p>\n<p>Q\u3001K \u548c V\u7684\u4e0a\u6295\u5f71\u662f\u4f7f\u7528\u4e13\u7528\u6743\u91cd\u77e9\u9635\u6267\u884c\u7684&#xff1a;<\/p>\n<ul>\n<li>W_UK\u200b \u8868\u793aK \u5411\u4e0a\u6295\u5f71<\/li>\n<li>W_UV\u200b \u7528\u4e8e\u5782\u76f4\u5411\u4e0a\u6295\u5f71<\/li>\n<li>W_UQ\u200b \u7528\u4e8eQ \u5411\u4e0a\u6295\u5f71<\/li>\n<\/ul>\n<p>\u8fd9\u91cc&#xff0c;\u201cU\u201d\u4ee3\u8868\u5411\u4e0a\u6295\u5f71&#xff0c;\u8868\u793a\u5c06\u538b\u7f29\u8868\u793a\u6269\u5c55\u56de\u66f4\u5927\u7ef4\u5ea6\u7684\u7a7a\u95f4\u4ee5\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/p>\n<p>\u6ce8\u610f&#xff1a;\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u8f93\u5165\u7ef4\u5ea6\u5c06\u8fdb\u884c\u8c03\u6574\u4ee5\u9002\u5e94\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165 (RoPE)\u3002\u6b64\u8c03\u6574\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u90e8\u5206\u4e2d\u53d8\u5f97\u66f4\u52a0\u6e05\u6670\u3002<\/p>\n<p>\u6b65\u9aa4 3&#xff1a;Q \u548c K \u4e2d\u7684 RoPE \u5d4c\u5165\u7528\u4e8e\u7f16\u7801\u4f4d\u7f6e\u4fe1\u606f<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010403-68059983d8450.png\" alt=\"\" \/><\/p>\n<p>\u6b64\u6b65\u9aa4\u7528\u4e8e\u8ba1\u7b97 RoPE \u5d4c\u5165\u6765\u7f16\u7801\u4f4d\u7f6e\u4fe1\u606f\u3002<\/p>\n<p>\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165&#xff08;RoPE&#xff09;\u7684\u7ed3\u5408&#xff1a;<\/p>\n<ul>\n<li>\u89e3\u8026 RoPE \u7b56\u7565&#xff1a;\u4e3a\u4e86\u6574\u5408\u4f4d\u7f6e\u4fe1\u606f&#xff0c;DeepSeek-V2&#xff08;\u968f\u540e\u662f DeepSeek-V3 \u548c DeepSeek-R1&#xff09;\u91c7\u7528\u4e86\u89e3\u8026 RoPE \u65b9\u6cd5\u3002\u8fd9\u6d89\u53ca\u521b\u5efa\u4e13\u95e8\u7528\u4e8e\u643a\u5e26\u4f4d\u7f6e\u4fe1\u606f\u7684\u989d\u5916\u67e5\u8be2 (Q) \u548c\u5bc6\u94a5 (K) \u5411\u91cf\u3002<\/li>\n<li>\u8fde\u63a5&#xff1a;\u8fd9\u4e9b RoPE \u589e\u5f3a\u7684 Q \u548c K \u5411\u91cf\u4e0e\u4e0a\u6295\u5f71\u7684 Q \u548c K \u5411\u91cf\u8fde\u63a5\u3002 \u8fd9\u662f MLA \u4e2d\u6bd4\u8f83\u68d8\u624b\u7684\u90e8\u5206\u3002<\/li>\n<\/ul>\n<p>\u6211\u5c06\u5c1d\u8bd5\u6309\u7167\u6211\u4ece DeepSeek \u7684\u6280\u672f\u62a5\u544a\u4e2d\u7406\u89e3\u7684\u65b9\u5f0f\u6765\u89e3\u91ca\u5b83\u3002<\/p>\n<p>\u5728\u4f20\u7edf\u7684 Transformer \u5c42\u4e2d&#xff0c;RoPE \u64cd\u4f5c\u76f4\u63a5\u4f5c\u7528\u4e8e Q \u548c K\u3002\u5b83\u4e0d\u6539\u53d8 Q \u548c K \u7684\u7ef4\u5ea6&#xff0c;\u800c\u662f\u6539\u53d8 Q \u548c K \u4e2d\u7684\u8bed\u4e49\u8868\u793a&#xff08;Q \u548c K \u4e2d\u7684\u6570\u503c&#xff09;\u6765\u7f16\u7801\u4f4d\u7f6e\u4fe1\u606f\u3002\u56e0\u6b64&#xff0c;\u5f97\u5230\u7684 Q \u548c K \u540c\u65f6\u5177\u6709\u8bed\u4e49\u548c\u4f4d\u7f6e\u4fe1\u606f\u3002<\/p>\n<p>\u4f46\u662f&#xff0c;\u5728\u5177\u6709 MLA \u7684transformer\u5c42\u4e2d&#xff0c;RoPE \u88ab\u5e94\u7528\u4e8e\u5206\u79bb\u65b0\u751f\u6210\u7684\u67e5\u8be2&#xff08;Q&#xff09;\u548c\u952e&#xff08;K&#xff09;\u5d4c\u5165&#xff0c;\u5e76\u5c06\u5176\u8fde\u63a5\u5230\u4e0a\u6295\u5f71\u7684 Q \u548c K\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010404-680599842f02a.png\" alt=\"\" \/><\/p>\n<p>\u6b65\u9aa4 3.1&#xff1a;\u4e3a Q \u751f\u6210 RoPE \u5d4c\u5165<\/p>\n<p>\u4f20\u7edf\u4e0a&#xff0c;RoPE&#xff08;\u65cb\u8f6c\u4f4d\u7f6e\u5d4c\u5165&#xff09;\u4f1a\u6839\u636e\u67e5\u8be2 (Q) \u548c\u952e (K) \u5411\u91cf\u5728\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e&#xff0c;\u5bf9\u5b83\u4eec\u5e94\u7528\u65cb\u8f6c\u77e9\u9635\u3002\u6b64\u8f6c\u6362\u4f1a\u76f4\u63a5\u5728 Q \u548c K \u4e2d\u7f16\u7801\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f&#xff0c;\u4ece\u800c\u65e0\u9700\u4f7f\u7528\u6b63\u5f26\u6216\u7edd\u5bf9\u7f16\u7801\u7b49\u663e\u5f0f\u4f4d\u7f6e\u5d4c\u5165\u3002<\/p>\n<p>\u4f46\u662f\u5728 MLA \u4e2d&#xff0c;\u4e0d\u662f\u5c06 RoPE \u5e94\u7528\u4e8e\u4e0a\u6295\u5f71\u7684 Q&#xff08;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          q <\/p>\n<p>          t <\/p>\n<p>          C <\/p>\n<p>        q_t^C <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">q<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0715em\">C<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff09;&#xff0c;\u800c\u662f\u4ece <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          c <\/p>\n<p>          t <\/p>\n<p>          Q <\/p>\n<p>        c_t^Q <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.205em;vertical-align: -0.2458em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">c<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.9592em\"><span class=\"\" style=\"top: -2.4542em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.1809em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">Q<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2458em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u751f\u6210\u65b0\u7684 Q \u5d4c\u5165&#xff08;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          q <\/p>\n<p>          t <\/p>\n<p>          R <\/p>\n<p>        q_t^R <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">q<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0077em\">R<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff09;\u5e76\u5bf9\u5176\u5e94\u7528 RoPE\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010404-6805998489f8c.png\" alt=\"\" \/><\/p>\n<p>\u901a\u8fc7\u5c06 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          c <\/p>\n<p>          t <\/p>\n<p>          Q <\/p>\n<p>        c_t^Q <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.205em;vertical-align: -0.2458em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">c<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.9592em\"><span class=\"\" style=\"top: -2.4542em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.1809em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">Q<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2458em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u4e0e\u6743\u91cd\u77e9\u9635 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          W <\/p>\n<p>           Q <\/p>\n<p>           R <\/p>\n<p>        W^{QR} <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8413em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">W<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0077em\">QR<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u76f8\u4e58&#xff0c;\u53ef\u4ee5\u751f\u6210\u5b8c\u5168\u72ec\u7acb\u7684\u67e5\u8be2\u5d4c\u5165\u3002\u8fd9\u4e9b\u65b0\u7684\u72ec\u7acb\u67e5\u8be2\u5d4c\u5165\u7ecf\u8fc7 RoPE \u8f6c\u6362&#xff0c;\u4e3a\u6211\u4eec\u63d0\u4f9b\u4f4d\u7f6e\u7f16\u7801\u67e5\u8be2\u5d4c\u5165 (<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          q <\/p>\n<p>          t <\/p>\n<p>          R <\/p>\n<p>        q_t^R <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">q<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0077em\">R<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010404-68059984b3671.png\" alt=\"\" \/><\/p>\n<p><span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          q <\/p>\n<p>          t <\/p>\n<p>          R <\/p>\n<p>        q_t^R <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">q<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0077em\">R<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u7684\u751f\u6210\u65b9\u5f0f\u662f&#xff0c;\u5b83\u4eec\u53ef\u4ee5\u8fde\u63a5\u5230\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u8f93\u5165\u67e5\u8be2\u5d4c\u5165&#xff0c;\u4ee5\u4fbf\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u90fd\u5177\u6709\u4f4d\u7f6e\u4fe1\u606f\u3002 [\u4ece\u7b49\u5f0f\u6765\u770b&#xff0c;\u8fd9\u4e2a\u8bf4\u6cd5\u4f3c\u4e4e\u662f\u6b63\u786e\u7684&#xff0c;\u4f46\u9700\u8981\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u3002]<\/p>\n<p>\u6b65\u9aa4 3.2&#xff1a;\u4e3a K \u751f\u6210 RoPE \u5d4c\u5165 \u7c7b\u4f3c\u5730&#xff0c;\u4e0d\u662f\u5c06 RoPE \u5e94\u7528\u4e8e\u4e0a\u6295\u5f71\u7684 K&#xff0c;\u800c\u662f\u751f\u6210\u65b0\u7684 K \u5d4c\u5165\u5e76\u5bf9\u5176\u5e94\u7528 RoPE\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010404-68059984e2371.png\" alt=\"\" \/><\/p>\n<p>\u4f46\u662f\u4e0e RoPE \u5d4c\u5165\u7684<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          q <\/p>\n<p>          t <\/p>\n<p>          R <\/p>\n<p>        q_t^R <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">q<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0077em\">R<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u6709\u4e24\u4e2a\u5173\u952e\u533a\u522b&#xff1a;<\/p>\n<p>\u65b0\u7684 K \u5d4c\u5165\u662f\u7531 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff08;\u8f93\u5165\u5d4c\u5165&#xff09;\u800c\u4e0d\u662f\u5411\u4e0b\u6295\u5f71\u7684 K&#xff08;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          c <\/p>\n<p>          t <\/p>\n<p>          K <\/p>\n<p>        c_t^K <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">c<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0715em\">K<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff09;\u751f\u6210\u7684\u3002 \u76f8\u540c\u7684 RoPE \u5d4c\u5165 K&#xff08;\u952e&#xff09;\u8fde\u63a5\u5230\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u8f93\u5165\u3002\u4f46\u662f&#xff0c;\u5355\u72ec\u7684 RoPE \u5d4c\u5165 Q&#xff08;\u67e5\u8be2&#xff09;\u88ab\u8ba1\u7b97\u5e76\u8fde\u63a5\u5230\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934&#xff0c;\u5982\u6b65\u9aa4 3.1 \u6240\u793a\u3002[\u4ece\u7b49\u5f0f\u6765\u770b&#xff0c;\u8fd9\u4e2a\u8bf4\u6cd5\u4f3c\u4e4e\u662f\u6b63\u786e\u7684&#xff0c;\u4f46\u9700\u8981\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u3002]<\/p>\n<p>\u4e3a\u4ec0\u4e48\u4e0d\u4ece\u5411\u4e0a\u6295\u5f71\u7684 K \u5373 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>           k <\/p>\n<p>           t <\/p>\n<p>           C <\/p>\n<p>         k_t^C <\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0883em;vertical-align: -0.247em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0315em\">k<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8413em\"><span class=\"\" style=\"top: -2.453em;margin-left: -0.0315em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0715em\">C<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.247em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u751f\u6210&#xff1f;<\/p>\n<p>DeepSeek-V2\u62a5\u544a\u4e2d\u7684\u63a8\u7406&#xff1a;<\/p>\n<p>\u201c\u5982\u679c\u6211\u4eec\u5c06 RoPE \u5e94\u7528\u4e8e\u952ek _&#x1d436;&#xff0c;&#x1d44a;&#x1d448;&#x1d43e; \u5c06\u4e0e\u4f4d\u7f6e\u654f\u611f\u7684 RoPE \u77e9\u9635\u8026\u5408\u3002\u8fd9\u6837&#xff0c;&#x1d44a;&#x1d448;&#x1d43e; \u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5c31\u4e0d\u80fd\u518d\u88ab\u5438\u6536\u5230&#x1d44a;_&#x1d444; \u4e2d&#xff0c;\u56e0\u4e3a\u4e0e\u5f53\u524d\u751f\u6210\u7684 token \u76f8\u5173\u7684 RoPE \u77e9\u9635\u5c06\u4f4d\u4e8e &#x1d44a;_&#x1d444; \u548c &#x1d44a;&#x1d448;&#x1d43e; \u4e4b\u95f4&#xff0c;\u800c\u77e9\u9635\u4e58\u6cd5\u4e0d\u9075\u5faa\u4ea4\u6362\u5f8b\u3002\u201d<\/p>\n<p>\u4ece\u4e0b\u9762\u7684\u89e3\u91ca\u622a\u56fe\u53ef\u4ee5\u66f4\u597d\u5730\u7406\u89e3\u8fd9\u4e00\u70b9&#xff1a; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010405-6805998512db7.png\" alt=\"\" \/><\/p>\n<p>K \u7684 RoPE \u5d4c\u5165&#xff1a;\u7b2c 1 \u90e8\u5206<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010405-6805998596895.png\" alt=\"\" \/><\/p>\n<p>K \u7684 RoPE \u5d4c\u5165&#xff1a;\u7b2c 2 \u90e8\u5206<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010406-6805998615124.png\" alt=\"\" \/><\/p>\n<p>K \u7684 RoPE \u5d4c\u5165&#xff1a;\u7b2c 3 \u90e8\u5206<\/p>\n<p>\u56e0\u6b64&#xff0c;\u4e3a\u4e86\u63d0\u9ad8\u63a8\u7406\u6548\u7387&#xff0c;\u4f4d\u7f6e\u5d4c\u5165\u7684 K&#xff08;\u5173\u952e&#xff09;\u5d4c\u5165\u662f\u4ece\u8f93\u5165\u5d4c\u5165 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u751f\u6210\u7684\u3002<\/p>\n<p>\u5728 MLA \u4e2d\u5f15\u5165\u989d\u5916\u7684\u6743\u91cd\u77e9\u9635\u4e0d\u4f1a\u5bfc\u81f4\u5185\u5b58\u548c\u8ba1\u7b97\u6548\u7387\u4f4e\u4e0b\u5417&#xff1f;<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u5f00\u9500&#xff0c;DeepSeek-V2 \u62a5\u544a&#xff1a;<\/p>\n<p>\u201c\u6b64\u5916&#xff0c;\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d&#xff0c;\u7531\u4e8e &#x1d44a;&#x1d448;&#x1d43e; \u53ef\u4ee5\u88ab\u5438\u6536\u5230 &#x1d44a;&#x1d444; \u4e2d&#xff0c;\u800c &#x1d44a;&#x1d448;&#x1d449; \u53ef\u4ee5\u88ab\u5438\u6536\u5230 &#x1d44a;&#x1d442; \u4e2d&#xff0c;\u6211\u4eec\u751a\u81f3\u4e0d\u9700\u8981\u8ba1\u7b97\u952e\u548c\u503c\u6765\u5f15\u8d77\u6ce8\u610f\u3002\u201d<\/p>\n<p>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u51cf\u5c11\u5185\u5b58\u6d88\u8017&#xff1a;<\/p>\n<p>\u201c\u6b64\u5916&#xff0c;\u4e3a\u4e86\u51cf\u5c11\u8bad\u7ec3\u671f\u95f4\u7684\u6fc0\u6d3b\u5185\u5b58&#xff0c;\u6211\u4eec\u8fd8\u5bf9\u67e5\u8be2\u6267\u884c\u4f4e\u79e9\u538b\u7f29&#xff0c;\u5373\u4f7f\u5b83\u4e0d\u80fd\u51cf\u5c11 KV \u7f13\u5b58\u201d<\/p>\n<p>\u6b65\u9aa44&#xff1a;\u8ba1\u7b97\u6ce8\u610f\u529b\u8f93\u51fa<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010406-6805998679a7c.png\" alt=\"\" \/><\/p>\n<p>\u8fde\u63a5\u8fc7\u7a0b\u4f1a\u589e\u52a0 Q \u548c K \u5411\u91cf\u7684\u7ef4\u6570\u3002\u4e3a\u4e86\u7ba1\u7406\u8fd9\u79cd\u589e\u52a0\u7684\u7ef4\u6570&#xff0c;\u6a21\u578b\u53ef\u4ee5\u91c7\u53d6\u4ee5\u4e0b\u4efb\u4e00\u65b9\u5f0f&#xff1a;<\/p>\n<ul>\n<li>\u589e\u52a0\u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf&#xff1a;\u8fd9\u5c06\u4fdd\u6301\u539f\u59cb\u7684\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6&#xff0c;\u4f46\u9700\u8981\u66f4\u591a\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/li>\n<li>\u8c03\u6574\u6bcf\u4e2a\u5934\u7684\u7ef4\u6570&#xff1a;\u4fdd\u6301\u5934\u7684\u6570\u91cf\u4e0d\u53d8&#xff0c;\u4f46\u589e\u52a0\u6bcf\u4e2a\u5934\u7684\u7ef4\u6570\u4ee5\u9002\u5e94\u8fde\u63a5\u7684\u5411\u91cf\u3002<\/li>\n<\/ul>\n<p>\u6ce8\u610f\u529b\u8f93\u51fa\u662f\u4f7f\u7528\u8fd9\u4e9b\u6807\u51c6\u6ce8\u610f\u529b\u65b9\u7a0b\u6765\u8ba1\u7b97\u7684&#xff1a;<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010406-68059986d5fe9.png\" alt=\"\" \/> O_t_i \u662f\u6ce8\u610f\u529b\u5f97\u5206&#xff0c;u_t \u662f\u6ce8\u610f\u529b\u8f93\u51fa\u3002W_o \u8868\u793a\u8f93\u51fa\u6295\u5f71\u6743\u91cd\u77e9\u9635\u3002\u8f93\u51fa\u88ab\u6295\u5f71\u56de\u4e0e\u8f93\u5165\u76f8\u540c\u7684\u7ef4\u5ea6&#xff08;\u5c31\u50cf\u5728\u6211\u4eec\u7684\u4f8b\u5b50\u4e2d&#xff1a;\u8fd9\u4e2a\u5f62\u72b6\u5c06\u662f input_sequence_length x 2000&#xff09;<\/p>\n<h3>&#x1f3ad; 6.\u6df7\u5408\u4e13\u5bb6&#xff08;MoE&#xff09;<\/h3>\n<p>\u4ec0\u4e48\u662f\u6df7\u5408\u4e13\u5bb6 (MoE)&#xff1f;<\/p>\n<p>\u4e3a\u4e86\u6e05\u695a\u5730\u7406\u89e3\u4ec0\u4e48\u662f MoE&#xff0c;\u9996\u5148\u8ba9\u6211\u4eec\u770b\u770b\u5b83\u5728 Transformer \u4e2d\u7684\u5177\u4f53\u7528\u9014\u4ee5\u53ca\u5b83\u7684\u67b6\u6784\u7b80\u4ecb\u3002\u6807\u51c6 Transformer \u5c42\u4e2d\u7684 FFN \u88ab MoE \u53d6\u4ee3\u3002<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21ggano0e0rc5.png\" alt=\"\" \/><\/p>\n<p>\u4ece\u672c\u8d28\u4e0a\u8bb2&#xff0c;MoE \u9075\u5faa\u6807\u51c6 Transformer \u8bbe\u8ba1&#xff0c;\u4f46\u901a\u8fc7\u5f15\u5165\u591a\u4e2a\u5e76\u884c\u4e13\u5bb6\u7f51\u7edc(FFN) \u800c\u4e0d\u662f\u5355\u4e2a\u5bc6\u96c6 FFN\u6765\u4fee\u6539\u524d\u9988\u5c42\u3002\u5176\u5de5\u4f5c\u539f\u7406\u5982\u4e0b&#xff1a;<\/p>\n<p>1. \u591a\u4e2a FFN&#xff08;\u800c\u975e\u4e00\u4e2a&#xff09;<\/p>\n<p>MoE \u4e0d\u4f7f\u7528\u5355\u4e2a\u5171\u4eab\u7684 FFN&#xff0c;\u800c\u662f\u4f7f\u7528\u591a\u4e2a\u5e76\u884c\u8bad\u7ec3\u7684FFN \u5c42&#xff08;\u4e13\u5bb6&#xff09; \u3002<\/p>\n<p>2. \u8f93\u5165\u5904\u7406\u548c\u4ee4\u724c\u8def\u7531<\/p>\n<ul>\n<li>\u6bcf\u4e2a token \u90fd\u50cf\u5f80\u5e38\u4e00\u6837\u7ecf\u8fc7 transformer\u81ea\u6ce8\u610f\u529b\u5c42\u3002<\/li>\n<li>\u5b83\u4e0d\u662f\u7531\u5355\u4e2a FFN \u5904\u7406&#xff0c;\u800c\u662f\u88ab\u53d1\u9001\u5230\u8def\u7531\u5668&#xff0c;\u7531\u8def\u7531\u5668\u51b3\u5b9a\u54ea\u4e9b\u4e13\u5bb6\u5e94\u8be5\u5904\u7406\u5b83\u3002<\/li>\n<\/ul>\n<p>3. \u901a\u8fc7\u8def\u7531\u5668\u9009\u62e9\u4e13\u5bb6<\/p>\n<ul>\n<li>\u4e00\u4e2a\u5c0f\u578b\u7684\u3001\u53ef\u8bad\u7ec3\u7684\u8def\u7531\u5668\u51b3\u5b9a\u54ea\u4e2a\u4e13\u5bb6\u5b50\u96c6&#xff08;FFN&#xff09;\u5e94\u8be5\u5904\u7406\u6bcf\u4e2a\u6807\u8bb0\u3002<\/li>\n<li>\u901a\u5e38&#xff0c;\u6bcf\u4e2a token\u4ec5\u9009\u62e9 1 \u6216 2 \u4e2a\u4e13\u5bb6\u4ee5\u4fdd\u6301\u6548\u7387&#xff08;\u4f8b\u5982&#xff0c; top-1 \u6216 top-2 \u95e8\u63a7&#xff09;\u3002DeepSeek -V3&#xff08;DeepSeek-R1&#xff09;\u4f7f\u7528 9 \u4e2a\u4e13\u5bb6&#xff0c;\u5176\u4e2d 1 \u4e2a\u662f\u5171\u4eab\u4e13\u5bb6&#xff0c;\u5176\u4ed6 8 \u4e2a\u662f\u8def\u7531\u4e13\u5bb6\u3002<\/li>\n<li>\u9009\u62e9\u901a\u5e38\u57fa\u4e8esoftmax \u8bc4\u5206\u673a\u5236&#xff0c;\u5176\u4e2d\u8def\u7531\u5668\u4e3a\u6bcf\u4e2a\u4e13\u5bb6\u5206\u914d\u6982\u7387\u3002\u5177\u4f53\u6765\u8bf4&#xff0c;\u5728 DeepSeek-V3 (DeepSeek-R1) \u4e2d&#xff0c;\u4f7f\u7528 Sigmoid \u800c\u4e0d\u662f softmax\u3002<\/li>\n<\/ul>\n<p>4. \u4e13\u5bb6\u7a00\u758f\u8ba1\u7b97<\/p>\n<ul>\n<li>\u53ea\u6709\u9009\u5b9a\u7684\u4e13\u5bb6\u624d\u80fd\u5904\u7406\u4ee4\u724c&#xff0c;\u800c\u5176\u4ed6\u4eba\u5219\u4fdd\u6301\u4e0d\u6d3b\u52a8\u72b6\u6001\u3002<\/li>\n<li>\u4e13\u5bb6\u8f93\u51fa\u4f7f\u7528\u52a0\u6743\u6c42\u548c\u8fdb\u884c\u7ec4\u5408&#xff0c;\u5e76\u4f20\u9012\u5230\u4e0b\u4e00\u4e2a Transformer \u5c42\u3002\u5728 DeepSeek-V3\/R1 \u4e2d&#xff0c;\u6743\u91cd\u662f\u5f52\u4e00\u5316\u7684 S \u578b\u8f93\u51fa\u3002<\/li>\n<li>\u8fd9\u79cd\u7a00\u758f\u6fc0\u6d3b\u53ef\u786e\u4fdd\u4efb\u4f55\u65f6\u5019\u4ec5\u4f7f\u7528\u6a21\u578b\u7684\u4e00\u5c0f\u90e8\u5206&#xff0c;\u4ece\u800c\u4fdd\u6301\u8ba1\u7b97\u53ef\u7ba1\u7406\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4ec0\u4e48\u8981\u7528 MoE \u53d6\u4ee3\u5355\u4e00 FFN&#xff1f;<\/p>\n<ul>\n<li>\u53ef\u6269\u5c55\u6027\u2014 MoE \u5141\u8bb8\u6a21\u578b\u4f7f\u7528\u66f4\u591a\u53c2\u6570\u8fdb\u884c\u6269\u5c55&#xff0c;\u800c\u65e0\u9700\u7ebf\u6027\u589e\u52a0\u8ba1\u7b97\u91cf\u3002<\/li>\n<li>\u9ad8\u6548\u5b66\u4e60\u2014 \u4e13\u5bb6\u4e13\u6ce8\u4e8e\u6570\u636e\u7684\u4e0d\u540c\u65b9\u9762&#xff0c;\u4ece\u800c\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n<li>\u8ba1\u7b97\u8282\u7701\u2014 \u7531\u4e8e\u6bcf\u4e2a token \u4ec5\u4f7f\u7528\u4e13\u5bb6\u5b50\u96c6&#xff0c;\u56e0\u6b64\u4e0e\u76f8\u540c\u5927\u5c0f\u7684\u5bc6\u96c6\u6a21\u578b\u76f8\u6bd4&#xff0c;MoE \u6a21\u578b\u7684\u8fd0\u884c\u6210\u672c\u66f4\u4f4e\u3002DeepSeek-V3\/R1 \u5171\u6709 6710 \u4ebf\u4e2a\u53c2\u6570&#xff0c;\u5176\u4e2d\u6bcf\u4e2a token \u6fc0\u6d3b 370 \u4ebf\u4e2a\u53c2\u6570\u3002<\/li>\n<\/ul>\n<p>MoE \u5728 DeepSeek-R1 \u4e2d\u5982\u4f55\u53d1\u6325\u4f5c\u7528&#xff1f;<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21n4ari0mkhhm.png\" alt=\"\" \/><\/p>\n<p>\u4ee5\u4e0b\u6765\u81ea DeepSeek-V3 \u6280\u672f\u62a5\u544a\u7684\u516c\u5f0f\u663e\u793a\u4e86\u6bcf\u4e2a MoE \u5c42\u4e2d\u7684\u8ba1\u7b97\u3002\u5728 DeepSeek \u7cfb\u5217\u6a21\u578b\u4e2d&#xff0c;MoE \u67b6\u6784\u9996\u6b21\u5728 DeepSeekMoE \u6a21\u578b\u4e2d\u5f15\u5165&#xff0c;\u5e76\u4e14\u6b63\u5728 DeepSeek-V2\u3001DeepSeek-V3 \u548c DeepSeek-R1 \u4e2d\u4f7f\u7528\u3002<\/p>\n<p>\u8def\u7531\u5668\u8ba1\u7b97&#xff1a;<\/p>\n<p>\u5728 DeepSeek-V3 \u3001 DeepSeek-R1 \u548c\u5176\u4ed6\u4e00\u4e9b\u73b0\u4ee3\u6df7\u5408\u4e13\u5bb6 (MoE) \u6a21\u578b\u4e2d&#xff0c;e_i\u200b \u8868\u793a\u5b66\u4e60\u5230\u7684\u8d28\u5fc3&#xff0c;\u6709\u52a9\u4e8e\u5c06\u8f93\u5165\u8def\u7531\u5230\u6b63\u786e\u7684\u4e13\u5bb6\u3002\u4e0e\u4f20\u7edf\u7684 MoE \u67b6\u6784\u4e2d\u57fa\u4e8e FFN \u7684\u8def\u7531\u5668\u8ba1\u7b97\u95e8\u63a7\u5206\u6570\u4e0d\u540c&#xff0c;\u6b64\u65b9\u6cd5\u9884\u5b9a\u4e49\u4e86\u4e00\u7ec4\u53ef\u5b66\u4e60\u5411\u91cfe_i\u200b&#xff0c;\u6bcf\u4e2a\u5411\u91cf\u5bf9\u5e94\u4e00\u4f4d\u4e13\u5bb6\u3002<\/p>\n<p>\u5173\u952e\u601d\u60f3&#xff1a;<\/p>\n<ul>\n<li>\u6bcf\u4e2a\u4e13\u5bb6 i \u90fd\u6709\u4e00\u4e2a\u76f8\u5173\u7684\u8d28\u5fc3\u5411\u91cf<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>           e <\/p>\n<p>           i <\/p>\n<p>         e_i <\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">e<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">i<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002<\/li>\n<li>\u6211\u4eec\u4e0d\u662f\u5c06\u8f93\u5165 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>           u <\/p>\n<p>           t <\/p>\n<p>          \u200b <\/p>\n<p>         u_t\u200b <\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">u<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mord\">\u200b<\/span><\/span><\/span><\/span><\/span> \u4f20\u9012\u7ed9 FFN \u6765\u83b7\u53d6\u4e13\u5bb6\u6982\u7387&#xff0c;\u800c\u662f\u901a\u8fc7\u70b9\u79ef\u6765\u8ba1<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>           u <\/p>\n<p>           t <\/p>\n<p>         u_t <\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">u<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u548c\u6bcf\u4e2a <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>           e <\/p>\n<p>           i <\/p>\n<p>          \u200b <\/p>\n<p>         e_i\u200b <\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">e<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">i<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mord\">\u200b<\/span><\/span><\/span><\/span><\/span> \u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6&#xff1a; <img decoding=\"async\" src=\"2025-04-21fdxaro2ramu.png\" alt=\"\" \/><\/li>\n<li>\u8be5\u5206\u6570\u51b3\u5b9a\u4e86\u4e13\u5bb6\u4e0e\u7ed9\u5b9a\u8f93\u5165\u7684\u76f8\u5173\u7a0b\u5ea6\u3002<\/li>\n<li>\u4ec5\u6fc0\u6d3b\u5177\u6709\u6700\u9ad8 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>           s <\/p>\n<p>            i <\/p>\n<p>            , <\/p>\n<p>            t <\/p>\n<p>         s_{i,t} <\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.7167em;vertical-align: -0.2861em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">s<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">i<\/span><span class=\"mpunct mtight\">,<\/span><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2861em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b \u503c\u7684Top-K\u4e13\u5bb6\u8fdb\u884c\u5904\u7406\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2025-04-21q1xbh1x14ls.png\" alt=\"\" \/><\/p>\n<ul>\n<li>\u5728 S \u578b\u8f93\u51fa\u4e2d\u6dfb\u52a0\u4e86\u4e00\u4e2a\u504f\u5dee\u9879&#xff0c;\u4ee5\u521b\u5efa\u65e0\u8f85\u52a9\u635f\u5931\u7684 MoE \u8d1f\u8f7d\u5e73\u8861\u3002<\/li>\n<\/ul>\n<p>DeepSeek-V3 \u8bba\u6587\u4e2d\u7684\u8fd9\u6bb5\u63cf\u8ff0\u8fdb\u4e00\u6b65\u9610\u660e\u4e86\u5b83\u7684\u7528\u9014\u4ee5\u53ca\u5b83\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u65b9\u5f0f&#xff1a;<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21mpwewhs50zx.png\" alt=\"\" \/><\/p>\n<ul>\n<li>\u4f7f\u7528\u9009\u5b9a\u7684 top-k \u503c\u5bf9\u8f93\u51fa\u503c\u8fdb\u884c\u6807\u51c6\u5316\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2025-04-215jdk0z3ofzy.png\" alt=\"\" \/> \u4e13\u5bb6\u8ba1\u7b97&#xff1a;<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21yg5k4isller.png\" alt=\"\" \/> <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          u <\/p>\n<p>          t <\/p>\n<p>        u_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">u<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u662f MoE \u5c42\u7684\u8f93\u5165\u3002\u7b49\u5f0f\u4e2d\u7684\u7b2c\u4e8c\u9879\u8868\u793a\u8f93\u5165\u4e0e\u5171\u4eab\u4e13\u5bb6\u76f8\u4e58\u3002\u6bcf\u4e2a\u4e13\u5bb6\u7531 FFN&#xff08;\u524d\u9988\u7f51\u7edc&#xff09;\u7ec4\u6210&#xff0c;\u56e0\u6b64\u7528\u201cFFN\u201d\u8868\u793a\u3002\u5728 DeepSeek-R1 \u4e2d&#xff0c;\u53ea\u6709 1 \u4e2a\u5171\u4eab\u4e13\u5bb6\u3002\u56e0\u6b64&#xff0c;Ns&#061;1\u3002\u540c\u6837&#xff0c;\u7b49\u5f0f\u4e2d\u7684\u7b2c\u4e09\u9879\u8868\u793a\u8f93\u5165\u4e0e\u6d3b\u8dc3\u7684\u4e2a\u4eba\u4e13\u5bb6\u76f8\u4e58\u3002\u5728 DeepSeek-R1 \u4e2d&#xff0c;\u603b\u5171\u6709 256 \u4e2a\u4e2a\u4eba\u4e13\u5bb6\u3002\u4f46\u6bcf\u4e2a token \u53ea\u6709 8 \u4e2a\u6d3b\u8dc3&#xff0c;\u56e0\u6b64 Nr&#061;8\u3002\u6bcf\u4e2a\u6d3b\u8dc3\u7684\u4e2a\u4eba\u4e13\u5bb6\u90fd\u5c06\u5177\u6709\u4e0e\u7b49\u5f0f 13 \u4e2d\u5173\u8054\u7684<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          g <\/p>\n<p>           i <\/p>\n<p>           , <\/p>\n<p>           t <\/p>\n<p>        g_{i,t} <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.7167em;vertical-align: -0.2861em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">g<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">i<\/span><span class=\"mpunct mtight\">,<\/span><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2861em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002\u5b83\u7528\u4e8e\u8ba1\u7b97\u7b2c\u4e09\u9879\u3002<\/p>\n<p>\u8f93\u51fa<\/p>\n<p><img decoding=\"async\" src=\"2025-04-210cytfu1hxac.png\" alt=\"\" \/> <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          h <\/p>\n<p>          t <\/p>\n<p>        h_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">h<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a MoE \u5c42\u7684\u8f93\u51fa\u3002<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          u <\/p>\n<p>          t <\/p>\n<p>        u_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">u<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f MoE \u5c42\u7684\u8f93\u5165\u3002\u4e13\u5bb6\u8ba1\u7b97\u7ed3\u679c\u6dfb\u52a0\u5230\u8f93\u5165 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          u <\/p>\n<p>          t <\/p>\n<p>        u_t <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5806em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">u<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2806em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">t<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u4e2d&#xff0c;\u5f97\u5230 MoE \u5c42\u7684\u8f93\u51fa\u3002<\/p>\n<h3>&#x1f522; 7. \u591a\u6807\u8bb0\u9884\u6d4b&#xff08;MTP&#xff09;<\/h3>\n<p>\u4ec0\u4e48\u662f\u591a\u6807\u8bb0\u9884\u6d4b&#xff1f;<\/p>\n<p>\u591a\u6807\u8bb0\u9884\u6d4b\u662f\u8bed\u8a00\u5efa\u6a21\u4e2d\u7684\u4e00\u79cd\u9ad8\u7ea7\u65b9\u6cd5&#xff0c;\u5176\u4e2d\u6a21\u578b\u4e0d\u662f\u4e00\u6b21\u9884\u6d4b\u4e00\u4e2a\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a\u5355\u8bcd&#xff0c;\u800c\u662f\u540c\u65f6\u9884\u6d4b\u591a\u4e2a\u672a\u6765\u6807\u8bb0\u3002\u6b64\u65b9\u6cd5\u4f7f\u6a21\u578b\u80fd\u591f\u5e76\u884c\u9884\u6d4b\u591a\u4e2a\u5373\u5c06\u5230\u6765\u7684\u5355\u8bcd&#xff0c;\u4ece\u800c\u63d0\u9ad8\u5b66\u4e60\u6548\u7387\u5e76\u52a0\u901f\u6587\u672c\u751f\u6210\u3002<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21uju5cg1twr0.png\" alt=\"\" \/><\/p>\n<p><img decoding=\"async\" src=\"2025-04-21gbspor3duei.png\" alt=\"\" \/> Meta \u5f15\u5165\u4e86\u4e00\u79cd\u591a\u6807\u8bb0\u9884\u6d4b\u67b6\u6784&#xff0c;\u53ef\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u540c\u65f6\u9884\u6d4b\u591a\u4e2a\u672a\u6765\u6807\u8bb0&#xff0c;\u4ece\u800c\u63d0\u9ad8\u91c7\u6837\u6548\u7387\u5e76\u52a0\u5feb\u63a8\u7406\u901f\u5ea6\u3002\u5728\u6b64\u6982\u5ff5\u7684\u57fa\u7840\u4e0a&#xff0c;DeepSeek-V3 \u6574\u5408\u4e86\u591a\u6807\u8bb0\u9884\u6d4b (MTP) \u76ee\u6807&#xff0c;\u4f7f\u6a21\u578b\u80fd\u591f\u540c\u65f6\u9884\u6d4b\u591a\u4e2a\u6807\u8bb0\u3002\u8fd9\u79cd\u65b9\u6cd5\u4f7f\u8bad\u7ec3\u4fe1\u53f7\u5bc6\u96c6\u5316&#xff0c;\u5e76\u80fd\u591f\u66f4\u597d\u5730\u9884\u5148\u89c4\u5212\u6807\u8bb0\u8868\u793a&#xff0c;\u4ece\u800c\u63d0\u9ad8\u590d\u6742\u57fa\u51c6\u6d4b\u8bd5\u7684\u6027\u80fd\u3002<\/p>\n<p>DeepSeek-V3\/R1 \u548c Meta \u7684\u591a\u4ee4\u724c\u9884\u6d4b\u6709\u4e24\u4e2a\u5173\u952e\u533a\u522b&#xff1a;<\/p>\n<p>\u201c\u4e0e Gloeckle \u7b49\u4eba&#xff08;2024 \u5e74&#xff09;[Meta Research] \u4f7f\u7528\u72ec\u7acb\u8f93\u51fa\u5934\u5e76\u884c\u9884\u6d4b &#x1d437; \u4e2a\u989d\u5916\u6807\u8bb0\u4e0d\u540c&#xff0c;\u6211\u4eec\u6309\u987a\u5e8f\u9884\u6d4b\u5176\u4ed6\u6807\u8bb0&#xff0c;\u5e76\u5728\u6bcf\u4e2a\u9884\u6d4b\u6df1\u5ea6\u4fdd\u7559\u5b8c\u6574\u7684\u56e0\u679c\u94fe\u3002\u201d \u2014 DeepSeek-V3<\/p>\n<li>Meta \u7684\u6a21\u578b\u9884\u6d4b\u4e864 \u4e2a token&#xff0c;\u800c DeepSeek-V3 \u9884\u6d4b\u4e862 \u4e2a token\u3002<\/li>\n<li>Meta \u6a21\u578b\u4e2d\u7684\u9884\u6d4b\u5934\u662f\u72ec\u7acb\u7684&#xff0c;\u800c DeepSeek-V3 \u7684\u9884\u6d4b\u5934\u662f\u987a\u5e8f\u8fde\u63a5\u7684\u3002<\/li>\n<p>MTP \u5728 DeepSeek-R1 \u4e2d\u5982\u4f55\u5de5\u4f5c&#xff1f;<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21icwgftdhhng.png\" alt=\"\" \/> \u8ba9\u6211\u4eec\u4e00\u6b65\u4e00\u6b65\u5730\u770b\u4e00\u4e0b\u8be5\u56fe\u8868\u3002<\/p>\n<p>\u5728\u8bad\u7ec3\u671f\u95f4&#xff0c;\u8f93\u5165\u6807\u8bb0&#xff08;\u4f4d\u4e8e\u5de6\u4e0b\u89d2&#xff09;\u7a7f\u8fc7\u5d4c\u5165\u5c42&#xff0c;\u7136\u540e\u4f20\u64ad\u5230\u6240\u6709\u53d8\u538b\u5668\u5757\/\u5c42\u3002<\/p>\n<p>\u7b2c\u4e00\u4e2a\u9884\u6d4b\u5934&#xff08;\u5305\u62ec\u8f93\u51fa\u5934&#xff09;\u76f4\u63a5\u8fde\u63a5\u5230\u4e3b\u6a21\u578b\u7684\u6700\u7ec8 Transformer \u5c42\u3002\u8f93\u51fa\u5934\u901a\u5e38\u662f\u524d\u9988\u7f51\u7edc (FFN)&#xff0c;\u5176\u8f93\u51fa\u7ef4\u5ea6\u4e0e\u6a21\u578b\u7684\u8bcd\u6c47\u91cf\u76f8\u5339\u914d\u3002\u8be5\u5934\u8d1f\u8d23\u6309\u987a\u5e8f\u9884\u6d4b\u4e0b\u4e00\u4e2a\u6807\u8bb0\u3002\u7ed9\u5b9a\u8f93\u5165\u6807\u8bb0t\u2081\u3001t\u2082\u3001t\u2083\u3001t\u2084&#xff0c;\u5b83\u4f1a\u9884\u6d4bt\u2082\u3001t\u2083\u3001t\u2084\u3001t\u2085 \u3002\u4f46\u662f&#xff0c;\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d&#xff0c;\u53ea\u8ba1\u7b97\u6700\u7ec8\u6807\u8bb0t\u2085 \u3002<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u9884\u6d4b\u5934\u901a\u8fc7\u6dfb\u52a0\u989d\u5916\u7684\u53ef\u5b66\u4e60\u5c42\u6269\u5c55\u4e86\u8fd9\u79cd\u65b9\u6cd5\u3002\u5b83\u4ece\u4e3b\u6a21\u578b\u7684\u6700\u7ec8 Transformer \u5c42\u83b7\u53d6\u8f93\u51fa&#xff0c;\u5e94\u7528 RMSNorm \u8fdb\u884c\u5f52\u4e00\u5316&#xff0c;\u7136\u540e\u5c06\u5176\u4e0e\u8f93\u5165\u5d4c\u5165\u8fde\u63a5\u8d77\u6765\u3002\u8fd9\u4e9b\u8f93\u5165\u5d4c\u5165\u662f\u4ece\u4e3b\u6a21\u578b\u4e2d\u4f7f\u7528\u7684\u76f8\u540c\u5d4c\u5165\u5c42\u83b7\u5f97\u7684\u3002\u4e0e\u7b2c\u4e00\u4e2a\u9884\u6d4b\u5934\u4e0d\u540c&#xff0c;\u8fd9\u4e2a\u5934\u4ecet\u2082\u800c\u4e0d\u662ft\u2081\u5f00\u59cb\u5904\u7406\u8f93\u5165\u6807\u8bb0\u3002\u7136\u540e\u4f7f\u7528\u7ebf\u6027\u6295\u5f71\u5c42\u5c06\u8fde\u63a5\u7684\u8f93\u51fa\u6295\u5f71\u5230\u5408\u9002\u7684\u5d4c\u5165\u5927\u5c0f&#xff0c;\u7136\u540e\u4f7f\u7528\u53ef\u5b66\u4e60\u7684 Transformer \u5757\/\u5c42\u8fdb\u884c\u8fdb\u4e00\u6b65\u5904\u7406\u3002\u5728\u8bad\u7ec3\u671f\u95f4&#xff0c;\u8fd9\u4e2a\u5934\u5c06t\u2083\u9884\u6d4b\u4e3at\u2086&#xff0c;\u4f46\u5728\u63a8\u7406\u4e2d&#xff0c;\u53ea\u8ba1\u7b97t\u2086 \u3002<\/p>\n<p>\u7c7b\u4f3c\u5730&#xff0c;\u7b2c\u4e09\u4e2a\u9884\u6d4b\u5934\u4ece\u7b2c\u4e8c\u4e2a\u9884\u6d4b\u5934\u7684transformer\u5668\u5757\/\u5c42\u83b7\u53d6\u8f93\u5165\u4ee5\u53ca\u76f8\u5e94\u7684\u8f93\u5165\u5d4c\u5165&#xff0c;\u73b0\u5728\u4ecet\u2083\u5f00\u59cb\u5230t\u2086\u3002\u5b83\u9075\u5faa\u4e0e\u524d\u51e0\u4e2a\u5934\u76f8\u540c\u7684\u7ed3\u6784&#xff0c;\u5728\u8bad\u7ec3\u671f\u95f4\u9884\u6d4bt\u2084\u5230t\u2087&#xff0c;\u4f46\u5728\u63a8\u7406\u671f\u95f4\u4ec5\u8ba1\u7b97t\u2087\u3002<\/p>\n<p>\u6bcf\u4e2a\u9884\u6d4b\u5934\u4f7f\u7528\u4ea4\u53c9\u71b5\u8ba1\u7b97\u635f\u5931\u3002\u7136\u540e&#xff0c;\u8fd9\u4e9b\u635f\u5931\u7528\u56e0\u5b50\u03bb\u52a0\u6743&#xff0c;\u53d6\u5176\u5e73\u5747\u503c\u4f5c\u4e3a\u6700\u7ec8\u635f\u5931\u503c\u3002<\/p>\n<p><img decoding=\"async\" src=\"2025-04-212irczd2eo5r.png\" alt=\"\" \/><\/p>\n<p>\u5355\u72ec\u9884\u6d4b\u5934\u635f\u5931<\/p>\n<p><img decoding=\"async\" src=\"2025-04-21qvk4nwpdn5o.png\" alt=\"\" \/><\/p>\n<p>\u6700\u7ec8\u635f\u5931<\/p>\n<p>\u5728 DeepSeek-V3 \u548c R1 \u4e2d&#xff0c;MTP \u4ec5\u5728\u8bad\u7ec3\u671f\u95f4\u4f7f\u7528&#xff0c;\u800c\u4e0d\u5728\u63a8\u7406\u671f\u95f4\u4f7f\u7528&#xff1a;<\/p>\n<p>\u201c\u63a8\u7406\u4e2d\u7684 MTP&#xff1a;\u6211\u4eec\u7684 MTP \u7b56\u7565\u4e3b\u8981\u662f\u4e3a\u4e86\u63d0\u9ad8\u4e3b\u6a21\u578b\u7684\u6027\u80fd&#xff0c;\u56e0\u6b64\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d&#xff0c;\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u4e22\u5f03 MTP \u6a21\u5757&#xff0c;\u4e3b\u6a21\u578b\u53ef\u4ee5\u72ec\u7acb\u6b63\u5e38\u8fd0\u884c\u3002\u201d\u2014\u2014DeepSeek-V3<\/p>\n<h3>\u53c2\u8003\u8d44\u6599<\/h3>\n<ul>\n<li>https:\/\/arxiv.org\/pdf\/2412.19437<\/li>\n<li>https:\/\/arxiv.org\/pdf\/2405.04434<\/li>\n<li>https:\/\/arxiv.org\/pdf\/2402.03300<\/li>\n<li>https:\/\/arxiv.org\/pdf\/2401.02954<\/li>\n<li>https:\/\/arxiv.org\/pdf\/2401.06066<\/li>\n<li>https:\/\/arxiv.org\/pdf\/2404.19737<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb6.8k\u6b21\uff0c\u70b9\u8d5e42\u6b21\uff0c\u6536\u85cf76\u6b21\u3002DeepSeek-R1 (DeepSeek-V3) \u603b\u5171\u5305\u542b 671B \u4e2a\u53c2\u6570\uff0c\u5176\u4e2d\u6bcf\u4e2a\u6807\u8bb0\u6fc0\u6d3b 37B \u4e2a\u3002\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\u3002_deepseek-r1\u67b6\u6784<\/p>\n","protected":false},"author":2,"featured_media":31222,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[68,50,132],"topic":[],"class_list":["post-31242","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-deepseek","tag-50","tag-132"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/31242.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb6.8k\u6b21\uff0c\u70b9\u8d5e42\u6b21\uff0c\u6536\u85cf76\u6b21\u3002DeepSeek-R1 (DeepSeek-V3) \u603b\u5171\u5305\u542b 671B \u4e2a\u53c2\u6570\uff0c\u5176\u4e2d\u6bcf\u4e2a\u6807\u8bb0\u6fc0\u6d3b 37B \u4e2a\u3002\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\u3002_deepseek-r1\u67b6\u6784\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/31242.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2025-04-21T01:04:08+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010400-68059980d42ec.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/31242.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/31242.html\",\"name\":\"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2025-04-21T01:04:08+00:00\",\"dateModified\":\"2025-04-21T01:04:08+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/31242.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/31242.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/31242.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/31242.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb6.8k\u6b21\uff0c\u70b9\u8d5e42\u6b21\uff0c\u6536\u85cf76\u6b21\u3002DeepSeek-R1 (DeepSeek-V3) \u603b\u5171\u5305\u542b 671B \u4e2a\u53c2\u6570\uff0c\u5176\u4e2d\u6bcf\u4e2a\u6807\u8bb0\u6fc0\u6d3b 37B \u4e2a\u3002\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\u3002_deepseek-r1\u67b6\u6784","og_url":"https:\/\/www.wsisp.com\/helps\/31242.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2025-04-21T01:04:08+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250421010400-68059980d42ec.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/31242.html","url":"https:\/\/www.wsisp.com\/helps\/31242.html","name":"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2025-04-21T01:04:08+00:00","dateModified":"2025-04-21T01:04:08+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/31242.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/31242.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/31242.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4e00\u6587\u6df1\u5165\u4e86\u89e3DeepSeek-R1\uff1a\u6a21\u578b\u67b6\u6784"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/31242","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=31242"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/31242\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/31222"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=31242"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=31242"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=31242"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=31242"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}