{"id":1647,"date":"2026-07-16T11:03:07","date_gmt":"2026-07-16T03:03:07","guid":{"rendered":"https:\/\/lingbo.online\/?p=1647"},"modified":"2026-07-16T11:04:11","modified_gmt":"2026-07-16T03:04:11","slug":"1647","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/1647\/","title":{"rendered":"FastMTP: Accelerating LLM Inference with Enhanced Multi-Token Prediction"},"content":{"rendered":"<h3>Meta Data<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4\uff1a2025-09-16<\/li>\n<li>\u4f5c\u8005\uff1aYuxuan Cai, Xiaozhuan Liang, Xinghua Wang, Jin Ma, Haijin Liang, Jinwen Luo, Xinyu Zuo, Lisheng Duan, Yuyang Yin, Xi Chen\uff08\u817e\u8baf\uff09<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2509.18362.pdf\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/pdf\/2509.18362.pdf<\/a><\/li>\n<li>arXiv \u9875\u9762\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2509.18362\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/abs\/2509.18362<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/Tencent-BAC\/FastMTP\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/Tencent-BAC\/FastMTP<\/a><\/li>\n<\/ul>\n<h3>FastMTP\uff1a\u901a\u8fc7\u589e\u5f3a\u591a\u8bcd\u5143\u9884\u6d4b\u52a0\u901f\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406<\/h3>\n<h4>\u6458\u8981<\/h4>\n<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Models, LLMs\uff09\u53d8\u5f97\u65e5\u76ca\u5f3a\u5927\uff0c\u81ea\u56de\u5f52\u751f\u6210\u7684\u987a\u5e8f\u7279\u6027\u9020\u6210\u4e86\u9650\u5236\u5b9e\u9645\u90e8\u7f72\u7684\u6839\u672c\u6027\u541e\u5410\u74f6\u9888\u3002\u5c3d\u7ba1\u591a\u8bcd\u5143\u9884\u6d4b\uff08Multi-Token Prediction, MTP\uff09\u5df2\u5728\u6a21\u578b\u8bad\u7ec3\u6548\u7387\u4e0e\u6027\u80fd\u65b9\u9762\u5c55\u73b0\u51fa\u663e\u8457\u6536\u76ca\uff0c\u4f46\u5176\u5728\u63a8\u7406\u52a0\u901f\u65b9\u9762\u7684\u56fa\u6709\u6f5c\u529b\u4ecd\u5927\u591a\u672a\u88ab\u5145\u5206\u63a2\u7d22\u3002\u672c\u6587\u63d0\u51fa FastMTP\uff0c\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u65b9\u6cd5\uff1a\u901a\u8fc7\u5c06 MTP \u8bad\u7ec3\u4e0e\u5176\u63a8\u7406\u6a21\u5f0f\u5bf9\u9f50\uff0c\u63d0\u5347\u591a\u6b65\u8349\u7a3f\u8d28\u91cf\uff0c\u4ece\u800c\u663e\u8457\u589e\u5f3a\u63a8\u6d4b\u89e3\u7801\u6027\u80fd\u3002\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u81ea\u84b8\u998f\u6570\u636e\u4e0a\u5bf9\u5355\u4e00\u3001\u4f4d\u7f6e\u5171\u4eab\u6743\u91cd\u7684 MTP \u5934\u8fdb\u884c\u5fae\u8c03\uff0c\u4f7f\u5176\u80fd\u591f\u6355\u83b7\u8fde\u7eed\u672a\u6765\u8bcd\u5143\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u5e76\u5728\u591a\u4e2a\u9012\u5f52\u8349\u7a3f\u6b65\u4e0a\u4fdd\u6301\u9ad8\u63a5\u53d7\u7387\u3002\u901a\u8fc7\u5c06\u8bed\u8a00\u611f\u77e5\u52a8\u6001\u8bcd\u8868\u538b\u7f29\u96c6\u6210\u5230 MTP \u5934\u4e2d\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u964d\u4f4e\u4e86\u8349\u7a3f\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u5f00\u9500\u3002\u5728\u4e03\u4e2a\u591a\u6837\u5316\u57fa\u51c6\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cFastMTP \u76f8\u5bf9\u6807\u51c6\u4e0b\u4e00\u8bcd\u5143\u9884\u6d4b\u5e73\u5747\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">2.03\\times<\/span> \u52a0\u901f\uff0c\u4e14\u8f93\u51fa\u8d28\u91cf\u65e0\u635f\uff0c\u76f8\u5bf9 vanilla MTP \u63d0\u5347 82%\u3002FastMTP \u4ec5\u9700\u8f7b\u91cf\u8bad\u7ec3\uff0c\u5e76\u53ef\u65e0\u7f1d\u96c6\u6210\u5230\u73b0\u6709\u63a8\u7406\u6846\u67b6\u4e2d\uff0c\u4e3a\u52a0\u901f LLM \u63a8\u7406\u63d0\u4f9b\u4e86\u4e00\u79cd\u5b9e\u7528\u4e14\u53ef\u5feb\u901f\u90e8\u7f72\u7684\u65b9\u6848\u3002<\/p>\n<h4>\u5f15\u8a00<\/h4>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5df2\u5728\u81ea\u4e3b\u667a\u80fd\u4f53\u3001\u4ee3\u7801\u751f\u6210\u4e0e\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7b49\u591a\u79cd\u5e94\u7528\u4e2d\u5c55\u73b0\u51fa\u5353\u8d8a\u80fd\u529b\u3002\u7136\u800c\uff0c\u5176\u5b9e\u7528\u90e8\u7f72\u9762\u4e34\u4e00\u4e2a\u6839\u672c\u6027\u6548\u7387\u74f6\u9888\uff1a\u8bcd\u5143\u751f\u6210\u7684\u81ea\u56de\u5f52\u672c\u8d28\u3002\u5f53\u524d LLM \u6309\u987a\u5e8f\u751f\u6210\u6587\u672c\uff0c\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u4ec5\u4ea7\u751f\u4e00\u4e2a\u8bcd\u5143\uff0c\u8fd9\u610f\u5473\u7740\u6574\u4f53\u751f\u6210\u65f6\u95f4\u968f\u5e8f\u5217\u957f\u5ea6\u7ebf\u6027\u589e\u957f\u3002\u8fd9\u5bf9\u9700\u8981\u5927\u91cf\u751f\u6210\u7684\u573a\u666f\u5c24\u4e3a\u68d8\u624b\uff0c\u4f8b\u5982\u6700\u5148\u8fdb\u7684\u5927\u578b\u63a8\u7406\u6a21\u578b\uff1a\u5b83\u4eec\u901a\u8fc7\u5728\u7ed9\u51fa\u6700\u7ec8\u7b54\u6848\u524d\u751f\u6210\u6269\u5c55\u7684\u3001\u7c7b\u4f3c\u4eba\u7c7b\u7684\u601d\u7ef4\u94fe\uff08Chain-of-Thoughts, CoTs\uff09\uff0c\u5728\u89e3\u51b3\u590d\u6742\u4e0e\u903b\u8f91\u5bc6\u96c6\u4efb\u52a1\u4e0a\u53d6\u5f97\u7a81\u7834\u6027\u8fdb\u5c55\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u6a21\u578b\u5c55\u73b0\u51fa\u5f3a\u5927\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5373\u4f7f\u5bf9\u7b80\u5355\u6837\u672c\u4e5f\u5e38\u5e38\u4ea7\u51fa\u8fc7\u957f\u7684\u63a8\u7406\u94fe\uff0c\u4e0d\u53ef\u907f\u514d\u5730\u5f15\u5165\u5927\u91cf\u8ba1\u7b97\u5f00\u9500\u3002\u56e0\u6b64\uff0c\u8feb\u5207\u9700\u8981\u6709\u6548\u7684\u52a0\u901f\u624b\u6bb5\u3002<\/p>\n<p>\u8fd1\u671f\u7814\u7a76\u63a2\u7d22\u4e86\u591a\u79cd\u52a0\u901f LLM \u63a8\u7406\u7684\u7b56\u7565\uff0c\u5305\u62ec\u9ad8\u6548\u6ce8\u610f\u529b\u673a\u5236\u4e0e\u6a21\u578b\u538b\u7f29\u4ee5\u964d\u4f4e\u8ba1\u7b97\u5f00\u9500\u3002\u53e6\u4e00\u4e9b\u5de5\u4f5c\u5219\u5173\u6ce8\u51cf\u5c11 CoT \u5197\u4f59\uff0c\u5305\u62ec\u5e26\u957f\u5ea6\u60e9\u7f5a\u7684\u5f3a\u5316\u5b66\u4e60\uff0c\u4ee5\u53ca\u5bf9\u53ef\u53d8\u957f\u5ea6 CoT \u6570\u636e\u7684\u76d1\u7763\u5fae\u8c03\u3002\u5728\u8fd9\u4e9b\u65b9\u6cd5\u4e2d\uff0c\u63a8\u6d4b\u89e3\u7801\u5df2\u6210\u4e3a\u4e00\u79cd\u6709\u524d\u666f\u7684\u6280\u672f\uff1a\u5b83\u5728\u4e0d\u635f\u5bb3\u8f93\u51fa\u4fdd\u771f\u5ea6\u7684\u524d\u63d0\u4e0b\u63d0\u5347\u89e3\u7801\u6548\u7387\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u4f7f\u7528\u4e00\u4e2a\u66f4\u5c0f\u7684\u6a21\u578b\uff08\u79f0\u4e3a\u8349\u7a3f\u6a21\u578b\uff09\u9884\u6d4b\u82e5\u5e72\u540e\u7eed\u8bcd\u5143\uff0c\u518d\u7531\u76ee\u6807 LLM \u5e76\u884c\u9a8c\u8bc1\uff0c\u4ece\u800c\u5b9e\u73b0\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u7684\u591a\u8bcd\u5143\u751f\u6210\u3002<\/p>\n<p>\u591a\u8bcd\u5143\u9884\u6d4b\uff08MTP\uff09\u6a21\u5757\u53ca\u5176\u8f85\u52a9\u8bad\u7ec3\u6700\u521d\u65e8\u5728\u6539\u8fdb\u8bad\u7ec3\uff0c\u5374\u4e5f\u4e3a\u901a\u8fc7\u63a8\u6d4b\u89e3\u7801\u5b9e\u73b0\u63a8\u7406\u52a0\u901f\u63d0\u4f9b\u4e86\u81ea\u7136\u673a\u4f1a\u3002MTP \u901a\u8fc7\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u540c\u65f6\u9884\u6d4b\u591a\u4e2a\u672a\u6765\u8bcd\u5143\uff0c\u6269\u5c55\u4e86\u4f20\u7edf\u7684\u4e0b\u4e00\u8bcd\u5143\u9884\u6d4b\u8303\u5f0f\u3002\u8be5\u65b9\u6cd5\u9f13\u52b1\u6a21\u578b\u63d0\u524d\u89c4\u5212\uff0c\u5e76\u5229\u7528\u66f4\u4e30\u5bcc\u7684\u76d1\u7763\u4fe1\u53f7\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0cDeepSeek-V3 \u4ee5\u7ea7\u8054 MTP \u6a21\u5757\u7684\u5e8f\u5217\u5316\u5b9e\u73b0\u6539\u8fdb\u4e86 MTP \u67b6\u6784\uff0c\u4fdd\u7559\u5b8c\u6574\u56e0\u679c\u94fe\u4ee5\u7ef4\u6301\u81ea\u56de\u5f52\u6027\u8d28\u3002<\/p>\n<p>\u5c3d\u7ba1 MTP \u5728\u8bad\u7ec3\u6539\u8fdb\u65b9\u9762\u7684\u91c7\u7528\u65e5\u76ca\u589e\u591a\uff0c\u5176\u5728\u63a8\u7406\u52a0\u901f\u65b9\u9762\u7684\u6f5c\u529b\u4ecd\u5927\u591a\u672a\u88ab\u5229\u7528\u3002\u5f53\u524d\u5b9e\u73b0\u8981\u4e48\u5728\u63a8\u7406\u65f6\u5b8c\u5168\u4e22\u5f03 MTP \u6a21\u5757\uff0c\u9000\u56de\u6807\u51c6\u4e0b\u4e00\u8bcd\u5143\u9884\u6d4b\uff1b\u8981\u4e48\u4ec5\u4fdd\u7559\u7b2c\u4e00\u4e2a MTP \u6a21\u5757\u7528\u4e8e\u591a\u8bcd\u5143\u9884\u6d4b\u3002\u63a8\u7406\u4e2d\u7684\u8fd9\u79cd\u672a\u5145\u5206\u5229\u7528\u53ef\u80fd\u6e90\u4e8e\u73b0\u6709 MTP \u5b9e\u73b0\u4e2d\u7684\u4e24\u4e2a\u6839\u672c\u6311\u6218\u3002\u7b2c\u4e00\uff0c\u5e8f\u5217\u5316 MTP \u67b6\u6784\u9700\u8981\u901a\u8fc7\u591a\u4e2a MTP \u6a21\u5757\u8fdb\u884c\u7ea7\u8054\u524d\u5411\u4f20\u64ad\uff0c\u6bcf\u4e2a\u6a21\u5757\u7ef4\u62a4\u72ec\u7acb\u6743\u91cd\u4e0e\u952e\u503c\u7f13\u5b58\uff0c\u5e26\u6765\u663e\u8457\u5185\u5b58\u5f00\u9500\u3002\u6b64\u5916\uff0c\u5c3d\u7ba1\u8bb8\u591a\u63a8\u7406\u6846\u67b6\u73b0\u5df2\u652f\u6301\u4f7f\u7528\u5355\u4e00\u8349\u7a3f\u6a21\u578b\u7684\u63a8\u6d4b\u89e3\u7801\uff0cMTP \u7684\u8bbe\u8ba1\u5374\u9700\u8981\u52a0\u8f7d\u5e76\u7f16\u6392\u591a\u4e2a\u8349\u7a3f\u6a21\u578b\u2014\u2014\u6bcf\u4e2a\u9884\u6d4b\u6b65\u4e00\u4e2a MTP \u6a21\u5757\u2014\u2014\u4ece\u800c\u8981\u6c42\u590d\u6742\u8c03\u5ea6\uff0c\u4e25\u91cd\u635f\u5bb3\u8ba1\u7b97\u6548\u7387\u3002\u8fd9\u6216\u8bb8\u53ef\u4ee5\u89e3\u91ca\u4e3a\u4f55\u4f7f\u7528\u591a\u5c42 MTP \u8bad\u7ec3\u7684\u6a21\u578b\u5728\u5f00\u6e90\u65f6\u5f80\u5f80\u53ea\u63d0\u4f9b\u5355\u4e00\u6a21\u5757\u7528\u4e8e\u63a8\u7406\u3002\u7b2c\u4e8c\uff0c\u8bd5\u56fe\u901a\u8fc7\u9012\u5f52\u590d\u7528\u5355\u4e00 MTP \u6a21\u5757\u6765\u89c4\u907f\u8fd9\u4e00\u590d\u6742\u6027\u65f6\uff0c\u5728\u7b2c\u4e00\u4e2a\u989d\u5916\u8bcd\u5143\u4e4b\u540e\u7684\u63a5\u53d7\u7387\u8f83\u5dee\uff0c\u56e0\u4e3a\u8be5\u6a21\u5757\u5e76\u672a\u88ab\u663e\u5f0f\u8bad\u7ec3\u7528\u4e8e\u9012\u5f52\u591a\u6b65\u9884\u6d4b\u6a21\u5f0f\uff0c\u56e0\u800c\u5bf9\u6269\u5c55\u8349\u7a3f\u751f\u6210\u6548\u679c\u4e0d\u4f73\u3002<\/p>\n<p>\u672c\u6587\u63d0\u51fa FastMTP\uff0c\u4e00\u79cd\u589e\u5f3a\u7684\u591a\u8bcd\u5143\u9884\u6d4b\u6846\u67b6\uff0c\u4f7f MTP \u6a21\u5757\u5728\u63a8\u7406\u65f6\u66f4\u6709\u6548\u3001\u66f4\u9ad8\u6548\u3001\u66f4\u6613\u90e8\u7f72\u3002\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u6240\u6709\u9884\u6d4b\u6b65\u4e0a\u5bf9\u5355\u4e00\u5171\u4eab\u6743\u91cd\u7684 MTP \u5934\u8fdb\u884c\u5fae\u8c03\uff0c\u4f7f\u5176\u5728\u4fdd\u6301\u56e0\u679c\u6027\u7684\u540c\u65f6\u6267\u884c\u591a\u8bcd\u5143\u751f\u6210\u3002\u8fd9\u4f7f\u6a21\u578b\u80fd\u591f\u6355\u83b7\u8fde\u7eed\u672a\u6765\u8bcd\u5143\u4e4b\u95f4\u7684\u4f9d\u8d56\uff0c\u4ece\u800c\u5728\u521d\u59cb\u8349\u7a3f\u4f4d\u7f6e\u4e4b\u540e\u83b7\u5f97\u66f4\u9ad8\u63a5\u53d7\u7387\uff0c\u5e76\u786e\u4fdd\u4e0e EAGLE \u98ce\u683c\u63a8\u6d4b\u89e3\u7801\u517c\u5bb9\uff0c\u4ece\u800c\u53ef\u65e0\u7f1d\u96c6\u6210\u5230 SGLang \u7b49\u73b0\u6709\u63a8\u7406\u6846\u67b6\u3002\u53d7 FR-Spec \u542f\u53d1\uff0c\u6211\u4eec\u96c6\u6210\u8bed\u8a00\u611f\u77e5\u52a8\u6001\u8bcd\u8868\u538b\u7f29\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u8349\u7a3f\u751f\u6210\u7684\u8ba1\u7b97\u6210\u672c\uff0c\u540c\u65f6\u5bf9\u591a\u6837\u5316\u4efb\u52a1\u4e0e\u8bed\u8a00\u4e0a\u7684\u63a5\u53d7\u7387\u5f71\u54cd\u53ef\u5ffd\u7565\u3002<\/p>\n<p>\u6211\u4eec\u7684\u4e3b\u8981\u8d21\u732e\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\n<p>\u6211\u4eec\u901a\u8fc7\u5728\u81ea\u84b8\u998f\u6570\u636e\u4e0a\u5fae\u8c03\uff0c\u4f7f\u5355\u4e00 MTP \u5934\u80fd\u591f\u8fdb\u884c\u6709\u6548\u7684\u9012\u5f52\u591a\u6b65\u8349\u7a3f\u751f\u6210\uff0c\u76f8\u5bf9 vanilla MTP \u590d\u7528\u7b56\u7565\uff0c\u663e\u8457\u5c06\u5e73\u5747\u63a5\u53d7\u7387\u4ece\u7b2c\u4e00\u8349\u7a3f\u8bcd\u5143\u7684 70% \u63d0\u5347\u5230 81%\uff0c\u7b2c\u4e8c\u8349\u7a3f\u8bcd\u5143\u4ece 11% \u63d0\u5347\u5230 56%\uff0c\u7b2c\u4e09\u8bcd\u5143\u4ece 2% \u63d0\u5347\u5230 36%\uff0c\u4ece\u800c\u91ca\u653e\u5176\u7528\u4e8e\u63a8\u7406\u52a0\u901f\u7684\u5168\u90e8\u6f5c\u529b\u3002<\/p>\n<\/li>\n<li>\n<p>\u6211\u4eec\u4e3a MTP \u5934\u91c7\u7528\u8bed\u8a00\u611f\u77e5\u52a8\u6001\u8bcd\u8868\u538b\u7f29\uff0c\u6839\u636e\u8f93\u5165\u4e0a\u4e0b\u6587\u964d\u4f4e\u8349\u7a3f\u9636\u6bb5\u8ba1\u7b97\u5f00\u9500\u3002\u5728\u8349\u7a3f\u751f\u6210\u4e09\u4e2a\u989d\u5916\u8bcd\u5143\u65f6\uff0c\u5b83\u8fdb\u4e00\u6b65\u5c06\u5e73\u5747\u8f93\u51fa\u541e\u5410\u63d0\u5347\u7ea6 16%\u3002<\/p>\n<\/li>\n<li>\n<p>\u6211\u4eec\u5728 7 \u4e2a\u57fa\u51c6\u4e0a\u8fdb\u884c\u4e86\u5927\u91cf\u5b9e\u9a8c\uff0c\u5728 7B \u6a21\u578b\u4e0a\u5c55\u793a\u5e73\u5747 <span class=\"katex-eq\" data-katex-display=\"false\">2.03\\times<\/span> \u52a0\u901f\u4e14\u751f\u6210\u8d28\u91cf\u65e0\u635f\uff0c\u663e\u8457\u4f18\u4e8e vanilla MTP \u590d\u7528\u7b56\u7565\uff08<span class=\"katex-eq\" data-katex-display=\"false\">1.21\\times<\/span>\uff09\u3002<\/p>\n<\/li>\n<\/ul>\n<h4>\u65b9\u6cd5<\/h4>\n<p>\u672c\u8282\u8be6\u7ec6\u63cf\u8ff0 FastMTP \u7684\u5b9e\u73b0\uff0c\u5176\u4e13\u95e8\u589e\u5f3a MTP\uff0c\u4ee5\u4fbf\u5728 LLM \u63a8\u7406\u4e2d\u8fdb\u884c\u66f4\u6709\u6548\u3001\u66f4\u9ad8\u6548\u7684\u63a8\u6d4b\u89e3\u7801\u3002\u56fe 1 \u5c55\u793a\u4e86\u5176\u8bad\u7ec3\u4e0e\u63a8\u7406\u6d41\u6c34\u7ebf\u3002<\/p>\n<p><center>\u56fe 1\uff1aFastMTP \u8bad\u7ec3\u4e0e\u63a8\u7406\u7b56\u7565\u793a\u610f\u56fe\u3002\u8bad\u7ec3\u9636\u6bb5\uff08\u5de6\uff09\uff1a\u7070\u8272\u5757\u8868\u793a\u51bb\u7ed3\u7684\u4e3b\u6a21\u578b\u6a21\u5757\uff0c\u6a59\u8272\u5757\u8868\u793a\u6743\u91cd\u5171\u4eab\u7684\u53ef\u8bad\u7ec3 MTP \u5934\uff0c\u84dd\u8272\u5757\u5c55\u793a\u5e26\u987a\u5e8f\u4f4d\u7f6e\u504f\u79fb\u7684\u8f93\u5165\u4e0e\u6807\u7b7e\u5e8f\u5217\u3002\u63a8\u7406\u9636\u6bb5\uff08\u53f3\uff09\uff1a\u4e3b\u6a21\u578b\u9884\u6d4b\u4e0b\u4e00\u8bcd\u5143\uff08\u7eff\u8272\uff09\uff0c\u518d\u9001\u5165 MTP \u5934\u9012\u5f52\u751f\u6210\u8349\u7a3f\u8bcd\u5143\uff08\u7c89\u8272\uff09\u5e76\u5e76\u884c\u9a8c\u8bc1\u3002\u7d2b\u8272\u5757\u8868\u793a\u9891\u7387\u6392\u5e8f\uff08FR\uff09\u8bed\u8a00\u611f\u77e5\u52a8\u6001\u8bcd\u8868\u538b\u7f29\uff0c\u7528\u4e8e\u52a0\u901f\u8349\u7a3f\u751f\u6210\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260716102113001.png\" alt=\"\u56fe 1\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<h5>\u5171\u4eab MTP \u5934\u7684\u8bad\u7ec3<\/h5>\n<h6>\u67b6\u6784\u8bbe\u8ba1<\/h6>\n<p>FastMTP \u91c7\u7528\u4e0e DeepSeek-V3 \u76f8\u540c\u7684 MTP \u67b6\u6784\u3002\u5173\u952e\u533a\u522b\u5728\u4e8e\u6211\u4eec\u5728\u6240\u6709\u9884\u6d4b\u6b65\u4f7f\u7528\u5355\u4e00\u3001\u6743\u91cd\u5171\u4eab\u7684 MTP \u5934\uff0c\u800c\u975e\u5bf9\u6bcf\u4e2a\u9884\u6d4b\u6df1\u5ea6\u4f7f\u7528\u591a\u4e2a\u72ec\u7acb\u6a21\u5757\u7684\u4f20\u7edf\u505a\u6cd5\u3002\u8fd9\u79cd\u5171\u4eab\u6743\u91cd\u8bbe\u8ba1\u4e0d\u4ec5\u964d\u4f4e\u5185\u5b58\u5360\u7528\uff0c\u66f4\u91cd\u8981\u7684\u662f\u8feb\u4f7f\u6a21\u578b\u6355\u83b7\u8fde\u7eed\u672a\u6765\u8bcd\u5143\u4e4b\u95f4\u7684\u4f9d\u8d56\uff0c\u4ee5\u5b9e\u73b0\u56e0\u679c\u591a\u8bcd\u5143\u9884\u6d4b\u3002<\/p>\n<h6>\u8bad\u7ec3\u673a\u5236<\/h6>\n<p>\u8bbe <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{F}<\/span> \u8868\u793a\u4e3b\u6a21\u578b\u7684 transformer \u5c42\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u8868\u793a MTP \u5934\u3002\u8003\u8651\u5904\u7406\u5e8f\u5217\u4e2d\u4f4d\u7f6e <span class=\"katex-eq\" data-katex-display=\"false\">i<\/span> \u7684\u8bcd\u5143\u3002\u8f93\u5165\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">t_{1:i}<\/span> \u9996\u5148\u901a\u8fc7\u5d4c\u5165\u5c42\u4e0e transformer \u5c42 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{F}<\/span> \u4ea7\u751f\u9690\u85cf\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">h_{1:i}<\/span>\u3002\u4e3a\u9884\u6d4b <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u4e2a\u989d\u5916\u672a\u6765\u8bcd\u5143\uff0cMTP \u5934 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u9012\u5f52\u8fd0\u884c\uff1a<\/p>\n<ul>\n<li>\n<p>\u5728\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">k=1<\/span>\uff1a <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u53d6\u81ea <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{F}<\/span> \u7684\u9690\u85cf\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">h_{i}<\/span>\uff0c\u4ee5\u53ca\u504f\u79fb\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">t_{i+1}<\/span>\uff08\u504f\u79fb 1 \u4e2a\u4f4d\u7f6e\uff09\u7684\u5d4c\u5165\uff0c\u4ee5\u9884\u6d4b\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{t}^{1}_{i+2}<\/span>\u3002<\/p>\n<\/li>\n<li>\n<p>\u5728\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">k\\gt 1<\/span>\uff1a <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u5904\u7406\u5176\u5728\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">k-1<\/span> \u4ea7\u751f\u7684\u81ea\u8eab\u8f93\u51fa\u9690\u85cf\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">h^{k-1}_{i+k-1}<\/span>\uff0c\u5e76\u7ed3\u5408\u504f\u79fb\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">t_{i+k}<\/span>\uff08\u504f\u79fb <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u4e2a\u4f4d\u7f6e\uff09\u7684\u5d4c\u5165\uff0c\u4ee5\u9884\u6d4b\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{t}^{k}_{i+k+1}<\/span>\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u671f\u95f4\uff0c\u8be5\u8fc7\u7a0b\u5e94\u7528\u4e8e\u5e8f\u5217\u4e2d\u6240\u6709\u6709\u6548\u4f4d\u7f6e\u3002\u5bf9\u957f\u5ea6\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">T<\/span> \u7684\u8bad\u7ec3\u5e8f\u5217\uff0c\u4f4d\u7f6e <span class=\"katex-eq\" data-katex-display=\"false\">i<\/span> \u7684\u8303\u56f4\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">1<\/span> \u5230 <span class=\"katex-eq\" data-katex-display=\"false\">T-K<\/span>\uff0c\u4ee5\u786e\u4fdd\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u4e2a\u9884\u6d4b\u6b65\u504f\u79fb\u540e\u6240\u6709\u7d22\u5f15\u4ecd\u6709\u6548\u3002<\/p>\n<h6>\u8bad\u7ec3\u76ee\u6807<\/h6>\n<p>\u6211\u4eec\u4f7f\u7528\u5e26\u6307\u6570\u8870\u51cf\u7684\u52a0\u6743\u4ea4\u53c9\u71b5\u635f\u5931\u4f18\u5316 MTP \u5934\uff0c\u4ee5\u5904\u7406\u66f4\u8fdc\u8bcd\u5143\u9884\u6d4b\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\mathcal{L}_{mtp}=\\sum_{k=1}^{K}\\alpha_{k}\\cdot\\mathcal{L}_{mtp}^{k}=\\sum_{k=1}^{K}\\alpha_{k}\\cdot\\operatorname{CE}\\,(\\hat{t}^{k}_{1+k\\,:\\,T+1},\\,t_{1+k\\,:\\,T+1})\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}_{mtp}^{k}<\/span> \u8868\u793a\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u4e2a\u9884\u6d4b\u6b65\u7684\u635f\u5931\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\operatorname{CE}(\\cdot,\\cdot)<\/span> \u8868\u793a\u9884\u6d4b\u4e0e\u771f\u503c\u8bcd\u5143\u4e4b\u95f4\u7684\u4ea4\u53c9\u71b5\u635f\u5931\u3002\u4f4d\u7f6e\u76f8\u5173\u6743\u91cd <span class=\"katex-eq\" data-katex-display=\"false\">\\alpha_{k}<\/span> \u9075\u5faa\u6307\u6570\u8870\u51cf\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\alpha_{k}=\\frac{\\beta^{k-1}}{\\sum_{j=1}^{K}\\beta^{j-1}}\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\beta<\/span> \u662f\u8870\u51cf\u56e0\u5b50\u3002\u8be5\u52a0\u6743\u7b56\u7565\u8003\u8651\u4e86\u9884\u6d4b\u4e0d\u786e\u5b9a\u6027\u968f\u6bcf\u4e00\u6b65\u989d\u5916\u9884\u6d4b\u800c\u7d2f\u79ef\u2014\u2014\u66f4\u8fdc\u8bcd\u5143\u4f9d\u8d56\u66f4\u591a\u4e2d\u95f4\u51b3\u7b56\uff0c\u56e0\u6b64\u9010\u6b65\u66f4\u96be\u9884\u6d4b\u3002\u6307\u6570\u8870\u51cf\u786e\u4fdd\u6a21\u578b\u4f18\u5148\u8fd1\u7a0b\u9884\u6d4b\uff0c\u540c\u65f6\u4ecd\u53d1\u5c55\u5bf9\u82e5\u5e72\u8fde\u7eed\u672a\u6765\u8bcd\u5143\u7684\u751f\u6210\u80fd\u529b\u3002<\/p>\n<p>\u4ec5\u5fae\u8c03 MTP \u5934 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span>\uff0c\u800c\u6240\u6709\u4e3b\u6a21\u578b\u7ec4\u4ef6\u4fdd\u6301\u51bb\u7ed3\uff0c\u66f4\u65b0\u53c2\u6570\u5c11\u4e8e 3%\uff0c\u65e2\u4fdd\u8bc1\u8ba1\u7b97\u6548\u7387\uff0c\u4e5f\u4fdd\u7559\u57fa\u5ea7\u6a21\u578b\u80fd\u529b\u3002<\/p>\n<h5>MTP \u8349\u7a3f\u4e2d\u7684\u8bed\u8a00\u611f\u77e5\u8bcd\u8868\u7a7a\u95f4\u538b\u7f29<\/h5>\n<p>\u4e3a\u52a0\u901f\u8349\u7a3f\u9636\u6bb5\uff0c\u6211\u4eec\u5c06\u57fa\u4e8e\u9891\u7387\u7684\u8bcd\u8868\u538b\u7f29\u4e0e MTP \u7ed3\u5408\uff0c\u964d\u4f4e MTP \u8f93\u51fa\u5934\u7684\u8ba1\u7b97\u5f00\u9500\u3002\u9075\u5faa FR-Spec \u4e2d\u7684\u5206\u6790\uff0c\u4e00\u5c0f\u90e8\u5206\u8bcd\u5143\u5360\u51fa\u73b0\u6b21\u6570\u7684\u7edd\u5927\u591a\u6570\uff0c\u800c\u5176\u4f59\u8bcd\u5143\u9891\u7387\u6781\u7a00\u758f\u2014\u2014\u5448\u73b0\u4e00\u81f4\u7684\u957f\u5c3e\u6a21\u5f0f\u3002\u8fd9\u4e00\u89c2\u5bdf\u4fc3\u4f7f\u6211\u4eec\u5728\u8349\u7a3f\u751f\u6210\u65f6\u5c06 MTP \u5934\u7684\u8f93\u51fa\u7a7a\u95f4\u9650\u5236\u4e3a\u9ad8\u9891\u8bcd\u5143\u3002<\/p>\n<p>\u7531\u4e8e\u539f\u59cb\u8bcd\u8868\u538b\u7f29\u65b9\u6cd5\u4e3b\u8981\u5728\u82f1\u8bed\u8bed\u6599\u4e0a\u5206\u6790\u8bcd\u5143\u5206\u5e03\uff0c\u5176\u538b\u7f29\u8bcd\u8868\u5bf9\u4e2d\u6587\u8bcd\u5143\u8868\u793a\u4e0d\u4f73\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u4f1a\u4e25\u91cd\u9650\u5236\u4e2d\u6587\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u3002\u4e3a\u514b\u670d\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u8ba1\u7b97\u8bed\u8a00\u7279\u5b9a\u9891\u7387\u7edf\u8ba1\uff0c\u5e76\u6839\u636e\u751f\u6210\u4e0a\u4e0b\u6587\u52a8\u6001\u8c03\u6574\u9ad8\u9891\u8bcd\u8868\uff0c\u786e\u4fdd\u4e0d\u540c\u8bed\u8a00\u7684\u9ad8\u9891\u8bcd\u5143\u5728\u538b\u7f29\u8bcd\u8868\u7a7a\u95f4\u4e2d\u5f97\u5230\u5145\u5206\u8868\u793a\u3002\u8fd9\u79cd\u8bed\u8a00\u611f\u77e5\u538b\u7f29\u5728\u4fdd\u6301\u8ba1\u7b97\u6548\u7387\u7684\u540c\u65f6\uff0c\u5bf9\u82f1\u6587\u4e0e\u4e2d\u6587\u751f\u6210\u90fd\u7ef4\u6301\u9ad8\u8349\u7a3f\u63a5\u53d7\u7387\u3002<\/p>\n<p>\u8bbe <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{V}<\/span> \u8868\u793a\u8bed\u8a00\u6a21\u578b\u7684\u5b8c\u6574\u8bcd\u8868\u3002\u5bf9\u6bcf\u79cd\u8bed\u8a00 <span class=\"katex-eq\" data-katex-display=\"false\">l<\/span>\uff0c\u6211\u4eec\u5b9a\u4e49 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{V}_{high}^{(l)}\\subset\\mathcal{V}<\/span> \u4e3a\u8be5\u8bed\u8a00\u7279\u5b9a\u7684\u9ad8\u9891\u8bcd\u5143\u5b50\u96c6\uff0c\u901a\u8fc7\u8bed\u6599\u7ea7\u7edf\u8ba1\u8bc6\u522b\u3002\u8349\u7a3f\u751f\u6210\u671f\u95f4\uff0cMTP \u5934\u6839\u636e\u5f53\u524d\u4e0a\u4e0b\u6587\u52a8\u6001\u9009\u62e9\u5408\u9002\u8bcd\u8868\u3002\u6362\u8a00\u4e4b\uff0c\u82e5\u4e3b\u6a21\u578b\u8f93\u51fa\u5934\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}\\in\\mathbb{R}^{|\\mathcal{V}|\\times d}<\/span>\uff0c\u6211\u4eec\u63d0\u53d6\u8bed\u8a00\u7279\u5b9a\u5b50\u77e9\u9635\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\tilde{\\mathbf{W}}^{(l)}[i,:]=\\mathbf{W}[\\mathcal{V}_{high}^{(l)}[i],\\,:],\\quad i=1,...,|\\mathcal{V}_{high}^{(l)}|\n\n<\/span>\n<p>\u5176\u4e2d\u8bed\u8a00 <span class=\"katex-eq\" data-katex-display=\"false\">l<\/span> \u7531\u8f93\u5165\u4e0a\u4e0b\u6587\u51b3\u5b9a\uff0c\u4ece\u800c\u5b9e\u73b0\u9488\u5bf9\u8bed\u8a00\u7279\u5b9a\u9ad8\u9891\u6a21\u5f0f\u7684\u805a\u7126\u538b\u7f29\u3002\u6709\u4e86\u63d0\u53d6\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\mathbf{W}}^{(l)}<\/span>\uff0cMTP \u5934\u4ec5\u5bf9 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{V}_{high}^{(l)}<\/span> \u4e2d\u7684\u8bcd\u5143\u8ba1\u7b97\u8f93\u51fa logits\uff0c\u800c\u975e\u5b8c\u6574\u8bcd\u8868\uff0c\u4ece\u800c\u5728\u964d\u4f4e\u8ba1\u7b97\u6210\u672c\u7684\u540c\u65f6\u7ef4\u6301\u4e0d\u540c\u8bed\u8a00\u4e0a\u7684\u9ad8\u63a5\u53d7\u7387\u3002\u6ce8\u610f\u8bcd\u8868\u538b\u7f29\u6280\u672f\u4ec5\u9650\u4e8e\u8349\u7a3f\u9636\u6bb5\uff1b\u9a8c\u8bc1\u9636\u6bb5\u4fdd\u7559\u5b8c\u6574\u8bcd\u8868\u7a7a\u95f4\uff0c\u4fdd\u8bc1\u65e0\u635f\u751f\u6210\u8d28\u91cf\u3002<\/p>\n<h5>EAGLE \u98ce\u683c MTP \u63a8\u7406<\/h5>\n<p>\u63a8\u7406\u8fc7\u7a0b\u4e0e\u8bad\u7ec3\u4e0d\u540c\uff1a\u4f7f\u7528\u5148\u524d\u6b65\u9aa4\u81ea\u56de\u5f52\u751f\u6210\u7684\u8bcd\u5143\uff0c\u800c\u975e teacher forcing\u3002\u7ed9\u5b9a\u8f93\u5165\u8bcd\u5143\u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">t_{1:i}<\/span>\uff0c\u63a8\u7406\u7ecf\u8fc7\u4e24\u4e2a\u9636\u6bb5\uff1a\u8349\u7a3f\u751f\u6210\u4e0e\u5e76\u884c\u9a8c\u8bc1\u3002<\/p>\n<h6>\u8349\u7a3f\u9636\u6bb5<\/h6>\n<p>\u5728\u8349\u7a3f\u751f\u6210\u9636\u6bb5\uff0c\u8f93\u5165\u9996\u5148\u901a\u8fc7\u4e3b\u6a21\u578b\u4ea7\u751f\u4e0b\u4e00\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{t}_{i+1}<\/span>\uff0c\u4f5c\u4e3a\u7b2c\u4e00\u4e2a\u5df2\u9a8c\u8bc1\u8bcd\u5143\u3002\u968f\u540e MTP \u5934 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u6309 EAGLE \u65b9\u6cd5\u81ea\u56de\u5f52\u751f\u6210 <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u4e2a\u8349\u7a3f\u8bcd\u5143\uff1a<\/p>\n<ul>\n<li>\n<p>\u5728\u521d\u59cb\u8349\u7a3f\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">k=1<\/span>\uff1a <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u53d6\u81ea transformer \u5c42 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{F}<\/span> \u7684\u6700\u540e\u9690\u85cf\u72b6\u6001\uff0c\u4ee5\u53ca\u62fc\u63a5\u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">[t_{2:i};\\hat{t}_{i+1}]<\/span> \u7684\u5d4c\u5165\uff0c\u5176\u4e2d\u65b0\u9884\u6d4b\u8bcd\u5143\u88ab\u8ffd\u52a0\u5230\u524d\u79fb\u4e00\u6b65\u7684\u8f93\u5165\u5e8f\u5217\u3002\u8fd9\u4ea7\u751f\u7b2c\u4e00\u4e2a\u8349\u7a3f\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{t}_{i+2}<\/span>\u3002<\/p>\n<\/li>\n<li>\n<p>\u5728\u540e\u7eed\u8349\u7a3f\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">k\\gt 1<\/span>\uff1a <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u81ea\u56de\u5f52\u8fd0\u884c\uff0c\u5904\u7406\u5176\u5728\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">k-1<\/span> \u7684\u81ea\u8eab\u8f93\u51fa\u9690\u85cf\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{h}_{i+k}<\/span>\uff0c\u5e76\u7ed3\u5408\u5148\u524d\u8349\u7a3f\u8bcd\u5143 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{t}_{i+k}<\/span> \u7684\u5d4c\u5165\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u7ecf\u8fc7 <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u4e2a\u9012\u5f52\u6b65\u540e\uff0c\u6211\u4eec\u5f97\u5230\u5b8c\u6574\u8349\u7a3f\u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">\\hat{t}_{i+2:i+K+1}<\/span>\u3002<\/p>\n<h6>\u9a8c\u8bc1\u9636\u6bb5<\/h6>\n<p>\u4e3b\u6a21\u578b\u968f\u540e\u5e76\u884c\u5904\u7406\u6240\u6709\u8349\u7a3f\u8bcd\u5143\uff0c\u540c\u65f6\u8ba1\u7b97\u4f4d\u7f6e <span class=\"katex-eq\" data-katex-display=\"false\">i+2<\/span> \u5230 <span class=\"katex-eq\" data-katex-display=\"false\">i+K+1<\/span> \u7684 logits\u3002\u6211\u4eec\u91c7\u7528\u6807\u51c6\u63a8\u6d4b\u89e3\u7801\u63a5\u53d7\u51c6\u5219\u51b3\u5b9a\u63a5\u53d7\u7684\u8349\u7a3f\u8bcd\u5143\u6570\u91cf\uff1a\u8bcd\u5143\u6309\u5e8f\u88ab\u63a5\u53d7\uff0c\u76f4\u5230\u7b2c\u4e00\u4e2a\u8349\u7a3f\u8bcd\u5143\u4e0e\u4e3b\u6a21\u578b\u672c\u4f1a\u91c7\u6837\u7684\u8bcd\u5143\u4e0d\u540c\u7684\u4f4d\u7f6e\u3002\u901a\u8fc7\u4e0d\u91c7\u7528\u8fd1\u4f3c\u7b56\u7565\u7684\u5e76\u884c\u9a8c\u8bc1\uff0c\u8fd9\u786e\u4fdd\u8f93\u51fa\u5206\u5e03\u4e0e\u539f\u6a21\u578b\u4fdd\u6301\u4e00\u81f4\uff0c\u8fd9\u4e00\u70b9\u5df2\u88ab\u7406\u8bba\u8bc1\u660e\uff1b\u540c\u65f6\u901a\u8fc7\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u63a5\u53d7\u591a\u4e2a\u8bcd\u5143\u5b9e\u73b0\u52a0\u901f\u3002<\/p>\n<h5>\u8bad\u7ec3\u6570\u636e<\/h5>\n<p>\u6211\u4eec\u91c7\u7528\u81ea\u84b8\u998f\u65b9\u6cd5\u8fdb\u884c\u8bad\u7ec3\u6570\u636e\u751f\u6210\uff0c\u7531\u4e3b\u6a21\u578b\u81ea\u8eab\u751f\u6210\u6240\u6709\u8bad\u7ec3\u54cd\u5e94\u3002\u53d7\u5148\u524d\u5de5\u4f5c\u542f\u53d1\uff0c\u8be5\u7b56\u7565\u786e\u4fdd MTP \u5934\u4e0e\u4e3b\u6a21\u578b\u5206\u5e03\u81ea\u7136\u5bf9\u9f50\uff0c\u4ece\u800c\u5728\u63a8\u6d4b\u89e3\u7801\u4e2d\u5e26\u6765\u66f4\u9ad8\u63a5\u53d7\u7387\u4e0e\u66f4\u6709\u6548\u7684\u8349\u7a3f\u751f\u6210\u3002<\/p>\n<p>\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u4ece\u8de8\u591a\u4e2a\u9886\u57df\u4e0e\u8bed\u8a00\u7684\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u4e2d\u6536\u96c6\u591a\u6837\u5316\u63d0\u793a\u3002\u5bf9\u6bcf\u4e2a\u539f\u59cb\u6837\u672c <span class=\"katex-eq\" data-katex-display=\"false\">(x_{n},y_{n})<\/span>\uff08\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">x_{n}<\/span> \u662f\u63d0\u793a\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">y_{n}<\/span> \u662f\u6e90\u6570\u636e\u96c6\u4e2d\u7684\u5bf9\u5e94\u54cd\u5e94\uff09\uff0c\u6211\u4eec\u4f7f\u7528\u4e3b\u6a21\u578b\u751f\u6210\u65b0\u54cd\u5e94 <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{y}_{n}<\/span>\u3002\u5f97\u5230\u7684\u81ea\u84b8\u998f\u6570\u636e\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\{(x_{n},\\tilde{y}_{n})\\}<\/span> \u6355\u83b7\u4e3b\u6a21\u578b\u7684\u8bed\u4e49\u7279\u5f81\u3001\u751f\u6210\u6a21\u5f0f\u4e0e\u504f\u597d\u3002\u5728\u8be5\u81ea\u84b8\u998f\u6570\u636e\u4e0a\u8bad\u7ec3 MTP \u5934\uff0c\u786e\u4fdd\u5176\u5b66\u4e60\u751f\u6210\u4e0e\u4e3b\u6a21\u578b\u884c\u4e3a\u4e00\u81f4\u7684\u8349\u7a3f\u8bcd\u5143\uff0c\u800c\u975e\u8bd5\u56fe\u6a21\u4eff\u53ef\u80fd\u5177\u6709\u4e0d\u540c\u5206\u5e03\u7684\u5916\u90e8\u6765\u6e90\u54cd\u5e94\u3002<\/p>\n<p>\u6211\u4eec\u7684\u6700\u7ec8\u8bad\u7ec3\u8bed\u6599\u5305\u542b 389.4K \u6761\u4e2d\u82f1\u6587\u6837\u672c\uff0c\u8986\u76d6\u901a\u7528\u77e5\u8bc6\u3001\u6570\u5b66\u63a8\u7406\u4e0e\u7f16\u7a0b\u4efb\u52a1\u3002\u8be6\u7ec6\u9886\u57df\u5206\u5e03\u4ee5\u53ca\u6570\u636e\u96c6\u6574\u7406\u4e0e\u8fc7\u6ee4\u6d41\u7a0b\u89c1\u9644\u5f55 A\u3002<\/p>\n<h4>\u5b9e\u9a8c<\/h4>\n<h5>\u5b9e\u9a8c\u8bbe\u7f6e<\/h5>\n<h6>\u6a21\u578b\u914d\u7f6e<\/h6>\n<p>FastMTP \u57fa\u4e8e\u9884\u8bad\u7ec3\u7684 MiMo-7B-RL \u68c0\u67e5\u70b9\u5b9e\u73b0\uff1a\u8fd9\u662f\u4e00\u4e2a\u7a20\u5bc6 7B \u53c2\u6570\u6a21\u578b\uff0c\u5177\u6709 36 \u4e2a\u89e3\u7801\u5668\u5c42\u3001\u5355\u5c42 MTP \u6a21\u5757\uff0c\u4ee5\u53ca 152K \u8bcd\u8868\u5927\u5c0f\uff08\u57fa\u4e8e Qwen2 tokenizer\uff09\u3002\u6211\u4eec\u91c7\u7528\u8ba1\u7b97\u9ad8\u6548\u7b56\u7565\uff1a\u51bb\u7ed3\u4e3b\u6a21\u578b\uff08\u5305\u62ec transformer \u5c42\u3001\u5d4c\u5165\u5c42\u4e0e\u8f93\u51fa\u5934\uff09\uff0c\u4ec5\u5fae\u8c03 MTP \u5934\u7684 210.8M \u53c2\u6570\uff0c\u5360\u603b\u53c2\u6570 7,833.4M \u7684\u4e0d\u5230 3%\u3002<\/p>\n<h6>\u8bad\u7ec3\u7ec6\u8282<\/h6>\n<p>MTP \u5934\u5728 389.4K \u6761\u81ea\u84b8\u998f\u6837\u672c\u4e0a\u8bad\u7ec3 3 \u4e2a epoch\uff08\u6570\u636e\u96c6\u7ec6\u8282\u89c1\u9644\u5f55 A\uff09\u3002\u8bad\u7ec3\u91c7\u7528\u4f59\u5f26\u5b66\u4e60\u7387\u8c03\u5ea6\uff0c\u5cf0\u503c\u5b66\u4e60\u7387\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">5\\times 10^{-5}<\/span>\uff0cwarmup \u6bd4\u4f8b\u4e3a 0.05\u3002AdamW \u4f18\u5316\u5668\u53c2\u6570\u8bbe\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">(\\beta_{1},\\beta_{2})=(0.9,\\,0.95)<\/span>\uff0c\u5168\u5c40 batch size \u4e3a 64\u3002\u5bf9 MTP \u7279\u5b9a\u8d85\u53c2\u6570\uff0c\u6211\u4eec\u4f7f\u7528\u635f\u5931\u6743\u91cd\u8870\u51cf\u56e0\u5b50 <span class=\"katex-eq\" data-katex-display=\"false\">\\beta=0.6<\/span> \u4e0e\u9884\u6d4b\u6df1\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span>\u3002\u8bad\u7ec3\u4f7f\u7528 ms-swift \u6846\u67b6\u5b8c\u6210\u3002\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u5728\u5355\u53f0 H20 \u670d\u52a1\u5668\u4e0a\u4e0d\u5230 1 \u5929\u5b8c\u6210\uff0c\u5c55\u793a\u4e86\u4f4e\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n<h6>\u8bc4\u4f30\u65b9\u6cd5<\/h6>\n<p>\u6211\u4eec\u5728\u6539\u7f16\u81ea Spec-Bench \u7684\u4e03\u4e2a\u4efb\u52a1\u4e0a\u8bc4\u4f30 FastMTP\uff1a\u7528\u4e8e\u591a\u8f6e\u5bf9\u8bdd\u7684 MT-Bench\u3001\u7528\u4e8e\u7f16\u7a0b\u7684 LiveCodeBench-v6\u3001\u7528\u4e8e\u6570\u5b66\u63a8\u7406\u7684 MATH-500\u3001\u7528\u4e8e RAG \u4e0e\u95ee\u7b54\u7684 Natural Questions\u3001\u7528\u4e8e\u6458\u8981\u7684 CNN\/Daily Mail\uff0c\u4ee5\u53ca\u7528\u4e8e\u4e2d\u6587\u77e5\u8bc6\u8bc4\u4f30\u7684 C-Eval\u3002\u8fd9\u4e9b\u8bc4\u4f30\u57fa\u51c6\u7684\u7ec6\u8282\u89c1\u9644\u5f55 B\u3002\u6240\u6709\u8bc4\u4f30\u91c7\u7528\u4e25\u683c\u7684\u63a8\u6d4b\u89e3\u7801\u63a5\u53d7\u51c6\u5219\uff0c\u786e\u4fdd\u751f\u6210\u8d28\u91cf\u65e0\u4e0b\u964d\u3002\u6240\u6709\u5b9e\u9a8c\u4f7f\u7528 SGLang \u63a8\u7406\u6846\u67b6\uff0c\u91c7\u7528\u5355 batch \u63a8\u7406\u3001\u8d2a\u5fc3\u89e3\u7801\uff08temperature 0\uff09\uff0c\u5e76\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u5c06\u6700\u5927\u751f\u6210\u957f\u5ea6\u8bbe\u4e3a 1024 \u8bcd\u5143\u3002<\/p>\n<h6>\u6307\u6807<\/h6>\n<p>\u6211\u4eec\u4f7f\u7528\u56db\u4e2a\u5e7f\u6cdb\u91c7\u7528\u7684\u8bc4\u4f30\u6307\u6807\uff1a\uff081\uff09\u5e73\u5747\u63a5\u53d7\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span>\uff1a\u4e3b\u6a21\u578b\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u5e73\u5747\u63a5\u53d7\u7684\u8bcd\u5143\u6570\u3002\uff082\uff09\u63a5\u53d7\u7387\uff1a\u9a8c\u8bc1\u671f\u95f4\u88ab\u63a5\u53d7\u7684\u8349\u7a3f\u8bcd\u5143\u767e\u5206\u6bd4\u3002\uff083\uff09\u5e73\u5747\u8f93\u51fa\u541e\u5410\uff1a\u5e73\u5747\u6bcf\u79d2\u8f93\u51fa\u8bcd\u5143\u6570\uff08token\/s\uff09\u3002\uff084\uff09\u52a0\u901f\u6bd4\uff1a\u76f8\u5bf9\u57fa\u7ebf\uff08\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\uff09\u7684\u76f8\u5bf9\u52a0\u901f\u3002<\/p>\n<h6>\u786c\u4ef6\u8bbe\u7f6e<\/h6>\n<p>\u6211\u4eec\u5728 NVIDIA A10 GPU\uff0824GB\uff09\u4e0a\u8fdb\u884c\u4e3b\u8981\u5b9e\u9a8c\uff0c\u8fd9\u662f\u751f\u4ea7\u73af\u5883\u4e2d\u5e7f\u6cdb\u90e8\u7f72\u7684\u52a0\u901f\u5668\u3002<\/p>\n<p>\u9700\u8981\u6ce8\u610f\uff0c\u63a8\u6d4b\u89e3\u7801\u5728\u7406\u8bba\u4e0a\u4fdd\u6301\u4e3b\u6a21\u578b\u7684\u8f93\u51fa\u5206\u5e03\u3002\u7531\u4e8e\u6211\u4eec\u91c7\u7528\u4e25\u683c\u63a5\u53d7\u51c6\u5219\u4e14\u4e0d\u8fdb\u884c\u4efb\u4f55\u653e\u677e\uff0c\u8f93\u51fa\u4e0e\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u76f8\u540c\uff0c\u56e0\u6b64\u65e0\u9700\u5355\u72ec\u8bc4\u4f30\u751f\u6210\u8d28\u91cf\u3002<\/p>\n<h5>\u4e3b\u8981\u7ed3\u679c<\/h5>\n<p>\u6211\u4eec\u8bc4\u4f30\u4e09\u79cd\u4e3b\u8981\u914d\u7f6e\uff1a\uff081\uff09\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\uff08\u57fa\u7ebf\uff09\uff1b\uff082\uff09\u672a\u7ecf\u5fae\u8c03\u7684\u539f\u59cb MTP \u68c0\u67e5\u70b9\uff08vanilla MTP\uff09\uff1b\uff083\uff09\u6240\u63d0\u51fa\u7684 FastMTP\u3002\u4e3a\u5206\u6790\u5404\u7ec4\u4ef6\u8d21\u732e\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u8bc4\u4f30 FastMTP \u53d8\u4f53\uff1a<\/p>\n<ul>\n<li>\n<p>Fixed-data FT\uff1a\u4f7f\u7528\u539f\u59cb\u6570\u636e\u6e90\u4e2d\u7684\u63d0\u793a\u4e0e\u54cd\u5e94\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n<\/li>\n<li>\n<p>Self-data FT\uff1a\u4f7f\u7528\u7531\u4e3b\u6a21\u578b\u751f\u6210\u7684\u81ea\u84b8\u998f\u54cd\u5e94\u3002<\/p>\n<\/li>\n<li>\n<p>Self-data FT + FR\uff1a\u989d\u5916\u5f15\u5165\u8bed\u8a00\u611f\u77e5\u8bcd\u8868\u538b\u7f29\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u6211\u4eec\u5728\u4e0d\u540c\u8349\u7a3f\u6df1\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">K\\in\\{0,1,2,3\\}<\/span> \u4e0a\u8bc4\u4f30\u6027\u80fd\uff0c\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">K=0<\/span> \u8868\u793a\u65e0\u8349\u7a3f\u7684\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span> \u8868\u793a\u4e09\u6b21\u9012\u5f52 MTP \u524d\u5411\u4f20\u64ad\u5e76\u751f\u6210\u4e09\u4e2a\u989d\u5916\u8349\u7a3f\u8bcd\u5143\u3002<\/p>\n<p><center>\u8868 1\uff1aMiMo-RL-7B \u4f7f\u7528\u4e0d\u540c\u65b9\u6cd5\u9884\u6d4b K \u4e2a\u8349\u7a3f\u8bcd\u5143\u65f6\u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\u4e0e\u89e3\u7801\u901f\u5ea6\uff08token\/s\uff09\u3002\u4efb\u52a1\u5305\u62ec\u591a\u8f6e\u5bf9\u8bdd\uff08MT.\uff09\u3001\u7f16\u7a0b\uff08Code\uff09\u3001\u6570\u5b66\u63a8\u7406\uff08Math\uff09\u3001\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u3001\u95ee\u7b54\uff08QA\uff09\u3001\u6458\u8981\uff08Summ.\uff09\u4e0e\u4e2d\u6587\u77e5\u8bc6\uff08ZH.\uff09\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260716102113002.png\" alt=\"\u8868 1\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<p>\u8868 1 \u4e0e\u56fe 2 \u5c55\u793a\u4e86 FastMTP \u5728\u4e03\u4e2a\u591a\u6837\u5316\u4efb\u52a1\u4e0a\u7684\u5168\u9762\u52a0\u901f\u6027\u80fd\u8bc4\u4f30\u3002\u5e26\u81ea\u84b8\u998f\u4e0e\u8bcd\u8868\u538b\u7f29\u7684 FastMTP\uff08Self-data FT + FR\uff09\u5728\u6240\u6709\u57fa\u51c6\u4e0a\u53d6\u5f97\u6700\u4f18\u6027\u80fd\uff0c\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span> \u65f6\u76f8\u5bf9\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u5e73\u5747\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">2.03\\times<\/span> \u52a0\u901f\u3002\u6027\u80fd\u6536\u76ca\u968f\u4efb\u52a1\u9886\u57df\u53d8\u5316\uff0c\u53cd\u6620\u5176\u4e0d\u540c\u7684\u751f\u6210\u7279\u6027\u3002\u6570\u5b66\u63a8\u7406\u8868\u73b0\u51fa\u6700\u9ad8\u52a0\u901f\u4e0e\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\uff08\u8bcd\u8868\u538b\u7f29\u524d\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span> \u65f6 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau=3.16<\/span>\uff09\uff0c\u8bc1\u660e MTP \u5934\u5728\u6355\u83b7\u7ed3\u6784\u5316\u63a8\u7406\u6a21\u5f0f\u65b9\u9762\u7684\u6709\u6548\u6027\u3002\u7f16\u7a0b\u4efb\u52a1\u8868\u73b0\u6b21\u4f18\uff0c\u53d7\u76ca\u4e8e\u666e\u904d\u7684\u56fa\u5b9a\u6a21\u677f\u4e0e\u91cd\u590d\u7f16\u7a0b\u7ed3\u6784\u3002\u95ee\u7b54\u7b49\u901a\u7528 NLP \u4efb\u52a1\u6539\u8fdb\u76f8\u5bf9\u8f83\u5c0f\uff08<span class=\"katex-eq\" data-katex-display=\"false\">1.84\\times<\/span>\u2013<span class=\"katex-eq\" data-katex-display=\"false\">2.07\\times<\/span>\uff09\uff0c\u53ef\u80fd\u56e0\u5176\u66f4\u9ad8\u7684\u8bed\u8a00\u591a\u6837\u6027\u4e0e\u66f4\u96be\u9884\u6d4b\u7684\u8bcd\u5143\u4f9d\u8d56\u3002\u56fe 2 \u53ef\u89c6\u5316\u7684\u8fd9\u79cd\u4efb\u52a1\u7279\u5b9a\u5dee\u5f02\u786e\u8ba4\uff1aFastMTP \u5728\u591a\u6837\u5316\u751f\u6210\u573a\u666f\u4e2d\u4fdd\u6301\u4e00\u81f4\u7684\u52a0\u901f\u6536\u76ca\u3002<\/p>\n<p><center>\u56fe 2\uff1a\u4e0d\u540c\u65b9\u6cd5\u5728\u5404\u5b50\u4efb\u52a1\u4e0a\u7684\u52a0\u901f\u6bd4\u8f83\uff0c\u5728\u5355\u5f20 A10 GPU \u4e0a\u8bc4\u4f30\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260716102113003.png\" alt=\"\u56fe 2\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<h5>\u8fdb\u4e00\u6b65\u5206\u6790<\/h5>\n<p>\u6211\u4eec\u7684\u8fdb\u4e00\u6b65\u5206\u6790\u63a2\u7d22 FastMTP \u8bbe\u8ba1\u4e0e\u6027\u80fd\u7684\u4e24\u4e2a\u5173\u952e\u65b9\u9762\uff1a<\/p>\n<p>\uff081\uff09\u6700\u4f18\u8349\u7a3f\u957f\u5ea6\uff1a\u5b9e\u73b0\u6700\u5927\u52a0\u901f\u7684\u6700\u4f18 MTP \u8349\u7a3f\u6b65\u6570\u662f\u591a\u5c11\uff1f\u6362\u8a00\u4e4b\uff0c\u5355\u4e00 MTP \u5934\u5728\u4fdd\u6301\u9ad8\u63a5\u53d7\u7387\u7684\u540c\u65f6\u80fd\u591f\u5b66\u4e60\u7684\u6709\u6548\u9884\u6d4b\u8ddd\u79bb\u662f\u591a\u5c11\uff1f<\/p>\n<p>\uff082\uff09\u8bcd\u8868\u538b\u7f29\u6743\u8861\uff1a\u4f55\u79cd\u7a0b\u5ea6\u7684\u8bcd\u8868\u538b\u7f29\u80fd\u5728\u8ba1\u7b97\u6548\u7387\u4e0e\u63a5\u53d7\u7387\u4e4b\u95f4\u53d6\u5f97\u6700\u4f18\u5e73\u8861\uff1f\u8be5\u6743\u8861\u5982\u4f55\u968f\u4e0d\u540c\u9886\u57df\u4e0e\u8bed\u8a00\u53d8\u5316\uff1f<\/p>\n<h6>\u6700\u4f18\u8349\u7a3f\u957f\u5ea6<\/h6>\n<p>\u4e3a\u786e\u5b9a\u6700\u4f18\u8349\u7a3f\u6b65\u6570\uff0c\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u80fd\u591f\u9884\u6d4b\u6700\u591a 7 \u4e2a\u989d\u5916\u8bcd\u5143\u7684 MTP \u5934\uff0c\u5e76\u5e94\u7528\u7b2c 2.1 \u8282\u6240\u8ff0\u76f8\u540c\u7684\u6307\u6570\u8870\u51cf\u635f\u5931\u52a0\u6743\u7b56\u7565\u3002\u968f\u540e\u6211\u4eec\u8bc4\u4f30\u4e0d\u540c\u8349\u7a3f\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u4e0b\u7684\u89e3\u7801\u901f\u5ea6\u4e0e\u63a5\u53d7\u957f\u5ea6\u3002<\/p>\n<p>\u4ece\u56fe 3 \u53ef\u89c1\uff0cFastMTP \u5728 <span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span> \u8fbe\u5230\u5cf0\u503c\u52a0\u901f\uff0c\u8fbe\u5230 140 token\/s\uff1b\u800c vanilla MTP \u68c0\u67e5\u70b9\u66f4\u65e9\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">K=2<\/span> \u8fbe\u5230\u5cf0\u503c\uff0c\u4e14\u8f93\u51fa\u541e\u5410\u66f4\u4f4e\u3002\u5bf9\u63a5\u53d7\u957f\u5ea6\u800c\u8a00\uff0c\u5e26\u81ea\u84b8\u998f\u5fae\u8c03\u7684 FastMTP \u968f <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u589e\u52a0\u4ece 1.0 \u6301\u7eed\u589e\u957f\u5230 3.2\uff0c\u8868\u660e\u6211\u4eec\u7684\u8bad\u7ec3\u7b56\u7565\u6210\u529f\u4f7f\u6a21\u578b\u5b66\u4f1a\u591a\u6b65\u9884\u6d4b\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cvanilla MTP \u4ece <span class=\"katex-eq\" data-katex-display=\"false\">K=2<\/span> \u8d77\u63a5\u53d7\u957f\u5ea6\u51e0\u4e4e\u5e73\u5766\u5730\u7ef4\u6301\u5728\u7ea6 1.8\uff0c\u8868\u660e\u5728\u7b2c\u4e00\u4f4d\u7f6e\u4e4b\u540e\u9884\u6d4b\u8d28\u91cf\u8f83\u5dee\u3002<\/p>\n<p><center>\u56fe 3\uff1a\u5728\u5355\u5f20 A100 \u4e0a\uff0c\u4e0d\u540c MTP \u8349\u7a3f\u8bcd\u5143\u6570\u91cf\u4e0b\u7684\u89e3\u7801\u901f\u5ea6\u4e0e\u63a5\u53d7\u957f\u5ea6\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260716102113004.png\" alt=\"\u56fe 3\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<p>\u5c3d\u7ba1\u63a5\u53d7\u957f\u5ea6\u5355\u8c03\u589e\u52a0\uff0c\u89e3\u7801\u901f\u5ea6\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span> \u8fbe\u5230\u5cf0\u503c\u540e\u9010\u6e10\u4e0b\u964d\u3002\u66f4\u957f\u8349\u7a3f\u5e8f\u5217\u4f1a\u63d0\u9ad8\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\uff0c\u4f46\u4e5f\u4f1a\u5e26\u6765\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u3002\u8d85\u8fc7 <span class=\"katex-eq\" data-katex-display=\"false\">K=3<\/span> \u540e\uff0c\u63a5\u53d7\u957f\u5ea6\u7684\u8fb9\u9645\u6536\u76ca\u65e0\u6cd5\u8865\u507f\u751f\u6210\u66f4\u957f\u8349\u7a3f\u7684\u589e\u957f\u8ba1\u7b97\u5f00\u9500\uff0c\u56e0\u4e3a\u66f4\u8fdc\u8bcd\u5143\u9010\u6b65\u66f4\u96be\u51c6\u786e\u9884\u6d4b\uff0c\u4ece\u800c\u964d\u4f4e\u6574\u4f53\u6548\u7387\u3002\u56e0\u6b64\uff0cFastMTP \u5728\u8349\u7a3f\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{K=3}<\/span> \u65f6\u5b9e\u73b0\u6700\u5927\u52a0\u901f\uff1a\u6b64\u65f6\u901a\u8fc7\u8bad\u7ec3\u83b7\u5f97\u7684\u63a5\u53d7\u957f\u5ea6\u63d0\u5347\u6700\u4f18\u5730\u8d85\u8fc7\u9012\u5f52\u8349\u7a3f\u7684\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n<h6>\u8bcd\u8868\u538b\u7f29\u6743\u8861<\/h6>\n<p>\u4e3a\u7814\u7a76\u8bcd\u8868\u538b\u7f29\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\uff0c\u6211\u4eec\u8bc4\u4f30\u4e86\u56db\u79cd\u8bcd\u8868\u5927\u5c0f\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">|\\mathcal{V}_{high}|=\\{8\\text{k},16\\text{k},32\\text{k},64\\text{k}\\}<\/span>\uff0c\u5e76\u4e0e\u5b8c\u6574 152k \u8bcd\u8868\u6bd4\u8f83\u3002\u6211\u4eec\u9009\u62e9\u4e24\u4e2a\u4ee3\u8868\u6027\u57fa\u51c6\uff1a\u7528\u4e8e\u4e2d\u6587\u77e5\u8bc6\u95ee\u7b54\u7684 C-Eval\uff0c\u4ee5\u53ca\u7528\u4e8e\u82f1\u6587\u591a\u8f6e\u5bf9\u8bdd\u7684 MT-Bench\u3002\u8fd9\u4e00\u9009\u62e9\u6709\u52a9\u4e8e\u6211\u4eec\u5728\u591a\u6837\u5316\u8bed\u8a00\u4e0a\u4e0b\u6587\u4e2d\u5206\u6790\u8bed\u8a00\u7279\u5b9a\u538b\u7f29\u6548\u679c\u3002<\/p>\n<p>\u8bcd\u8868\u9891\u7387\u6a21\u5f0f\u88ab\u89c2\u5bdf\u5230\u9ad8\u5ea6\u4f9d\u8d56\u8bed\u8a00\uff0c\u5370\u8bc1\u4e86 FR-Spec \u7684\u4e0a\u4e0b\u6587\u76f8\u5173\u52a0\u901f\u8303\u5f0f\u3002\u539f\u59cb\u5b9e\u73b0\u4f7f\u7528\u6765\u81ea SlimPajama-627B\uff08\u4ee5\u82f1\u8bed\u4e3a\u4e3b\u7684\u8bed\u6599\uff09\u7684\u8bcd\u5143\u7edf\u8ba1\uff0c\u7531\u4e8e\u9891\u7387\u5206\u5e03\u4e0d\u5339\u914d\uff0c\u5728\u771f\u5b9e\u4e16\u754c\u4e2d\u6587\u4efb\u52a1\u4e0a\u63a5\u53d7\u7387\u4f4e\u3001\u6027\u80fd\u5dee\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u9650\u5236\uff0c\u6211\u4eec\u4e3a\u6bcf\u4e2a\u76ee\u6807\u9886\u57df\u6784\u5efa\u8bed\u8a00\u7279\u5b9a\u538b\u7f29\u8bcd\u8868\u3002\u5bf9\u4e2d\u6587\u4efb\u52a1\uff0c\u6211\u4eec\u9009\u62e9 Chinese-DeepSeek-R1-Distill-110k-SFT \u6570\u636e\u96c6\u2014\u2014\u5b83\u5305\u542b\u4ece\u5f3a\u5927\u7684 DeepSeek-R1 \u6a21\u578b\u84b8\u998f\u5f97\u5230\u7684\u9ad8\u8d28\u91cf\u6570\u5b66\u63a8\u7406\u4e0e\u591a\u6837\u5316\u4e2d\u6587\u6307\u4ee4\u8ddf\u968f\u6837\u4f8b\u2014\u2014\u4ee5\u8bc6\u522b\u4e2d\u6587\u751f\u6210\u6a21\u5f0f\u7279\u6709\u7684\u9ad8\u9891\u8bcd\u5143\u3002\u63a8\u7406\u671f\u95f4\uff0cFastMTP \u6839\u636e\u8f93\u5165\u4e0a\u4e0b\u6587\u52a8\u6001\u5207\u6362\u5230\u5408\u9002\u7684\u538b\u7f29\u8bcd\u8868\u3002<\/p>\n<p>\u8868 2 \u5c55\u793a\u4e86\u4e0d\u540c\u8bcd\u8868\u914d\u7f6e\u4e0b\u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\u3001\u89e3\u7801\u901f\u5ea6\u4e0e\u52a0\u901f\u6bd4\u3002\u7ed3\u679c\u8868\u660e\uff0c\u5373\u4f7f\u6fc0\u8fdb\u8bcd\u8868\u538b\u7f29\u4e5f\u4ec5\u5e26\u6765\u63a5\u53d7\u957f\u5ea6\u7684\u6700\u5c0f\u9000\u5316\uff0c\u540c\u65f6\u5e26\u6765\u53ef\u89c2\u52a0\u901f\u6536\u76ca\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6700\u4f18\u538b\u7f29\u6bd4\u968f\u8bed\u8a00\u53d8\u5316\u3002\u5bf9 MT-Bench\uff0c32k \u8bcd\u8868\u914d\u7f6e\u8fbe\u5230\u5cf0\u503c\u6027\u80fd\uff0c\u52a0\u901f <span class=\"katex-eq\" data-katex-display=\"false\">2.028\\times<\/span>\uff0c\u76f8\u5bf9\u5b8c\u6574\u8bcd\u8868\u4ec5\u727a\u7272 0.068 \u4e2a\u63a5\u53d7\u957f\u5ea6\u8bcd\u5143\u3002\u8fd9\u4e0e\u5148\u524d\u4ee5\u82f1\u8bed\u4e3a\u4e2d\u5fc3\u538b\u7f29\u5de5\u4f5c\u7684\u53d1\u73b0\u4e00\u81f4\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cC-Eval \u5728\u66f4\u6fc0\u8fdb\u7684 16k \u8bcd\u8868\u4e0a\u8fbe\u5230\u6700\u4f18\u6027\u80fd\uff0c\u52a0\u901f <span class=\"katex-eq\" data-katex-display=\"false\">1.990\\times<\/span>\uff0c\u63a5\u53d7\u957f\u5ea6\u4e0b\u964d\u751a\u81f3\u66f4\u5c0f\uff0c\u4ec5\u4e3a 0.026 \u4e2a\u8bcd\u5143\u3002\u8fd9\u8868\u660e\u4e2d\u6587\u6587\u672c\u751f\u6210\u53ef\u80fd\u5448\u73b0\u66f4\u96c6\u4e2d\u7684\u8bcd\u5143\u4f7f\u7528\u6a21\u5f0f\uff0c\u56e0\u800c\u4ece\u8bcd\u8868\u538b\u7f29\u7b56\u7565\u4e2d\u6536\u76ca\u66f4\u5927\u3002\u603b\u4e4b\uff0c\u6700\u4f18\u8bcd\u8868\u5927\u5c0f\u56e0\u8bed\u8a00\u800c\u5f02\u2014\u2014\u4e2d\u6587 16k\u3001\u82f1\u6587 32k\u2014\u2014\u63ed\u793a\u4e86\u4e0d\u540c\u7684\u8bcd\u5143\u5206\u5e03\u6a21\u5f0f\u3002<\/p>\n<p><center>\u8868 2\uff1a\u4e0d\u540c FR-Spec \u914d\u7f6e\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260716102113005.png\" alt=\"\u8868 2\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<h5>\u6d88\u878d\u5b9e\u9a8c<\/h5>\n<p>\u6211\u4eec\u8fdb\u884c\u5168\u9762\u6d88\u878d\u5b9e\u9a8c\uff0c\u4ee5\u9a8c\u8bc1 FastMTP \u4e2d\u5404\u7ec4\u4ef6\u7684\u8d21\u732e\uff0c\u5e76\u5206\u6790\u8bad\u7ec3\u540e\u6bcf\u4e2a\u8349\u7a3f\u4f4d\u7f6e\u7684\u63a5\u53d7\u7387\u3002<\/p>\n<h6>\u8bbe\u8ba1\u9009\u62e9\u7684\u5f71\u54cd<\/h6>\n<p>\u4e3a\u91cf\u5316\u8bbe\u8ba1\u9009\u62e9\u7684\u5404\u81ea\u8d21\u732e\uff0c\u6211\u4eec\u76f8\u5bf9 vanilla MTP \u57fa\u7ebf\u8bc4\u4f30\u4e09\u79cd\u914d\u7f6e\uff1a\u56fa\u5b9a\u6570\u636e\u5fae\u8c03\u3001\u81ea\u84b8\u998f\u6570\u636e\u5fae\u8c03\uff0c\u4ee5\u53ca\u5e26\u8bcd\u8868\u538b\u7f29\u7684\u5b8c\u6574 FastMTP\u3002\u5982\u8868 1 \u6240\u793a\uff0cFastMTP \u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">2.03\\times<\/span> \u52a0\u901f\uff0c\u76f8\u5bf9 vanilla MTP\uff08<span class=\"katex-eq\" data-katex-display=\"false\">1.21\\times<\/span>\uff09\u63d0\u5347 82%\uff0c\u76f8\u5bf9\u56fa\u5b9a\u6570\u636e\u5fae\u8c03\uff08<span class=\"katex-eq\" data-katex-display=\"false\">1.67\\times<\/span>\uff09\u63d0\u5347 36%\uff0c\u76f8\u5bf9\u65e0\u8bcd\u8868\u538b\u7f29\u7684\u81ea\u84b8\u998f\u6570\u636e\u5fae\u8c03\uff08<span class=\"katex-eq\" data-katex-display=\"false\">1.81\\times<\/span>\uff09\u63d0\u5347 22%\u3002\u8fd9\u4e9b\u589e\u91cf\u6539\u8fdb\u9a8c\u8bc1\u4e86\u5404\u7ec4\u4ef6\u7684\u6709\u6548\u6027\uff1a\u5fae\u8c03\u5171\u4eab\u6743\u91cd MTP \u5934\u4f7f\u63a8\u7406\u65f6\u591a\u8bcd\u5143\u8349\u7a3f\u66f4\u51c6\u786e\u3001\u66f4\u5feb\uff1b\u81ea\u84b8\u998f\u5b9e\u73b0\u8349\u7a3f\u4e0e\u4e3b\u6a21\u578b\u4e4b\u95f4\u66f4\u597d\u7684\u5206\u5e03\u5bf9\u9f50\u4ee5\u83b7\u5f97\u66f4\u9ad8\u63a5\u53d7\u957f\u5ea6\uff1b\u8bcd\u8868\u538b\u7f29\u964d\u4f4e\u8ba1\u7b97\u5f00\u9500\u4ee5\u63d0\u5347\u8f93\u51fa\u541e\u5410\uff0c\u540c\u65f6\u5bf9\u63a5\u53d7\u7387\u5f71\u54cd\u6700\u5c0f\u3002<\/p>\n<h6>\u63a5\u53d7\u7387\u5206\u6790<\/h6>\n<p>\u56fe 4(a) \u8fdb\u4e00\u6b65\u5c55\u793a\u5fae\u8c03\u5728\u5b9e\u73b0\u6709\u6548\u591a\u8bcd\u5143\u8349\u7a3f\u4e2d\u7684\u5173\u952e\u4f5c\u7528\u3002Vanilla MTP \u5728\u7b2c\u4e00\u8349\u7a3f\u6b65\u4e4b\u540e\u8868\u73b0\u4e25\u91cd\u9000\u5316\uff1a\u63a5\u53d7\u7387\u4ece <span class=\"katex-eq\" data-katex-display=\"false\">k=1<\/span> \u65f6\u7ea6 70% \u6025\u5267\u4e0b\u964d\u5230 <span class=\"katex-eq\" data-katex-display=\"false\">k=2<\/span> \u65f6\u4ec5\u7ea6 10%\uff0c\u5e76\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">k=3<\/span> \u63a5\u8fd1\u96f6\u3002\u8fd9\u79cd\u5d29\u6e83\u66b4\u9732\u4e86\u5c06 vanilla MTP \u68c0\u67e5\u70b9\u90e8\u7f72\u7528\u4e8e\u63a8\u6d4b\u89e3\u7801\u4e2d\u9012\u5f52\u591a\u8bcd\u5143\u9884\u6d4b\u7684\u56fa\u6709\u5c40\u9650\u2014\u2014\u5b83\u4ec5\u88ab\u8bad\u7ec3\u7528\u4e8e\u5355\u6b65\u8d85\u524d\u9884\u6d4b\uff0c\u56e0\u800c\u5728\u5e94\u7528\u4e8e\u66f4\u6df1\u8349\u7a3f\u4f4d\u7f6e\u65f6\u96be\u4ee5\u7ef4\u6301\u9ad8\u63a5\u53d7\u7387\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u6211\u4eec\u5fae\u8c03\u540e\u7684 MTP \u5728\u6240\u6709\u8349\u7a3f\u6b65\u4e0a\u8fbe\u5230\u663e\u8457\u66f4\u9ad8\u7684\u63a5\u53d7\u7387\uff1a\u5e73\u5747\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">k=1<\/span> \u4e3a 80%\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">k=2<\/span> \u4e3a 56%\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">k=3<\/span> \u4e3a 36%\u3002\u6570\u5b66\u63a8\u7406\u6536\u76ca\u6700\u9ad8\uff0c\u5373\u4f7f\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">k=3<\/span> \u4ecd\u4fdd\u6301\u8d85\u8fc7 50% \u7684\u63a5\u53d7\u7387\uff0c\u800c vanilla MTP \u4ec5\u7ea6 3%\u3002\u5728\u5168\u90e8\u4e03\u4e2a\u8bc4\u4f30\u4efb\u52a1\u4e0a\u7684\u4e00\u81f4\u4f18\u52bf\u9a8c\u8bc1\uff1a\u4f7f\u7528\u81ea\u84b8\u998f\u6570\u636e\u5fae\u8c03\u6709\u6548\u4f7f MTP \u5934\u5177\u5907\u6355\u83b7\u8fde\u7eed\u672a\u6765\u8bcd\u5143\u4f9d\u8d56\u7684\u80fd\u529b\uff0c\u800c\u8fd9\u6b63\u662f\u6709\u6548\u591a\u6b65\u9884\u6d4b\u6240\u5fc5\u9700\u7684\u3002\u56fe 4(b) \u5c55\u793a\u8bad\u7ec3\u635f\u5931\u66f2\u7ebf\u3002\u635f\u5931\u5728\u8bad\u7ec3\u65e9\u671f\u9636\u6bb5\uff08\u7ea6 0.5 \u4e2a epoch\uff09\u6025\u5267\u4e0b\u964d\uff0c\u968f\u540e\u7f13\u6162\u4f46\u7a33\u5b9a\u5730\u6539\u8fdb\u76f4\u81f3\u6536\u655b\u3002\u5982\u9884\u671f\uff0c\u9884\u6d4b\u968f\u8bcd\u5143\u4f4d\u7f6e\u589e\u52a0\u800c\u66f4\u4e0d\u51c6\u786e\uff0c\u66f4\u6df1\u4f4d\u7f6e\u4e0d\u53ef\u907f\u514d\u5730\u663e\u793a\u66f4\u9ad8\u635f\u5931\uff0c\u8fdb\u4e00\u6b65\u8bf4\u660e\u66f4\u8fdc\u8bcd\u5143\u9010\u6b65\u66f4\u96be\u51c6\u786e\u9884\u6d4b\u3002<\/p>\n<p><center>\u56fe 4\uff1a(a) \u81ea\u84b8\u998f\u6570\u636e\u96c6\u5fae\u8c03 MTP \u76f8\u5bf9 vanilla MTP \u5728\u4e0d\u540c\u8349\u7a3f\u6b65\u4e0a\u7684\u63a5\u53d7\u7387\u63d0\u5347\u3002(b) \u4e0d\u540c\u8349\u7a3f\u6b65\u4e0a MTP \u7684\u8bad\u7ec3\u635f\u5931\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260716102113006.png\" alt=\"\u56fe 4\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<h4>\u76f8\u5173\u5de5\u4f5c<\/h4>\n<h5>LLM \u63a8\u7406\u52a0\u901f<\/h5>\n<p>LLM \u4e0d\u65ad\u589e\u957f\u7684\u8ba1\u7b97\u9700\u6c42\u63a8\u52a8\u4e86\u5bf9\u63a8\u7406\u52a0\u901f\u6280\u672f\u7684\u5e7f\u6cdb\u7814\u7a76\uff0c\u5927\u4f53\u53ef\u5206\u4e3a\u56db\u4e2a\u4e3b\u8981\u65b9\u5411\uff1a\u67b6\u6784\u521b\u65b0\u3001\u6a21\u578b\u538b\u7f29\u3001\u6846\u67b6\u4f18\u5316\u4e0e\u63a8\u6d4b\u89e3\u7801\u3002<\/p>\n<p>\u67b6\u6784\u521b\u65b0\u805a\u7126\u4e8e\u91cd\u65b0\u8bbe\u8ba1\u6a21\u578b\u7ec4\u4ef6\u4ee5\u63d0\u5347\u6548\u7387\u3002\u4ee3\u8868\u6027\u5de5\u4f5c\u5305\u62ec\u964d\u4f4e\u4e8c\u6b21\u590d\u6742\u5ea6\u7684\u9ad8\u6548\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4f8b\u5982\u5c06 softmax \u6ce8\u610f\u529b\u8fd1\u4f3c\u4e3a\u7ebf\u6027\u8fd0\u7b97\u7684\u7ebf\u6027\u6ce8\u610f\u529b\u3001\u57fa\u4e8e\u56fa\u5b9a\u6216\u52a8\u6001\u7a00\u758f\u6a21\u5f0f\u5c06\u8ba1\u7b97\u9650\u5236\u5230\u9009\u62e9\u6027 key \u5b50\u96c6\u7684\u7a00\u758f\u6ce8\u610f\u529b\uff0c\u4ee5\u53ca\u901a\u8fc7\u7ebf\u6027\u6295\u5f71\u8fdb\u884c\u7ef4\u5ea6\u538b\u7f29\u7684\u4f4e\u79e9\u6ce8\u610f\u529b\u3002\u6a21\u578b\u538b\u7f29\u6280\u672f\u964d\u4f4e\u8ba1\u7b97\u4e0e\u5185\u5b58\u9700\u6c42\uff0c\u5305\u62ec\u91cf\u5316\u3001\u526a\u679d\u4e0e\u77e5\u8bc6\u84b8\u998f\u3002\u6846\u67b6\u4f18\u5316\u901a\u8fc7\u7cfb\u7edf\u7ea7\u5de5\u7a0b\u63d0\u5347\u90e8\u7f72\u6548\u7387\u3002\u663e\u8457\u8d21\u732e\u5305\u62ec\u5c06\u6574\u4e2a\u6ce8\u610f\u529b\u8fd0\u7b97\u878d\u5408\u4e3a\u5355\u4e00\u7b97\u5b50\u7684 FlashAttention\u3001\u5e26 PagedAttention \u7684 vLLM \u7528\u4e8e\u5185\u5b58\u7ba1\u7406\u3001\u5e26 RadixAttention \u7684 SGLang \u7528\u4e8e KV cache \u590d\u7528\uff0c\u4ee5\u53ca\u5185\u7f6e\u591a\u79cd\u5e76\u884c\u7b56\u7565\u4e0e\u9ad8\u7ea7\u7279\u6027\u652f\u6301\u7684 TensorRT-LLM\u3002\u63a8\u6d4b\u89e3\u7801\u5229\u7528\u8349\u7a3f-\u518d\u9a8c\u8bc1\u8303\u5f0f\u52a0\u901f\u89e3\u7801\u3002<\/p>\n<h5>\u63a8\u6d4b\u89e3\u7801\u4f18\u5316<\/h5>\n<p>\u5728\u4e0a\u8ff0\u52a0\u901f\u6280\u672f\u4e2d\uff0c\u63a8\u6d4b\u89e3\u7801\u72ec\u7279\u5730\u5728\u4fdd\u6301\u8f93\u51fa\u8d28\u91cf\u7684\u540c\u65f6\u5b9e\u73b0\u663e\u8457\u52a0\u901f\u3002\u63a8\u6d4b\u89e3\u7801\u7684\u4f18\u5316\u52aa\u529b\u4e3b\u8981\u9488\u5bf9\u4e24\u4e2a\u65b9\u9762\uff1a\u83b7\u5f97\u66f4\u9ad8\u63a5\u53d7\u7387\uff0c\u4ee5\u53ca\u964d\u4f4e\u8349\u7a3f\u751f\u6210\u5f00\u9500\u3002<\/p>\n<p>\u4e3a\u83b7\u5f97\u66f4\u9ad8\u63a5\u53d7\u7387\uff0c\u5148\u524d\u7814\u7a76\u63a2\u7d22\u4e86\u4e24\u4e2a\u4e3b\u8981\u65b9\u5411\u3002\u7b2c\u4e00\u662f\u901a\u8fc7\u4e13\u95e8\u6a21\u578b\u8bbe\u8ba1\u4e0e\u8bad\u7ec3\u63d0\u5347\u8349\u7a3f\u8d28\u91cf\u3002\u82e5\u5e72\u5de5\u4f5c\u901a\u8fc7\u5728\u76ee\u6807\u6a21\u578b\u4e0a\u6dfb\u52a0\u989d\u5916\u9884\u6d4b\u5934\u6765\u53d1\u5c55\u81ea\u8349\u7a3f\u673a\u5236\u3002Medusa \u4f7f\u7528\u989d\u5916 MLP \u5934\uff0c\u590d\u7528\u76ee\u6807\u6a21\u578b\u7684\u6700\u540e\u9690\u85cf\u72b6\u6001\u4ee5\u5e76\u884c\u9884\u6d4b\u63a5\u4e0b\u6765\u82e5\u5e72\u8bcd\u5143\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cEAGLE \u540c\u65f6\u7eb3\u5165\u6700\u540e\u9690\u85cf\u72b6\u6001\u4e0e\u5148\u524d\u8bcd\u5143\uff0c\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u8fdb\u884c\u8349\u7a3f\uff0c\u663e\u8457\u63d0\u5347\u8349\u7a3f\u7a33\u5b9a\u6027\u4e0e\u51c6\u786e\u6027\uff0c\u5e76\u786e\u7acb\u5176\u4f5c\u4e3a\u5f53\u524d\u6700\u5148\u8fdb\u65b9\u6cd5\u7684\u5730\u4f4d\u3002\u53e6\u4e00\u4e9b\u5de5\u4f5c\u4f7f\u7528\u540c\u4e00\u6a21\u578b\u7cfb\u5217\u4e2d\u7684\u66f4\u5c0f\u6a21\u578b\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\uff0c\u5229\u7528\u76f8\u4f3c\u6027\u83b7\u5f97\u66f4\u597d\u5bf9\u9f50\u3002\u7b2c\u4e8c\u65b9\u5411\u7565\u5fae\u653e\u677e\u5339\u914d\u8981\u6c42\uff0c\u66f4\u4fe1\u4efb\u8349\u7a3f\u7ed3\u679c\uff0c\u4ece\u800c\u63a5\u53d7\u66f4\u591a\u8349\u7a3f\u8bcd\u5143\u3002\u4f8b\u5982\uff0cSpecDec \u4ec5\u8981\u6c42\u8349\u7a3f\u8bcd\u5143\u843d\u5728 top-<span class=\"katex-eq\" data-katex-display=\"false\">\\beta<\/span> \u5019\u9009\u4e2d\uff0c\u5e76\u4e0e top-1 \u7ed3\u679c\u4e4b\u95f4\u5b58\u5728\u53ef\u5bb9\u5fcd\u7684\u5206\u6570\u5dee\u8ddd\u3002<\/p>\n<p>\u4e3a\u964d\u4f4e\u8349\u7a3f\u751f\u6210\u5f00\u9500\uff0c\u8fd1\u671f\u5de5\u4f5c\u63a2\u7d22\u4e86\u591a\u79cd\u4f18\u5316\u7b56\u7565\u3002\u4e00\u7c7b\u5de5\u4f5c\u805a\u7126\u52a8\u6001\u8349\u7a3f\u6811\u6784\u9020\u3002BiLD \u4e0e Kangaroo \u57fa\u4e8e\u8349\u7a3f\u6a21\u578b\u7f6e\u4fe1\u5ea6\u5b9e\u73b0\u65e9\u505c\u673a\u5236\u4ee5\u63a7\u5236\u6811\u6df1\u5ea6\uff0c\u800c EAGLE-2 \u8fdb\u4e00\u6b65\u5229\u7528\u7f6e\u4fe1\u5ea6\u5206\u6570\u8fd1\u4f3c\u63a5\u53d7\u7387\u5e76\u76f8\u5e94\u8c03\u6574\u8349\u7a3f\u6811\u7ed3\u6784\u3002\u53e6\u4e00\u7c7b\u5de5\u4f5c\u9488\u5bf9\u8349\u7a3f\u8fc7\u7a0b\u672c\u8eab\u7684\u8ba1\u7b97\u74f6\u9888\u3002TriForce \u901a\u8fc7 KV-cache \u538b\u7f29\u52a0\u901f\u957f\u4e0a\u4e0b\u6587\u8349\u7a3f\uff0c\u800c Ouroboros \u901a\u8fc7\u9002\u914d lookahead decoding \u6280\u672f\u63d0\u5347\u6548\u7387\u3002\u9488\u5bf9\u8bcd\u8868\u76f8\u5173\u5f00\u9500\uff0cFR-Spec \u8bc6\u522b\u5927\u8bcd\u8868 LM \u5934\u7684\u8ba1\u7b97\u74f6\u9888\uff0c\u5e76\u5c06\u8349\u7a3f\u7a7a\u95f4\u9650\u5236\u4e3a\u9ad8\u9891\u8bcd\u5143\u5b50\u96c6\u4ee5\u4f7f\u8349\u7a3f\u6a21\u578b\u66f4\u5feb\u3002<\/p>\n<h4>\u7ed3\u8bba<\/h4>\n<p>\u672c\u6587\u63d0\u51fa FastMTP\uff0c\u4f5c\u4e3a\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u4e2d\u63a8\u6d4b\u89e3\u7801\u573a\u666f\u4e0b vanilla \u591a\u8bcd\u5143\u9884\u6d4b\u7684\u6539\u8fdb\u3002FastMTP \u5f15\u5165\u4e24\u9879\u5173\u952e\u5b9e\u73b0\u3002\u7b2c\u4e00\uff0c\u6211\u4eec\u901a\u8fc7\u81ea\u84b8\u998f\u8bad\u7ec3\u5355\u4e00\u5171\u4eab\u6743\u91cd\u7684 MTP \u5934\uff0c\u6d88\u9664\u5bf9\u591a\u4e2a\u72ec\u7acb MTP \u6a21\u5757\u7684\u9700\u6c42\uff0c\u540c\u65f6\u63d0\u5347\u591a\u6b65\u9884\u6d4b\u80fd\u529b\u3002\u7b2c\u4e8c\uff0c\u6211\u4eec\u91c7\u7528\u8bed\u8a00\u611f\u77e5\u8bcd\u8868\u538b\u7f29\u4ee5\u964d\u4f4e\u8349\u7a3f\u751f\u6210\u671f\u95f4\u7684\u8ba1\u7b97\u6210\u672c\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\u6240\u63d0\u65b9\u6cd5\u7684\u4f18\u8d8a\u6027\uff1a\u5728\u4e00\u7cfb\u5217\u573a\u666f\u4e2d\uff0c\u8349\u7a3f\u8bcd\u5143\u8d28\u91cf\u4e0e\u6574\u4f53\u8f93\u51fa\u52a0\u901f\u5747\u53ef\u5f97\u5230\u63d0\u5347\u3002<\/p>\n<h4>\u9644\u5f55 A\uff1a\u8bad\u7ec3\u6570\u636e\u96c6\u7ec6\u8282<\/h4>\n<p>\u6211\u4eec\u4ece\u591a\u79cd\u6570\u636e\u6e90\u6536\u96c6\u7528\u4e8e MTP \u5934\u5fae\u8c03\u7684\u6570\u636e\u96c6\u3002\u4e3a\u786e\u4fdd\u5206\u5e03\u5bf9\u9f50\uff0c\u6211\u4eec\u91c7\u7528\u81ea\u84b8\u998f\uff0c\u4f7f\u6240\u6709\u54cd\u5e94\u5747\u7531\u4e3b\u6a21\u578b\u81ea\u8eab\u751f\u6210\u3002\u5bf9\u4ece\u6e90\u6570\u636e\u96c6\u63d0\u53d6\u7684\u6bcf\u4e2a\u63d0\u793a <span class=\"katex-eq\" data-katex-display=\"false\">x_{n}<\/span>\uff0c\u6211\u4eec\u4f7f\u7528\u4e3b\u6a21\u578b\u751f\u6210\u5bf9\u5e94\u54cd\u5e94 <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{y}_{n}<\/span>\uff0c\u751f\u6210\u914d\u7f6e\u4e3a\uff1atemperature 0.6\u3001top-k 20\u3001top-p 0.95\u3001\u6700\u5927\u957f\u5ea6 4096 \u8bcd\u5143\u3002\u6574\u4e2a\u6570\u636e\u84b8\u998f\u8fc7\u7a0b\u4f7f\u7528 SGLang \u63a8\u7406\u6846\u67b6\u5b8c\u6210\u3002<\/p>\n<p>\u84b8\u998f\u540e\uff0c\u6211\u4eec\u5e94\u7528\u53bb\u91cd\u3001\u6570\u636e\u6e05\u6d17\u4e0e\u6df7\u5408\u7b56\u7565\u4ee5\u6574\u7406\u9ad8\u8d28\u91cf\u8bcd\u5143\uff1a<\/p>\n<ul>\n<li>\n<p>\u53bb\u91cd\uff1a\u6211\u4eec\u5728\u5355\u4e2a\u6570\u636e\u6e90\u5185\u90e8\u4ee5\u53ca\u6574\u4e2a\u6570\u636e\u96c6\u8303\u56f4\u5185\u6267\u884c\u5168\u5c40 MinHash \u53bb\u91cd\uff0c\u4ee5\u79fb\u9664\u8fd1\u91cd\u590d\u6837\u672c\u3002<\/p>\n<\/li>\n<li>\n<p>\u6570\u636e\u6e05\u6d17\uff1a\u6211\u4eec\u5f00\u53d1\u542f\u53d1\u5f0f\u89c4\u5219\u8fc7\u6ee4\u4f4e\u8d28\u91cf\u6837\u672c\u3002\u542f\u53d1\u5f0f\u793a\u4f8b\u5305\u62ec\uff1a\uff081\uff09\u63a8\u7406\u94fe\u4e0d\u5b8c\u6574\u6216\u88ab\u622a\u65ad\u7684\u6837\u672c\uff1b\uff082\uff09\u5305\u542b\u8fc7\u591a\u91cd\u590d\u5185\u5bb9\u7684\u6837\u672c\uff1b\uff083\uff09\u843d\u5728\u671f\u671b\u957f\u5ea6\u8303\u56f4\u4e4b\u5916\u7684\u6837\u672c\u3002<\/p>\n<\/li>\n<li>\n<p>\u6570\u636e\u6df7\u5408\uff1a\u6211\u4eec\u7684\u6700\u7ec8\u6570\u636e\u96c6\u5305\u542b\u56db\u4e2a\u4e3b\u8981\u7c7b\u522b\uff0c\u5e76\u7ecf\u4ed4\u7ec6\u5e73\u8861\u6bd4\u4f8b\uff1a\u7ea6 42% \u7684\u8bcd\u5143\u5bf9\u5e94\u901a\u7528\u77e5\u8bc6\u4e0e\u4efb\u52a1\uff0c18% \u5bf9\u5e94\u6570\u5b66\u4e0e\u63a8\u7406\u5185\u5bb9\uff0c13% \u5bf9\u5e94\u4ee3\u7801\uff0c27% \u5bf9\u5e94\u4e2d\u6587\u6587\u672c\uff0c\u6700\u7ec8\u5f97\u5230 389.4K \u6761\u9ad8\u8d28\u91cf\u8bad\u7ec3\u6837\u4f8b\u3002<\/p>\n<\/li>\n<\/ul>\n<h4>\u9644\u5f55 B\uff1a\u8bc4\u4f30\u57fa\u51c6\u7ec6\u8282<\/h4>\n<p>\u6211\u4eec\u5728\u4e03\u4e2a\u591a\u6837\u5316\u4efb\u52a1\u4e0a\u8bc4\u4f30 FastMTP\uff0c\u4ee5\u786e\u4fdd\u5168\u9762\u8986\u76d6\u771f\u5b9e\u4e16\u754c\u5e94\u7528\uff1a\u591a\u8f6e\u5bf9\u8bdd\u3001\u4ee3\u7801\u751f\u6210\u3001\u6570\u5b66\u63a8\u7406\u3001\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u3001\u95ee\u7b54\u3001\u6458\u8981\u4e0e\u4e2d\u6587\u77e5\u8bc6\u8bc4\u4f30\u3002\u6539\u7f16\u81ea Spec-Bench \u8bc4\u4f30\u6846\u67b6\uff0c\u6211\u4eec\u4ed4\u7ec6\u9009\u62e9\u4ee3\u8868\u4e0d\u540c\u751f\u6210\u6a21\u5f0f\u4e0e\u6311\u6218\u7684\u57fa\u51c6\u3002\u4e3a\u786e\u4fdd\u516c\u5e73\u8bc4\u4f30\uff0c\u6211\u4eec\u4ece\u6bcf\u4e2a\u4efb\u52a1\u968f\u673a\u91c7\u6837 80 \u4e2a\u5b9e\u4f8b\uff08C-Eval \u4e3a 102 \u4e2a\uff09\uff0c\u5e76\u907f\u514d\u53ef\u80fd\u5728 MTP \u8bad\u7ec3\u671f\u95f4\u66b4\u9732\u7684\u57fa\u51c6\u3002<\/p>\n<p>\u6211\u4eec\u7684\u57fa\u51c6\u5957\u4ef6\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p>MT-Bench\uff1a\u4e00\u7cfb\u5217\u5f00\u653e\u5f0f\u95ee\u9898\uff0c\u8bc4\u4f30\u804a\u5929\u673a\u5668\u4eba\u7684\u591a\u8f6e\u5bf9\u8bdd\u4e0e\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u3002MT-bench \u4e5f\u88ab\u4ed4\u7ec6\u6784\u9020\u4ee5\u57fa\u4e8e\u6838\u5fc3\u80fd\u529b\uff08\u5982\u63a8\u7406\u4e0e\u6570\u5b66\uff09\u533a\u5206\u804a\u5929\u673a\u5668\u4eba\u3002<\/p>\n<\/li>\n<li>\n<p>LiveCodeBench-v6\uff1a\u5bf9 LLM \u4ee3\u7801\u80fd\u529b\u7684\u5168\u9762\u4e14\u65e0\u6c61\u67d3\u8bc4\u4f30\uff0c\u968f\u65f6\u95f4\u4ece\u4e09\u4e2a\u7ade\u8d5b\u5e73\u53f0\uff08LeetCode\u3001AtCoder \u4e0e CodeForces\uff09\u6536\u96c6\u65b0\u95ee\u9898\u3002\u6211\u4eec\u9009\u62e9\u8be5\u6570\u636e\u96c6\u7684\u7b2c\u516d\u7248\uff0c\u5305\u542b 2023 \u5e74 5 \u6708\u81f3 2025 \u5e74 4 \u6708\u95f4\u53d1\u5e03\u7684 1055 \u9053\u9898\u76ee\u3002<\/p>\n<\/li>\n<li>\n<p>MATH-500\uff1a\u5305\u542b 500 \u9053\u6311\u6218\u6027\u6570\u5b66\u95ee\u9898\u7684\u6570\u636e\u96c6\uff0c\u7528\u4e8e\u6d4b\u8bd5\u9ad8\u7ea7\u6570\u5b66\u63a8\u7406\u4e0e\u89e3\u9898\u6280\u80fd\u3002\u5b83\u5305\u62ec\u4ee3\u6570\u3001\u5fae\u79ef\u5206\u3001\u51e0\u4f55\u4e0e\u6570\u8bba\u7b49\u9886\u57df\u95ee\u9898\uff0c\u4e3b\u8981\u9762\u5411\u9ad8\u4e2d\u4e0e\u672c\u79d1\u65e9\u671f\u6c34\u5e73\u3002<\/p>\n<\/li>\n<li>\n<p>Natural Questions\uff1a\u5927\u89c4\u6a21\u95ee\u7b54\u6570\u636e\u96c6\uff0c\u5305\u542b\u771f\u5b9e\u7528\u6237\u67e5\u8be2\u4ee5\u53ca\u6765\u81ea Wikipedia \u6587\u6863\u7684\u9ad8\u8d28\u91cf\u6807\u6ce8\u3002\u6211\u4eec\u5728\u4e24\u4e2a\u5b50\u4efb\u52a1\u4e2d\u4f7f\u7528\u8be5\u57fa\u51c6\uff1a\u95ee\u7b54\u4e0e\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u3002<\/p>\n<\/li>\n<li>\n<p>CNN\/Daily Mail\uff1a\u82f1\u8bed\u6570\u636e\u96c6\uff0c\u5305\u542b CNN \u4e0e Daily Mail \u8bb0\u8005\u64b0\u5199\u7684\u903e 30 \u4e07\u7bc7\u72ec\u7279\u65b0\u95fb\u6587\u7ae0\uff0c\u6bcf\u7bc7\u914d\u6709 3\u20134 \u6761\u603b\u7ed3\u6587\u7ae0\u5185\u5bb9\u7684\u8981\u70b9\u3002\u8be5\u57fa\u51c6\u8bc4\u4f30\u6a21\u578b\u5728\u4fdd\u7559\u5173\u952e\u4fe1\u606f\u7684\u540c\u65f6\u5bf9\u65b0\u95fb\u6587\u7ae0\u751f\u6210\u62bd\u53d6\u5f0f\u4e0e\u62bd\u8c61\u5f0f\u6458\u8981\u7684\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p>C-Eval\uff1a\u5168\u9762\u7684\u4e2d\u6587\u8bc4\u4f30\u5957\u4ef6\uff0c\u7528\u4e8e\u8bc4\u4f30\u57fa\u7840\u6a21\u578b\u5728\u4e2d\u6587\u8bed\u5883\u4e0b\u7684\u9ad8\u7ea7\u77e5\u8bc6\u4e0e\u63a8\u7406\u80fd\u529b\u3002C-Eval \u5305\u542b\u56db\u4e2a\u96be\u5ea6\u7ea7\u522b\u7684\u9009\u62e9\u9898\uff1a\u521d\u4e2d\u3001\u9ad8\u4e2d\u3001\u5927\u5b66\u4e0e\u4e13\u4e1a\u3002\u95ee\u9898\u8986\u76d6 52 \u4e2a\u591a\u6837\u5316\u5b66\u79d1\uff0c\u4ece\u4eba\u6587\u5230\u79d1\u5b66\u4e0e\u5de5\u7a0b\u3002<\/p>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Meta Data \u53d1\u8868\u65f6\u95f4\uff1a2025-09-16 \u4f5c\u8005\uff1aYuxuan Cai, Xiaozhuan Liang, Xinghua &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-1647","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1647","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1647"}],"version-history":[{"count":2,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1647\/revisions"}],"predecessor-version":[{"id":1649,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1647\/revisions\/1649"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1647"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1647"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1647"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}