{"id":39291,"date":"2025-03-08T08:02:22","date_gmt":"2025-03-08T00:02:22","guid":{"rendered":"https:\/\/17aitech.com\/?p=39291"},"modified":"2025-03-08T08:02:22","modified_gmt":"2025-03-08T00:02:22","slug":"deepseek%e7%9a%84mla%ef%bc%8c%e4%bb%bb%e6%84%8f%e5%a4%a7%e6%a8%a1%e5%9e%8b%e9%83%bd%e8%83%bd%e8%bd%bb%e6%9d%be%e8%bf%81%e7%a7%bb%e4%ba%86","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39291","title":{"rendered":"DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-06-16\" target=\"_blank\">DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png\"><\/a><\/section>\n<p><strong>\u590d\u65e6 NLP \u5b9e\u9a8c\u5ba4\u535a\u58eb\u540e\u7eaa\u7118\u662f\u8fd9\u7bc7\u6587\u7ae0\u7684\u7b2c\u4e00\u4f5c\u8005\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u5927\u6a21\u578b\u9ad8\u6548\u63a8\u7406\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\uff0c\u8fd1\u671f\u4ee3\u8868\u5de5\u4f5c\u4e3a\u9996\u4e2aNoPE\u5916\u63a8HeadScale\u3001\u6ce8\u610f\u529b\u5206\u5757\u5916\u63a8LongHeads\u3001\u591a\u89c6\u89c9\u4e13\u5bb6\u5927\u6a21\u578bMouSi\uff0c\u53d1\u8868ACL\u3001ICLR\u3001EMNLP\u7b49\u9876\u4f1a\u9876\u520a\u8bba\u6587 20 \u4f59\u7bc7\u3002<\/strong><\/p>\n<p>DeepSeek-R1 \u4f5c\u4e3a AI \u4ea7\u4e1a\u98a0\u8986\u5f0f\u521b\u65b0\u7684\u4ee3\u8868\u8f70\u52a8\u4e86\u4e1a\u754c\uff0c\u7279\u522b\u662f\u5176\u8bad\u7ec3\u4e0e\u63a8\u7406\u6210\u672c\u4ec5\u4e3a\u540c\u7b49\u6027\u80fd\u5927\u6a21\u578b\u7684\u6570\u5341\u5206\u4e4b\u4e00\u3002\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\u7f51\u7edc\uff08Multi-head Latent Attention, MLA\uff09\u662f\u5176\u7ecf\u6d4e\u63a8\u7406\u67b6\u6784\u7684\u6838\u5fc3\u4e4b\u4e00\uff0c\u901a\u8fc7\u5bf9\u952e\u503c\u7f13\u5b58\u8fdb\u884c\u4f4e\u79e9\u538b\u7f29\uff0c\u663e\u8457\u964d\u4f4e\u63a8\u7406\u6210\u672c [1]\u3002<\/p>\n<p>\u7136\u800c\uff0c\u73b0\u6709\u4e3b\u6d41\u5927\u6a21\u578b\u4ecd\u7136\u57fa\u4e8e\u6807\u51c6\u6ce8\u610f\u529b\u67b6\u6784\u53ca\u5176\u53d8\u79cd\uff08e.g., MHA, GQA, MQA\uff09\uff0c\u63a8\u7406\u6210\u672c\u76f8\u6bd4 MLA \u5448\u73b0\u663e\u8457\u52a3\u52bf\u3002\u4f7f\u9884\u8bad\u7ec3\u7684\u4efb\u610f LLMs \u5feb\u901f\u8fc1\u79fb\u81f3 MLA \u67b6\u6784\u800c\u65e0\u9700\u4ece\u5934\u9884\u8bad\u7ec3\uff0c\u8fd9\u65e2\u6709\u91cd\u5927\u610f\u4e49\u53c8\u5177\u6709\u6311\u6218\u6027\u3002<\/p>\n<p><strong>\u590d\u65e6 NLP \u5b9e\u9a8c\u5ba4\u3001\u534e\u4e1c\u5e08\u5927\u3001\u4e0a\u6d77 AI Lab\u3001\u6d77\u5eb7\u5a01\u89c6\u8054\u5408\u63d0\u51fa MHA2MLA \u6846\u67b6<\/strong>\uff0c\u901a\u8fc7\u90e8\u5206 <strong>RoPE \u4fdd\u7559\uff08Partial-RoPE\uff09<\/strong>\u548c<strong>\u952e\u503c\u8054\u5408\u8868\u793a\u4f4e\u79e9\u8fd1\u4f3c\uff08Low-rank Approximation\uff09<\/strong>\u4e24\u4e2a\u5173\u952e\u6b65\u9aa4\uff0c\u6210\u529f\u5c06\u4efb\u610f MHA\/GQA \u67b6\u6784\u8fc1\u79fb\u5230 MLA\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e47c2a79598039b760c219ce9b17f1cd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e47c2a79598039b760c219ce9b17f1cd.png\"><\/a><\/section>\n<section>\u76ee\u524d\uff0cMHA2MLA \u5df2\u4f4d\u5217\ud83d\ude80alphaXiv \u70ed\u5ea6\u699c\ud83d\udd25<\/section>\n<section><\/section>\n<section>\u590d\u65e6 NLP \u5b9e\u9a8c\u5ba4\u535a\u58eb\u540e\u7eaa\u7118\u4e3a\u7b2c\u4e00\u4f5c\u8005\uff0c\u526f\u7814\u7a76\u5458\u6842\u97ec\u4e3a\u901a\u8baf\u4f5c\u8005\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-a491e9c6fff018d54ab97512c241ef76.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-a491e9c6fff018d54ab97512c241ef76.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aTowards Economical Inference: Enabling \u00a0DeepSeek&#8217;s Multi-Head Latent Attention in Any Transformer-based LLMs<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2502.14837<\/section>\n<\/li>\n<li>\n<section>\u5f00\u6e90\u4ee3\u7801\uff1ahttps:\/\/github.com\/JT-Ushio\/MHA2MLA<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u8bba\u6587\u6982\u89c8<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u6587\u805a\u7126\u5982\u4f55\u5c06\u9884\u8bad\u7ec3\u7684\u57fa\u4e8e MHA\/GQA \u7684\u5927\u8bed\u8a00\u6a21\u578b\u9ad8\u6548\u8fc1\u79fb\u5230 DeepSeek \u63d0\u51fa\u7684\u7ecf\u6d4e\u63a8\u7406\u67b6\u6784 \u2014\u2014\u00a0<strong>\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\uff08MLA\uff09<\/strong>\u3002<\/section>\n<section><\/section>\n<section>MHA \u4e0e MLA \u5728\u591a\u5904\u5b58\u5728\u5dee\u5f02\uff0c\u4f7f\u5f97 MHA2MLA \u6781\u5177\u6311\u6218\uff1a<\/section>\n<section><\/section>\n<ol>\n<li>\n<section><strong>\u4f4d\u7f6e\u7f16\u7801\u4e0d\u540c<\/strong>\uff1aMHA \u91c7\u7528\u5168\u7ef4\u5ea6\u4f4d\u7f6e\u7f16\u7801\uff08PE\uff09\uff0cMLA \u4ec5\u5c11\u91cf\u7ef4\u5ea6\u91c7\u7528 PE\uff0c\u5269\u4f59\u7ef4\u5ea6\u5219 PE \u65e0\u5173<\/section>\n<\/li>\n<li>\n<section><strong>\u7f13\u5b58\u5bf9\u8c61\u4e0d\u540c<\/strong>\uff1aMHA \u7f13\u5b58\u5206\u79bb\u7684\u952e\u5411\u91cf\u53ca\u503c\u5411\u91cf\uff0cMLA \u7f13\u5b58\u5e26 PE \u7684\u952e\u5411\u91cf\u53ca PE \u65e0\u5173\u7684\u952e\u503c\u8054\u5408\u4f4e\u7ef4\u8868\u793a\u5411\u91cf<\/section>\n<\/li>\n<li>\n<section><strong>\u53c2\u6570\u77e9\u9635\u4e0d\u540c<\/strong>\uff1aMHA \u5305\u542b\u67e5\u8be2\u3001\u952e\u3001\u503c\u4e09\u4e2a\u7ebf\u6027\u53d8\u6362\u77e9\u9635\uff0cMLA \u5219\u66f4\u52a0\u590d\u6742\u3001\u591a\u8fbe\u4e03\u4e2a\u76ee\u7684\u4e0d\u540c\u7684\u7ebf\u6027\u53d8\u6362\u77e9\u9635<\/section>\n<\/li>\n<li>\n<section><strong>\u8fd0\u7b97\u5f62\u5f0f\u4e0d\u540c<\/strong>\uff1aMHA \u7684\u8fd0\u7b97\u53d7\u9650\u4e8e\u8bbf\u5b58\u74f6\u9888\uff0cMLA \u5219\u80fd\u901a\u8fc7\u77e9\u9635\u5438\u6536\u7b49\u4f18\u5316\u5b9e\u73b0\u66f4\u9ad8\u7684\u8bbf\u5b58\u6548\u7387<\/section>\n<\/li>\n<\/ol>\n<section><\/section>\n<section>\u672c\u6587\u63d0\u51fa\u7684 MHA2MLA \u4e3a\u4e86\u6700\u5927\u5316\u5229\u7528 MHA \u9884\u8bad\u7ec3\u53c2\u6570\u77e9\u9635\u5e76\u5bf9\u9f50 MLA \u7684\u7f13\u5b58\u5bf9\u8c61\u548c\u8fd0\u7b97\u5f62\u5f0f\uff0c\u9996\u5148\u901a\u8fc7\u90e8\u5206 RoPE \u4fdd\u7559\uff08Partial-RoPE\uff09\u5206\u79bb\u51fa PE \u76f8\u5173\u8868\u793a\uff08\u5c11\u91cf\u7ef4\u5ea6\uff0c\u5982 1\/8\uff09\u548c PE \u65e0\u5173\u8868\u793a\uff08\u5927\u91cf\u7ef4\u5ea6\uff09\uff0c\u5176\u4e2d <strong>PE \u76f8\u5173\u7684\u952e\u5411\u91cf\u5bf9\u9f50 MLA<\/strong>\u3002\u5176\u6b21\u62fc\u63a5\u503c\u7684\u53d8\u6362\u77e9\u9635\uff08W_v\uff09\u548c PE \u65e0\u5173\u7684\u952e\u7684\u53d8\u6362\u77e9\u9635\uff08W_{k, nope}\uff09\uff0c\u5e76\u8fdb\u884c SVD \u5206\u89e3\u5f97\u5230\u964d\u7ef4\u53d8\u6362\u77e9\u9635\u548c\u5347\u7ef4\u53d8\u5316\u77e9\u9635\uff0c\u4e2d\u95f4\u7684<strong>\u952e\u503c\u8054\u5408\u4f4e\u79e9\u8868\u793a\u5bf9\u9f50 MLA<\/strong>\uff0c\u5b8c\u6210\u4e86\u7f13\u5b58\u5bf9\u8c61\u7684\u5bf9\u9f50\u4ee5\u53ca\u8fd0\u7b97\u5f62\u5f0f\u7684\u5bf9\u9f50\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-cbf0a23841fbaadf2b3191e9539b35ca.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-cbf0a23841fbaadf2b3191e9539b35ca.png\"><\/a><\/section>\n<section>\u5728 135M~7B \u4e0a\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u4ec5\u9700\u4f7f\u7528\u9884\u8bad\u7ec3\u6570\u636e\u7684 0.3% \u5230 0.6% \u8fdb\u884c\u9ad8\u6548\u5fae\u8c03\uff0c\u5373\u53ef\u57fa\u672c\u8fd8\u539f\u67b6\u6784\u8fc1\u79fb\u5e26\u6765\u7684\u6027\u80fd\u635f\u5931\u3002\u5e76\u4e14 MHA2MLA \u8fd8\u80fd\u7ed3\u5408\u5176\u4ed6\u9ad8\u6548\u63a8\u7406\u6280\u672f\uff0c\u4f8b\u5982\u7ed3\u5408 4-bit KV \u7f13\u5b58\u91cf\u5316\uff0c<strong>Llama2-7B \u51cf\u5c11\u4e86 92.19% KV \u7f13\u5b58\uff0c\u800c LongBench \u4e0a\u7684\u6027\u80fd\u4ec5\u4e0b\u964d 0.5%\u3002<\/strong><\/section>\n<section><\/section>\n<section><strong>\u90e8\u5206 RoPE \u4fdd\u7559\uff08Partial-RoPE\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u5b9e\u73b0\u4ece\u6807\u51c6\u7684 MHA\uff08\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff09\u5230 MLA\uff08\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\u673a\u5236\uff09\u7684\u8fc1\u79fb\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u90e8\u5206 RoPE \u5fae\u8c03\uff08partial-RoPE finetuning\uff09\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u901a\u8fc7\u4ece\u5927\u91cf\u7ef4\u5ea6\u4e2d\u79fb\u9664 RoPE\uff08\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff09\u5e76\u5c06\u5176\u8f6c\u6362\u4e3a NoPE\uff08\u65e0\u4f4d\u7f6e\u7f16\u7801\uff09\u6765\u89e3\u51b3 MLA \u548c RoPE \u51b2\u7a81\u7684\u95ee\u9898\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-46ac0ad2468e7b909e5152b28e6ee1fc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-46ac0ad2468e7b909e5152b28e6ee1fc.png\"><\/a><\/section>\n<section>\u4f5c\u8005\u4e3b\u8981\u5c1d\u8bd5\u4e86\u56db\u79cd\u79fb\u9664 RoPE \u7684\u7b56\u7565\uff1a1\uff09\u4fdd\u7559\u9ad8\u9891\u4f4d\u7f6e\u4fe1\u606f S_high\uff0c\u8be5\u65b9\u6cd5\u6700\u7b80\u5355\u76f4\u63a5\uff0c\u4fdd\u7559\u4e86\u5c40\u90e8\u8bed\u4e49\u7279\u5f81\u76f8\u5173\u7684\u9ad8\u9891\u7279\u5f81 [2]\uff1b2\uff09\u4fdd\u7559\u4f4e\u9891\u4f4d\u7f6e\u4fe1\u606f S_low\uff0c\u4e0e\u4fdd\u7559\u9ad8\u9891\u4f4d\u7f6e\u4fe1\u606f\u7684\u7b56\u7565\u5f62\u6210\u5bf9\u6bd4\uff0c\u68c0\u9a8c\u4f4e\u9891\u6210\u5206\u5728\u8bed\u4e49\u7406\u89e3\u4efb\u52a1\u4e2d\u7684\u6f5c\u5728\u4f5c\u7528\uff1b3\uff09\u5747\u5300\u91c7\u6837\u7b56\u7565 S_uniform\uff0c\u7b49\u95f4\u9694\u5747\u5300\u91c7\u6837\u9891\u7387\u4fdd\u7559\u4f4d\u7f6e\u9891\u7387\uff1b4\uff09\u4f7f\u7528\u67e5\u8be2\u3001\u952e\u5411\u91cf\u8303\u6570\u4e58\u79ef (2-norm) \u8fd1\u4f3c\u6ce8\u610f\u529b\u8d21\u732e\u503c [2] \u7684\u7b5b\u9009\u7b56\u7565 S_{2-norm}\uff0c\u9488\u5bf9\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\uff0c\u8ba1\u7b97\u6240\u6709\u9891\u7387\u7684\u5e73\u5747 2-norm \u5206\u6570\uff0c\u968f\u540e\u9009\u62e9\u5f97\u5206\u8f83\u9ad8\u7684\u9891\u7387\u4fdd\u7559\u4f4d\u7f6e\u4fe1\u606f\u3002\u8be5\u7b56\u7565\u80fd\u81ea\u9002\u5e94\u8bc6\u522b\u5bf9\u6a21\u578b\u6027\u80fd\u5173\u952e\u7684\u7279\u5f81\u9891\u7387\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-fd60b5a0800bb4efa99bd5e7e0d9d034.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-fd60b5a0800bb4efa99bd5e7e0d9d034.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2069ebe0acce667c903bbcd5b51d0fea.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2069ebe0acce667c903bbcd5b51d0fea.png\"><\/a><\/section>\n<section>Partial-RoPE \u7684\u6d88\u878d\u5b9e\u9a8c\u8868\u660e\uff1a1\uff09\u4fdd\u7559\u4f4e\u9891\u4f4d\u7f6e\u4fe1\u606f\u7684 S_low \u5bfc\u81f4\u4e86\u6700\u5927\u7684\u6027\u80fd\u635f\u5931\uff0c\u4fdd\u7559\u9ad8\u9891\u4f4d\u7f6e\u4fe1\u606f\u7684 S_high \u5bfc\u81f4\u7684\u6027\u80fd\u635f\u5931\u660e\u663e\u5c0f\u4e8e\u4fdd\u7559\u4f4e\u9891\uff0c\u8bf4\u660e\u4e86\u9ad8\u9891\u7ef4\u5ea6\u7684\u91cd\u8981\u6027\uff1b2\uff09S_uniform \u548c S_{2-norm}\u00a0\u5747\u5c55\u73b0\u51fa\u66f4\u4f18\u7684\u6027\u80fd\uff0c\u5206\u522b\u5728 135M \u6a21\u578b\u548c 1.7B \u6a21\u578b\u4e0a\u53d6\u5f97\u4e86\u6700\u5c11\u7684\u6027\u80fd\u635f\u5931\u3002\u6700\u7ec8\u4f5c\u8005\u9009\u62e9 S_{2-norm}\u00a0\u4f5c\u4e3a\u9ed8\u8ba4\u914d\u7f6e\uff0c\u662f\u56e0\u4e3a\u6ce8\u610f\u529b\u8d21\u732e\u5206\u6570\u8f83\u4f4e\u7684\u7ef4\u5ea6\u5728\u7ed3\u5408\u4f4e\u79e9\u8fd1\u4f3c\u65f6\u635f\u5931\u66f4\u5c11\u3002<\/section>\n<section><\/section>\n<section><strong>\u952e\u503c\u8054\u5408\u8868\u793a\u4f4e\u79e9\u8fd1\u4f3c<\/strong><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-3dab7e0c5cfc5768b1eb96c205990da7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-3dab7e0c5cfc5768b1eb96c205990da7.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9b3f3172054d6a56c8ac649d713249b0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9b3f3172054d6a56c8ac649d713249b0.png\"><\/a><\/section>\n<section>\u79fb\u9664\u4e86\u5927\u91cf\u7ef4\u5ea6\u7684 RoPE \u4e4b\u540e\uff0cMHA2MLA \u5c31\u53ef\u4ee5\u5bf9\u503c\u5411\u91cf\u548c PE \u65e0\u5173\u7684\u952e\u5411\u91cf\u8fdb\u884c\u4f4e\u79e9\u8fd1\u4f3c\uff0c\u4ece\u800c\u5927\u5e45\u51cf\u5c11\u7f13\u5b58\u7a7a\u95f4\u3002\u4e3a\u6700\u5927\u5316\u4fdd\u7559\u9884\u8bad\u7ec3\u77e5\u8bc6\uff0c\u672c\u6587\u63d0\u51fa\u4e24\u79cd\u57fa\u4e8e\u5947\u5f02\u503c\u5206\u89e3 (SVD) \u7684\u6295\u5f71\u77e9\u9635\u521d\u59cb\u5316\u7b56\u7565\uff1a1\uff09SVD_split\uff0c\u5206\u522b\u5bf9\u77e9\u9635\u8fdb\u884c\u4f4e\u79e9\u5206\u89e3\uff0c\u4fdd\u6301\u5404\u81ea\u7684\u8868\u5f81\u7279\u6027\uff1b2\uff09SVD_joint\uff0c\u8003\u8651\u952e\u503c\u77e9\u9635\u4e4b\u95f4\u7684\u5173\u8054\u6027\uff0c\u53c2\u6570\u77e9\u9635\u62fc\u63a5\u540e\u6574\u4f53\u8fdb\u884c\u4f4e\u79e9\u5206\u89e3\u3002<\/section>\n<section><\/section>\n<section>\u6d88\u878d\u5b9e\u9a8c\u8868\u660e\uff1a\u65e0\u8bba\u662f\u5728 GQA \u57fa\u5ea7\u8fd8\u662f MHA \u57fa\u5ea7\u4e0a\uff0cSVD_joint \u65b9\u6cd5\u59cb\u7ec8\u4f18\u4e8e SVD_split \u65b9\u6cd5\u3002<\/section>\n<section><\/section>\n<section><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4766031c1e3c961c54815ecf4c9c5239.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4766031c1e3c961c54815ecf4c9c5239.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-081db19aef1e59466ea33e8a8bdde68f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-081db19aef1e59466ea33e8a8bdde68f.png\"><\/a><\/section>\n<section>\u4f5c\u8005\u5728\u591a\u79cd\u89c4\u6a21\u7684\u8bed\u8a00\u6a21\u578b\uff08SmolLM-135M\/360M\/1B7 \u548c Llama2-7B\uff09\u4ee5\u53ca\u4e0d\u540c\u538b\u7f29\u6bd4\u4f8b\u7684\u914d\u7f6e\u4e0b\u8bc4\u4f30\u4e86\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u3002\u5b9e\u9a8c\u8868\u660e\uff1a1\uff09\u76f8\u540c\u5fae\u8c03\u8bbe\u7f6e\u4e0b\uff0c\u538b\u7f29\u6bd4\u4f8b\u8d8a\u9ad8\uff0c\u6027\u80fd\u635f\u5931\u8d8a\u5927\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u4e24\u4e2a GQA \u6a21\u578b\uff1b2\uff09\u76f8\u540c\u538b\u7f29\u6bd4\u4f8b\u4e0b\uff0c\u539f\u59cb\u6a21\u578b\u53c2\u6570\u8d8a\u591a\uff0c\u6027\u80fd\u635f\u5931\u8d8a\u5c0f\uff0c\u63ed\u793a\u4e86 MHA2MLA \u7684<strong>\u6f5c\u5728 scaling law<\/strong>\u30023\uff09MHA2MLA \u7684\u5fae\u8c03\u6570\u636e\u91cf\u4ec5\u9700\u9884\u8bad\u7ec3\u6570\u636e\u7684 0.3%~0.6%\uff0c\u907f\u514d\u4e86\u4ece\u5934\u9884\u8bad\u7ec3 MLA \u6a21\u578b\u7684\u9ad8\u6602\u6210\u672c\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d183cefee7b24ccda82359f147258f08.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d183cefee7b24ccda82359f147258f08.png\"><\/a><\/section>\n<section>\u4f5c\u8005\u5728 LongBench \u957f\u6587\u672c\u751f\u6210\u4efb\u52a1\u4e2d\u8bc4\u4f30\u4e86\u7ed3\u6784\u8fc1\u79fb\u540e\u7684 Llama2-7B \u6a21\u578b\uff0c\u5c06 KV \u7f13\u5b58\u91cf\u5316\u4f5c\u4e3a\u57fa\u51c6\u5bf9\u6bd4\u65b9\u6848\u3002\u5b9e\u9a8c\u8868\u660e\uff0cMHA2MLA \u80fd\u5728 d_{kv}=16 \u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u4e0e 2-bit \u91cf\u5316\u76f8\u540c\u7684\u538b\u7f29\u6bd4\u4f8b\uff0887.5%\uff09\uff0c\u540c\u65f6\u4ec5\u635f\u5931\u4e00\u534a\u7684\u6027\u80fd\uff08-3.0% vs. -6.2%\uff09\uff1b\u8fdb\u4e00\u6b65\u7ed3\u5408 4-bit \u91cf\u5316\u540e\uff0c\u4e0d\u4ec5\u538b\u7f29\u6bd4\u4f8b\u8d85\u8fc7 2-bit \u91cf\u5316\uff0c\u6027\u80fd\u635f\u5931\u4e5f\u90fd\u4f18\u4e8e\u6240\u6709 2-bit \u7684\u57fa\u7ebf\u65b9\u6cd5\uff0c\u4f8b\u5982 92.19% \u538b\u7f29\u6bd4\u4f8b\u4ec5\u6389 0.5%\uff0c96.87% \u538b\u7f29\u6bd4\u4f8b\u4ec5\u6389 3.2%\uff0c\u8bc1\u660e\u4e86 MHA2MLA \u80fd\u663e\u8457\u51cf\u5c11\u63a8\u7406\u65f6\u7684\u8bbf\u5b58\u74f6\u9888\u3002<\/section>\n<section><\/section>\n<section><strong>\u603b\u7ed3\u4e0e\u5c55\u671b<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u6587\u4e3b\u8981\u7814\u7a76\u5982\u4f55\u5c06\u57fa\u4e8e MHA \u7684\u9884\u8bad\u7ec3 LLMs\uff08\u6216\u5176\u53d8\u4f53\uff09\u9002\u914d\u4e3a KV \u7f13\u5b58\u9ad8\u6548\u7684 MLA \u67b6\u6784\uff0c\u4ee5\u663e\u8457\u964d\u4f4e\u63a8\u7406\u65f6\u7684\u8bbf\u5b58\u74f6\u9888\u3002\u901a\u8fc7\u7cbe\u5fc3\u7684\u67b6\u6784\u8bbe\u8ba1\uff0cMHA2MLA \u4ec5\u9700 0.3% \u81f3 0.6% \u9884\u8bad\u7ec3\u6570\u636e\u3002\u8be5\u6846\u67b6\u5c55\u73b0\u4e86\u4e0e\u73b0\u6709\u538b\u7f29\u6280\u672f\u7684\u5f3a\u517c\u5bb9\u6027\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u5e38\u8bc6\u63a8\u7406\u548c\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u80fd\u529b\uff0c\u4e3a\u90e8\u7f72\u8d44\u6e90\u9ad8\u6548\u7684 LLMs \u63d0\u4f9b\u4e86\u4e00\u6761\u5b9e\u7528\u8def\u5f84\u3002<\/section>\n<section><\/section>\n<section>\u4f5c\u8005\u63d0\u5230\u8be5\u7814\u7a76\u53d7\u9650\u4e8e\u786c\u4ef6\u6761\u4ef6\uff0c\u5f53\u524d\u5b9e\u9a8c\u672a\u80fd\u8986\u76d6 Llama3 \u7b49\u9700 128K \u957f\u4e0a\u4e0b\u6587\u5fae\u8c03\u7684\u6a21\u578b\uff0c\u4e5f\u672a\u7a81\u7834 7B \u53c2\u6570\u89c4\u6a21\u7684\u9a8c\u8bc1\u74f6\u9888\u3002\u6269\u5c55\u81f3\u66f4\u591a\u7684\u57fa\u5ea7\u5c06\u4f5c\u4e3a\u672a\u6765\u5de5\u4f5c\u4e4b\u4e00\u3002\u4f5c\u8005\u8fd8\u8ba1\u5212\u7ed3\u5408\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u7b56\u7565\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u67b6\u6784\u8fc1\u79fb\u8fc7\u7a0b\u4e2d\u7684\u53c2\u6570\u66f4\u65b0\u89c4\u6a21\u3002<\/section>\n<section><\/section>\n<section><em><sup>\u53c2\u8003\u6587\u732e\uff1a<\/sup><\/em><\/section>\n<section><sup><em>[1] DeepSeek-AI, Aixin Liu, Bei Feng et al.DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model. \u00a0ArXiv preprint.<\/em><\/sup><\/section>\n<section><em><sup>[2] Federico Barbero, Alex Vitvitskyi, Christos Perivolaropoulos, Razvan Pascanu, Petar Veli\u010dkovi\u0107. Round and Round We Go! What makes Rotary Positional Encodings useful? CoRR 2024<\/sup><\/em><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-06-16\" target=\"_blank\">DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:DeepSeek\u7684ML [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[73,68],"class_list":["post-39291","post","type-post","status-publish","format-standard","hentry","category-news","tag-73","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86 - \u4e00\u8d77AI\u6280\u672f<\/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:\/\/17aitech.com\/?p=39291\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=39291\",\"url\":\"https:\/\/17aitech.com\/?p=39291\",\"name\":\"DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=39291#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=39291#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png\",\"datePublished\":\"2025-03-08T00:02:22+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=39291#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=39291\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=39291#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=39291#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86 - \u4e00\u8d77AI\u6280\u672f","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:\/\/17aitech.com\/?p=39291","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39291","url":"https:\/\/17aitech.com\/?p=39291","name":"DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39291#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39291#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png","datePublished":"2025-03-08T00:02:22+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39291#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39291"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39291#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab74c42c4ddcc896cc3670282de96bba.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39291#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"DeepSeek\u7684MLA\uff0c\u4efb\u610f\u5927\u6a21\u578b\u90fd\u80fd\u8f7b\u677e\u8fc1\u79fb\u4e86"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39291","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=39291"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39291\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39291"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39291"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39291"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}