{"id":1608,"date":"2026-06-10T11:56:08","date_gmt":"2026-06-10T03:56:08","guid":{"rendered":"https:\/\/lingbo.online\/?p=1608"},"modified":"2026-06-10T16:14:22","modified_gmt":"2026-06-10T08:14:22","slug":"1608","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/1608\/","title":{"rendered":"LongSpec: Long-Context Lossless Speculative Decoding with Efficient Drafting and Verification"},"content":{"rendered":"<h3>Meta Data<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4\uff1a2025-02-24\uff1b\u6700\u65b0 arXiv \u4fee\u8ba2\uff1a2026-04-08<\/li>\n<li>\u4f5c\u8005\uff1aPenghui Yang, Cunxiao Du, Fengzhuo Zhang, Haonan Wang, Tianyu Pang, Chao Du, Bo An<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2502.17421.pdf\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/pdf\/2502.17421.pdf<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/sail-sg\/LongSpec\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/sail-sg\/LongSpec<\/a><\/li>\n<\/ul>\n<h3>LongSpec\uff1a\u9762\u5411\u957f\u4e0a\u4e0b\u6587\u7684\u65e0\u635f\u63a8\u6d4b\u89e3\u7801\uff0c\u901a\u8fc7\u9ad8\u6548\u8349\u7a3f\u751f\u6210\u4e0e\u9a8c\u8bc1\u5b9e\u73b0\u52a0\u901f<\/h3>\n<h4>\u6458\u8981<\/h4>\n<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Models, LLMs\uff09\u73b0\u5728\u80fd\u591f\u5904\u7406\u6781\u957f\u4e0a\u4e0b\u6587\uff0c\u5728\u8fd9\u4e9b\u6269\u5c55\u8f93\u5165\u4e0a\u7684\u9ad8\u6548\u63a8\u7406\u53d8\u5f97\u8d8a\u6765\u8d8a\u91cd\u8981\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e LLM agents \u7b49\u9ad8\u5ea6\u4f9d\u8d56\u8fd9\u4e00\u80fd\u529b\u7684\u65b0\u5174\u5e94\u7528\u3002\u4e0e\u91cf\u5316\u548c\u6a21\u578b\u7ea7\u8054\u7b49\u6709\u635f\u66ff\u4ee3\u65b9\u6848\u76f8\u6bd4\uff0c\u63a8\u6d4b\u89e3\u7801\uff08Speculative Decoding, SD\uff09\u63d0\u4f9b\u4e86\u4e00\u79cd\u5f88\u6709\u524d\u666f\u7684\u65e0\u635f\u52a0\u901f\u6280\u672f\u3002\u7136\u800c\uff0c\u5927\u591a\u6570\u6700\u5148\u8fdb\u7684 SD \u65b9\u6cd5\u662f\u5728\u77ed\u6587\u672c\u4e0a\u8bad\u7ec3\u7684\uff08\u901a\u5e38\u5c11\u4e8e 4k tokens\uff09\uff0c\u8fd9\u4f7f\u5b83\u4eec\u4e0d\u9002\u5408\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5c06\u8fd9\u4e9b\u65b9\u6cd5\u9002\u914d\u5230\u957f\u4e0a\u4e0b\u6587\u4f1a\u9762\u4e34\u4e09\u4e2a\u5173\u952e\u6311\u6218\uff1a\uff081\uff09\u7531\u4e8e\u5927\u578b Key-Value\uff08KV\uff09\u7f13\u5b58\uff0c\u8349\u7a3f\u6a21\u578b\u4f1a\u5e26\u6765\u8fc7\u9ad8\u7684\u5185\u5b58\u9700\u6c42\uff1b\uff082\uff09\u77ed\u4e0a\u4e0b\u6587\u8bad\u7ec3\u4e0e\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u4e4b\u95f4\u7684\u4e0d\u5339\u914d\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\uff1b\uff083\uff09\u5728\u7ba1\u7406\u957f token \u5e8f\u5217\u65f6\uff0c\u6811\u6ce8\u610f\u529b\u673a\u5236\u6548\u7387\u4f4e\u4e0b\u3002\u672c\u6587\u63d0\u51fa\u4e86 <strong>LongSpec<\/strong>\uff0c\u8fd9\u662f\u4e00\u4e2a\u901a\u8fc7\u4e09\u9879\u6838\u5fc3\u521b\u65b0\u6765\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\u7684\u6846\u67b6\uff1a\u5177\u6709\u5e38\u6570\u5927\u5c0f KV \u7f13\u5b58\u7684\u5185\u5b58\u9ad8\u6548\u8349\u7a3f\u6a21\u578b\uff1b\u7f13\u89e3\u8bad\u7ec3-\u63a8\u7406\u4e0d\u5339\u914d\u7684\u65b0\u578b\u4f4d\u7f6e\u7d22\u5f15\uff1b\u4ee5\u53ca\u4e00\u79cd\u5c06\u5feb\u901f\u524d\u7f00\u8ba1\u7b97\u4e0e\u6807\u51c6\u6811\u6ce8\u610f\u529b\u7ed3\u5408\u8d77\u6765\u3001\u4ece\u800c\u5b9e\u73b0\u9ad8\u6548\u89e3\u7801\u7684\u6ce8\u610f\u529b\u805a\u5408\u7b56\u7565\u3002\u5b9e\u9a8c\u7ed3\u679c\u8bc1\u5b9e\u4e86 LongSpec \u7684\u6709\u6548\u6027\uff1a\u5728\u4e94\u4e2a\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u6570\u636e\u96c6\u4e0a\uff0c\u76f8\u6bd4\u5f3a\u5927\u7684 Flash Attention \u57fa\u7ebf\uff0c\u6700\u9ad8\u5b9e\u73b0 3.26<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\uff1b\u540c\u65f6\uff0c\u5728\u4f7f\u7528 QwQ \u6a21\u578b\u7684\u56db\u4e2a\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e0a\uff0c\u5899\u949f\u65f6\u95f4\u51cf\u5c11 2.34<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\uff0c\u5c55\u793a\u51fa\u9762\u5411\u957f\u4e0a\u4e0b\u6587\u5e94\u7528\u7684\u663e\u8457\u5ef6\u8fdf\u6539\u8fdb\u3002<\/p>\n<h3>\u5f15\u8a00<\/h3>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610035935554.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 1\uff1a\u5f53\u524d\u6700\u5148\u8fdb\u7684 SD \u65b9\u6cd5 EAGLE \u7684\u8bad\u7ec3\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a 2048\uff0c\u663e\u8457\u77ed\u4e8e\u73b0\u4ee3 LLM \u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/center><\/p>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5df2\u7ecf\u5c55\u793a\u51fa\u5353\u8d8a\u80fd\u529b\uff0c\u800c\u5b83\u4eec\u5904\u7406\u6269\u5c55\u4e0a\u4e0b\u6587\u7684\u80fd\u529b\u6b63\u5728\u6210\u4e3a LLM agents \u548c\u957f\u63a8\u7406\u4efb\u52a1\u7b49\u65b0\u5174\u5e94\u7528\u7684\u5173\u952e\u80fd\u529b\uff0c\u8fd9\u4e9b\u5e94\u7528\u73b0\u5728\u8fd0\u884c\u5728\u53ef\u6269\u5c55\u5230\u6570\u767e\u4e07 tokens \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e4b\u4e0a\u3002\u5728\u8fd9\u4e9b\u8981\u6c42\u5f88\u9ad8\u7684\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\uff0c\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u7684\u9ad8\u63a8\u7406\u5ef6\u8fdf\u4f1a\u6210\u4e3a\u660e\u663e\u74f6\u9888\u3002\u867d\u7136\u91cf\u5316\u3001\u7a00\u758f\u6ce8\u610f\u529b\u548c\u6a21\u578b\u7ea7\u8054\u7b49\u591a\u79cd\u52a0\u901f\u6280\u672f\u5df2\u7ecf\u88ab\u63d0\u51fa\u7528\u4e8e\u7f13\u89e3\u8fd9\u4e00\u95ee\u9898\uff0c\u4f46\u5b83\u4eec\u5f80\u5f80\u4f1a\u727a\u7272\u8f93\u51fa\u8d28\u91cf\uff0c\u56e0\u6b64\u5c5e\u4e8e\u6709\u635f\u65b9\u6848\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u63a8\u6d4b\u89e3\u7801\uff08SD\uff09\u901a\u8fc7\u4f7f\u7528\u4e00\u4e2a\u66f4\u5c0f\u7684\u8349\u7a3f\u6a21\u578b\u6765\u63d0\u51fa token \u5e8f\u5217\uff0c\u518d\u7531\u66f4\u5927\u7684\u76ee\u6807\u6a21\u578b\u5e76\u884c\u9a8c\u8bc1\u8fd9\u4e9b\u5e8f\u5217\uff0c\u4ece\u800c\u63d0\u4f9b\u4e00\u79cd<strong>\u65e0\u635f<\/strong>\u52a0\u901f\u7b56\u7565\u3002<\/p>\n<p>\u7136\u800c\uff0c\u6700\u5148\u8fdb\u7684 SD \u65b9\u6cd5\uff08\u4f8b\u5982 EAGLE\uff09\u901a\u5e38\u4f9d\u8d56\u4e00\u4e2a\u5c0f\u578b\u4e14\u72ec\u7acb\u7684\u8349\u7a3f\u6a21\u578b\uff0c\u4e3b\u8981\u9762\u5411\u77ed\u4e0a\u4e0b\u6587\u6570\u636e\u8fdb\u884c\u8bbe\u8ba1\u548c\u8bc4\u4f30\uff0c\u5178\u578b\u5e8f\u5217\u957f\u5ea6\u5c11\u4e8e 4k tokens\u3002\u867d\u7136\u4e00\u4e9b\u73b0\u6709 SD \u65b9\u6cd5\u53ef\u4ee5\u6269\u5c55\u5230\u66f4\u957f\u4e0a\u4e0b\u6587\uff0c\u4f46\u5b83\u4eec\u901a\u5e38\u4f7f\u7528\u5b8c\u6574\u76ee\u6807\u6a21\u578b\u5e76\u914d\u5408\u538b\u7f29\u7684 Key-Value\uff08KV\uff09\u7f13\u5b58\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\u3002\u8fd9\u4e9b\u65b9\u6cd5\u907f\u514d\u4e86\u8bad\u7ec3\u4e13\u7528\u8349\u7a3f\u6a21\u578b\u7684\u5f00\u9500\uff0c\u4f46\u5b83\u4eec\u4f9d\u8d56\u5b8c\u6574\u76ee\u6807\u6a21\u578b\uff0c\u800c\u5b8c\u6574\u76ee\u6807\u6a21\u578b\u5e76\u4e0d\u591f\u8f7b\u91cf\uff0c\u8fd9\u9650\u5236\u4e86\u8349\u7a3f\u751f\u6210\u901f\u5ea6\u3002\u56e0\u6b64\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u80fd\u4e0d\u5982\u6700\u5148\u8fdb\u7684\u77ed\u4e0a\u4e0b\u6587 SD \u6280\u672f\u8868\u73b0\u597d\u3002<\/p>\n<p>\u8fd9\u79cd\u5dee\u5f02\u63d0\u51fa\u4e86\u4e00\u4e2a\u5173\u952e\u95ee\u9898\uff1a<\/p>\n<p><center><i>\u4e3a\u4ec0\u4e48\u9762\u5411\u77ed\u4e0a\u4e0b\u6587\u7684\u6700\u5148\u8fdb SD \u65b9\u6cd5\u4e0d\u80fd\u76f4\u63a5\u5e94\u7528\u5230\u957f\u5e8f\u5217\uff1f<\/i><\/center><\/p>\n<p>\u5bf9\u4e8e\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5c06\u6709\u6548\u77ed\u4e0a\u4e0b\u6587 SoTA SD \u6280\u672f\u96be\u4ee5\u76f4\u63a5\u9002\u914d\u5230\u957f\u4e0a\u4e0b\u6587\u8bbe\u7f6e\u5f52\u56e0\u4e8e\u4e09\u4e2a\u65b0\u51fa\u73b0\u7684\u6311\u6218\uff1a<\/p>\n<ol>\n<li><strong>\u67b6\u6784\uff1a<\/strong> \u5728 SoTA SD \u65b9\u6cd5\u4e2d\uff08\u4f8b\u5982 EAGLE\uff09\uff0c\u8349\u7a3f\u6a21\u578b\u7684 KV cache \u4ecd\u7136\u4f1a\u968f\u4e0a\u4e0b\u6587\u957f\u5ea6\u7ebf\u6027\u589e\u957f\u3002\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u589e\u52a0\uff0c\u8fd9\u79cd\u7ebf\u6027\u589e\u957f\u4f1a\u6210\u4e3a\u96be\u4ee5\u627f\u53d7\u7684\u5185\u5b58\u74f6\u9888\u3002<\/li>\n<li><strong>\u8bad\u7ec3\uff1a<\/strong> \u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u901a\u5e38\u4f9d\u8d56\u5927\u91cf\u77ed\u5e8f\u5217\u6570\u636e\uff0c\u800c\u957f\u5e8f\u5217\u6570\u636e\u76f8\u5bf9\u7a00\u7f3a\u3002\u8bad\u7ec3\u6570\u636e\u7684\u4e0d\u5e73\u8861\u4f7f\u6a21\u578b\u96be\u4ee5\u6cdb\u5316\u5230\u66f4\u957f\u4e0a\u4e0b\u6587\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u70b9\uff0c\u8bad\u7ec3\u957f\u4e0a\u4e0b\u6587 LLM \u7684\u4f20\u7edf\u7ecf\u9a8c\u901a\u5e38\u91c7\u7528\u957f\u5ea6\u5916\u63a8\uff0c\u5c24\u5176\u662f\u901a\u8fc7\u6269\u5c55 Rotary Position Embedding\uff08RoPE\uff09\u7684 base \u6765\u5bb9\u7eb3\u66f4\u957f\u4e0a\u4e0b\u6587\u3002\u7136\u800c\uff0c\u8fd9\u4e00\u65b9\u6848\u4e0d\u80fd\u76f4\u63a5\u5e94\u7528\u5230 SoTA SD \u8349\u7a3f\u6a21\u578b\uff0c\u56e0\u4e3a\u5b83\u4eec\u7684 RoPE base \u5fc5\u987b\u4e0e\u76ee\u6807\u6a21\u578b\u5339\u914d\uff1b\u800c\u76ee\u6807\u6a21\u578b\u7684 RoPE base \u662f\u56fa\u5b9a\u7684\uff0c\u5e76\u4e14\u5df2\u7ecf\u4e3a\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u8fdb\u884c\u4e86\u7f29\u653e\u3002SoTA SD \u6280\u672f\u901a\u5e38\u8981\u6c42\u8349\u7a3f\u6a21\u578b\u4f7f\u7528\u6765\u81ea\u76ee\u6807\u6a21\u578b\u7684\u4e2d\u95f4\u7279\u5f81\uff08\u4f8b\u5982 hidden states \u6216 KV cache\uff09\uff0c\u8fd9\u4e9b\u4fe1\u606f\u5bf9\u4e8e\u8ba9\u8349\u7a3f\u6a21\u578b\u66f4\u597d\u5730\u5bf9\u9f50\u5e76\u9884\u6d4b\u76ee\u6807\u6a21\u578b\u8f93\u51fa\u975e\u5e38\u5173\u952e\u3002<\/li>\n<li><strong>\u63a8\u7406\uff1a<\/strong> \u6811\u6ce8\u610f\u529b\u9a8c\u8bc1\u5728\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\u7684\u6709\u6548\u6027\u4f1a\u4e0b\u964d\u3002\u5c24\u5176\u662f\uff0c\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\u7684\u5e38\u89c1\u63a8\u7406\u4f18\u5316\u4e3b\u8981\u9762\u5411\u89c4\u5219\u3001\u7ed3\u6784\u5316\u7684 attention mask \u8bbe\u8ba1\uff0c\u5e76\u6ca1\u6709\u9488\u5bf9\u4efb\u610f\u6216\u975e\u7ed3\u6784\u5316 attention mask \u4f18\u5316\u3002\u56e0\u6b64\uff0c\u63a8\u6d4b\u53ef\u80fd\u5e26\u6765\u7684\u6f5c\u5728\u52a0\u901f\u4f4e\u4e8e\u9884\u671f\u3002<\/li>\n<\/ol>\n<p>\u4e3a\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u6211\u4eec\u63d0\u51fa LongSpec\uff0c\u4e00\u4e2a\u7528\u4e8e\u9ad8\u6548\u957f\u4e0a\u4e0b\u6587\u65e0\u635f\u63a8\u6d4b\u89e3\u7801\u7684\u7efc\u5408\u6846\u67b6\u3002 LongSpec \u901a\u8fc7\u4e09\u9879\u5173\u952e\u521b\u65b0\u514b\u670d\u4e0a\u8ff0\u969c\u788d\uff1a<\/p>\n<ol>\n<li><strong>\u5185\u5b58\u9ad8\u6548\u67b6\u6784\u3002<\/strong> \u6211\u4eec\u63d0\u51fa\u4e00\u79cd\u65e0\u8bba\u4e0a\u4e0b\u6587\u957f\u5ea6\u5982\u4f55\u90fd\u4fdd\u6301\u5e38\u6570\u5185\u5b58\u4f7f\u7528\u91cf\u7684\u8349\u7a3f\u6a21\u578b\u67b6\u6784\uff0c\u4ece\u800c\u6709\u6548\u89e3\u51b3\u5df2\u6709 SoTA \u81ea\u56de\u5f52\u8349\u7a3f\u6a21\u578b\u7684\u53ef\u6269\u5c55\u6027\u9650\u5236\u3002<\/li>\n<li><strong>\u6709\u6548\u8bad\u7ec3\u673a\u5236\u3002<\/strong> \u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u79cd\u6d89\u53ca Anchor-Offset Indices \u7684\u65b0\u578b\u8bad\u7ec3\u7b56\u7565\uff0c\u4f7f\u5728\u77ed\u5e8f\u5217\u4e0a\u8bad\u7ec3\u7684\u8349\u7a3f\u6a21\u578b\u80fd\u591f\u5728\u63a8\u7406\u65f6\u7a33\u5065\u5730\u6cdb\u5316\u5230\u957f\u5f97\u591a\u7684\u4e0a\u4e0b\u6587\u3002<\/li>\n<li><strong>\u5feb\u901f\u6811\u6ce8\u610f\u529b\u3002<\/strong> \u6211\u4eec\u63d0\u51fa Hybrid Tree Attention\uff0c\u8fd9\u662f\u4e00\u79cd\u65b0\u7684\u8ba1\u7b97\u65b9\u6cd5\uff0c\u901a\u8fc7\u5206\u89e3\u6ce8\u610f\u529b\u8ba1\u7b97\u5e76\u5229\u7528\u4f18\u5316\u7684 Triton kernels\uff0c\u663e\u8457\u52a0\u901f\u6811\u9a8c\u8bc1\u3002<\/li>\n<\/ol>\n<p>\u5728\u4f7f\u7528\u4e94\u4e2a LLM \u4f5c\u4e3a\u76ee\u6807\u6a21\u578b\u3001\u8986\u76d6\u4e94\u4e2a\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u6570\u636e\u96c6\u7684\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u7684 LongSpec \u80fd\u591f\u663e\u8457\u964d\u4f4e\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u5ef6\u8fdf\uff1a\u76f8\u6bd4\u4f7f\u7528 <code>Flash Attention<\/code> \u7684\u5f3a\u57fa\u7ebf\u6700\u9ad8\u5b9e\u73b0 3.26<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\uff1b\u76f8\u6bd4\u4f7f\u7528 HuggingFace \u5b9e\u73b0\u7684\u5e38\u89c1\u57fa\u7ebf\u6700\u9ad8\u5b9e\u73b0 7<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\u3002\u5728\u4f7f\u7528\u957f\u63a8\u7406\u6a21\u578b QwQ \u7684\u56db\u4e2a\u6570\u5b66\u63a8\u7406\u6570\u636e\u96c6\u4e0a\u7684\u989d\u5916\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86 LongSpec \u7684\u6709\u6548\u6027\uff0c\u5899\u949f\u65f6\u95f4\u52a0\u901f\u8fbe\u5230 2.34<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\u3002\u6b64\u5916\uff0c\u6211\u4eec\u63d0\u51fa\u7684 Anchor-Offset Indices \u4f7f\u6a21\u578b\u8fbe\u5230\u76f8\u540c loss \u6c34\u5e73\u7684\u901f\u5ea6\u63d0\u5347 3.93<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\uff0c\u800c Hybrid Tree Attention \u76f8\u6bd4\u6807\u51c6 HuggingFace \u5b9e\u73b0\u5c06\u6ce8\u610f\u529b\u8ba1\u7b97\u5ef6\u8fdf\u964d\u4f4e\u7ea6 75%\u3002<\/p>\n<h3>\u76f8\u5173\u5de5\u4f5c<\/h3>\n<p>\u63a8\u6d4b\u89e3\u7801\u63d0\u4f9b\u4e86\u4e00\u79cd\u5728\u4e0d\u635f\u5bb3 LLM \u8f93\u51fa\u8d28\u91cf\u7684\u60c5\u51b5\u4e0b\u52a0\u901f LLM \u7684\u6709\u524d\u666f\u65b9\u6cd5\u3002\u672c\u6587\u5173\u6ce8\u539f\u59cb\u7684\u65e0\u635f\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\uff1b\u4e00\u4e9b\u8fd1\u671f\u5de5\u4f5c\u63a2\u7d22\u4e86\u6709\u635f\u63a8\u6d4b\u89e3\u7801\uff0c\u9644\u5f55\u4e2d\u7ed9\u51fa\u7b80\u8981\u6982\u89c8\u3002\u65e9\u671f\u5de5\u4f5c\u4f9d\u8d56\u73b0\u6709\u5c0f\u578b LLM \u6765\u751f\u6210\u8349\u7a3f\u5e8f\u5217\u3002\u53e6\u4e00\u4e9b\u65b9\u6cd5\u5219\u8bd5\u56fe\u6539\u8fdb\u8fd9\u4e9b\u65e9\u671f\u65b9\u6848\u3002\u4e5f\u6709\u4e00\u4e9b\u5de5\u4f5c\u4f7f\u7528\u76ee\u6807\u6a21\u578b\u7684\u4e00\u90e8\u5206\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\u3002\u57fa\u4e8e\u68c0\u7d22\u7684\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u5219\u901a\u8fc7\u4f7f\u7528 <span class=\"katex-eq\" data-katex-display=\"false\">N<\/span>-gram \u5339\u914d\u800c\u4e0d\u662f\u4f9d\u8d56\u5c0f\u6a21\u578b\u6765\u63d0\u4f9b\u53e6\u4e00\u79cd\u66ff\u4ee3\u65b9\u6848\u3002\u8fd9\u4e9b\u65b9\u6cd5\u7ed5\u5f00\u4e86\u989d\u5916\u6a21\u578b\u8bad\u7ec3\u9700\u6c42\uff0c\u5229\u7528\u5df2\u6709\u6570\u636e\u6a21\u5f0f\u9ad8\u6548\u6784\u9020\u8349\u7a3f\u5e8f\u5217\u3002<\/p>\n<p>\u66f4\u8fd1\u671f\u7684\u8fdb\u5c55\u5728\u8fd9\u4e9b\u57fa\u7840\u4e0a\u8fdb\u884c\u4e86\u6269\u5c55\uff1a\u5b83\u4eec\u8bbe\u8ba1\u4e13\u7528\u8349\u7a3f\u6a21\u578b\uff0c\u5e76\u5f15\u5165\u6811\u63a8\u6d4b\u4e0e\u9a8c\u8bc1\u6280\u672f\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5229\u7528\u4e3a\u63a8\u6d4b\u89e3\u7801\u5b9a\u5236\u7684\u8349\u7a3f\u6a21\u578b\uff0c\u5b9e\u73b0\u66f4\u9ad8\u6548\u7387\u548c\u6027\u80fd\u3002\u6b64\u5916\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u91c7\u7528\u7684\u57fa\u4e8e\u6811\u7684\u65b9\u6848\u5141\u8bb8\u66f4\u81ea\u9002\u5e94\u3001\u66f4\u6613\u5e76\u884c\u7684\u89e3\u7801\u8fc7\u7a0b\uff0c\u4e3a\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\u7b49\u771f\u5b9e\u7cfb\u7edf\u4e2d\u7684\u66f4\u5e7f\u6cdb\u5e94\u7528\u94fa\u5e73\u9053\u8def\u3002<\/p>\n<p>\u5c3d\u7ba1\u63a8\u6d4b\u89e3\u7801\u5728\u5e38\u89c4\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0a\u5df2\u7ecf\u53d6\u5f97\u663e\u8457\u8fdb\u5c55\uff0c\u4f46\u53ea\u6709\u5c11\u6570\u73b0\u6709\u8bba\u6587\u5173\u6ce8\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\u7684\u65e0\u635f\u63a8\u6d4b\u89e3\u7801\u3002TriForce \u5f15\u5165\u4e86\u4e00\u4e2a\u53ef\u6269\u5c55\u5230\u957f\u5e8f\u5217\u751f\u6210\u7684\u4e09\u5c42\u63a8\u6d4b\u89e3\u7801\u7cfb\u7edf\u3002MagicDec \u4f7f\u7528\u63a8\u6d4b\u89e3\u7801\u540c\u65f6\u63d0\u5347 LLM \u63a8\u7406\u541e\u5410\u548c\u5ef6\u8fdf\u3002QuantSpec \u5bf9\u8349\u7a3f\u6a21\u578b\u91c7\u7528\u5206\u5c42 4-bit \u91cf\u5316 KV cache \u548c 4-bit \u91cf\u5316\u6743\u91cd\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u4e3b\u8981\u4f7f\u7528\u5e26\u7a00\u758f KV cache \u7684\u76ee\u6807\u6a21\u578b\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\u3002\u8ba1\u7b97\u5bc6\u96c6\u578b\u8349\u7a3f\u6a21\u578b\u9650\u5236\u4e86\u8fd9\u4e9b\u65b9\u6cd5\u5728\u4e0d\u540c batch size \u4e0b\u7684\u5b9e\u9645\u4f7f\u7528\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u6211\u4eec\u7684\u5de5\u4f5c\u5173\u6ce8\u5982\u4f55\u9ad8\u6548\u6784\u5efa\u4e00\u4e2a\u4ec5\u5305\u542b\u4e00\u4e2a transformer block \u7684\u8349\u7a3f\u6a21\u578b\uff0c\u4ece\u800c\u5728\u4e0d\u540c\u573a\u666f\u4e2d\u5b9e\u73b0\u66f4\u6709\u6548\u7684\u6027\u80fd\u3002<\/p>\n<h3>\u65b9\u6cd5<\/h3>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u7528\u4e8e\u957f\u4e0a\u4e0b\u6587\u63a8\u6d4b\u89e3\u7801\u7684 LongSpec \u6846\u67b6\u3002\u5b83\u901a\u8fc7\u4ee5\u4e0b\u4e09\u70b9\u89e3\u51b3\u4e09\u4e2a\u5173\u952e\u6311\u6218\uff1a\uff081\uff09\u8bbe\u8ba1\u4e00\u4e2a\u5177\u6709\u5e38\u6570\u5927\u5c0f\u5185\u5b58\u5f00\u9500\u7684\u8f7b\u91cf\u8349\u7a3f\u6a21\u578b\u67b6\u6784\uff1b\uff082\uff09\u8bbe\u8ba1\u5e26 anchor-offset indices \u7684\u8bad\u7ec3\u7b56\u7565\uff0c\u4ee5\u6709\u6548\u5904\u7406\u957f\u4e0a\u4e0b\u6587\uff1b\uff083\uff09\u5b9e\u73b0\u4e00\u79cd\u5feb\u901f\u6ce8\u610f\u529b\u805a\u5408\u673a\u5236\uff0c\u5229\u7528\u57fa\u4e8e\u6811\u7684\u63a8\u6d4b\u4e0e\u9a8c\u8bc1\u6765\u652f\u6301\u5b9e\u9645\u4f7f\u7528\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257002.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 2\uff1a\u5185\u5b58\u9ad8\u6548\u8349\u7a3f\u6a21\u578b\u3001Anchor-Offset Indices \u4e0e Hybrid Tree Attention \u7684\u793a\u610f\u56fe\u3002\uff08a\uff09\u6211\u4eec\u4f7f\u7528\u6ed1\u52a8\u7a97\u53e3\u81ea\u6ce8\u610f\u529b\u5c42\u6355\u83b7\u5c40\u90e8\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u5e76\u4f7f\u7528\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u805a\u5408\u957f\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\uff08b\uff09vanilla indexing \u4e0e Anchor-Offset Indices \u7684\u5dee\u5f02\u3002\u901a\u8fc7\u5f15\u5165\u968f\u673a\u9009\u62e9\u7684 offset \u548c\u82e5\u5e72 anchor indices\uff0cAnchor-Offset Indices \u4f7f\u77ed\u4e0a\u4e0b\u6587\u8bad\u7ec3\u9636\u6bb5\u80fd\u591f\u4e0e\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u9636\u6bb5\u65e0\u7f1d\u7ed3\u5408\u3002\uff08c\uff09Hybrid Tree Attention \u7ed3\u5408\u4e86 <code>Flash Attention<\/code> \u4e0e\u6211\u4eec\u57fa\u4e8e Triton \u5b9e\u73b0\u7684 Attention \u7684\u4f18\u52bf\u3002<\/center><\/p>\n<h4>\u5185\u5b58\u9ad8\u6548\u67b6\u6784<\/h4>\n<p>\u5728\u5148\u524d\u5de5\u4f5c\u4e2d\uff0cSoTA \u6a21\u578b EAGLE \u7684\u6210\u529f\u4f9d\u8d56\u4e24\u4e2a\u5173\u952e\u56e0\u7d20\uff1a\uff081\uff09\u76ee\u6807\u6a21\u578b\u63d0\u4f9b\u7684 hidden states\uff1b\uff082\uff09\u5b83\u7684\u81ea\u56de\u5f52\u7ed3\u6784\u3002\u7136\u800c\uff0c\u81ea\u56de\u5f52\u8349\u7a3f\u6a21\u578b\u4e0d\u53ef\u907f\u514d\u5730\u9700\u8981\u7ef4\u62a4\u81ea\u5df1\u7684 KV cache\uff0c\u8fd9\u4f1a\u5728\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u671f\u95f4\u5f15\u5165\u989d\u5916\u5f00\u9500\uff0c\u5e76\u9700\u8981\u5927\u91cf GPU \u5185\u5b58\uff0c\u5c24\u5176\u662f\u5728 LLM agents \u548c\u957f\u63a8\u7406\u7b49\u4f1a\u4ea7\u751f\u5927\u91cf\u8f93\u51fa\u7684\u4efb\u52a1\u4e2d\u3002<\/p>\n<p>\u4e3a\u907f\u514d\u8fd9\u79cd\u989d\u5916\u5185\u5b58\u5f00\u9500\uff0c\u6211\u4eec\u63d0\u51fa\u4e00\u79cd\u5185\u5b58\u4f7f\u7528\u91cf\u4e0e\u4e0a\u4e0b\u6587\u957f\u5ea6\u65e0\u5173\u3001\u4fdd\u6301\u5e38\u6570\u7684\u8349\u7a3f\u6a21\u578b\u3002\u5982\u56fe 2\uff08a\uff09\u6240\u793a\uff0c\u6211\u4eec\u7684\u6a21\u578b\u7531\u4e24\u4e2a\u7ec4\u4ef6\u6784\u6210\uff1aself-attention \u6a21\u5757\uff0c\u4ee5\u53ca\u5176\u540e\u7684 cross-attention \u6a21\u5757\u3002self-attention \u6a21\u5757\u4e13\u6ce8\u4e8e\u5efa\u6a21\u5c40\u90e8\u4e0a\u4e0b\u6587\uff0c\u800c cross-attention \u6a21\u5757\u6355\u83b7\u957f\u7a0b\u4f9d\u8d56\u3002\u4e3a\u9650\u5236\u5185\u5b58\u4f7f\u7528\uff0c\u6211\u4eec\u5bf9 self-attention \u6a21\u5757\u5e94\u7528\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u673a\u5236\uff0c\u8fd9\u662f\u73b0\u4ee3 LLM \u4e2d\u5e7f\u6cdb\u91c7\u7528\u7684\u6280\u672f\u3002\u56e0\u6b64\uff0c\u5728\u63a8\u7406\u671f\u95f4\uff0cself-attention \u4e0d\u4f1a\u8d85\u8fc7\u7a97\u53e3\u5927\u5c0f\uff1b\u672c\u6587\u5c06\u7a97\u53e3\u5927\u5c0f\u8bbe\u4e3a 512\u3002<\/p>\n<p>\u5bf9\u4e8e cross-attention \u7ec4\u4ef6\uff0c\u53d7 GliDe \u542f\u53d1\uff0c\u6211\u4eec\u5229\u7528\u76ee\u6807\u6a21\u578b\u7684 KV cache\u3002\u8be5\u8bbe\u8ba1\u4e0d\u4ec5\u4f7f\u6a21\u578b\u80fd\u66f4\u597d\u5730\u5efa\u6a21\u5386\u53f2\u4fe1\u606f\uff0c\u8fd8\u5b8c\u5168\u6d88\u9664\u4e86\u957f\u4e0a\u4e0b\u6587\u7684\u989d\u5916\u5b58\u50a8\u5f00\u9500\uff0c\u56e0\u4e3a\u65e0\u8bba\u662f\u5426\u91c7\u7528\u63a8\u6d4b\u89e3\u7801\uff0c\u5927\u6a21\u578b\u7684 KV cache \u90fd\u5fc5\u987b\u88ab\u5b58\u50a8\u3002\u4e0d\u540c\u4e8e GliDe\uff0c\u6211\u4eec\u8fd8\u5728\u76ee\u6807\u6a21\u578b\u4e0e\u8349\u7a3f\u6a21\u578b\u4e4b\u95f4\u5171\u4eab Embedding Layer \u548c LM Head \u7684\u6743\u91cd\uff0c\u8fd9\u663e\u8457\u964d\u4f4e\u4e86 LLaMA-3\uff08\u8bcd\u8868\u5927\u5c0f 128,256\uff09\u548c Qwen-2.5\uff08\u8bcd\u8868\u5927\u5c0f 152,064\uff09\u7b49\u5927\u8bcd\u8868 LLM \u7684\u5185\u5b58\u6d88\u8017\u3002<\/p>\n<h4>\u6709\u6548\u8bad\u7ec3\u673a\u5236<\/h4>\n<p><strong>Anchor-Offset Indices\u3002<\/strong> \u4f7f\u7528 vanilla position indices \u65f6\uff0c\u4f4d\u7f6e\u7d22\u5f15\u7531\u4ece <span class=\"katex-eq\" data-katex-display=\"false\">0<\/span> \u5f00\u59cb\u7684\u8fde\u7eed\u6574\u6570\u7ec4\u6210\uff0c\u5e8f\u5217\u4e2d\u8f83\u65e9\u51fa\u73b0\u7684\u7d22\u5f15\u4f1a\u6bd4\u8f83\u5927\u7684\u4f4d\u7f6e\u7d22\u5f15\u51fa\u73b0\u5f97\u66f4\u9891\u7e41\uff0c\u5982\u56fe 2\uff08b\uff09\u4e0a\u534a\u90e8\u5206\u6240\u793a\u3002\u56e0\u6b64\uff0c\u8f83\u5927\u7684\u4f4d\u7f6e\u7d22\u5f15\u83b7\u5f97\u7684\u8bad\u7ec3\u66f4\u65b0\u4e0d\u8db3\uff0c\u4ece\u800c\u5bfc\u81f4\u8bad\u7ec3-\u63a8\u7406\u5dee\u5f02\u3002\u6b63\u5982\u6211\u4eec\u5728\u5f15\u8a00\u4e2d\u6307\u51fa\u7684\uff0c\u5e38\u89c1\u7684\u57fa\u4e8e RoPE \u7684\u5916\u63a8\u4e0d\u80fd\u76f4\u63a5\u5728\u8fd9\u91cc\u4f7f\u7528\uff0c\u56e0\u4e3a\u4e00\u65e6\u76ee\u6807\u6a21\u578b\u88ab\u9009\u5b9a\uff0cRoPE base \u5c31\u56fa\u5b9a\u4e86\u3002\u4e3a\u5229\u7528\u76ee\u6807\u6a21\u578b\u7684 KV cache\uff0c\u8349\u7a3f\u6a21\u578b\u5fc5\u987b\u4fdd\u6301\u4e0e\u76ee\u6807\u6a21\u578b\u76f8\u540c\u7684 RoPE base\u3002<\/p>\n<p>\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u6211\u4eec\u53ea\u80fd\u5229\u7528\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u7d22\u5f15\u3002\u8fd9\u4e9b\u7d22\u5f15\u5fc5\u987b\u786e\u4fdd\uff1a\uff081\uff09\u8349\u7a3f\u6a21\u578b\u4e2d\u7684\u4f4d\u7f6e\u7d22\u5f15\u80fd\u591f\u7528\u77ed\u4e0a\u4e0b\u6587\u6570\u636e\u5f97\u5230\u5145\u5206\u8bad\u7ec3\uff1b\uff082\uff09\u8fd9\u4e9b\u7d22\u5f15\u4e0d\u4f1a\u8ba9\u76ee\u6807\u6a21\u578b\u8868\u73b0\u51fa\u5206\u5e03\u5916\u884c\u4e3a\uff0c\u56e0\u4e3a\u8bad\u7ec3\u671f\u95f4\u76ee\u6807\u6a21\u578b\u4e0e\u8349\u7a3f\u6a21\u578b\u5171\u4eab\u76f8\u540c\u7d22\u5f15\u3002<\/p>\n<p>\u4e3a\u6ee1\u8db3\u8fd9\u4e9b\u7ea6\u675f\uff0c\u6211\u4eec\u63d0\u51fa Anchor-Offset Indices \u7b56\u7565\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u524d\u56db\u4e2a\u4f4d\u7f6e <span class=\"katex-eq\" data-katex-display=\"false\">[0,1,2,3]<\/span> \u4fdd\u7559\u4e3a <em>attention sink<\/em> tokens\uff1b\u968f\u540e\u5c06\u6240\u6709 token \u5206\u914d\u5230\u4ece\u968f\u673a offset \u5f00\u59cb\u7684\u5927\u578b\u8fde\u7eed\u7d22\u5f15\u4e0a\uff0c\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">[0,1,2,3,8192,8193,8194,\\dots]<\/span>\u3002\u6839\u636e attention sink \u73b0\u8c61\uff0cLLM \u5728\u5904\u7406\u957f\u6587\u672c\u65f6\uff0c\u6ce8\u610f\u529b\u6743\u91cd\u4e3b\u8981\u96c6\u4e2d\u5728\u524d\u56db\u4e2a token \u548c\u6700\u8fd1 token \u4e0a\u3002\u5229\u7528\u8fd9\u4e00\u73b0\u8c61\uff0c\u6211\u4eec\u8ba4\u4e3a Anchor-Offset Indices \u80fd\u81ea\u7136\u5730\u5f15\u5bfc\u76ee\u6807\u6a21\u578b\u8868\u73b0\u51fa\u5206\u5e03\u5185\u884c\u4e3a\u3002anchor indices \u4e0e\u968f\u673a offset \u786e\u4fdd\u6bcf\u4e2a\u4f4d\u7f6e\u7d22\u5f15\u90fd\u80fd\u5f97\u5230\u5145\u5206\u8bad\u7ec3\uff0c\u89e3\u51b3 vanilla \u65b9\u6cd5\u53cd\u590d\u53ea\u8bad\u7ec3\u8f83\u5c0f\u7d22\u5f15\u7684\u95ee\u9898\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0c\u5728\u76ee\u6807\u6a21\u578b\u4e2d\u91c7\u7528\u8fd9\u4e9b\u7d22\u5f15\u53ea\u4f1a\u4f7f loss \u589e\u52a0\u7ea6 0.001\uff0c\u8bf4\u660e\u76ee\u6807\u6a21\u578b\u786e\u5b9e\u975e\u5e38\u9002\u5408\u8fd9\u79cd\u6539\u53d8\u3002\u4f2a\u4ee3\u7801\u89c1\u9644\u5f55\u3002<\/p>\n<p><strong>Flash Noisy Training\u3002<\/strong> \u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u8349\u7a3f\u6a21\u578b\u5229\u7528\u6765\u81ea\u5927\u6a21\u578b\u7684 KV cache\uff1b\u4f46\u5728\u63a8\u7406\u671f\u95f4\uff0c\u8fd9\u4e2a KV cache \u5e76\u4e0d\u603b\u662f\u53ef\u89c1\u3002\u8fd9\u662f\u56e0\u4e3a\u5927\u6a21\u578b\u53ea\u6709\u5728\u9a8c\u8bc1\u5b8c\u6210\u65f6\u624d\u66f4\u65b0\u5176 KV cache\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u8349\u7a3f\u6a21\u578b\u4e2d\u7684\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">t<\/span> \u4e2a cross-attention query <span class=\"katex-eq\" data-katex-display=\"false\">Q_t<\/span>\uff0c\u6211\u4eec\u53ea\u80fd\u4fdd\u8bc1\u8bbf\u95ee\u6ee1\u8db3\u4e0b\u5f0f\u7684\u5bf9\u5e94 key-value states <span class=\"katex-eq\" data-katex-display=\"false\">K_{\\lt t^{\\prime}}<\/span>\u3001<span class=\"katex-eq\" data-katex-display=\"false\">V_{\\lt t^{\\prime}}<\/span>\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">1\\le|t^{\\prime} - t|\\lt \\gamma<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span> \u662f\u63a8\u6d4b\u6b65\u6570\u3002<\/p>\n<p>\u4e3a\u786e\u4fdd\u8bad\u7ec3\u4e0e\u63a8\u7406\u4e00\u81f4\uff0c\u4e00\u4e2a\u76f4\u63a5\u65b9\u6848\u662f\u6dfb\u52a0 attention mask\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e0e <code>Flash Attention<\/code> \u4e0d\u517c\u5bb9\uff0c\u4f1a\u663e\u8457\u964d\u4f4e\u8bad\u7ec3\u901f\u5ea6\uff0c\u5e76\u9020\u6210\u4e0d\u53ef\u63a5\u53d7\u7684\u5185\u5b58\u5f00\u9500\uff0c\u5c24\u5176\u662f\u5728\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u573a\u666f\u4e2d\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u63d0\u51fa\u4e00\u79cd\u79f0\u4e3a <strong>flash noisy training<\/strong> \u7684\u6280\u672f\u3002\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u6211\u4eec\u4ee5 <span class=\"katex-eq\" data-katex-display=\"false\">1 \\le j \\lt \\gamma<\/span> \u968f\u673a\u79fb\u52a8 queries \u4e0e key-value states \u7684\u7d22\u5f15\u3002\u5047\u8bbe\u5e8f\u5217\u957f\u5ea6\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">l<\/span>\uff0c\u5219\u8ba1\u7b97\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\nO_{\\geq j} = \\operatorname{att} \\bigl(Q_{\\geq j}, \\,K_{\\lt l-j}, \\,V_{\\lt l-j}\\bigr).\n\n<\/span>\n<p>\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6211\u4eec\u6709\u6548\u6a21\u62df\u4e86\u63a8\u7406\u9636\u6bb5\u76f8\u540c\u7684\u53ef\u89c1\u6027\u7ea6\u675f\uff0c\u5373 <span class=\"katex-eq\" data-katex-display=\"false\">1\\le|t^{\\prime} - t|\\lt \\gamma<\/span>\uff0c\u4ece\u800c\u5bf9\u9f50\u8bad\u7ec3\u65f6\u884c\u4e3a\u4e0e\u63a8\u7406\u884c\u4e3a\u3002\u4f7f\u7528 Flash Noisy Training \u65f6\uff0c\u6211\u4eec\u89c2\u5bdf\u5230 acceptance length \u76f8\u6bd4\u4e0d\u4f7f\u7528\u5b83\u8bad\u7ec3\u65f6\u63d0\u5347 14.7%\uff0c\u63d0\u5347\u6700\u96c6\u4e2d\u5728\u6700\u540e\u7684\u63a8\u6d4b tokens \u4e0a\u3002\u8fd9\u7a81\u51fa\u4e86\u5b83\u5728\u7f13\u89e3\u8bad\u7ec3-\u63a8\u7406 gap \u4e2d\u7684\u4f5c\u7528\u3002\u4f2a\u4ee3\u7801\u89c1\u9644\u5f55\u3002<\/p>\n<h4>\u5feb\u901f\u6811\u6ce8\u610f\u529b<\/h4>\n<p>Tree Speculative Decoding \u5229\u7528 speculation trees \u548c LLM \u7684\u56e0\u679c\u7ed3\u6784\uff0c\u4f7f\u8349\u7a3f\u6a21\u578b\u80fd\u591f\u63d0\u51fa\u591a\u4e2a\u5019\u9009\u5e8f\u5217\uff0c\u800c\u76ee\u6807\u6a21\u578b\u53ea\u9700\u9a8c\u8bc1\u4e00\u6b21\uff0c\u5e76\u4e14\u4e0d\u4f1a\u6539\u53d8\u6700\u7ec8\u7ed3\u679c\u3002\u5728\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0c<em>Tree Attention<\/em> \u5728\u786e\u4fdd\u6b63\u786e\u6027\u4e0e\u6548\u7387\u65b9\u9762\u53d1\u6325\u5173\u952e\u4f5c\u7528\u3002\u65e9\u671f\u5de5\u4f5c\u5c06\u7531 prefix trees \u5f97\u5230\u7684 attention masks \u5e94\u7528\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">QK^\\mathsf{T}<\/span> attention matrix\uff0c\u4ece\u800c\u7981\u7528\u63a8\u6d4b tokens \u4e4b\u95f4\u7684\u9519\u8bef\u7ec4\u5408\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u53ea\u80fd\u8fd0\u884c\u5728 PyTorch eager execution mode \u4e0a\uff0c\u65e0\u6cd5\u4f7f\u7528 <code>Flash Attention<\/code> \u7b49\u66f4\u5148\u8fdb\u7684 attention kernels\u3002\u56e0\u6b64\uff0c\u968f\u7740\u5e8f\u5217\u957f\u5ea6\u589e\u52a0\uff0c\u63a8\u7406\u901f\u5ea6\u4f1a\u663e\u8457\u4e0b\u964d\u3002<\/p>\n<p>\u4e3a\u89e3\u51b3\u8fd9\u4e9b\u6027\u80fd\u74f6\u9888\uff0c\u6211\u4eec\u63d0\u51fa <strong>Hybrid Tree Attention<\/strong> \u673a\u5236\uff0c\u5982\u56fe 2\uff08c\uff09\u6240\u793a\u3002\u6211\u4eec\u7684\u65b9\u6cd5\u57fa\u4e8e\u4e24\u4e2a\u5173\u952e\u89c2\u5bdf\uff1a\uff081\uff09\u6267\u884c Tree Attention \u65f6\uff0cqueries \u4e0e\u7f13\u5b58\u7684 key-value pairs <span class=\"katex-eq\" data-katex-display=\"false\">\\{K_{\\mathrm{cache}}, V_{\\mathrm{cache}}\\}<\/span> \u4e0d\u9700\u8981\u989d\u5916 masks\uff1b\uff082\uff09\u53ea\u6709 queries \u4e0e\u6765\u81ea\u5f53\u524d speculative tokens \u7684 key-value pairs <span class=\"katex-eq\" data-katex-display=\"false\">\\{K_{\\mathrm{specs}}, V_{\\mathrm{specs}}\\}<\/span> \u9700\u8981 masking\uff0c\u800c\u8fd9\u7c7b speculative tokens \u7684\u6570\u91cf\u901a\u5e38\u8f83\u5c0f\u3002\u57fa\u4e8e\u8fd9\u4e9b\u89c2\u5bdf\uff0c\u6211\u4eec\u91c7\u7528 divide and aggregate \u65b9\u6cd5\uff0c\u5c06 attention \u8ba1\u7b97\u62c6\u6210\u4e24\u90e8\u5206\uff0c\u5e76\u5728\u4e4b\u540e\u5408\u5e76\u3002<\/p>\n<p><strong>\u62c6\u5206 Key-Value Pairs\u3002<\/strong> \u6211\u4eec\u5c06\u6240\u6709 key-value pairs \u5206\u6210\u4e24\u7ec4\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">\\{K_{\\mathrm{cache}}, V_{\\mathrm{cache}}\\}<\/span>\uff0c\u5373\u4e3b\u5e8f\u5217\u7684\u7f13\u5b58\u90e8\u5206\uff0c\u4e0d\u9700\u8981 attention mask\uff1b\u4ee5\u53ca <span class=\"katex-eq\" data-katex-display=\"false\">\\{K_{\\mathrm{specs}}, V_{\\mathrm{specs}}\\}<\/span>\uff0c\u5373\u63a8\u6d4b\u9636\u6bb5\u90e8\u5206\uff0c\u9700\u8981 attention masks\u3002\u5bf9\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">\\{K_{\\mathrm{cache}}, V_{\\mathrm{cache}}\\}<\/span>\uff0c\u6211\u4eec\u8c03\u7528\u9ad8\u6548\u7684 <code>Flash Attention<\/code> kernel\u3002\u5bf9\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">\\{K_{\\mathrm{specs}}, V_{\\mathrm{specs}}\\}<\/span>\uff0c\u6211\u4eec\u4f7f\u7528\u81ea\u5b9a\u4e49 Triton kernel <code>fused_mask_attn<\/code>\uff0c\u5b83\u5728 KV \u7ef4\u5ea6\u4e2d\u5e94\u7528 blockwise loading \u4e0e masking\uff0c\u4ece\u800c\u5b9e\u73b0\u5feb\u901f attention \u8ba1\u7b97\u3002\u8be5\u6b65\u9aa4\u4ea7\u751f\u4e24\u7ec4 attention outputs <span class=\"katex-eq\" data-katex-display=\"false\">\\{O_{\\mathrm{cache}}, O_{\\mathrm{specs}}\\}<\/span>\uff0c\u4ee5\u53ca\u5b83\u4eec\u5bf9\u5e94\u7684 denominators\uff08\u5373\u6240\u6709 attention scores \u7684 log-sum-exp\uff09<span class=\"katex-eq\" data-katex-display=\"false\">\\{\\mathrm{LSE}_{\\mathrm{cache}}, \\mathrm{LSE}_{\\mathrm{specs}}\\}<\/span>\u3002<\/p>\n<p><strong>\u805a\u5408\u3002<\/strong> \u7136\u540e\u6211\u4eec\u901a\u8fc7 log-sum-exp trick \u5c06\u8fd9\u4e24\u90e8\u5206\u5408\u5e76\u6210\u6700\u7ec8 attention output <span class=\"katex-eq\" data-katex-display=\"false\">O_{\\mathrm{merge}}<\/span>\u3002\u9996\u5148\u8ba1\u7b97\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\begin{aligned}\n\n\\mathrm{LSE}_{\\mathrm{merge}}\n\n&amp;= \\log\\Bigl(\\exp\\bigl(\\mathrm{LSE}_{\\mathrm{cache}}\\bigr) + \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{specs}}\\bigr)\\Bigr),\n\n\\end{aligned}\n\n<\/span>\n<p>\u7136\u540e\u5bf9\u4e24\u4e2a outputs \u5e94\u7528\u52a0\u6743\u6c42\u548c\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\begin{aligned}\n\nO_{\\mathrm{merge}}\n\n=\\; &amp;O_{\\mathrm{cache}} \\cdot \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{cache}} - \\mathrm{LSE}_{\\mathrm{merge}}\\bigr)\\\\\n\n+\\; &amp;O_{\\mathrm{specs}} \\cdot \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{specs}} - \\mathrm{LSE}_{\\mathrm{merge}}\\bigr).\n\n\\end{aligned}\n\n<\/span>\n<p>\u7406\u8bba\u4fdd\u8bc1\u89c1\u9644\u5f55\u3002\u5982\u4e0a\u6240\u8ff0\uff0c\u8fd9\u79cd hybrid \u65b9\u6cd5\u5728\u957f\u5e8f\u5217\u63a8\u7406\u7684\u5927\u90e8\u5206\u8ba1\u7b97\u4e2d\u4f7f\u7528\u9ad8\u5ea6\u9ad8\u6548\u7684 <code>Flash Attention<\/code> kernel\uff0c\u5e76\u4e14\u53ea\u5bf9\u5c11\u91cf speculative tokens \u4f7f\u7528\u81ea\u5b9a\u4e49 masking attention <code>fused_mask_attn<\/code>\u3002kernel <code>fused_mask_attn<\/code> \u9075\u5faa <code>Flash Attention 2<\/code> \u7684\u8bbe\u8ba1\u54f2\u5b66\uff0c\u5c06 <span class=\"katex-eq\" data-katex-display=\"false\">Q<\/span>\u3001<span class=\"katex-eq\" data-katex-display=\"false\">K_{\\text{specs}}<\/span> \u4e0e <span class=\"katex-eq\" data-katex-display=\"false\">V_{\\text{specs}}<\/span> \u62c6\u5206\u6210\u5c0f block\u3002\u8be5\u7b56\u7565\u51cf\u5c11\u5168\u5c40\u5185\u5b58 I\/O\uff0c\u5e76\u5145\u5206\u5229\u7528 GPU streaming multiprocessors\u3002\u6b64\u5916\uff0c\u5728\u8ba1\u7b97 <span class=\"katex-eq\" data-katex-display=\"false\">QK_{\\text{specs}}^\\top<\/span> \u7684\u6bcf\u4e2a block \u65f6\uff0cmask matrix \u4f1a\u88ab\u52a0\u8f7d\u5e76\u7528\u4e8e\u5e94\u7528 masking \u64cd\u4f5c\u3002Hybrid Tree Attention \u5728\u591a\u4e2a\u5206\u652f\u7684\u5e76\u884c\u9a8c\u8bc1\u4e0e\u66f4\u9ad8\u63a8\u7406\u901f\u5ea6\u4e4b\u95f4\u5b9e\u73b0\u4e86\u6709\u6548\u5e73\u8861\uff0c\u540c\u65f6\u4e0d\u635f\u5bb3\u6b63\u786e\u6027\u3002<\/p>\n<h3>\u5b9e\u9a8c<\/h3>\n<h4>\u8bbe\u7f6e<\/h4>\n<p><strong>\u76ee\u6807\u6a21\u578b\u4e0e\u8349\u7a3f\u6a21\u578b\u3002<\/strong> \u6211\u4eec\u9009\u62e9\u56db\u7c7b\u5e7f\u6cdb\u4f7f\u7528\u7684\u957f\u4e0a\u4e0b\u6587 LLM \u4f5c\u4e3a\u76ee\u6807\u6a21\u578b\uff1aVicuna\uff08\u5305\u62ec 7B \u548c 13B\uff09\u3001LongChat\uff08\u5305\u62ec 7B \u548c 13B\uff09\u3001LLaMA-3.1-8B-Instruct\uff0c\u4ee5\u53ca QwQ-32B\u3002\u4e3a\u4f7f\u8349\u7a3f\u6a21\u578b\u4e0e\u76ee\u6807\u6a21\u578b\u66f4\u52a0\u517c\u5bb9\uff0c\u6211\u4eec\u7684\u8349\u7a3f\u6a21\u578b\u5728\u591a\u79cd\u53c2\u6570\u4e0a\u4e0e\u76ee\u6807\u6a21\u578b\u4fdd\u6301\u4e00\u81f4\uff0c\u4f8b\u5982 KV heads \u6570\u91cf\u3002<\/p>\n<p><strong>\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/strong> \u6211\u4eec\u9996\u5148\u5728 SlimPajama-6B \u9884\u8bad\u7ec3\u6570\u636e\u96c6\u4e0a\u4f7f\u7528 Anchor-Offset Indices \u8bad\u7ec3\u8349\u7a3f\u6a21\u578b\u3002\u5bf9\u4e8e Vicuna \u6a21\u578b\u548c LongChat-7B\uff0crandom offset \u8bbe\u4e3a 0 \u5230 15k \u4e4b\u95f4\u7684\u968f\u673a\u6574\u6570\uff1b\u5bf9\u4e8e\u53e6\u5916\u4e09\u4e2a\u6a21\u578b\uff0c\u7531\u4e8e\u5b83\u4eec\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u66f4\u957f\uff0crandom offset \u8bbe\u4e3a 0 \u5230 30k\u3002\u7136\u540e\u6211\u4eec\u5728 Prolong-64k \u957f\u4e0a\u4e0b\u6587\u6570\u636e\u96c6\u7684\u4e00\u4e2a\u5c0f\u5b50\u96c6\u4e0a\u8bad\u7ec3\u6a21\u578b\uff0c\u4ee5\u83b7\u5f97\u5904\u7406\u957f\u6587\u672c\u7684\u80fd\u529b\u3002\u6700\u540e\uff0c\u6211\u4eec\u5728\u81ea\u5efa\u957f\u4e0a\u4e0b\u6587 supervised-finetuning\uff08SFT\uff09\u6570\u636e\u96c6\u4e0a\u5fae\u8c03\u6a21\u578b\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u6027\u80fd\u3002\u540e\u4e24\u4e2a\u9636\u6bb5\u7684\u4f4d\u7f6e\u7d22\u5f15\u91c7\u7528 vanilla indexing policy\uff0c\u56e0\u4e3a\u8bad\u7ec3\u6570\u636e\u5df2\u7ecf\u8db3\u591f\u957f\u3002\u6211\u4eec\u5728\u6240\u6709\u4e09\u4e2a\u9636\u6bb5\u90fd\u5e94\u7528 flash noisy training\uff0c\u4ee5\u7f13\u89e3\u8bad\u7ec3\u4e0e\u63a8\u7406\u4e0d\u4e00\u81f4\u95ee\u9898\uff0c\u5e76\u4e14 flash noisy training \u7684\u989d\u5916\u5f00\u9500\u53ef\u4ee5\u5ffd\u7565\u3002\u66f4\u591a\u6a21\u578b\u8bad\u7ec3\u7ec6\u8282\u89c1\u9644\u5f55\u3002<\/p>\n<p><center>\u8868 1\uff1a\u4e0d\u540c\u6a21\u578b\u4e0e\u8bbe\u7f6e\u4e0b\u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span>\u3001\u89e3\u7801\u901f\u5ea6\uff08tokens\/s\uff09\u548c\u52a0\u901f\u6bd4\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u201cVanilla HF\u201d \u6307 HuggingFace \u57fa\u4e8e PyTorch \u7684 attention \u5b9e\u73b0\uff0c\u800c \u201cVanilla FA\u201d \u4f7f\u7528 <code>Flash Attention<\/code>\u3002\u52a0\u901f\u6bd4\u7edf\u8ba1\u7684\u662f\u76f8\u5bf9 Vanilla HF \u65b9\u6cd5\u7684\u52a0\u901f\u6bd4\u4f8b\u3002\u6240\u6709\u7ed3\u679c\u5747\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">T=0<\/span> \u4e0b\u8ba1\u7b97\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257008.png\" alt=\"\u8868 1\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<p><strong>\u6d4b\u8bd5\u57fa\u51c6\u3002<\/strong> \u5bf9\u4e8e\u5e38\u89c4\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u4efb\u52a1\uff0c\u6211\u4eec\u4ece LongBench benchmark \u4e2d\u9009\u62e9\u9700\u8981\u751f\u6210\u8f83\u957f\u8f93\u51fa\u7684\u4efb\u52a1\uff0c\u56e0\u4e3a\u8f93\u51fa\u8f83\u77ed\u7684\u4efb\u52a1\uff08\u4f8b\u5982 document-QA\uff09\u4f1a\u8ba9\u63a8\u6d4b\u89e3\u7801\u7684\u52a0\u901f\u6bd4\u96be\u4ee5\u516c\u5e73\u8861\u91cf\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u5173\u6ce8\u957f\u6587\u6863\u6458\u8981\u548c\u4ee3\u7801\u8865\u5168\u4efb\u52a1\uff0c\u5e76\u5728\u4e94\u4e2a\u6570\u636e\u96c6\u4e0a\u6d4b\u8bd5\uff1aGovReport\u3001QMSum\u3001Multi-News\u3001LCC \u548c RepoBench-P\u3002\u5bf9\u4e8e\u6570\u5b66\u63a8\u7406\u4efb\u52a1\uff0c\u6211\u4eec\u5728\u56db\u4e2a\u6570\u5b66\u63a8\u7406\u6570\u636e\u96c6\u4e0a\u6d4b\u8bd5 QwQ-32B\uff1aAIME24\u3001AMC\u3001MATH500 \u548c Minerva Math\u3002<\/p>\n<p>\u6211\u4eec\u5c06\u65b9\u6cd5\u4e0e\u539f\u59cb\u76ee\u6807\u6a21\u578b\u3001PLD \u548c MagicDec \u6bd4\u8f83\u3002PLD \u662f\u6700\u6d41\u884c\u7684\u57fa\u4e8e\u68c0\u7d22\u7684\u65b9\u6cd5\uff0c\u4e5f\u88ab\u79f0\u4e3a vLLM \u4e2d\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">n<\/span>-gram SD\uff1bMagicDec \u662f TriForce \u7684\u7b80\u5355\u539f\u578b\u3002\u4e3a\u7a81\u51fa <code>Flash Attention<\/code> \u5728\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\u7684\u91cd\u8981\u6027\uff0c\u6211\u4eec\u8fd8\u5c55\u793a\u4e86\u539f\u59cb\u76ee\u6807\u6a21\u578b\u5206\u522b\u4f7f\u7528 HuggingFace eager attention \u548c <code>Flash Attention<\/code> \u7684\u6027\u80fd\u3002\u4e3a\u516c\u5e73\u6bd4\u8f83\uff0cMagicDec \u57fa\u7ebf\u4e5f\u4f7f\u7528 <code>Flash Attention<\/code>\u3002\u63a8\u6d4b\u89e3\u7801\u6700\u91cd\u8981\u7684\u6307\u6807\u662f <em>walltime speedup ratio<\/em>\uff0c\u5373\u76f8\u5bf9\u4e8e vanilla autoregressive decoding \u7684\u5b9e\u9645\u6d4b\u8bd5\u52a0\u901f\u6bd4\u3002\u6211\u4eec\u4e5f\u6d4b\u8bd5 <em>average acceptance length<\/em> <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span>\uff0c\u5373\u76ee\u6807 LLM \u6bcf\u6b21 forward pass \u5e73\u5747\u63a5\u53d7\u7684 tokens \u6570\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257003.png\" alt=\"T1\" \/><\/p>\n<p><center>\u56fe 3\uff1a\u4e0d\u540c\u6a21\u578b\u4e0e\u8bbe\u7f6e\u4e0b\u7684\u89e3\u7801\u901f\u5ea6\uff08tokens\/s\uff09\u3002\u6240\u6709\u7ed3\u679c\u5747\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">T=1<\/span> \u4e0b\u8ba1\u7b97\u3002\u6a2a\u8f74\u4e0a\u7684\u5b57\u6bcd G\u3001Q\u3001M\u3001L \u548c R \u5206\u522b\u8868\u793a GovReport\u3001QMSum\u3001Multi-News\u3001LCC \u548c RepoBench-P \u6570\u636e\u96c6\u3002<\/center><\/p>\n<h4>\u4e3b\u8981\u7ed3\u679c<\/h4>\n<p>\u8868 1 \u548c\u56fe 3 \u5c55\u793a\u4e86\u4e94\u4e2a\u8bc4\u4f30\u6570\u636e\u96c6\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">T=0<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">T=1<\/span> \u4e0b\u7684\u89e3\u7801\u901f\u5ea6\u4e0e\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\uff0c\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">T<\/span> \u8868\u793a LLM sampling \u4f7f\u7528\u7684 temperature\u3002\u6211\u4eec\u63d0\u51fa\u7684\u65b9\u6cd5\u5728\u6458\u8981\u4efb\u52a1\u548c\u4ee3\u7801\u8865\u5168\u4efb\u52a1\u4e0a\u90fd\u663e\u8457\u4f18\u4e8e\u6240\u6709\u5176\u4ed6\u65b9\u6cd5\u3002\u5f53 <span class=\"katex-eq\" data-katex-display=\"false\">T=0<\/span> \u65f6\uff0c\u5728\u6458\u8981\u4efb\u52a1\u4e0a\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u53ef\u5b9e\u73b0\u7ea6 3.5 \u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\u548c\u6700\u9ad8 2.67<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\uff1b\u5728\u4ee3\u7801\u8865\u5168\u4efb\u52a1\u4e0a\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u53ef\u5b9e\u73b0\u7ea6 4 \u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\u548c\u6700\u9ad8 3.26<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\u3002\u8fd9\u7a81\u51fa\u4e86\u6211\u4eec\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u7684\u7a33\u5065\u6027\u548c\u6cdb\u5316\u80fd\u529b\uff0c\u5c24\u5176\u662f\u5728\u957f\u6587\u672c\u751f\u6210\u4efb\u52a1\u4e2d\u3002\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">T=1<\/span> \u65f6\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5b9e\u73b0\u7ea6 2.5<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\uff0c\u4ecd\u663e\u8457\u9886\u5148 MagicDec\u3002\u8fd9\u8868\u660e\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u4e0d\u540c temperature \u8bbe\u7f6e\u4e0b\u90fd\u5f88\u7a33\u5065\uff0c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u5176\u5408\u7406\u6027\u548c\u6548\u7387\u3002<\/p>\n<p>\u867d\u7136 PLD \u80fd\u5728\u8bb8\u591a\u6570\u636e\u96c6\u4e0a\u52a0\u901f\u751f\u6210\uff0c\u4f46\u5b83\u4ecd\u4e0d\u53ca\u6211\u4eec\u63d0\u51fa\u7684 LongSpec\u3002\u5728\u67d0\u4e9b\u573a\u666f\u4e2d\uff08\u4f8b\u5982 retrieval \u5f88\u5c11\u65f6\uff09\uff0cPLD \u751a\u81f3\u53ef\u80fd\u4ea7\u751f\u8d1f\u52a0\u901f\u3002\u5bf9\u4e8e\u53e6\u4e00\u4e2a\u57fa\u7ebf MagicDec\uff0c\u5c3d\u7ba1\u5b83\u5c55\u793a\u51fa\u4e0e LongSpec \u76f8\u6bd4\u5177\u6709\u7ade\u4e89\u529b\u7684 acceptance rates\uff0c\u4f46\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\u5176\u52a0\u901f\u660e\u663e\u66f4\u4f4e\u3002\u8fd9\u662f\u56e0\u4e3a MagicDec \u4e3b\u8981\u4e3a\u5927 batch size \u548c tensor parallelism \u573a\u666f\u8bbe\u8ba1\u3002\u5728\u4f4e batch size \u8bbe\u7f6e\u4e2d\uff0c\u5b83\u7684\u8349\u7a3f\u6a21\u578b\u4f7f\u7528\u5e26\u7a00\u758f KV cache \u7684\u76ee\u6807\u6a21\u578b\u5168\u90e8\u53c2\u6570\uff0c\u56e0\u800c\u53d8\u5f97\u8fc7\u91cd\u3002\u8fd9\u4e00\u8bbe\u8ba1\u9009\u62e9\u5bfc\u81f4\u4f4e\u6548\uff0c\u56e0\u4e3a\u8349\u7a3f\u6a21\u578b\u7684\u8ba1\u7b97\u5f00\u9500\u8d85\u8fc7\u4e86\u5176\u63a8\u6d4b\u6536\u76ca\u3002\u6211\u4eec\u7684\u7ed3\u679c\u663e\u793a\uff0c\u5f53 guess length <span class=\"katex-eq\" data-katex-display=\"false\">\\gamma=2<\/span> \u65f6\uff0cMagicDec \u53ea\u5728\u90e8\u5206\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">\\gt \\!1<\/span> \u7684\u52a0\u901f\u6bd4\uff1b\u5f53 <span class=\"katex-eq\" data-katex-display=\"false\">\\gamma\\geq3<\/span> \u65f6\uff0c\u5b83\u6301\u7eed\u8868\u73b0\u51fa\u7ea6 0.7<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u7684\u8d1f\u52a0\u901f\uff0c\u8fdb\u4e00\u6b65\u7a81\u51fa\u4e86\u8be5\u65b9\u6cd5\u5728\u8fd9\u4e9b\u914d\u7f6e\u4e0b\u7684\u5c40\u9650\u6027\u3002MagicDec \u5728\u66f4\u5927 batch size \u4e0b\u7684\u6027\u80fd\u89c1\u541e\u5410\u91cf\u5c0f\u8282\u3002<\/p>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u53d1\u73b0 attention \u5b9e\u73b0\u5bf9\u957f\u4e0a\u4e0b\u6587\u63a8\u6d4b\u89e3\u7801\u6027\u80fd\u8d77\u7740\u5173\u952e\u4f5c\u7528\u3002\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u201cVanilla HF\u201d \u6307 HuggingFace \u7684 attention \u5b9e\u73b0\uff0c\u800c \u201cVanilla FA\u201d \u4f7f\u7528 <code>Flash Attention<\/code>\u3002\u540e\u8005\u5373\u4f7f\u4f5c\u4e3a\u72ec\u7acb\u7ec4\u4ef6\uff0c\u4e5f\u76f8\u6bd4\u524d\u8005\u5c55\u793a\u51fa\u8fd1 2<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\uff0c\u800c\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u4ee3\u7801\u8865\u5168\u6570\u636e\u96c6\u4e0a\u76f8\u6bd4 HF Attention \u53ef\u5b9e\u73b0\u6700\u9ad8 6<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u52a0\u901f\u3002\u8be5\u7ed3\u679c\u5f3a\u8c03\uff0c\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u5fc5\u987b\u4e0e <code>Flash Attention<\/code> \u7b49\u4f18\u5316 attention \u673a\u5236\u517c\u5bb9\uff0c\u5c24\u5176\u662f\u5728\u957f\u6587\u672c\u8bbe\u7f6e\u4e2d\u3002\u6211\u4eec\u7684 hybrid tree attention \u65b9\u6cd5\u5b9e\u73b0\u4e86\u8fd9\u79cd\u517c\u5bb9\u6027\uff0c\u4f7f\u6211\u4eec\u80fd\u5145\u5206\u5229\u7528 <code>Flash Attention<\/code> \u7684\u4f18\u52bf\u5e76\u8fdb\u4e00\u6b65\u52a0\u901f\u3002<\/p>\n<h4>\u6d88\u878d\u7814\u7a76<\/h4>\n<p><center>\u8868 2\uff1a\u5728 Multi-News \u548c RepoBench-P \u6570\u636e\u96c6\u4e0a\uff0c\u6709\u65e0 Anchor-Offset Indices \u7684\u6027\u80fd\u6bd4\u8f83\u3002\u5e26 Anchor-Offset Indices \u7684\u6a21\u578b\u5b9e\u73b0\u66f4\u9ad8\u8f93\u51fa\u901f\u5ea6\u548c\u66f4\u5927 acceptance length\uff0c\u7a81\u51fa\u5176\u6548\u7387\u548c\u6709\u6548\u6027\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257009.png\" alt=\"\u8868 2\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<p><strong>Anchor-Offset Indices\u3002<\/strong> \u5b9e\u9a8c\u7ed3\u679c\u5c55\u793a\u4e86\u5f15\u5165 Anchor-Offset Indices \u7684\u663e\u8457\u6536\u76ca\u3002\u56fe 4 \u8868\u660e\uff0c\u5728\u771f\u5b9e\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u65f6\uff0c\u4f7f\u7528 Anchor-Offset Indices \u8bad\u7ec3\u7684\u6a21\u578b\u76f8\u6bd4\u4e0d\u4f7f\u7528\u5b83\u7684\u6a21\u578b\u5177\u6709\u66f4\u4f4e\u7684 initial loss \u548c final loss\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4f7f\u7528 Anchor-Offset Indices \u521d\u59cb\u5316\u7684\u6a21\u578b\u8fbe\u5230\u76f8\u540c loss \u6c34\u5e73\u7684\u901f\u5ea6\u6bd4\u5bf9\u5e94\u6a21\u578b\u5feb 3.93<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\u3002\u8868 2 \u8fdb\u4e00\u6b65\u5c55\u793a\u4e86\u5b83\u5728\u4e24\u4e2a\u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\u63d0\u5347\uff1a\u4e00\u4e2a\u6458\u8981\u6570\u636e\u96c6 Multi-News\uff0c\u4ee5\u53ca\u4e00\u4e2a\u4ee3\u7801\u8865\u5168\u6570\u636e\u96c6 RepoBench-P\u3002\u5e26 Anchor-Offset Indices \u7684\u6a21\u578b\u5c55\u73b0\u51fa\u66f4\u5feb\u8f93\u51fa\u901f\u5ea6\u548c\u66f4\u5927\u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span>\u3002\u8fd9\u4e9b\u7ed3\u679c\u5f3a\u8c03\u4e86 Anchor-Offset Indices \u5728\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u548c\u6a21\u578b\u6027\u80fd\u65b9\u9762\u7684\u6709\u6548\u6027\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257004.png\" alt=\"ablation1\" \/><\/p>\n<p><center>\u56fe 4\uff1a\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u4e0a\u7684\u8bad\u7ec3 loss \u66f2\u7ebf\u3002\u4f7f\u7528 Anchor-Offset Indices \u7684\u9884\u8bad\u7ec3\u6a21\u578b\u5c55\u793a\u51fa\u66f4\u4f4e\u7684 initial loss \u548c final loss\uff0c\u5e76\u4e14\u76f8\u6bd4\u4e0d\u4f7f\u7528 Anchor-Offset Indices \u7684\u6a21\u578b\uff0c\u8fbe\u5230\u76f8\u540c loss \u6c34\u5e73\u7684\u901f\u5ea6\u5feb 3.93<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\u3002<\/center><\/p>\n<p><strong>Hybrid Tree Attention\u3002<\/strong> \u56fe 5 \u6240\u793a\u7ed3\u679c\u7a81\u51fa\u4e86\u6240\u63d0\u51fa Hybrid Tree Attention \u7684\u6709\u6548\u6027\uff0c\u5b83\u7ed3\u5408\u4e86 <code>Flash Attention<\/code> \u4e0e Triton kernel <code>fused_mask_attn<\/code>\u3002\u867d\u7136\u4e24\u79cd\u65b9\u6cd5\u5728\u8349\u7a3f\u6a21\u578b forward pass \u548c\u76ee\u6807\u6a21\u578b FFN \u8ba1\u7b97\u4e0a\u7684\u8017\u65f6\u76f8\u8fd1\uff0c\u4f46 hybrid \u65b9\u6cd5\u5728\u76ee\u6807\u6a21\u578b attention layer\uff08\u9ec4\u8272\u90e8\u5206\uff09\u4e0a\u663e\u8457\u964d\u4f4e\u4e86\u5ef6\u8fdf\u3002\u5177\u4f53\u6765\u8bf4\uff0cattention computation latency \u4ece HF \u5b9e\u73b0\u4e2d\u7684 49.92 ms \u964d\u81f3 hybrid \u65b9\u6cd5\u4e2d\u7684 12.54 ms\uff0c\u5e26\u6765\u7ea6 75% \u7684\u6539\u8fdb\u3002\u9a8c\u8bc1\u6b65\u9aa4\u7684\u65f6\u95f4\u5dee\u5f02\u5f88\u5c0f\uff0c\u8fdb\u4e00\u6b65\u5de9\u56fa\u4e86\u4e3b\u8981\u6027\u80fd\u6536\u76ca\u6765\u81ea\u4f18\u5316 attention mechanism \u7684\u7ed3\u8bba\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257005.png\" alt=\"ablation2\" \/><\/p>\n<p><center>\u56fe 5\uff1a\u5355\u4e2a\u63a8\u6d4b\u89e3\u7801\u5faa\u73af\u7684\u5ef6\u8fdf\u5206\u89e3\uff0c\u5bf9\u6bd4 EAGLE \u5b9e\u73b0\u4e0e\u6240\u63d0\u51fa\u7684 Hybrid Tree Attention\u3002\u4f7f\u7528\u6211\u4eec\u7684\u65b9\u6cd5\u65f6\uff0c\u76ee\u6807\u6a21\u578b attention layer\uff08\u9ec4\u8272\u90e8\u5206\uff09\u51fa\u73b0\u663e\u8457\u5ef6\u8fdf\u964d\u4f4e\u3002<\/center><br \/>\n<center>\u8868 3\uff1a\u6211\u4eec\u7684\u65b9\u6cd5\u5728 QwQ-32B \u6a21\u578b\u4e0a\u3001\u56db\u4e2a\u6570\u5b66\u63a8\u7406\u6570\u636e\u96c6\u4e2d\u7684\u6027\u80fd\uff0c\u6700\u5927\u8f93\u51fa\u957f\u5ea6\u4e3a 32k tokens\u3002\u8be5\u8868\u5c55\u793a\u6bcf\u79d2\u751f\u6210 tokens \u6570\u548c\u5e73\u5747 accepted tokens \u6570 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span>\uff1b\u6211\u4eec\u7684\u65b9\u6cd5\u76f8\u6bd4\u57fa\u7ebf\u5e73\u5747\u7ea6\u5feb 2.34<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\uff0c\u5e73\u5747\u63a5\u53d7 tokens \u6570\u4e3a 3.81\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257010.png\" alt=\"\u8868 3\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<h4>\u957f\u63a8\u7406\u52a0\u901f<\/h4>\n<p>\u957f\u63a8\u7406\u4efb\u52a1\u6700\u8fd1\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\uff0c\u56e0\u4e3a\u5b83\u4eec\u80fd\u591f\u8ba9\u6a21\u578b\u5728\u6269\u5c55\u8f93\u51fa\u4e0a\u6267\u884c\u590d\u6742\u63a8\u7406\u548c\u95ee\u9898\u6c42\u89e3\u3002\u5728\u8fd9\u4e9b\u4efb\u52a1\u4e2d\uff0c\u867d\u7136\u524d\u7f00\u8f93\u5165\u901a\u5e38\u76f8\u5bf9\u8f83\u77ed\uff0c\u4f46\u751f\u6210\u8f93\u51fa\u53ef\u80fd\u6781\u957f\uff0c\u4ece\u800c\u5728\u6548\u7387\u548c token acceptance \u65b9\u9762\u5e26\u6765\u72ec\u7279\u6311\u6218\u3002\u6211\u4eec\u7684\u65b9\u6cd5\u7279\u522b\u9002\u5408\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u80fd\u591f\u6709\u6548\u5904\u7406\u957f\u8f93\u51fa\u573a\u666f\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cMagicDec \u4e0d\u9002\u7528\u4e8e\u8fd9\u79cd\u957f\u8f93\u51fa\u573a\u666f\uff0c\u56e0\u4e3a\u957f\u63a8\u7406\u4efb\u52a1\u7684\u521d\u59cb\u63a8\u7406\u9636\u6bb5\u4e0e\u4f20\u7edf\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e0d\u540c\u3002\u5728\u957f\u63a8\u7406\u4efb\u52a1\u4e2d\uff0cprefix \u76f8\u5bf9\u8f83\u77ed\uff0cMagicDec \u4e2d\u7684\u8349\u7a3f\u6a21\u578b\u4f1a\u5b8c\u5168\u9000\u5316\u4e3a\u76ee\u6807\u6a21\u578b\uff0c\u65e0\u6cd5\u5b9e\u73b0\u52a0\u901f\u3002<\/p>\n<p>\u6211\u4eec\u5728 QwQ-32B \u6a21\u578b\u4e0a\u4f7f\u7528\u56db\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684 benchmark \u8bc4\u4f30\u65b9\u6cd5\uff0c\u6700\u5927\u8f93\u51fa\u957f\u5ea6\u8bbe\u4e3a 32k tokens\u3002\u8868 3 \u6240\u793a\u7ed3\u679c\u5c55\u793a\u4e86\u751f\u6210\u901f\u5ea6\u548c\u5e73\u5747\u63a5\u53d7 tokens \u6570\u65b9\u9762\u7684\u663e\u8457\u63d0\u5347\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5b9e\u73b0\u7ea6 45 tokens\/s \u7684\u751f\u6210\u7387\uff0c\u6bd4\u5f3a <code>Flash Attention<\/code> \u57fa\u7ebf\u9ad8 2.34<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\uff0c\u5e76\u5177\u6709\u5e73\u5747 3.81 \u4e2a accepted tokens\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5e26 LongSpec \u7684 QwQ-32B \u751a\u81f3\u6bd4\u5e26 <code>Flash Attention<\/code> \u7684\u6807\u51c6 7B \u6a21\u578b\u5ef6\u8fdf\u66f4\u4f4e\uff0c\u8868\u660e\u6211\u4eec\u7684\u65b9\u6cd5\u6709\u6548\u52a0\u901f\u4e86\u957f\u63a8\u7406\u6a21\u578b\u3002\u8fd9\u4e9b\u53d1\u73b0\u4e0d\u4ec5\u7a81\u51fa\u4e86\u6211\u4eec\u65b9\u6cd5\u5728\u957f\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\uff0c\u4e5f\u4e3a o1-like \u6a21\u578b\u7684\u65e0\u635f\u63a8\u7406\u52a0\u901f\u63d0\u4f9b\u4e86\u65b0\u89c1\u89e3\u3002\u6211\u4eec\u76f8\u4fe1\uff0c\u63a8\u6d4b\u89e3\u7801\u672a\u6765\u5c06\u5728\u52a0\u901f\u8fd9\u7c7b\u6a21\u578b\u4e2d\u53d1\u6325\u5173\u952e\u4f5c\u7528\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257006.png\" alt=\"longcot\" \/><\/p>\n<p><center>\u56fe 6\uff1a\u957f\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u5c55\u793a\u3002<\/center><\/p>\n<h4>\u541e\u5410\u91cf<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257007.png\" alt=\"throughput\" \/><\/p>\n<p><center>\u56fe 7\uff1aVanilla\u3001MagicDec \u548c LongSpec \u7684\u541e\u5410\u91cf\u5bf9\u6bd4\u3002<\/center><\/p>\n<p>\u5982\u56fe 7 \u6240\u793a\uff0cVicuna-7B \u5728 RepoBench-P \u6570\u636e\u96c6\u4e0a\u7684\u541e\u5410\u91cf\u7ed3\u679c\u8868\u660e\uff0cLongSpec \u5728\u6240\u6709 batch sizes \u4e0b\u90fd\u6301\u7eed\u4f18\u4e8e Vanilla \u548c MagicDec\u3002\u5f53 batch size \u4e3a 8 \u65f6\uff0cLongSpec \u5b9e\u73b0 561.32 tokens\/s \u7684\u541e\u5410\u91cf\uff0c\u7ea6\u4e3a MagicDec\uff08310.58 tokens\/s\uff09\u7684 1.8<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\uff0c\u5e76\u4e14\u63a5\u8fd1 Vanilla\uff08286.96 tokens\/s\uff09\u7684 2<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span>\u3002MagicDec \u4ee5\u541e\u5410\u4f18\u5316\u4e3a\u76ee\u6807\u8fdb\u884c\u8bbe\u8ba1\uff0c\u56e0\u6b64\u968f\u7740 batch size \u589e\u52a0\u4f1a\u8d85\u8fc7 Vanilla\uff0c\u53cd\u6620\u51fa\u5176\u9488\u5bf9\u6027\u6539\u8fdb\u3002\u7136\u800c\uff0cLongSpec \u4ecd\u7136\u4fdd\u6301\u4f18\u52bf\uff0c\u5728\u6240\u6709\u6d4b\u8bd5 batch sizes \u4e0b\u7ef4\u6301\u66f4\u9ad8\u541e\u5410\u91cf\u3002<\/p>\n<h3>\u7ed3\u8bba<\/h3>\n<p>\u672c\u6587\u63d0\u51fa\u4e86 LongSpec\uff0c\u4e00\u4e2a\u65e8\u5728\u589e\u5f3a\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e0b\u65e0\u635f\u63a8\u6d4b\u89e3\u7801\u7684\u65b0\u6846\u67b6\u3002\u4e0d\u540c\u4e8e\u6b64\u524d\u4e3b\u8981\u5173\u6ce8\u77ed\u4e0a\u4e0b\u6587\u8bbe\u7f6e\u7684\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\uff0cLongSpec \u76f4\u63a5\u5904\u7406\u4e09\u4e2a\u5173\u952e\u6311\u6218\uff1a\u8fc7\u9ad8\u5185\u5b58\u5f00\u9500\u3001\u5927\u4f4d\u7f6e\u7d22\u5f15\u8bad\u7ec3\u4e0d\u8db3\uff0c\u4ee5\u53ca\u4f4e\u6548\u7684\u6811\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u4e3a\u7f13\u89e3\u5185\u5b58\u7ea6\u675f\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u9ad8\u6548\u8349\u7a3f\u6a21\u578b\u67b6\u6784\uff0c\u5b83\u901a\u8fc7\u7ed3\u5408\u6ed1\u52a8\u7a97\u53e3 self-attention \u4e0e\u65e0\u7f13\u5b58 cross-attention\uff0c\u5c06\u5185\u5b58\u5360\u7528\u4fdd\u6301\u4e3a\u5e38\u6570\u3002\u4e3a\u89e3\u51b3\u77ed\u4e0a\u4e0b\u6587\u6570\u636e\u76f8\u5173\u7684\u8bad\u7ec3\u9650\u5236\uff0c\u6211\u4eec\u63d0\u51fa Anchor-Offset Indices\uff0c\u786e\u4fdd\u5373\u4f7f\u5728\u77ed\u5e8f\u5217\u6570\u636e\u96c6\u4e2d\uff0c\u5927\u4f4d\u7f6e\u7d22\u5f15\u4e5f\u80fd\u5f97\u5230\u5145\u5206\u8bad\u7ec3\u3002\u6700\u540e\uff0c\u6211\u4eec\u63d0\u51fa Hybrid Tree Attention\uff0c\u5b83\u5c06\u57fa\u4e8e\u6811\u7684\u63a8\u6d4b\u89e3\u7801\u4e0e <code>Flash Attention<\/code> \u9ad8\u6548\u7ed3\u5408\u3002\u5927\u91cf\u5b9e\u9a8c\u5c55\u793a\u4e86 LongSpec \u5728\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u4efb\u52a1\u548c\u771f\u5b9e\u957f\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\u3002\u6211\u4eec\u7684\u53d1\u73b0\u5f3a\u8c03\u4e86\u4e3a\u957f\u4e0a\u4e0b\u6587\u8bbe\u7f6e\u4e13\u95e8\u8bbe\u8ba1\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u7684\u91cd\u8981\u6027\uff0c\u5e76\u6307\u51fa\u4e86\u9ad8\u6548\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u672a\u6765\u7814\u7a76\u7684\u6709\u524d\u666f\u65b9\u5411\u3002<\/p>\n<h3>\u9644\u5f55<\/h3>\n<h4>\u5173\u4e8e\u6709\u635f\u63a8\u6d4b\u89e3\u7801\u7684\u76f8\u5173\u5de5\u4f5c<\/h4>\n<p>\u867d\u7136\u539f\u59cb\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u4e3b\u8981\u662f\u65e0\u635f\u7684\uff0c\u4f46\u4e00\u4e9b\u8fd1\u671f\u5de5\u4f5c\u5c1d\u8bd5\u653e\u5bbd\u7ea6\u675f\u5e76\u63a2\u7d22\u6709\u635f\u63a8\u6d4b\u89e3\u7801\u3002\u4f8b\u5982\uff0cBiLD \u4f7f\u7528\u4e00\u4e2a\u5c0f\u6a21\u578b\u8fdb\u884c\u81ea\u56de\u5f52\u6587\u672c\u751f\u6210\uff0c\u5e76\u5076\u5c14\u4ee5\u975e\u81ea\u56de\u5f52\u65b9\u5f0f\u8c03\u7528\u4e00\u4e2a\u66f4\u5927\u7684\u6a21\u578b\u6765\u4fee\u6b63\u4e0d\u51c6\u786e\u9884\u6d4b\uff0c\u4ece\u800c\u5728\u8d28\u91cf\u9000\u5316\u5f88\u5c0f\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u52a0\u901f\u3002Narasimhan \u7b49\u4eba\u63d0\u51fa speculative cascading\uff0c\u8fd9\u662f\u4e00\u79cd\u5c06 cascade-style deferral rules \u4e0e speculative execution \u96c6\u6210\u7684\u65b9\u6cd5\uff0c\u76f8\u6bd4\u5355\u72ec\u4f7f\u7528\u4efb\u4e00\u65b9\u6cd5\u90fd\u80fd\u83b7\u5f97\u66f4\u597d\u7684 cost-quality trade-off\u3002\u53e6\u4e00\u79cd\u65b9\u6cd5 MTAD \u4f7f\u7528\u4e00\u4e2a\u8f83\u5c0f\u8f85\u52a9\u6a21\u578b\u6765\u8fd1\u4f3c\u5927\u6a21\u578b\u7684 multi-token joint distribution\uff0c\u901a\u8fc7\u63a5\u53d7\u8fd9\u79cd\u8fd1\u4f3c\u4e2d\u7684\u6709\u754c\u8bef\u5dee\uff0c\u540c\u65f6\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u4e0e\u8f93\u51fa\u6709\u6548\u6027\u3002\u4e3a\u89e3\u51b3\u9ad8\u8d28\u91cf\u4f46\u672a\u5bf9\u9f50 draft tokens \u88ab\u62d2\u7edd\u7684\u95ee\u9898\uff0cBachmann \u7b49\u4eba\u63d0\u51fa\u901a\u8fc7\u8bad\u7ec3\u4e00\u4e2a\u7d27\u51d1\u7684 \u201cjudge\u201d \u6a21\u5757\u6765\u8bc6\u522b\u6709\u6548 continuation\uff0c\u4ece\u800c\u9002\u914d verification step\uff0c\u5373\u4f7f\u6ca1\u6709\u5b8c\u7f8e\u76ee\u6807\u6a21\u578b\u5bf9\u9f50\u4e5f\u80fd\u663e\u8457\u63d0\u9ad8 acceptance rates \u548c\u901f\u5ea6\u3002RSD \u5f15\u5165 process reward model \u6765\u8bc4\u4f30\u4e2d\u95f4\u89e3\u7801\u6b65\u9aa4\uff0c\u52a8\u6001\u51b3\u5b9a\u4f55\u65f6\u8c03\u7528\u76ee\u6807\u6a21\u578b\uff0c\u5e76\u5f15\u5165\u671d\u5411\u9ad8\u5956\u52b1\u8f93\u51fa\u7684\u53d7\u63a7\u504f\u7f6e\uff0c\u4ee5\u4f18\u5316 cost-quality trade-off\u3002RAPID \u4f7f\u7528\u57fa\u4e8e RAG \u7684\u65b9\u6cd5\u5728\u7f29\u77ed\u4e0a\u4e0b\u6587\u4e0a\u4f5c\u4e3a drafter\u3002TokenSwift \u7efc\u5408\u4f7f\u7528\u5e26\u90e8\u5206 KV cache \u7684 LLM \u548c <span class=\"katex-eq\" data-katex-display=\"false\">N<\/span>-gram tables \u6765\u52a0\u901f\u8d85\u957f\u5e8f\u5217\u751f\u6210\uff08\u6700\u9ad8 100k tokens\uff09\uff0c\u540c\u65f6\u5c06\u8ba1\u7b97\u65f6\u95f4\u4ece\u6570\u5c0f\u65f6\u51cf\u5c11\u5230\u6570\u5206\u949f\u3002<\/p>\n<h4>\u4e3a\u4ec0\u4e48 KV Cache \u80fd\u63d0\u4f9b\u5e2e\u52a9\u7684\u76f4\u89c9<\/h4>\n<p>KV cache \u5b58\u50a8\u4e86\u6a21\u578b\u5728\u5904\u7406\u5148\u524d tokens \u65f6\u79ef\u7d2f\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u5728\u9884\u6d4b\u4e0b\u4e00\u4e2a token \u65f6\uff0c\u76ee\u6807\u6a21\u578b\u4f9d\u8d56\u4e09\u4e2a\u7ec4\u4ef6\uff1aKV cache\uff08\u4e0a\u4e0b\u6587\u8bb0\u5fc6\uff09\u3001\u8f93\u5165\u8bcd\u5d4c\u5165\uff0c\u4ee5\u53ca\u6a21\u578b\u53c2\u6570\u3002<\/p>\n<p>\u5728\u6211\u4eec\u7684\u65b9\u6cd5\u4e2d\uff0c\u8349\u7a3f\u6a21\u578b\u5df2\u7ecf\u4e0e\u76ee\u6807\u6a21\u578b\u5171\u4eab\u8f93\u5165\u5d4c\u5165\uff0c\u56e0\u6b64\u4e8c\u8005\u9884\u6d4b\u4e2d\u7684\u4e3b\u8981\u5dee\u5f02\u6765\u81ea KV cache \u548c\u5185\u90e8\u53c2\u6570\u3002\u901a\u8fc7\u5141\u8bb8\u8349\u7a3f\u6a21\u578b\u4f7f\u7528\u7531\u76ee\u6807\u6a21\u578b\u751f\u6210\u7684 KV cache\uff0c\u6211\u4eec\u6d88\u9664\u4e86\u53e6\u4e00\u4e2a\u53d8\u5316\u6765\u6e90\u3002\u56e0\u6b64\uff0c\u5b83\u4eec\u9884\u6d4b\u4e4b\u95f4\u552f\u4e00\u5269\u4e0b\u7684\u5dee\u5f02\u6765\u81ea\u6a21\u578b\u53c2\u6570\u3002\u8fd9\u79cd\u5171\u4eab\u4f7f\u8349\u7a3f\u6a21\u578b\u7684\u9884\u6d4b\u66f4\u63a5\u8fd1\u76ee\u6807\u6a21\u578b\u7684\u9884\u6d4b\uff0c\u56e0\u4e3a\u5b83\u79fb\u9664\u4e86\u7531\u4e0d\u540c\u4e0a\u4e0b\u6587\u8868\u793a\u9020\u6210\u7684\u5dee\u5f02\u3002<\/p>\n<h4>\u6ce8\u610f\u529b\u805a\u5408\u7684\u6b63\u786e\u6027<\/h4>\n<p>\u56e0\u4e3a query matrix <span class=\"katex-eq\" data-katex-display=\"false\">Q<\/span> \u53ef\u4ee5\u5206\u89e3\u6210\u82e5\u5e72\u884c\uff0c\u6bcf\u4e00\u884c\u4ee3\u8868\u4e00\u4e2a\u5355\u72ec query <span class=\"katex-eq\" data-katex-display=\"false\">q<\/span>\uff0c\u6240\u4ee5\u6211\u4eec\u53ea\u9700\u8003\u8651\u6bcf\u4e00\u884c <span class=\"katex-eq\" data-katex-display=\"false\">q<\/span> \u4e0e KV \u8ba1\u7b97 attention \u540e\u7684\u8f93\u51fa\u3002\u8fd9\u6837\u53ef\u4ee5\u5047\u8bbe\u53c2\u4e0e\u8ba1\u7b97\u7684 KV \u5df2\u7ecf\u7ecf\u8fc7 tree mask \u5904\u7406\uff0c\u4ece\u800c\u7b80\u5316\u8bc1\u660e\u3002\u6211\u4eec\u53ea\u9700\u8bc1\u660e\u7531\u6bcf\u4e2a\u5355\u72ec <span class=\"katex-eq\" data-katex-display=\"false\">q<\/span> \u5f97\u5230\u7684\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">o<\/span> \u6ee1\u8db3\u8981\u6c42\uff0c\u8fd9\u5373\u53ef\u8bf4\u660e\u6574\u4e2a\u77e9\u9635 <span class=\"katex-eq\" data-katex-display=\"false\">Q<\/span> \u7684\u6574\u4f53\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">O<\/span> \u4e5f\u6ee1\u8db3\u8981\u6c42\u3002<\/p>\n<p><strong>\u547d\u9898\u3002<\/strong> \u8bb0 merged attention \u7684 log-sum-exp \u4e3a\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\mathrm{LSE}_{\\mathrm{merge}} = \\log\\Bigl(\\exp\\bigl(\\mathrm{LSE}_{\\mathrm{cache}}\\bigr) \\;+\\; \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{specs}}\\bigr)\\Bigr),\n\n<\/span>\n<p>\u5219 merged attention output \u53ef\u5199\u4e3a\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\no_{\\mathrm{merge}} =\n\no_{\\mathrm{cache}} \\cdot \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{cache}} - \\mathrm{LSE}_{\\mathrm{merge}}\\bigr)\n\n+ o_{\\mathrm{specs}} \\cdot\\exp\\bigl(\\mathrm{LSE}_{\\mathrm{specs}} - \\mathrm{LSE}_{\\mathrm{merge}}\\bigr).\n\n<\/span>\n<p><strong>\u8bc1\u660e\u3002<\/strong> \u5bf9\u5927\u5c0f\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">d_{qk}<\/span> \u7684 <span class=\"katex-eq\" data-katex-display=\"false\">q<\/span>\uff0c\u5bf9 <span class=\"katex-eq\" data-katex-display=\"false\">K_{\\mathrm{merge}}<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">V_{\\mathrm{merge}}<\/span> \u6267\u884c\u6807\u51c6 scaled dot-product attention\uff0c\u5176\u4e2d\u4e8c\u8005\u5927\u5c0f\u5206\u522b\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">(M+N) \\times d_{qk}<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">(M+N) \\times d_v<\/span>\uff0c\u53ef\u5199\u4f5c\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\begin{aligned}\n\no_{\\mathrm{merge}} &amp;= \\operatorname{mha}\\left(q, K_{\\mathrm{merge}}, V_{\\mathrm{merge}}\\right) \\\\\n\n&amp;= \\operatorname{softmax}\\left(qK_{\\mathrm{merge}}^\\top\/\\sqrt{d_{qk}}\\right) V_{\\mathrm{merge}}.\n\n\\end{aligned}\n\n<\/span>\n<p>\u7531\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">K<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">V<\/span> \u7531 <span class=\"katex-eq\" data-katex-display=\"false\">\\left(K_{\\mathrm{specs}}, K_{\\mathrm{cache}}\\right)<\/span> \u4e0e <span class=\"katex-eq\" data-katex-display=\"false\">\\left(V_{\\mathrm{specs}}, V_{\\mathrm{cache}}\\right)<\/span> \u5806\u53e0\u5f62\u6210\uff0c\u6211\u4eec\u76f8\u5e94\u62c6\u5206 logit matrix\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\nq K_{\\mathrm{merge}}^\\top \/ \\sqrt{d_{qk}} =\n\n\\operatorname{concat}\\Bigl(\n\nq K_{\\mathrm{cache}}^\\top \/ \\sqrt{d_{qk}},\n\nq K_{\\mathrm{specs}}^\\top \/ \\sqrt{d_{qk}}\n\n\\Bigr).\n\n<\/span>\n<p>\u8bb0\u8fd9\u4e9b sub-logit matrices \u4e3a\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\nZ_{\\mathrm{cache}} = q K_{\\mathrm{cache}}^\\top \/ \\sqrt{d_{qk}},\\quad\n\nZ_{\\mathrm{specs}} = q K_{\\mathrm{specs}}^\\top \/ \\sqrt{d_{qk}}.\n\n<\/span>\n<p><span class=\"katex-eq\" data-katex-display=\"false\">Z_{\\mathrm{specs}}<\/span> \u7684\u6bcf\u4e00\u884c\u5bf9\u5e94 <span class=\"katex-eq\" data-katex-display=\"false\">q<\/span> \u4e2d\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">i<\/span> \u4e2a query \u4e0e <span class=\"katex-eq\" data-katex-display=\"false\">K_{\\mathrm{specs}}<\/span> \u4e2d\u6240\u6709\u884c\u7684\u70b9\u79ef\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">Z_{\\mathrm{cache}}<\/span> \u7684\u884c\u5219\u5bf9\u5e94\u540c\u4e00\u4e2a query \u4e0e <span class=\"katex-eq\" data-katex-display=\"false\">K_{\\mathrm{cache}}<\/span>\u3002<\/p>\n<p>\u4e3a\u7ec4\u5408 partial attentions\uff0c\u6211\u4eec\u8bb0\u5f55\u6bcf\u4e2a sub-logit set \u7684 exponentials \u4e4b\u548c\u7684\u5bf9\u6570\u3002\u5177\u4f53\u5b9a\u4e49\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\begin{aligned}\n\n\\mathrm{LSE}_{\\mathrm{cache}} &amp;= \\log\\left(\\sum\\nolimits_{j=1}^{N} \\exp\\left(Z_{\\mathrm{cache}}^{(j)}\\right)\\right),\\\\\n\n\\mathrm{LSE}_{\\mathrm{specs}} &amp;= \\log\\left(\\sum\\nolimits_{j=1}^{M} \\exp\\left(Z_{\\mathrm{specs}}^{(j)}\\right)\\right),\n\n\\end{aligned}\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">Z_{\\mathrm{specs}}^{(j)}<\/span> \u8868\u793a\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">j<\/span> \u4e2a\u5143\u7d20\u7684 logit\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">Z_{\\mathrm{cache}}^{(j)}<\/span> \u7c7b\u4f3c\u3002<\/p>\n<p>\u4e8e\u662f <span class=\"katex-eq\" data-katex-display=\"false\">o_{\\mathrm{cache}}<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">o_{\\mathrm{specs}}<\/span> \u53ef\u5199\u4e3a\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\begin{aligned}\n\no_{\\mathrm{cache}} &amp;= \\frac{\\sum_{j=1}^{N} \\exp\\left(Z_{\\mathrm{cache}}^{(j)}\\right) V_{\\mathrm{cache}}^{(j)}}{\\exp\\left(\\mathrm{LSE}_{\\mathrm{cache}}\\right)},\\\\\n\no_{\\mathrm{specs}} &amp;= \\frac{\\sum_{j=1}^{M} \\exp\\left(Z_{\\mathrm{specs}}^{(j)}\\right) V_{\\mathrm{specs}}^{(j)}}{\\exp\\left(\\mathrm{LSE}_{\\mathrm{specs}}\\right)}.\n\n\\end{aligned}\n\n<\/span>\n<p>\u6574\u4e2a attention score \u53ef\u5199\u4e3a\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\begin{aligned}\n\nN_{\\mathrm{num}} &amp;=\n\n\\sum_{j=1}^{N} \\exp\\bigl(Z_{\\mathrm{cache}}^{(j)}\\bigr) V_{\\mathrm{cache}}^{(j)}\n\n+ \\sum_{j=1}^{M} \\exp\\bigl(Z_{\\mathrm{specs}}^{(j)}\\bigr) V_{\\mathrm{specs}}^{(j)},\\\\\n\nD_{\\mathrm{den}} &amp;=\n\n\\exp\\bigl(\\mathrm{LSE}_{\\mathrm{cache}}\\bigr)\n\n+ \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{specs}}\\bigr),\\\\\n\no_{\\mathrm{merge}} &amp;= \\frac{N_{\\mathrm{num}}}{D_{\\mathrm{den}}}.\n\n\\end{aligned}\n\n<\/span>\n<p>\u5c06 split attention \u805a\u5408\u8fdb whole attention\uff0c\u5373\u53ef\u5f97\u5230\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\no_{\\mathrm{merge}} =\n\no_{\\mathrm{cache}} \\cdot\\exp\\bigl(\\mathrm{LSE}_{\\mathrm{cache}} - \\mathrm{LSE}_{\\mathrm{merge}}\\bigr)\n\n+ o_{\\mathrm{specs}} \\cdot \\exp\\bigl(\\mathrm{LSE}_{\\mathrm{specs}} - \\mathrm{LSE}_{\\mathrm{merge}}\\bigr).\n\n<\/span>\n<p>\u8bc1\u6bd5\u3002<\/p>\n<h4>\u5b9e\u9a8c\u7ec6\u8282<\/h4>\n<p>\u6240\u6709\u6a21\u578b\u5747\u4f7f\u7528 8 \u5f20 A100 80GB GPU \u8bad\u7ec3\u3002\u5bf9\u4e8e\u5728\u77ed\u4e0a\u4e0b\u6587\u6570\u636e\u4e0a\u8bad\u7ec3\u7684 7B\u30018B \u548c 13B \u76ee\u6807\u6a21\u578b\uff0c\u6211\u4eec\u4f7f\u7528\u5e26 ZeRO-1 \u7684 LongSpec\u3002\u5bf9\u4e8e\u5728\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u4e0a\u8bad\u7ec3\u7684 7B\u30018B \u548c 13B \u6a21\u578b\uff0c\u4ee5\u53ca\u6240\u6709 33B \u76ee\u6807\u6a21\u578b\u8bbe\u7f6e\uff0c\u6211\u4eec\u4f7f\u7528 ZeRO-3\u3002<\/p>\n<p>\u6807\u51c6 cross-entropy \u7528\u4e8e\u4f18\u5316\u8349\u7a3f\u6a21\u578b\uff0c\u540c\u65f6\u76ee\u6807\u6a21\u578b\u53c2\u6570\u4fdd\u6301\u51bb\u7ed3\u3002\u4e3a\u7f13\u89e3 logits \u8ba1\u7b97\u9020\u6210\u7684 VRAM \u5cf0\u503c\uff0c\u6211\u4eec\u4f7f\u7528\u7531 Liger Kernel \u5b9e\u73b0\u7684 fused-linear-and-cross-entropy loss\uff0c\u5b83\u5c06 LM head \u4e0e softmax function \u4e00\u8d77\u8ba1\u7b97\uff0c\u5e76\u80fd\u5927\u5e45\u7f13\u89e3\u8be5\u95ee\u9898\u3002<\/p>\n<p>\u5bf9\u4e8e SlimPajama-6B \u6570\u636e\u96c6\uff0c\u6211\u4eec\u5c06 batch size\uff08\u5305\u62ec accumulation\uff09\u914d\u7f6e\u4e3a 2048\uff0c\u5c06\u6700\u5927 learning rate \u8bbe\u4e3a 5e-4\uff0c\u4f7f\u7528 cosine learning rate schedule\uff0c\u5e76\u7528 AdamW \u4f18\u5316\u8349\u7a3f\u6a21\u578b\u3002\u5728\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u65f6\uff0c\u6211\u4eec\u91c7\u7528 batch size 256 \u548c\u6700\u5927 learning rate 5e-6\u3002\u8349\u7a3f\u6a21\u578b\u5728\u6240\u6709\u6570\u636e\u96c6\u4e0a\u90fd\u53ea\u8bad\u7ec3\u4e00\u4e2a epoch\u3002<\/p>\n<p>\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u4e3b\u8981\u8ba1\u7b97\u6210\u672c\u6765\u81ea\u5bf9\u76ee\u6807\u6a21\u578b\u8fdb\u884c forward \u4ee5\u83b7\u5f97 KV cache\u3002\u8fd1\u671f\uff0c\u4e00\u4e9b\u516c\u53f8\u5f15\u5165\u4e86\u4e00\u79cd\u79f0\u4e3a context caching \u7684\u670d\u52a1\uff0c\u6d89\u53ca\u5b58\u50a8\u5927\u91cf KV cache\u3002\u56e0\u6b64\uff0c\u5728\u771f\u5b9e\u90e8\u7f72\u4e2d\uff0c\u8fd9\u4e9b\u9884\u5b58\u50a8 KV caches \u53ef\u76f4\u63a5\u4f5c\u4e3a\u8bad\u7ec3\u6570\u636e\u4f7f\u7528\uff0c\u4ece\u800c\u663e\u8457\u52a0\u901f\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/p>\n<p>\u5bf9\u4e8e LongSpec \u7684 tree decoding\uff0c\u6211\u4eec\u91c7\u7528 dynamic beam search \u6765\u6784\u9020\u6811\u3002\u5148\u524d\u7814\u7a76\u8868\u660e\uff0cbeam search \u867d\u7136\u80fd\u5b9e\u73b0\u9ad8 acceptance rates\uff0c\u4f46\u5728\u63a8\u6d4b\u89e3\u7801\u4e2d\u5904\u7406\u901f\u5ea6\u8f83\u6162\u3002\u6211\u4eec\u7684\u7814\u7a76\u53d1\u73b0\uff0c\u8be5 slowdown \u4e3b\u8981\u7531 KV cache movement \u5f15\u8d77\u3002\u5728\u4f20\u7edf beam search \u4e2d\uff0c\u4e0d\u5c5e\u4e8e top-<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> likelihood \u7684 nodes \u4f1a\u88ab\u4e22\u5f03\uff0c\u8fd9\u4e00\u6b65\u9700\u8981\u79fb\u52a8 KV cache\u3002\u7136\u800c\uff0c\u5728\u63a8\u6d4b\u89e3\u7801\u4e2d\uff0c\u4e22\u5f03\u8fd9\u4e9b nodes \u5e76\u65e0\u5fc5\u8981\uff0c\u56e0\u4e3a draft sequences \u4e0d\u9700\u8981\u4fdd\u6301\u7edf\u4e00\u957f\u5ea6\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u53ef\u4ee5\u7b80\u5355\u5730\u505c\u6b62\u5bf9\u4f4e likelihood branches \u7684 descendant nodes \u8fdb\u884c\u8ba1\u7b97\uff0c\u800c\u4e0d\u5b8c\u5168\u79fb\u9664\u5b83\u4eec\u3002\u91c7\u7528\u8fd9\u4e00\u65b9\u6cd5\u540e\uff0cbeam search \u80fd\u83b7\u5f97\u5f3a\u6027\u80fd\u800c\u4e0d\u4f1a\u5e26\u6765\u8fc7\u9ad8\u8ba1\u7b97\u5f00\u9500\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0c\u6bcf\u4e2a\u63a8\u6d4b\u6b65\u9aa4\u7684 beam width \u8bbe\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">[4, 16, 16, 16, 16]<\/span>\u3002\u672c\u6587\u6240\u6709\u63a8\u7406\u5b9e\u9a8c\u5747\u5728\u5355\u5f20 A100 80GB GPU \u4e0a\u4f7f\u7528 float16 precision \u8fdb\u884c\u3002<\/p>\n<h4>EAGLE \u4e0e Token Recycling \u5728\u957f\u4e0a\u4e0b\u6587\u63a8\u6d4b\u89e3\u7801\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c<\/h4>\n<p>\u5728\u8868 4 \u4e2d\uff0c\u6211\u4eec\u6bd4\u8f83\u4e24\u4e2a\u6a21\u578b\u5728\u4e94\u79cd\u8bbe\u7f6e\u4e0b\u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span> \u548c\u89e3\u7801\u901f\u5ea6\uff08tokens\/s\uff09\uff1a\u6765\u81ea HuggingFace \u7684 baseline PyTorch \u5b9e\u73b0\uff08\u201cVanilla HF\u201d\uff09\u3001\u4f7f\u7528 <code>Flash Attention<\/code> \u7684\u76f8\u540c\u6a21\u578b\uff08\u201cVanilla FA\u201d\uff09\u3001Token Recycling\uff08\u201cTR\u201d\uff0c\u4e00\u79cd SoTA retrieval-based \u65b9\u6cd5\uff09\u3001EAGLE\uff08\u4f7f\u7528 anchor offset indices \u8bad\u7ec3\uff0c\u5e76\u4f7f\u7528 HuggingFace \u63a8\u7406\uff09\uff0c\u4ee5\u53ca\u6211\u4eec\u5e26 hybrid tree attention \u7684 LongSpec\u3002\u5728\u4e94\u4e2a\u6570\u636e\u96c6\uff08GovReport\u3001QMSum\u3001MultiNews\u3001LCC \u548c RB-P\uff09\u4e0a\uff0cVanilla HF \u7684\u89e3\u7801\u901f\u5ea6\u88ab\u9650\u5236\u5728 14 \u5230 30 tokens\/s \u4e4b\u95f4\uff0c\u800c\u5207\u6362\u5230 <code>Flash Attention<\/code> \u540e\u901f\u5ea6\u63d0\u5347\u5230\u7ea6 50 tokens\/s\uff0c\u5e26\u6765\u8d85\u8fc7 2.5<span class=\"katex-eq\" data-katex-display=\"false\">\\times<\/span> \u7684\u52a0\u901f\u3002<\/p>\n<p>EAGLE \u5c06 acceptance length \u6269\u5c55\u5230\u7ea6 2\uff0c\u5e76\u5b9e\u73b0 26-40 tokens\/s\uff0c\u76f8\u6bd4 Vanilla HF \u83b7\u5f97 30-50% \u7684\u52a0\u901f\u3002\u7136\u800c\uff0c\u7531\u4e8e EAGLE \u65e0\u6cd5\u5229\u7528 <code>Flash Attention<\/code>\uff0c\u5b83\u5728\u6bcf\u79cd\u8bbe\u7f6e\u4e0b\u7684\u89e3\u7801\u901f\u5ea6\u90fd\u663e\u8457\u4f4e\u4e8e Vanilla FA\u3002\u81f3\u4e8e TR\uff0c\u867d\u7136\u5b83\u5c06 acceptance length \u6269\u5c55\u5230\u7ea6 3\uff08\u8fdc\u5927\u4e8e EAGLE\uff09\uff0c\u5e76\u5728\u8bb8\u591a\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e2d\u7b49\u7a0b\u5ea6\u52a0\u901f\uff0c\u4f46\u5b83\u5728\u6574\u4f53\u4e0a\u59cb\u7ec8\u4e0d\u5982 LongSpec\u3002<\/p>\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5e26 hybrid tree attention \u7684 LongSpec \u5728\u6240\u6709\u6a21\u578b\u548c\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u7ea6 100 tokens\/s \u7684\u9ad8\u5f97\u591a\u7684\u89e3\u7801\u901f\u5ea6\u3002\u8fd9\u8868\u660e EAGLE \u4e0e <code>Flash Attention<\/code> \u7684\u4e0d\u517c\u5bb9\u4ece\u6839\u672c\u4e0a\u9650\u5236\u4e86\u5b83\u7684\u89e3\u7801\u6027\u80fd\u3002\u6211\u4eec\u7684 hybrid tree attention \u4fdd\u6301\u4e86\u4e0e <code>Flash Attention<\/code> \u7684\u517c\u5bb9\u6027\uff0c\u56e0\u6b64\u91ca\u653e\u51fa\u663e\u8457\u66f4\u9ad8\u7684\u89e3\u7801\u901f\u5ea6\uff0c\u51f8\u663e\u4e86\u5c06 tree-structured attention \u4e0e <code>Flash Attention<\/code> \u7b49 SoTA \u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u6280\u672f\u7ed3\u5408\u7684\u91cd\u8981\u6027\u3002<\/p>\n<p><center>\u8868 4\uff1a\u4e0d\u540c\u6a21\u578b\u4e0e\u8bbe\u7f6e\u4e0b\u7684\u5e73\u5747\u63a5\u53d7\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span> \u548c\u89e3\u7801\u901f\u5ea6\uff08tokens\/s\uff09\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u201cVanilla HF\u201d \u6307 HuggingFace \u57fa\u4e8e PyTorch \u7684 attention \u5b9e\u73b0\uff0c\u800c \u201cVanilla FA\u201d \u4f7f\u7528 <code>Flash Attention<\/code>\u3002\u6240\u6709\u7ed3\u679c\u5747\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">T=0<\/span> \u4e0b\u8ba1\u7b97\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257011.png\" alt=\"\u8868 4\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<h4>\u4e0d\u540c Prefill \u957f\u5ea6\u4e0b\u7684\u6027\u80fd\u5206\u6790<\/h4>\n<p><center>\u8868 5\uff1a\u968f\u7740 prefill length \u589e\u52a0\u7684\u8be6\u7ec6\u6027\u80fd\u5206\u89e3\uff0c\u8bbe\u7f6e\u4e3a LongChat-7B on GovReport\u3002<\/center><br \/>\n<center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/06\/20260610160257012.png\" alt=\"\u8868 5\" style=\"max-width:100%; height:auto;\"><\/center><\/p>\n<p>\u5728\u8868 5 \u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u968f\u7740 prefill length \u589e\u52a0\u65f6\u7684\u8be6\u7ec6\u6027\u80fd\u5206\u89e3\uff0c\u8bbe\u7f6e\u4e3a LongChat-7B on GovReport\u3002\u5728\u6240\u6709 token ranges \u4e2d\uff0c\u751f\u6210\u901f\u5ea6\u90fd\u4fdd\u6301\u975e\u5e38\u7a33\u5b9a\uff0c\u53ea\u5728 25k-32k \u8303\u56f4\u5185\u7565\u6709\u4e0b\u964d\u3002\u5e73\u5747 acceptance length \u5728\u6240\u6709 ranges \u4e2d\u4fdd\u6301\u4e00\u81f4\uff0c\u8fd9\u8868\u660e\u7cfb\u7edf\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u9009\u62e9\u4fdd\u7559\u7684 tokens \u6570\u8868\u73b0\u7a33\u5b9a\u3002\u8fd9\u79cd\u7a33\u5b9a\u6027\u8bf4\u660e draft quality \u4e0d\u53d7 prefill \u957f\u5ea6\u5f71\u54cd\uff0c\u5e76\u80fd\u7ef4\u6301\u4e00\u81f4\u7684\u8f93\u51fa\u52a8\u6001\u3002<\/p>\n<p>\u5728\u5ef6\u8fdf\u65b9\u9762\uff0cdraft time \u53ea\u5c0f\u5e45\u589e\u52a0\uff0c\u4ece\u6700\u77ed\u4e0a\u4e0b\u6587\u8303\u56f4\u4e2d\u7684 8.91 ms \u589e\u81f3\u6700\u957f\u4e0a\u4e0b\u6587\u8303\u56f4\u4e2d\u7684 9.25 ms\uff1b\u800c target time \u4ece 25.63 ms \u9010\u6b65\u589e\u52a0\u5230 30.89 ms\uff0c\u53cd\u6620\u51fa\u7ba1\u7406\u66f4\u5927\u4e0a\u4e0b\u6587\u6240\u589e\u52a0\u7684\u8ba1\u7b97\u8d1f\u8f7d\u3002Verify time \u5728\u6240\u6709 ranges \u4e2d\u51e0\u4e4e\u4fdd\u6301\u5e38\u6570\uff0c\u53ea\u4ece 6.18 ms \u7565\u589e\u81f3 6.28 ms\u3002<\/p>\n<p>\u603b\u4f53\u800c\u8a00\uff0c\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0c\u7cfb\u7edf\u80fd\u591f\u968f\u7740\u66f4\u957f\u8f93\u5165\u4e0a\u4e0b\u6587\u6709\u6548\u6269\u5c55\uff0c\u5728\u5ef6\u8fdf\u4ec5\u9002\u5ea6\u589e\u52a0\u7684\u60c5\u51b5\u4e0b\uff0c\u4fdd\u6301\u9ad8\u541e\u5410\u548c\u4e00\u81f4\u7684 drafting quality\u3002\u8fd9\u7a81\u51fa\u4e86\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u6d89\u53ca\u6269\u5c55\u8f93\u5165\u5e8f\u5217\u7684\u771f\u5b9e\u5e94\u7528\u4e2d\u7684\u5b9e\u7528\u6027\u548c\u7a33\u5065\u6027\u3002<\/p>\n<h4>\u4f2a\u4ee3\u7801<\/h4>\n<p>\u8fd9\u91cc\u7ed9\u51fa Anchor-Offset Indexing \u548c Flash Noisy Training \u7684\u4f2a\u4ee3\u7801\u3002<\/p>\n<p><center>\u7b97\u6cd5 1\uff1aAnchor-Offset Indexing<\/center><\/p>\n<pre><code class=\"language-text\">\u8f93\u5165\uff1a\u5e8f\u5217\u957f\u5ea6 N\uff1b\u6700\u5927\u957f\u5ea6 MAX_LEN\uff1bQuery states q_s\u3002\n\u8f93\u51fa\uff1a\u4f7f\u7528\u4fee\u6539\u540e\u7d22\u5f15\u5e94\u7528 RoPE \u7684 Query states\u3002\n\nP &lt;- {0, 1, ..., N-1}                         \/\/ \u521d\u59cb\u4f4d\u7f6e\u7d22\u5f15\no &lt;- RandomInt(0, MAX_LEN - N)                 \/\/ \u751f\u6210\u968f\u673a offset\nP[4:] += o                                     \/\/ \u5bf9\u524d 4 \u4e2a anchors \u4e4b\u540e\u7684\u7d22\u5f15\u5e94\u7528 offset\n\/\/ \u4f8b\u5982\uff0c\u5f53 N=128, o=16257 \u65f6\uff0cP \u53d8\u4e3a [0, 1, 2, 3, 16261, ..., 16385]\nreturn RoPE(q_s, P)<\/code><\/pre>\n<p><center>\u7b97\u6cd5 2\uff1aFlash Noisy Training<\/center><\/p>\n<pre><code class=\"language-text\">\u8f93\u5165\uff1aQueries Q\uff0cKey cache K\uff0cValue cache V\u3002\n\u8f93\u51fa\uff1a\u6700\u7ec8 attention output\u3002\n\nj &lt;- RandomInt(1, 4)                           \/\/ \u968f\u673a\u9009\u62e9\u8981\u4e22\u5f03\u7684 tokens \u6570\n\n\/\/ \u5728\u5207\u7247\u8f93\u5165\u4e0a\u6267\u884c attention\nQ&#039; &lt;- Q[j:]                                    \/\/ \u4e22\u5f03\u524d j \u4e2a queries\nK&#039; &lt;- K[:-j]                                   \/\/ \u4ece cache \u4e2d\u4e22\u5f03\u6700\u540e j \u4e2a keys\nV&#039; &lt;- V[:-j]                                   \/\/ \u4ece cache \u4e2d\u4e22\u5f03\u6700\u540e j \u4e2a values\nattn_out &lt;- FlashAttention(Q&#039;, K&#039;, V&#039;)\n\n\/\/ \u5bf9 output \u8fdb\u884c padding\uff0c\u4ee5\u5339\u914d\u539f\u59cb query length\npadded_out &lt;- Concat(Zeros(j), attn_out)\nreturn OutputProjection(padded_out)<\/code><\/pre>\n<h4>\u6848\u4f8b\u7814\u7a76<\/h4>\n<p>\u8fd9\u91cc\u5c55\u793a LongChat-7B \u6a21\u578b\u5728 GovReport \u4e0a\u7684\u4e00\u4e9b\u8bf4\u660e\u6027\u6848\u4f8b\uff0c\u5176\u4e2d\u84dd\u8272\u6807\u8bb0\u7684 tokens \u8868\u793a\u88ab\u76ee\u6807\u6a21\u578b\u63a5\u53d7\u7684 draft tokens\u3002\u7531\u4e8e\u7bc7\u5e45\u9650\u5236\uff0c\u8bba\u6587\u672a\u5c55\u793a\u5b8c\u6574\u7b54\u6848\u3002<\/p>\n<p>\u539f\u8bba\u6587\u5728\u8be5\u9644\u5f55\u4e2d\u4f7f\u7528 token \u7ea7\u989c\u8272\u6807\u6ce8\u5c55\u793a\u591a\u6bb5\u82f1\u6587\u751f\u6210\u8f93\u51fa\u3002\u4e3a\u4fdd\u6301 token \u63a5\u53d7\u60c5\u51b5\u7684\u65e0\u635f\u6027\uff0c\u793a\u4f8b\u8f93\u51fa\u672c\u8eab\u4fdd\u7559\u82f1\u6587\u539f\u6587\uff0c\u5e76\u5df2\u5c06\u989c\u8272\u6807\u6ce8\u673a\u68b0\u8f6c\u6362\u4e3a HTML span\u3002\u5c55\u5f00\u7248\u4fdd\u5b58\u5728\u672c\u5730\uff1a<\/p>\n<p><code>LongSpec\/case_study.md<\/code><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Meta Data \u53d1\u8868\u65f6\u95f4\uff1a2025-02-24\uff1b\u6700\u65b0 arXiv \u4fee\u8ba2\uff1a2026-04-08 \u4f5c\u8005\uff1aPenghui Yang, &#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-1608","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1608","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=1608"}],"version-history":[{"count":8,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1608\/revisions"}],"predecessor-version":[{"id":1622,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1608\/revisions\/1622"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1608"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1608"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1608"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}